number,title,url,primary_bucket,attention_score,attention_reason,maintainer_next_action,review_decision,merge_state_status,is_draft,failed_status_check_count,pending_status_check_count,failed_status_check_names,pending_status_check_names,closing_issue_numbers,linked_issue_numbers,referenced_issue_numbers,referenced_pr_numbers,supersedes_numbers,superseded_by_numbers,review_comment_count,commit_count,relationship_edge_count,patch_hash,diff_hash,label_text,created_at,updated_at,stale_days,body_excerpt
189096,[aot_autograd] Make CPU-side stream syncs real cross-stream barriers,https://github.com/pytorch/pytorch/pull/189096,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,189095,,,0,1,3,,,"module: inductor, ciflow/inductor, release notes: inductor",2026-07-06T21:34:22Z,2026-07-09T02:18:21Z,0,"Stack from ghstack (oldest at bottom): -> #189096 #189095 synchronize_stream / synchronize_device / synchronize_event block the CPU until a stream (or event) drains, so every subsequent kernel launch is host-ordered after them. Unlike device-side waits (wait_event / wait_stream), nothing on a consumer's stream waits..."
189024,[Dynamo] Support exception attribute access (StopIteration/AttributeError/NameError),https://github.com/pytorch/pytorch/pull/189024,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,0,1,,statusCheckRollup,,,,"157149, 187690, 187744, 188004, 188638, 188639, 188824, 188825, 188834",,,0,8,30,55b2a9c4b73c3cb2f348855a62c912fd5c207e8285992d5a37c7b04b3eea6630,43fb37d4a6ab4d189c6b782d1b86acdda16c4e0761e7d3e65c162599968e952a,"open source, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, merging, ciflow/torchtitan",2026-07-06T13:42:25Z,2026-07-09T02:17:19Z,0,"Stack from ghstack (oldest at bottom): -> #189024 #188825 #188834 #188824 #157149 #188639 #188638 #188004 #187744 #187690 Dynamo wrapped exceptions in a single ExceptionVariable that only tracked args/context/cause/traceback. Reading exception-specific attributes (StopIteration.value, AttributeError.name/.obj, NameE..."
189318,Bump pip from 26.0.1 to 26.1.2 in /.ci/docker,https://github.com/pytorch/pytorch/pull/189318,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing, python, dependencies, ciflow/docker",2026-07-08T23:28:56Z,2026-07-09T02:16:34Z,0,Bumps pip from 26.0.1 to 26.1.2. Changelog Sourced from pip's changelog. 26.1.2 (2026-05-31) Bug Fixes Reject console_scripts and gui_scripts entry points whose name would install a script outside the scripts directory. ([#14000](https://github.com/pypa/pip/issues/14000) <https://github.com/pypa/pip/issues/14000>_)...
189329,[DEBUG] Profile the XPU SYCL device-link bottleneck,https://github.com/pytorch/pytorch/pull/189329,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,1,,statusCheckRollup,,,,,,,0,1,0,,,ciflow/xpu,2026-07-09T02:15:39Z,2026-07-09T02:15:47Z,0,"Do not merge - temporary diagnostic. Why linux-noble-xpu-n-py3.10 / build takes ~35-45 min vs 5-15 min for the other builds in the same workflow. Profiling pull run 28986110660 showed the Build step is ~33 min even with sccache at 99.93% hit rate - because sccache caches only host C/C++, not SYCL device code. The si..."
188597,[ROCm] Add initial support for gfx1250,https://github.com/pytorch/pytorch/pull/188597,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/binaries, ciflow/docker, ciflow/h100","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,188242,,,0,22,2,,,"module: rocm, triaged, open source, Merged, Reverted, ciflow/binaries, ciflow/trunk, release notes: releng, module: inductor, ciflow/inductor, merging, ci-no-td, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, ciflow/docker",2026-06-30T20:28:11Z,2026-07-09T02:14:29Z,0,"This PR adds initial ROCm support for gfx1250 (CDNA5). What this PR enables: Gates gfx1250-specific behavior on ROCm 7.14+ (the support floor), applied consistently across CUDABlas.cpp, ScaledBlas.cpp, CUDAHooks.cpp, and the test helpers. Adds gfx1250 to the hipBLASLt preferred/supported arch lists and the hipSparse..."
188963,[XPU] Enable XPU testing for 50 test classes across 40 test files,https://github.com/pytorch/pytorch/pull/188963,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"open source, release notes: quantization, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, keep-going, ciflow/xpu, ci-no-td, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-07-05T06:42:54Z,2026-07-09T02:13:43Z,0,"Summary Enable XPU (allow_xpu=True) for test classes that are confirmed working on Intel GPU (XPU). Each class was validated by running pytest -k ""<ClassName> and xpu"" in an XPU-enabled PyTorch environment. Changes by file File Classes Enabled test/complex_tensor/test_complex_tensor.py TestComplexTensor, TestComplex..."
189173,torchfuzz: detach batch_norm running stats,https://github.com/pytorch/pytorch/pull/189173,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, merging",2026-07-07T17:20:25Z,2026-07-09T02:13:41Z,0,"The fuzzer marks float tensor args requires_grad=True for the backward pass, but torch.nn.functional.batch_norm is not differentiable w.r.t. running_mean / running_var, so when those inputs are grad-requiring args the generated program raises RuntimeError: ... 'running_mean' ... cannot have requires_grad True in eag..."
189111,[torchcomms hash update] update the pinned torchcomms hash,https://github.com/pytorch/pytorch/pull/189111,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,0,35,,"statusCheckRollup, macos-py3-arm64 / test (default, 1, 3, macos-m1-stable), macos-py3-arm64 / test (default, 2, 3, macos-m1-stable), macos-py3-arm64 / test (mps, 1, 1, macos-m1-14), macos-py3-arm64 / test (default, 3, 3, macos-m1-stable), win-vs2022-cpu-py3 / test (default, 2, 4, windows.4xlarge.nonephemeral), win-vs2022-cpu-py3 / test (default, 4, 4, windows.4xlarge.nonephemeral), win-vs2022-cpu-py3 / test (default, 3, 4, windows.4xlarge.nonephemeral), win-vs2022-cpu-py3 / test (default, 1, 4, windows.4xlarge.nonephemeral), linux-jammy-cuda13.0-py3.10-gcc11 / test (distributed, 1, 3, mt-l-x86iavx512-45-172-t4-4), linux-jammy-cuda13.0-py3.10-gcc11 / test (default, 1, 5, mt-l-x86aavx2-29-113-l4), linux-jammy-cuda13.0-py3.10-gcc11 / test (default, 3, 5, mt-l-x86aavx2-29-113-l4), linux-jammy-cuda13.0-py3.10-gcc11 / test (distributed, 3, 3, mt-l-x86iavx512-45-172-t4-4), linux-jammy-cuda13.0-py3.10-gcc11 / test (default, 2, 5, mt-l-x86aavx2-29-113-l4), linux-jammy-cuda13.0-py3.10-gcc11 / test (default, 4, 5, mt-l-x86aavx2-29-113-l4), linux-jammy-cuda13.0-py3.10-gcc11 / test (distributed, 2, 3, mt-l-x86iavx512-45-172-t4-4), linux-jammy-cuda13.0-py3.10-gcc11 / test (default, 5, 5, mt-l-x86aavx2-29-113-l4), unit-test / inductor-test / test (inductor_cpp_wrapper, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test (inductor_cpp_wrapper, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), linux-jammy-rocm-py3.10-mi350 / test (default, 4, 8, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (default, 1, 8, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (inductor, 2, 2, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (default, 6, 8, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (default, 5, 8, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (distributed, 3, 3, linux.rocm.gpu.gfx950.2), linux-jammy-rocm-py3.10-mi350 / test (distributed, 2, 3, linux.rocm.gpu.gfx950.2), linux-jammy-rocm-py3.10-mi350 / test (default, 3, 8, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (default, 7, 8, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (distributed, 1, 3, linux.rocm.gpu.gfx950.2), linux-jammy-rocm-py3.10-mi350 / test (inductor, 1, 2, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (default, 8, 8, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (default, 2, 8, linux.rocm.gpu.gfx950.1), unit-test / inductor-cpu-test / test (inductor_amx, 1, 2, mt-l-x86iamx-8-64)",,,,,,,0,1,0,9050a51eeaf0bb3c6a74ce7986b0afc9a60a89ac3786ba1b7a2445e39922a219,beb304d9b5ee4a9b27872c4b59aa54d6492bc57a94138872b9e340c25d80ddb8,"open source, ciflow/trunk, topic: not user facing, ciflow/inductor, merging, ci-no-td",2026-07-07T00:49:13Z,2026-07-09T02:13:38Z,0,This PR is auto-generated nightly by this action. Update the pinned torchcomms hash.
184218,Support decorator/curried calling convention for torch.utils.checkpoint.checkpoint,https://github.com/pytorch/pytorch/pull/184218,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,0,1,,statusCheckRollup,,,,,,,0,14,0,,,"ciflow/trunk, release notes: distributed (fsdp), topic: new features, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-18T15:06:12Z,2026-07-09T02:13:23Z,0,"Stack from ghstack (oldest at bottom): -> #184218 checkpoint() can now be called without a function argument to produce a decorator, keeping checkpoint configuration separate from the checkpointed function's arguments: checkpointed_fn = checkpoint(use_reentrant=False)(fn) out = checkpointed_fn(*args, **kwargs) This..."
188980,Fix inconsistent accelerator guard in checkpoint staging test,https://github.com/pytorch/pytorch/pull/188980,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,0,1,,statusCheckRollup,188790,188790,188790,,,,0,1,2,2864cb694b19924211a41a32b77aab4a367c86bdc826c798f2a80d7f9df9cf1e,97823578d961826aa39a10cff53ef426d7206881613e579299792e15281f670d,"open source, ciflow/trunk, topic: not user facing, release notes: distributed (checkpoint), merging",2026-07-05T23:22:39Z,2026-07-09T02:09:58Z,0,"Fixes #188790 Summary In test/distributed/checkpoint/_experimental/test_staging.py, the block that appends the async-staging and non-blocking-copy test cases was guarded by torch.cuda.is_available(), while the config values constructed inside that same block already use the backend-generic torch.accelerator.is_avail..."
187940,[xpu][inductor] fix the fallback behavior of maxpool backward for performance,https://github.com/pytorch/pytorch/pull/187940,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk, ciflow/xpu",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,0,0,,,,,,182619,,,0,11,2,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/xpu, ciflow/torchtitan",2026-06-23T12:41:47Z,2026-07-09T02:08:16Z,0,"Fix max_pool2d_backward fallback to eager op behavior. Previously, the decomposition pass was executed before lowering, causing max_pool2d_backward to be decomposed before the fallback mechanism could be applied(described in #182619). As a result, the eager fallback path was bypassed, and the operation was fused int..."
189328,Add retry with backoff to download_mnist.py,https://github.com/pytorch/pytorch/pull/189328,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,0,1,,statusCheckRollup,,,,,,,0,1,0,,,topic: not user facing,2026-07-09T02:04:52Z,2026-07-09T02:05:04Z,0,"Stack from ghstack (oldest at bottom): -> #189328 Note from human: this is not a big deal but it's also the kind of thing coding agents eat for lunch, so whatever. MNIST download in CI hits a single S3 mirror via urlretrieve with no retry, so a transient network blip fails the whole job. This adds up to 3 attempts p..."
189284,Add torch.autograd.graph.node_creation_hook,https://github.com/pytorch/pytorch/pull/189284,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,0,54,,"statusCheckRollup, linux-jammy-py3.10-clang18-asan / test (default, 7, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (openreg, 1, 1, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (default, 3, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (default, 6, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (default, 1, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (default, 2, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (default, 4, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (default, 5, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test (default, 2, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test (distributed, 2, 3, mt-l-x86iamx-8-64), linux-jammy-py3.10-gcc11 / test (distributed, 1, 3, mt-l-x86iamx-8-64), linux-jammy-py3.10-gcc11 / test (distributed, 3, 3, mt-l-x86iamx-8-64), linux-jammy-py3.10-gcc11 / test (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test (default, 3, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test (default, 5, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test (docs_test, 1, 1, mt-l-x86iavx512-8-64), linux-jammy-py3.10-gcc11 / test (default, 1, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (dynamo_wrapped, 3, 3, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (dynamo_wrapped, 2, 3, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (dynamo_wrapped, 1, 3, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (default, 5, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (crossref, 2, 2, mt-l-x86iavx512-8-64), linux-jammy-py3.14t-clang18 / test (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (default, 2, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (default, 1, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (crossref, 1, 2, mt-l-x86iavx512-8-64), linux-jammy-py3.14t-clang18 / test (default, 3, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test (dynamo_wrapped, 2, 3, mt-l-x86iavx512-8-64), linux-jammy-py3.10-clang18 / test (dynamo_wrapped, 3, 3, mt-l-x86iavx512-8-64), linux-jammy-py3.10-clang18 / test (default, 2, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test (dynamo_wrapped, 1, 3, mt-l-x86iavx512-8-64), linux-jammy-py3.10-clang18 / test (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test (crossref, 2, 2, mt-l-x86iavx512-8-64), linux-jammy-py3.10-clang18 / test (default, 1, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test (crossref, 1, 2, mt-l-x86iavx512-8-64), linux-jammy-py3.10-clang18 / test (default, 3, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test (default, 5, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test (default, 1, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test (dynamo_wrapped, 3, 3, mt-l-x86iavx512-8-64), linux-jammy-py3.14-clang18 / test (dynamo_wrapped, 2, 3, mt-l-x86iavx512-8-64), linux-jammy-py3.14-clang18 / test (default, 5, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test (crossref, 1, 2, mt-l-x86iavx512-8-64), linux-jammy-py3.14-clang18 / test (default, 2, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test (crossref, 2, 2, mt-l-x86iavx512-8-64), linux-jammy-py3.14-clang18 / test (default, 3, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test (dynamo_wrapped, 1, 3, mt-l-x86iavx512-8-64), linux-jammy-py3.14-clang18 / test (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-aarch64-py3.10 / test (default, 4, 5, mt-l-arm64g4-16-62), linux-jammy-aarch64-py3.10 / test (default, 1, 5, mt-l-arm64g4-16-62), linux-jammy-aarch64-py3.10 / test (default, 5, 5, mt-l-arm64g4-16-62), linux-jammy-aarch64-py3.10 / test (default, 3, 5, mt-l-arm64g4-16-62), linux-jammy-aarch64-py3.10 / test (default, 2, 5, mt-l-arm64g4-16-62), linux-noble-xpu-n-py3.10 / build",,,,,,,14,4,2,d35d3fbb255637545c6feecea0682da5d7efac570426ae85ee864764429287e3,d5ca0c017080d27b72b666874733e5b7f2751798a2e696e8a3dae7d3b22befa5,"release notes: autograd, topic: new features",2026-07-08T17:32:25Z,2026-07-09T02:04:54Z,0,"Stack from ghstack (oldest at bottom): -> #189284 Adds a public context manager that registers a thread-local hook fired once for every autograd Node created within it, with the freshly created Node as argument. This gives users a creation-time dual of saved_tensors_hooks: the hook can inspect the node, record it, s..."
181680,[triton hash update] update the pinned triton hash,https://github.com/pytorch/pytorch/pull/181680,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,1,1,,,"open source, ciflow/trunk, topic: not user facing, ciflow/inductor, ci-no-td, ciflow/docker",2026-04-28T00:49:13Z,2026-07-09T01:54:30Z,0,This PR is auto-generated nightly by this action. Update the pinned triton hash.
189217,[vision hash update] update the pinned vision hash,https://github.com/pytorch/pytorch/pull/189217,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/inductor, ci-no-td",2026-07-08T00:42:06Z,2026-07-09T01:54:24Z,0,This PR is auto-generated nightly by this action. Update the pinned vision hash.
187288,[vllm hash update] update the pinned vllm hash,https://github.com/pytorch/pytorch/pull/187288,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/inductor, ci-no-td, ciflow/vllm, ci-refresh-hf-cache",2026-06-15T01:06:43Z,2026-07-09T01:54:21Z,0,This PR is auto-generated nightly by this action. Update the pinned vllm hash.
188262,[CI] fix OpenBLAS build for non-AArch64,https://github.com/pytorch/pytorch/pull/188262,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td, ciflow/docker",2026-06-26T09:32:33Z,2026-07-09T01:54:09Z,0,"TARGET=ARMV8 was hardcoded, breaking non-aarch64 builds. Now it's only set on aarch64; other platforms (like riscv64) rely on DYNAMIC_ARCH instead. Also forcibly enable the float/double/complex precision variants and pass build flags to make install consistently."
179651,[torchtitan hash update] update the pinned torchtitan hash,https://github.com/pytorch/pytorch/pull/179651,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-04-08T00:42:38Z,2026-07-09T01:54:03Z,0,This PR is auto-generated nightly by this action. Update the pinned torchtitan hash.
187600,[ROCm][Inductor] Enable NHWC convs by default on CDNA with an Inductor layout-opt gate,https://github.com/pytorch/pytorch/pull/187600,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi350",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,149039,,,0,4,1,,,"module: rocm, triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, keep-going, release notes: inductor, ciflow/rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/inductor-rocm-mi300, ciflow/torchtitan, ciflow/docker",2026-06-17T20:02:50Z,2026-07-09T01:54:02Z,0,"Enable NHWC convs by default on CDNA (gfx9xx) and gate Inductor layout-opt so grouped convs that hit MIOpen's slow naive kernel stay NCHW. Builds on / supersedes #149039 (closed): keeps its NHWC-on-CDNA enablement, fixes the arch-guard bug, makes ConvUtils.h ATEN_CPU-lint-clean via a runtime CUDAHooks hook, adds a c..."
189083,[profiler][cupti] Drive periodic cuptiActivityFlushAll from the decode thread,https://github.com/pytorch/pytorch/pull/189083,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,0,1,,statusCheckRollup,,,,"188621, 189109, 189185",,,0,12,3,,,topic: not user facing,2026-07-06T20:10:32Z,2026-07-09T01:53:56Z,0,Stack from ghstack (oldest at bottom): #189185 -> #189083 #188621 #189109 The CUPTI monitor's background flush loop woke on a Python thread every background_flush_period_s just to call cuptiActivityFlushAll -- a pure-C call that needs no GIL -- so it held the GIL for no reason and contended with the training thread....
189185,[profiler][cupti] Make CuptiMonitor a singleton configured via a global configure(),https://github.com/pytorch/pytorch/pull/189185,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,0,1,,statusCheckRollup,,,,"188621, 189083, 189109",,,0,7,3,,,,2026-07-07T19:18:48Z,2026-07-09T01:53:53Z,0,"Stack from ghstack (oldest at bottom): -> #189185 #189083 #188621 #189109 Ports the config-only change from commit 4a24d3b207 (drop the TORCH_CUPTI_MONITOR_* env vars) and turns CuptiMonitor into a process-wide singleton, mirroring PmSampler. buffer_size, background_flush_period_s, and background_drain_period_s were..."
188621,"[profiler][cupti] Generate CUPTI field-id catalogs from the ABI as Field(id, ctype)",https://github.com/pytorch/pytorch/pull/188621,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,0,1,,statusCheckRollup,,,,"189083, 189109, 189185",,,0,12,3,,,topic: not user facing,2026-07-01T00:07:00Z,2026-07-09T01:53:52Z,0,"Stack from ghstack (oldest at bottom): #189185 #189083 -> #188621 #189109 The v2 CUPTI user-defined-record path selects records by field id (CUpti_ActivityFieldIds), which cupti-python does not expose. Instead of hand-maintaining those ids, generate them from cupti_activity.h at build time, and model each field as a..."
189229,[DO NOT MERGE][ROCm] Use torch.version.hip to detect ROCm when /opt/rocm is absent,https://github.com/pytorch/pytorch/pull/189229,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/dtensor, ciflow/inductor, ciflow/rocm-nightly","identify whether block is CI, merge conflict, or review gate",,DIRTY,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,20,0,,,"module: rocm, open source, release notes: releng, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), keep-going, ci-no-td, ciflow/rocm-nightly, ciflow/dtensor, ciflow/docker",2026-07-08T03:31:45Z,2026-07-09T01:53:50Z,0,cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coc...
188100,Support callable msg lambda in TestCase.assert* methods beyond assertEqual(),https://github.com/pytorch/pytorch/pull/188100,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188299, 188573, 189286",,,0,5,3,,,"ciflow/trunk, topic: not user facing, mergedog, make tests great again",2026-06-24T21:03:34Z,2026-07-09T01:53:44Z,0,"Stack from ghstack (oldest at bottom): #189286 #188573 #188299 -> #188100 Allows for lazy building of error messages only when tests fail! Saves on overhead, especially when serialized tensors are involved."
189275,Consolidate CD wheel build pipelines under .ci/wheel/ and unify the numpy pin,https://github.com/pytorch/pytorch/pull/189275,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"open source, release notes: releng, ciflow/binaries_wheel",2026-07-08T16:17:40Z,2026-07-09T01:53:40Z,0,"Summary The three CD wheel-build pipelines (Linux, Windows, macOS) grew as three independent Python ports and drifted apart: copy-pasted helpers, three different numpy build-pin tables, and different directory homes. This PR consolidates them under a common .ci/wheel/ root with a single shared _common.py, and reconc..."
115316,Automated submodule update: FBGEMM,https://github.com/pytorch/pytorch/pull/115316,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2023-12-07T01:42:28Z,2026-07-09T01:53:38Z,0,This is an automated pull request to update the first-party submodule for pytorch/FBGEMM. New submodule commit: pytorch/FBGEMM@a75b43f Test Plan: Ensure that CI jobs succeed on GitHub before landing.
189175,torchfuzz: fix clamp_min/max emitting None bound,https://github.com/pytorch/pytorch/pull/189175,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-07-07T17:21:36Z,2026-07-09T01:53:33Z,0,"ClampMinOperator/ClampMaxOperator stored the random bound as mutable instance state in fuzz_inputs_specs and read it back in codegen. The same operator instance serves multiple clamp nodes, so during code generation the first codegen resets the stored bound to None and later clamp nodes emit clamp_min(x, min=None) o..."
167224,[FlopCounter] Add skip_unsupported parameter for graceful handling of unsupported HOPs,https://github.com/pytorch/pytorch/pull/167224,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,134385,134385,134385,,,,0,1,5,,,"triaged, open source, release notes: python_frontend",2025-11-06T16:18:23Z,2026-07-09T01:53:29Z,0,"Fixes #134385 FlopCounterMode returns NotImplemented when it encounters a Higher Order Operator it does not handle, which makes it hard to use for benchmarking real-world models that mix standard PyTorch ops with custom HOPs or kernels. What this PR does Adds a skip_unsupported flag (default False) to FlopCounterMod..."
188299,Lazily build error messages for assert*() calls across tests to reduce overhead,https://github.com/pytorch/pytorch/pull/188299,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/h100-symm-mem","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188100, 188573, 189286",,,0,3,3,,,"release notes: quantization, topic: not user facing, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), ciflow/rocm-mi300, ciflow/h100, ciflow/h100-symm-mem, ciflow/b200, ciflow/dtensor, mergedog, make tests great again",2026-06-26T20:52:23Z,2026-07-09T01:53:28Z,0,Stack from ghstack (oldest at bottom): #189286 #188573 -> #188299 #188100 Apply new lazy error message feature to call-sites throughout test files. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aak...
189192,[MPS] Add `geqrf` and refactor `linalg_qr`,https://github.com/pytorch/pytorch/pull/189192,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/mps","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, release notes: mps, ciflow/mps, ciflow/inductor, release notes: inductor (aoti)",2026-07-07T20:17:38Z,2026-07-09T01:53:27Z,0,Stack from ghstack (oldest at bottom): -> #189192 This PR splits out the geqrf code from the linalg_qr kernel and refactors linalg_qr for MPS to call the device-agnostic linalg_qr_out function that CPU and CUDA use.
188573,Cleanup dead code across tests,https://github.com/pytorch/pytorch/pull/188573,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188100, 188299, 189286",,,0,4,3,,,"release notes: quantization, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/dtensor, mergedog, make tests great again",2026-06-30T18:03:26Z,2026-07-09T01:53:25Z,0,"Stack from ghstack (oldest at bottom): #189286 -> #188573 #188299 #188100 I don't expect huge perf improvements, but why not? cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @..."
189264,"[ROCm][CI] Make rocm-nightly workflow fully mi300-specific (name, ciflow tag, filename)",https://github.com/pytorch/pytorch/pull/189264,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: rocm, triaged, open source, ciflow/binaries, topic: not user facing, ciflow/rocm-mi300",2026-07-08T14:40:50Z,2026-07-09T01:53:23Z,0,Summary Make the rocm-nightly workflow fully arch-specific for mi300: Renamed the workflow file to .github/workflows/rocm-nightly-mi300.yml. Set the workflow display name to rocm-nightly-mi300 and its push-trigger ciflow tag to ciflow/rocm-nightly-mi300/*. Job name and build-environment are linux-noble-rocm-nightly-...
189193,[XCCL] Allow XCCL to be set as default in `_new_process_group_helper`,https://github.com/pytorch/pytorch/pull/189193,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,4098,4098,,,,,0,1,0,,,"open source, ciflow/trunk, release notes: distributed (c10d), ciflow/torchtitan",2026-07-07T20:23:13Z,2026-07-09T01:53:22Z,0,"Fixes intel/torch-xpu-ops#4098 For DeviceMesh which doesn't specify the backend type, _new_process_group_helper is called with the ""undefined"" backend, which is set based on available devices. NCCL backend can be set as default, as well as CUSTOM backends. This PR also allows the XCCL backend to be set as default if..."
187465,[Symm_mem] Support expandable segements inside PyTorch Symmetric memory for NCCL backend,https://github.com/pytorch/pytorch/pull/187465,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,189088,,,0,13,1,,,"open source, ciflow/trunk, release notes: distributed (c10d), ciflow/h100-symm-mem, ciflow/torchtitan",2026-06-16T17:24:46Z,2026-07-09T01:53:15Z,0,"Stack from ghstack (oldest at bottom): -> #187465 #189088 This PR allows the NCCL symmetric-memory allocator to back its allocations with the CUDA caching allocator's expandable segments; when expandable_segments is enabled. In that mode, NCCLSymmetricMemoryAllocator now allocates/frees the symmetric buffer via CUDA..."
189302,[ROCm][CI] Make ROCM_SO_FILES a single shared source of truth for wheel bundling + preload,https://github.com/pytorch/pytorch/pull/189302,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, triaged, open source, ciflow/binaries_wheel, ciflow/rocm-mi300",2026-07-08T20:43:15Z,2026-07-09T01:53:12Z,0,"Summary The ROCm shared-library basename list was duplicated in two places that had drifted apart: .ci/manywheel/repair_wheel.py::ROCM_SO_FILES (27 entries) — the libs bundled into the wheel. torch/__init__.py::_rocm_core_libs (22 entries) — the libs preloaded RTLD_GLOBAL, leaf-first, before import torch._C (added i..."
189314,flex_gemm: recognize 2-D block local-reduce contracts,https://github.com/pytorch/pytorch/pull/189314,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188112, 188468, 188469, 188470, 188739, 189188, 189190, 189315, 189316",,,0,1,9,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-08T23:12:36Z,2026-07-09T01:53:11Z,0,"Stack from ghstack (oldest at bottom): #189316 #189315 -> #189314 #189190 #189188 #188739 #188112 #188470 #188469 #188468 Recognize structural 2-D block local-reduce aux outputs of the form acc.view(M // bm, bm, N // bn, bn) reduced over both grouped dims. This threads block geometry through the output plan, templat..."
189315,flex_gemm: store 2-D block local reductions,https://github.com/pytorch/pytorch/pull/189315,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188112, 188468, 188469, 188470, 188739, 189188, 189190, 189314, 189316",,,0,1,19,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-08T23:12:40Z,2026-07-09T01:53:10Z,0,"Stack from ghstack (oldest at bottom): #189316 -> #189315 #189314 #189190 #189188 #188739 #188112 #188470 #188469 #188468 Enable the first QUACK-backed 2-D block local-reduce STORE path for 128x128 block-scale aux outputs. The compiler now threads block geometry into runtime dispatch with explicit block kwargs, requ..."
187699,Add a debug log when we skip sync_decision_cross_ranks (#187699),https://github.com/pytorch/pytorch/pull/187699,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-18T22:20:02Z,2026-07-09T01:53:09Z,0,"Summary: When there is collectives in the graph but there is no sync decision cross ranks, it could potentially have issue of inconsistent decisions on different ranks on what to save. So adding this debug log to facilitate debug when the inconsistent issue happens. Test Plan: Existing tests already cover it Reviewe..."
189316,flex_gemm: support mx block-scale stores,https://github.com/pytorch/pytorch/pull/189316,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188112, 188468, 188469, 188470, 188739, 189188, 189190, 189314, 189315",,,0,1,9,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-08T23:12:43Z,2026-07-09T01:53:08Z,0,Stack from ghstack (oldest at bottom): -> #189316 #189315 #189314 #189190 #189188 #188739 #188112 #188470 #188469 #188468 Compose 2-D block local-reduce stores with physical-only finalizers such as mx_e8m0_scale. Block aux plan classification now traces a shape-preserving pointwise finalizer back to its base block r...
189320,[inductor] Enable nested Inductor configs for invoke_subgraph regions,https://github.com/pytorch/pytorch/pull/189320,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,0,1,,statusCheckRollup,,,,189321,,,0,1,1,,,"module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-09T00:48:43Z,2026-07-09T01:53:07Z,0,"Stack from ghstack (oldest at bottom): #189321 -> #189320 The normal Inductor invoke_subgraph path compiled each nested region under the surrounding graph's Inductor config, so per-region config patches were dropped and nested options had no effect on the region. This threads the nested-region config through the inv..."
189307,[CI/CD] Copy newer CUPTI headers into manywheel binary-build images,https://github.com/pytorch/pytorch/pull/189307,PR blocked,108,failed checks or blocked/unstable merge state; approved with no detected failing/pending checks; CI lane labels: ciflow/docker,final maintainer merge/release decision,APPROVED,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"topic: not user facing, ciflow/docker",2026-07-08T21:41:31Z,2026-07-09T01:53:06Z,0,"#189214 added install_cupti_headers, which stages the newer standalone CUPTI redist headers into /usr/local/cupti-headers-<major.minor> for the CUDA-13 build containers. However, the manywheel binary-build images only COPY /usr/local/cuda-<ver> out of the cuda build stage, so the staged headers never reach the final..."
189324,[T278829272][Claude Code] Gate TLX flex_attention choices on tlx_mode,https://github.com/pytorch/pytorch/pull/189324,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,1,,statusCheckRollup,,,174929,,,,0,1,1,,,"module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-07-09T01:25:35Z,2026-07-09T01:53:05Z,0,"Summary: X-link: facebookexperimental/triton#1961 15 caffe2/test/inductor:flex_attention tests broke on Hopper/Blackwell CI, root-caused to a missing gate against tlx-mode in flex_attention code. TLXInductorChoices.append_flex_attention_choices was the only method on the class that did not gate on config.triton.tlx_..."
189321,[inductor] Enable per-region cudagraphs for invoke_subgraph via auto graph partition,https://github.com/pytorch/pytorch/pull/189321,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,0,1,,statusCheckRollup,,,,189320,,,0,1,1,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-09T00:48:48Z,2026-07-09T01:53:04Z,0,"Stack from ghstack (oldest at bottom): -> #189321 #189320 Building on nested Inductor configs, a region can set triton.cudagraphs to opt into or out of cudagraphs independently of the enclosing graph. When a region's preference differs from the enclosing graph and graph partition is globally off, compile_fx enables..."
189266,[ROCm][CI] Fix py3.15 ROCm manywheel numpy source build (pkg-config python3),https://github.com/pytorch/pytorch/pull/189266,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,1,,statusCheckRollup,,,,,,,0,2,0,,,"module: rocm, triaged, open source, ciflow/binaries_wheel",2026-07-08T14:53:37Z,2026-07-09T01:53:02Z,0,"Summary Python 3.15 has no prebuilt numpy wheel on PyPI yet, so pip builds numpy from source for every arch. meson's Cython sanity check resolves the python3 pkg-config dependency to get Python include/link flags. The ROCm manylinux2_28-builder image ships the system Python 3.6 pkg-config file on the default search..."
189291,Add an accurate Metal erfc for MPS (eager + inductor),https://github.com/pytorch/pytorch/pull/189291,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,1,,statusCheckRollup,,,174929,189234,,,0,1,2,,,"triaged, open source, ciflow/mps, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-08T19:16:55Z,2026-07-09T01:53:01Z,0,"Splits the MPS backend portion out of #189234 (gh-187806) per review. Problem c10::metal::erfc is 1.0 - erf(x): once erf saturates in fp32 (x ~ 3.9), erfc returns 0.0, i.e. 100% relative error over the entire upper tail. Everything consuming it on MPS inherits the flush: eager torch.erfc / torch.special.erfc igammac..."
188862,[inductor] Fix bool minimum/maximum crash in index propagation (#188862) (#188862),https://github.com/pytorch/pytorch/pull/188862,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,0,12,,"statusCheckRollup, macos-py3-arm64 / test (default, 1, 3, macos-m1-stable), macos-py3-arm64 / test (openreg, 1, 1, macos-m1-stable), macos-py3-arm64 / test (default, 3, 3, macos-m1-stable), macos-py3-arm64 / test (default, 2, 3, macos-m1-stable), unit-test / inductor-test / test (inductor_cpp_wrapper, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test (inductor_cpp_wrapper, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), linux-jammy-rocm-py3.10-mi350 / test (inductor, 2, 2, linux.rocm.gpu.gfx950.1), linux-jammy-rocm-py3.10-mi350 / test (inductor, 1, 2, linux.rocm.gpu.gfx950.1), Import Status",,,188230,,,,3,1,1,0089bd2a6ca6aaf1554f5f5831f10f4cb41ad71662af351064abe5c439f30365,e542668cc9c3ba92e7fc970111ed9636aff3bc3ae5d58126be31654e71cc5cab,"module: cpu, ciflow/trunk, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-07-02T23:47:01Z,2026-07-09T01:53:01Z,0,"Summary: Context torch.minimum/torch.maximum on boolean tensors crash under torch.compile(backend=""inductor"") with ValueError: The argument 'False' is not comparable., while eager mode works. Reported in #188230 (repro: threshold -> torch.eq -> torch.empty_like -> torch.minimum). The root cause is that inductor feed..."
188643,[AOTI] Add external constants API for zero-copy weight sharing (#188643),https://github.com/pytorch/pytorch/pull/188643,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,0,2,,"statusCheckRollup, linux-jammy-cuda13.0-py3.10-gcc11 / test (distributed, 3, 3, lf-l-x86iavx512-45-172-t4-4)",,,,,,,0,1,2,7f6f50af9093301979ed636c8a993d35fbfea47142a6dc49a86a6fb62904e8a6,b6edc9e8a276f153953468b71a9cf2b4a2358a406381db35aa58afd8fdea4065,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-07-01T03:54:07Z,2026-07-09T01:52:59Z,0,"Summary: Add a new AOTI C API, AOTInductorModelContainerCreateWithExternalConstants, that creates a model container with caller-provided weight tensors instead of loading from the .so file. The caller retains ownership (user_managed=true), enabling zero-copy weight sharing via CUDA IPC. The new ExternalConstantsTag..."
189109,[profiler][cupti] Resolve a version-gated CUPTI header and conditionally install libclang for the field-id codegen,https://github.com/pytorch/pytorch/pull/189109,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,0,2,,"statusCheckRollup, linux-jammy-py3.10-clang18-asan / test (default, 6, 7, mt-l-x86iavx512-16-128)",,,,"188621, 189083, 189185",,,9,8,11,92f9e41beaebe8e14977e5688b03f180a8e3f0bf107fff4823205c738292d767,049fe5bc188aa845f8f3bbceb4b100ea5f729895fb03bec33e6f6fbd3dbcd93f,topic: not user facing,2026-07-06T23:30:20Z,2026-07-09T01:52:57Z,0,"Stack from ghstack (oldest at bottom): #189185 #189083 #188621 -> #189109 The CUPTI field-id codegen (tools/gen_cupti_stubs.py, stacked on top) parses the CUPTI ABI header cupti_activity.h with libclang to emit the _cupti_stubs field-id catalogs. This commit adds the header resolver + the conditional libclang build..."
181559,[XPU] Enable _flash_attention_forward/_flash_attention_backward,https://github.com/pytorch/pytorch/pull/181559,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,2442,2442,,,,,0,6,0,,,"open source, intel priority, ciflow/trunk, ciflow/inductor, ciflow/xpu, release notes: inductor (aoti)",2026-04-27T08:33:28Z,2026-07-09T01:52:50Z,0,Fix intel/torch-xpu-ops#2442 intel/torch-xpu-ops#2853 at::_flash_attention_forward and at::_flash_attention_backward are two ops which is one-one mapping to Tridao's FlashAttention2 API. at::_scaled_dot_product_flash_attention and at::_scaled_dot_product_flash_attention_backward are two ops of SDPA FlashAttention ba...
181726,[xpu][1/4]Implement scaled_mm_v2 for MXFP8/MXFP4/NVFP4 on XPU,https://github.com/pytorch/pytorch/pull/181726,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181727, 181728, 187315, 187318",,,0,10,4,,,"module: cpu, triaged, open source, intel priority",2026-04-28T07:26:31Z,2026-07-09T01:51:44Z,0,"Extend the XPU scaled_mm_v2 dispatch to support MXFP8 (BlockWise1x32 with e8m0fnu scales), MXFP4 (BlockWise1x32 with e8m0fnu scales on packed FP4 data), and NVFP4(BlockWise1x16 with float8_e4m3fn scales on packed FP4 data) Updated after blockwise FP8 support #173630 is merged. PR Stack: Since I don't have ghstack pe..."
189051,[dynamo] Reject keyword args to set()/frozenset() constructors,https://github.com/pytorch/pytorch/pull/189051,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"185424, 189021, 189022, 189052, 189053",,,0,3,5,,,"open source, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-06T17:56:18Z,2026-07-09T01:51:36Z,0,"Stack from ghstack (oldest at bottom): #185424 #189053 #189052 #189022 -> #189051 #189021 set(a=1) / set().init(a=1) silently returned an empty set under Dynamo instead of raising TypeError. BuiltinVariable.call_set / call_frozenset checked the keyword arguments only after a len(args) == 0 early-return, so the zero-..."
189317,[doc] Add accelerator test reuse guide (#181474),https://github.com/pytorch/pytorch/pull/189317,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,181474,181474,181474,,,,0,1,2,,,"open source, topic: not user facing",2026-07-08T23:25:07Z,2026-07-09T01:41:05Z,0,Adds a documentation page for the Accelerator Integration series explaining how out-of-tree backend developers can reuse PyTorch's built-in test cases. Covers: bypass_device_restrictions for @OnlyOn decorator bypass (#177248) op_skips / op_decorators for operator-level skipping (#177256) skipped_testcases for test c...
188632,Fix Dynamo opaque object staticmethod guards,https://github.com/pytorch/pytorch/pull/188632,ready for maintainer decision,112,"needs-review label present; approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,188544,188544,,,,0,2,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-07-01T01:45:33Z,2026-07-09T01:33:58Z,0,"Stack from ghstack (oldest at bottom): -> #188632 Value opaque objects are represented as TorchScriptObjectVariable, whose constructor accepted a source but did not pass it to UserDefinedObjectVariable. That left the base class without the instance source it needs to derive cls_source for descriptor lookup. When Dyn..."
189325,[Dynamo] Fix self-referential bw_compiler on backend re-entry,https://github.com/pytorch/pytorch/pull/189325,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-09T01:29:19Z,2026-07-09T01:32:21Z,0,"Why When one aot_autograd backend compiles a second, distinct graph, AotAutograd.__call__ re-enters with bw_compiler already wrapped. The _is_wrapped_bw_compiler branch then ran bw_compiler.compiler_fn = bw_compiler, grafting a self-referential attribute (and needing a type: ignore) instead of leaving the wrapper un..."
187404,[inductor][cutlass] Normalize reshaped external reads in EVT epilogue fusion,https://github.com/pytorch/pytorch/pull/187404,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,3,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-16T01:48:05Z,2026-07-09T01:25:28Z,0,"Stack from ghstack (oldest at bottom): -> #187404 Motivation Running the MobileBertForMaskedLM HuggingFace benchmark on XPU with CUTLASS epilogue fusion crashed during codegen: RuntimeError: Dimension mismatch between accum(1, 16384, 128), buf19(128, 128, 128) CUTLASS EVT treats the 2D GEMM output (M, N) = (16384, 1..."
179286,Use C++20 concepts where it improves readability,https://github.com/pytorch/pytorch/pull/179286,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,176662,,,,0,4,1,,,"open source, Merged, NNC, Reverted, ciflow/trunk, release notes: jit, topic: not user facing, ciflow/inductor, ci-no-td, no-runner-experiments, ciflow/vllm",2026-04-03T19:00:30Z,2026-07-09T01:05:38Z,0,"C++20 introduces concepts which can simplify some template code. In this PR, I have changed some instances of enable_if with concepts where it's not hard to verify correctness. #176662 cc @EikanWang @jgong5"
178393,[DO NOT MERGE] Dynamo 3.15 support,https://github.com/pytorch/pytorch/pull/178393,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"179094, 187631, 188029, 189306",,,0,37,4,,,"open source, release notes: releng, module: dynamo, ciflow/inductor, ciflow/dynamo, ciflow/torchtitan",2026-03-25T16:19:08Z,2026-07-09T01:05:02Z,0,Stack from ghstack (oldest at bottom): -> #178393 #179094 #188029 #187631 #189306
189249,[Dynamo] Remove dead code in tvm and common backends,https://github.com/pytorch/pytorch/pull/189249,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-08T09:07:55Z,2026-07-09T01:04:57Z,0,"Why Two branches in torch/_dynamo/backends can never execute. tvm() assigns options a default when it is None, then immediately re-checks it for None. AotAutograd.__call__ assigns use_fallback = False and then branches on it, so the fallback and its not_ok counter increment are dead. How Drop the redundant if option..."
188742,[ROCm][inductor] Enable scaled grouped GEMM Triton lowering,https://github.com/pytorch/pytorch/pull/188742,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,1,,,"module: rocm, triaged, open source, ciflow/trunk, release notes: rocm, release notes: linalg_frontend, module: inductor, ciflow/inductor, release notes: inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/rocm-mi355, ciflow/b200, ciflow/torchtitan, ciflow/rocm, ciflow/rocm-mi350",2026-07-01T22:11:37Z,2026-07-09T01:04:51Z,0,"Summary Build on the ROCm grouped GEMM Triton lowering change so legacy aten._scaled_grouped_mm can use the shared scaled grouped Triton template on ROCm. Gate ROCm scaled grouped GEMM Triton selection to FP8-capable hardware and the expected FP8 encoding (fnuz on gfx94, OCP FP8 on newer FP8-capable arches). Enable..."
189299,[MemoryViz] Decimate over-long stacked-area polygons so the summarize…,https://github.com/pytorch/pytorch/pull/189299,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,,2026-07-08T20:25:48Z,2026-07-09T01:04:20Z,0,"…d band renders The ""Active Memory Timeline"" / ""Allocated Memory (incl. Private Pools)"" views build one stacked-area per allocation series and advance a point onto each live series at every alloc/free action. A series that stays live across the whole trace therefore gains one point per timestep. The global summarize..."
189311,[wip] Reduce hackery for Dynamo appeasement,https://github.com/pytorch/pytorch/pull/189311,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-08T21:56:25Z,2026-07-09T01:04:10Z,0,Stack from ghstack (oldest at bottom): (to be filled) cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
186055,Add CUDA graph support for torch.while_loop,https://github.com/pytorch/pytorch/pull/186055,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186056,,,0,3,3,,,"open source, ciflow/trunk, release notes: cuda",2026-06-03T09:06:10Z,2026-07-09T01:04:09Z,0,"Stack from ghstack (oldest at bottom): #186056 -> #186055 Motivation is to allow us to handle some forms of dynamic shapes inside of a cuda graph. This does not necessarily improve performance, but it can make your entire workload cuda graph compatible, which is good for developer experience For example, if you want..."
189088,[SymmMem] Place the signal pad at the front of symmetric memory allocations,https://github.com/pytorch/pytorch/pull/189088,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200-symm-mem, ciflow/h100-symm-mem","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,0,2,,"statusCheckRollup, linux-jammy-cuda13.0-py3.12-gcc11-sm100-symm / test (b200-symm-mem, 1, 1, linux.dgx.b200.8)",,,,187465,,,26,7,5,bc1b16c9d47e2538058df44995cad9fd0f5296f3e10d53a218da8c0ef8647770,115decc56f0cc353bc13fa99a991d9624baab8cd2f061eb366426d301804cf79,"open source, release notes: distributed (c10d), release notes: distributed (symm_mem), ciflow/h100-symm-mem, ciflow/b200-symm-mem",2026-07-06T21:08:18Z,2026-07-09T01:04:08Z,0,"Stack from ghstack (oldest at bottom): #187465 -> #189088 We place the signal pad at the front of symmetric memory allocations in this PR. The purpose is to fix the potential signal pad pollution when we reuse the same VA for one block. For example, within one block if we first use that VA for a large tensor and the..."
187631,[dynamo] Fix segfault in list comprehension graph break,https://github.com/pytorch/pytorch/pull/187631,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"178393, 179094, 188029, 189306",,,0,12,9,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-18T01:29:31Z,2026-07-09T01:04:06Z,0,Stack from ghstack (oldest at bottom): #178393 #179094 #188029 -> #187631 #189306 Reusing the name of the iterator variable in a list comprehension that graph breaks to store the result causes a segfault (the stack underflows in the generated comprehension function). The current approach using bytecode templates is...
189305,"Reapply ""Remove CUDA sync in torch.combinations to improve `torch.compile` support (#186595)""",https://github.com/pytorch/pytorch/pull/189305,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,,,,,0,2,0,,,"open source, release notes: cuda, module: inductor, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-07-08T21:26:05Z,2026-07-09T01:04:00Z,0,Reapplies #186595 after Meta internal revert. Let me know if there are any failures that require fixing from my end. /cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo...
188029,[dynamo] Add namespace for cell variables that collide with locals,https://github.com/pytorch/pytorch/pull/188029,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"178393, 179094, 187631, 189306",,,0,9,4,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-24T05:39:54Z,2026-07-09T01:03:53Z,0,"Stack from ghstack (oldest at bottom): #178393 #179094 -> #188029 #187631 #189306 Cell variables that share a name with a local variable are stored separately in localsplus in cpython. This distinction is lost in dynamo, where all variables are stored in symbolic_locals. This doesn't matter for most purposes, but be..."
188521,Fix OrderingBarrier additional_buffer_deps key mismatch in control_deps lowering,https://github.com/pytorch/pytorch/pull/188521,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,188513,188513,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-30T02:55:14Z,2026-07-09T01:03:51Z,0,"Summary Fix a key mismatch in control_deps_op_lowering that caused OrderingBarrier's additional_buffer_deps to be silently dropped by the scheduler, breaking stream synchronization ordering. OrderingBarrier inherits from OperationBuffer, which has distinct buffer and operation names (assigned by register_buffer() an..."
183328,[special] Add arbitrary-order modified Bessel functions,https://github.com/pytorch/pytorch/pull/183328,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,76324,,,,0,8,1,,,"module: cpu, triaged, open source, release notes: cpp, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-11T16:27:19Z,2026-07-09T00:43:43Z,0,"Issue Partially addresses #76324. Summary Adds arbitrary-order modified Bessel functions: torch.special.modified_bessel_i(x, nu) torch.special.modified_bessel_k(x, nu) This follows the existing torch.special pointwise operator pattern with: ATen operator registration TensorIterator CPU/CUDA kernels CUDA Jiterator su..."
186927,[MPS] Gemv kernels,https://github.com/pytorch/pytorch/pull/186927,PR blocked,108,failed checks or blocked/unstable merge state; approved with no detected failing/pending checks; CI lane labels: ciflow/mps,final maintainer merge/release decision,APPROVED,BLOCKED,False,0,0,,,,,,,,,0,17,0,,,"triaged, open source, topic: improvements, release notes: mps, ciflow/mps",2026-06-10T11:17:16Z,2026-07-09T00:28:47Z,0,Gemv kernels. Draft till I put perf gains below on various machines and add tests. TODOs: Add tests Heuristics M2 base Heuristics M3 Ultra Heuristics M4 base Heuristics M5 Pro More heuristics (need help) Add perf benchmarks on M5 Pro Add perf benchmarks on M4 Add perf benchmarks on M3 Ultra Add perf benchmarks on M2...
179094,[DO NOT MERGE] Add docker image for python 3.15,https://github.com/pytorch/pytorch/pull/179094,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"178393, 187631, 188029, 189306",,,0,35,4,,,"open source, topic: not user facing, no-runner-experiments, ciflow/docker",2026-04-02T00:01:37Z,2026-07-09T00:21:49Z,0,Stack from ghstack (oldest at bottom): #178393 -> #179094 #188029 #187631 #189306 Authored with Claude.
186252,[inductor] Fix truncdiv off-by-one on CUDA with _div_rn,https://github.com/pytorch/pytorch/pull/186252,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,184408,184408,184408,,,,0,1,2,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-04T18:45:29Z,2026-07-09T00:21:46Z,0,"Fixes #184408 Triton's approximate fp32 division can produce results slightly below the true quotient, causing trunc(a / b) to be off by one when the quotient is near an integer. Use _div_rn (IEEE round-to-nearest) instead of plain div, matching the existing fix for the floor path. Change: trunc(div(a, b)) -> trunc(..."
189313,Support aten::cumprod ONNX export via CumProd (opset 26),https://github.com/pytorch/pytorch/pull/189313,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,102948,102948,102948,,,,0,2,1,,,"open source, release notes: onnx",2026-07-08T22:57:16Z,2026-07-09T00:21:13Z,0,"Fixing an Issue Before submitting, please review: The Ultimate Guide to PyTorch Contributions AI-Assisted Development policy Summary torch.cumprod could not be exported to ONNX (UnsupportedOperatorError: aten::cumprod). ONNX added the CumProd operator in opset 26, so this registers an implementation in the dynamo ex..."
189286,Fix DeviceContext mode leaks from checkpoint recompute and set_default_device restore,https://github.com/pytorch/pytorch/pull/189286,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,189133,189133,"188100, 188299, 188573",,,0,2,9,,,"release notes: autograd, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, mergedog, make tests great again",2026-07-08T18:12:08Z,2026-07-09T00:20:54Z,0,Stack from ghstack (oldest at bottom): -> #189286 #188573 #188299 #188100 Fixes #189133 This PR addresses flaky tests by fixing some DeviceContext mode leaks that may randomly affect other tests (e.g. the vmap test disabled by #189133). There are 3 issues addressed by this PR: TestAutograd and TestCompiledAutogradOp...
189306,"[dynamo, 3.15] Fix test_bytecode_utils",https://github.com/pytorch/pytorch/pull/189306,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"178393, 179094, 187631, 188029",,,0,2,4,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-07-08T21:26:36Z,2026-07-09T00:20:53Z,0,"Stack from ghstack (oldest at bottom): #178393 #179094 #188029 #187631 -> #189306 Two changes to update this for 3.15 co_lnotab has been deprecated since 3.10, and was removed in 3.15. It was replaced by co_lines, which we can use instead LOAD_COMMON_CONSTANT can load None in 3.15, so is now used for functions that..."
189297,[CI] Pin quack to torch's vendored SHA to fix B200 smoke cutlass import crash,https://github.com/pytorch/pytorch/pull/189297,draft/noise,13,"PR is draft; CI lane labels: ciflow/b200, ciflow/trunk",wait for author to mark ready,,CLEAN,True,0,0,,,,,188477,,,,0,1,1,,,"ciflow/trunk, topic: not user facing, ciflow/b200",2026-07-08T19:49:36Z,2026-07-09T00:20:52Z,0,"Stack from ghstack (oldest at bottom): -> #189297 The B200 smoke job's test_flex_flash::test_hierarchical_indexing_4d crashes at runtime with ModuleNotFoundError: No module named 'cutlass._mlir_helpers'. The compiled flex-attention CuTeDSL kernel imports flash_attn.cute, which pulls in quack, whose recent releases d..."
189319,Enable flake8-bugbear B007 (unused loop control variable),https://github.com/pytorch/pytorch/pull/189319,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/h100-symm-mem, ciflow/inductor, ciflow/inductor-pallas","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,106571,,,,0,1,1,,,"open source, release notes: quantization, release notes: releng, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), module: compiled autograd, ciflow/h100-symm-mem, ciflow/inductor-pallas, ciflow/torchtitan, ciflow/dtensor",2026-07-09T00:00:10Z,2026-07-09T00:09:57Z,0,"Summary Enables flake8-bugbear B007 (unused loop control variable), one of the remaining bugbear codes still suppressed in the ignore list of pyproject.toml. Part of #106571. This removes B007 from the ignore list and fixes all 169 violations across 110 files so the lint is enforced going forward. Approach Most fixe..."
188600,[ROCm][inductor] Support grouped GEMM Triton lowering,https://github.com/pytorch/pytorch/pull/188600,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,1,,,"module: rocm, triaged, open source, release notes: rocm, release notes: linalg_frontend, module: inductor, ciflow/inductor, release notes: inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-06-30T21:15:59Z,2026-07-08T23:51:01Z,0,Summary Allow Inductor's grouped GEMM Triton lowering on ROCm instead of gating it to NVIDIA SM90+ only. Match ROCm ATen extern output layout by using contiguous grouped GEMM output strides on HIP while preserving CUDA's TMA-aligned padded strides. Enable existing grouped GEMM compile tests on ROCm. Test plan python...
181781,[dynamo] turn canonicalize_output_graph_node_order on by default,https://github.com/pytorch/pytorch/pull/181781,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181780, 188175, 188176",,,0,39,3,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-04-28T19:51:27Z,2026-07-08T23:50:03Z,0,Stack from ghstack (oldest at bottom): -> #181781 #188176 #188175 #181780 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo @azahed98
188137,[inductor] Decompose shard_dim_alltoall via regular decomposition,https://github.com/pytorch/pytorch/pull/188137,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,2,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: distributed (dtensor), release notes: inductor, ciflow/torchtitan, ciflow/dtensor",2026-06-25T06:32:10Z,2026-07-08T23:49:43Z,0,"Summary: add an opt-in regular Inductor decomposition for _dtensor.shard_dim_alltoall into layout ops plus _c10d_functional.all_to_all_single/wait_tensor decompose only when the dims are valid, the dtype is supported, the process group size is resolvable, and the local shard dim is evenly divisible by the process gr..."
189312,[CUDA][cuBLAS] Change cuBLAS default workspace size for SM 11.0 to 32 MiB,https://github.com/pytorch/pytorch/pull/189312,PR blocked,108,failed checks or blocked/unstable merge state; approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"module: cuda, module: cublas, open source, topic: not user facing",2026-07-08T22:11:09Z,2026-07-08T23:49:19Z,0,"Thor is closest to datacenter blackwell, and current defaults cause CUBLAS_STATUS_NOT_SUPPORTED in e.g., python test/test_linalg.py TestLinalgCUDA.test_matmul_mv_cuda_float32 I assume there are basically internal users of SM 11.0 that would complain? CC @ngimel authored with codex cc @ptrblck @msaroufim @tinglvv @nW..."
189301,[ROCm] Bump AOTriton to 0.13b,https://github.com/pytorch/pytorch/pull/189301,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/rocm-mi300,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,,,,,,,0,1,0,,,"module: rocm, open source, topic: not user facing, ciflow/rocm-mi300",2026-07-08T20:36:00Z,2026-07-08T23:49:17Z,0,"Notable new features compared with 0.12.1b: Debug printing can be enabled with env var AOTRITON_DEBUG_LEVEL, without re-compiling We use the same logging level as of AMD_LOG_LEVEL CAVEAT: AOTriton itself does not read AMD_LOG_LEVEL Set AOTRITON_DEBUG_LEVEL=5 to print everything BREAKING .aks2 files are packed into u..."
189022,[dynamo] Support itertools.count keyword args and repr,https://github.com/pytorch/pytorch/pull/189022,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"185424, 189021, 189051, 189052, 189053",,,0,3,5,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-06T13:39:55Z,2026-07-08T23:49:13Z,0,"Stack from ghstack (oldest at bottom): #185424 #189053 #189052 -> #189022 #189051 #189021 Two itertools.count object-protocol gaps, both mirroring CPython Modules/itertoolsmodule.c: Construction: the count branch had a not kwargs guard, so the keyword forms count(start=..)/count(step=..) fell through to a graph brea..."
188006,[ROCm] Fix torch.pdist correctness on ROCm; re-enable test_pdist_norm_large,https://github.com/pytorch/pytorch/pull/188006,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,0,0,,,168868,168868,168868,,,,0,4,3,,,"module: rocm, triaged, open source, release notes: rocm",2026-06-24T00:31:41Z,2026-07-08T23:43:22Z,0,"Fixes #168868 Investigation summary Test environment PyTorch version: 2.13.0a0+git75d18bb Hip version: 7.2.53211 GPU 0 name: AMD Instinct MI355X Command to reproduce $ PYTORCH_TEST_WITH_ROCM=1 pytest test/test_torch.py -k ""test_pdist_norm_large"" -v Default behaviour test/test_torch.py::TestTorchDeviceTypeCPU::test_p..."
189092,[dynamo] reduce call_obj_hasattr overrides,https://github.com/pytorch/pytorch/pull/189092,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"187468, 187469, 187531, 187532, 187707, 189091",,,0,3,7,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-06T21:22:16Z,2026-07-08T23:23:53Z,0,Stack from ghstack (oldest at bottom): -> #189092 #189091 #187707 #187532 #187531 #187469 #187468 Fix object_generic_getattr step 7 to raise ObservedAttributeError instead of _UnhandledDescriptorError when an attribute genuinely doesn't exist after exhaustive MRO search and getattr fallback. This is semantically cor...
187707,[dynamo] remove GetAttrVariable,https://github.com/pytorch/pytorch/pull/187707,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"187468, 187469, 187531, 187532, 189091, 189092",,,0,18,7,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-19T02:19:28Z,2026-07-08T23:23:45Z,0,"Stack from ghstack (oldest at bottom): #189092 #189091 -> #187707 #187532 #187531 #187469 #187468 GetAttrVariable was a catch-all fallback that deferred attribute access by wrapping it in a proxy. It made it impossible to distinguish ""attribute not found"" from ""attribute found but unhandled,"" masking real errors (e...."
189091,[dynamo] guard getattr() builtin constant-fold fallback,https://github.com/pytorch/pytorch/pull/189091,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"187468, 187469, 187531, 187532, 187707, 189092",,,0,3,7,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-06T21:22:10Z,2026-07-08T23:23:44Z,0,"Stack from ghstack (oldest at bottom): #189092 -> #189091 #187707 #187532 #187531 #187469 #187468 GetAttrBuiltinVariable.call_function has a constant-fold fallback that fires when getattro_impl raises Unsupported and all args are python constants. It was calling VariableTracker.build(tx, result) without a source, so..."
188978,"swap isinstance(t, FakeTensor) to is_fake",https://github.com/pytorch/pytorch/pull/188978,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,188694,,188979,,,0,6,1,,,"ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, release notes: export, release notes: inductor (aoti), ciflow/torchtitan, ciflow/docker",2026-07-05T22:10:54Z,2026-07-08T23:23:41Z,0,"closed #188694 bc need to stack smth on top copying pr summary over... in preparation for C++ FakeTensor migration, i am landing this change first to change isinstance(t, FakeTensor) to is_fake_tensor and adding a check for C++ FakeTensor to the is_fake_tensor function. but since C++ faketensor is not being used yet..."
187531,[dynamo] enable generic attribute mutation on non-UDOV VTs,https://github.com/pytorch/pytorch/pull/187531,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"187468, 187469, 187532, 187707, 189091, 189092",,,0,19,7,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-17T00:31:31Z,2026-07-08T23:23:35Z,0,Stack from ghstack (oldest at bottom): #189092 #189091 #187707 #187532 -> #187531 #187469 #187468 Adds a get_value_for_setattr() opt-in hook on the base VariableTracker that enables attribute mutation (setattr/getattr/hasattr) for VTs beyond UserDefinedObjectVariable. The mechanism: VTs override get_value_for_setatt...
187532,[dynamo] fix __getattr__ and __getattribute__ dunder semantics,https://github.com/pytorch/pytorch/pull/187532,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"187468, 187469, 187531, 187707, 189091, 189092",,,0,18,7,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-17T00:31:37Z,2026-07-08T23:23:32Z,0,"Stack from ghstack (oldest at bottom): #189092 #189091 #187707 -> #187532 #187531 #187469 #187468 Previously, both obj.__getattr__(""x"") and obj.__getattribute__(""x"") in call_method dispatched through getattro_impl, which does full GenericGetAttr including the __getattr__ fallback. This was semantically wrong for bot..."
187468,[dynamo] add break_graph_if_unsupported to LOAD_ATTR,https://github.com/pytorch/pytorch/pull/187468,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"187469, 187531, 187532, 187707, 189091, 189092",,,0,20,7,,,"topic: not user facing, module: dynamo, ciflow/inductor, keep-going, ciflow/dynamo, ciflow/torchtitan",2026-06-16T17:59:46Z,2026-07-08T23:23:28Z,0,"Stack from ghstack (oldest at bottom): #189092 #189091 #187707 #187532 #187531 #187469 -> #187468 LOAD_ATTR previously relied on the step() fallback for graph breaks, which restarts from the most recent empty-stack checkpoint. This is less precise than other opcodes like STORE_ATTR and CALL that use break_graph_if_u..."
187469,[dynamo] extract UDOV hooks: lookup_instance_dict and call_getattr_fallback,https://github.com/pytorch/pytorch/pull/187469,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"187468, 187531, 187532, 187707, 189091, 189092",,,0,19,7,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-16T17:59:54Z,2026-07-08T23:23:14Z,0,"Stack from ghstack (oldest at bottom): #189092 #189091 #187707 #187532 #187531 -> #187469 #187468 Extract two chunks of UDOV's generic_getattr into hook overrides, matching the hook interface on base VariableTracker: lookup_instance_dict (step 3): instance dict lookup including side-effects awareness (INSTANCE_DICT..."
181780,[dynamo] set canonicalize_output_graph_node_order=True for export tests,https://github.com/pytorch/pytorch/pull/181780,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181781, 188175, 188176",,,0,39,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, keep-going, ciflow/torchtitan, ciflow/dtensor",2026-04-28T19:51:22Z,2026-07-08T23:22:50Z,0,Stack from ghstack (oldest at bottom): #181781 #188176 #188175 -> #181780 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
188175,[dynamo] fix autograd cache and split_module tests under canonicalization,https://github.com/pytorch/pytorch/pull/188175,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"181780, 181781, 188176",,,0,14,3,,,"module: inductor, module: dynamo, ciflow/inductor",2026-06-25T18:58:02Z,2026-07-08T23:22:38Z,0,Stack from ghstack (oldest at bottom): #181781 #188176 -> #188175 #181780 Two tests opted out of canonicalize_output_graph_node_order because they were sensitive to placeholder ordering. This PR fixes both so the opt-outs can be removed. test_multi_graph_specialization failed because canonicalization reorders placeh...
189292,Fix dead ExecuTorch XNNPACK doc link in mobile_optimizer,https://github.com/pytorch/pytorch/pull/189292,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, topic: not user facing",2026-07-08T19:19:11Z,2026-07-08T23:22:32Z,0,Stack from ghstack (oldest at bottom): -> #189292 The nightly Link checks / lint-urls / lint job fails on a genuine 404: the XNNPACK backend doc referenced from mobile_optimizer.md was renamed by ExecuTorch from native-delegates-executorch-xnnpack-delegate.html (now 404) to backends-xnnpack.html (200). Point the lin...
188176,[dynamo] fix test_perf.py tests under canonicalization,https://github.com/pytorch/pytorch/pull/188176,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"181780, 181781, 188175",,,0,14,3,,,"topic: not user facing, module: inductor, ciflow/inductor",2026-06-25T18:58:08Z,2026-07-08T23:22:21Z,0,Stack from ghstack (oldest at bottom): #181781 -> #188176 #188175 #181780 Two tests in test_perf.py opted out of canonicalize_output_graph_node_order unnecessarily. This removes both opt-outs. test_cat_pointwise passes deterministically with canonicalization -- the numel counts are unchanged because canonicalization...
189190,flex_gemm: extend feed-main to trailing pointwise and axis-1 fragments,https://github.com/pytorch/pytorch/pull/189190,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188112, 188468, 188469, 188470, 188739, 189188, 189314, 189315, 189316",,,0,3,9,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T20:08:41Z,2026-07-08T23:12:53Z,0,"Stack from ghstack (oldest at bottom): #189316 #189315 #189314 -> #189190 #189188 #188739 #188112 #188470 #188469 #188468 The feed-main matcher now recurses through trailing shape-preserving pointwise nodes after the un-grouping view, so quant casts like .to(float8_e4m3fn) compose with physical axis-0 feeds. Axis-1..."
189188,flex_gemm: support scalar captured epilogue args,https://github.com/pytorch/pytorch/pull/189188,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"188112, 188468, 188469, 188470, 188739, 189190, 189314, 189315, 189316",,,0,1,19,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T19:24:45Z,2026-07-08T23:12:52Z,0,"Stack from ghstack (oldest at bottom): #189316 #189315 #189314 #189190 -> #189188 #188739 #188112 #188470 #188469 #188468 Add a fourth captured-epilogue-arg kind ""scalar"" beside tile/row/col for [1, 1] read-only tensors. This unlocks native float8 tensorwise quant epilogues, e.g. (acc * s).to(torch.float8_e4m3fn) wi..."
188739,flex_gemm: add quant lowerings,https://github.com/pytorch/pytorch/pull/188739,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,"188112, 188468, 188469, 188470, 189188, 189190, 189314, 189315, 189316",,,0,10,9,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-01T21:48:41Z,2026-07-08T23:12:51Z,0,Stack from ghstack (oldest at bottom): #189316 #189315 #189314 #189190 #189188 -> #188739 #188112 #188470 #188469 #188468
188112,flex_gemm: feed local reductions into epilogues,https://github.com/pytorch/pytorch/pull/188112,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,"188468, 188469, 188470, 188739, 189188, 189190, 189314, 189315, 189316",,,0,32,9,,,"topic: not user facing, skip-pr-sanity-checks, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-24T23:35:26Z,2026-07-08T23:12:51Z,0,"Stack from ghstack (oldest at bottom): #189316 #189315 #189314 #189190 #189188 #188739 -> #188112 #188470 #188469 #188468 Let dense aten.mm FlexGEMM epilogues feed a grouped local reduction back into the main output expression (and optionally store it), e.g. acc / acc.view(-1, group, N).sum(1) or centered-mean/varia..."
188470,flex_gemm: extend local reduce aux groups,https://github.com/pytorch/pytorch/pull/188470,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,"188112, 188468, 188469, 188739, 189188, 189190, 189314, 189315, 189316",,,0,20,9,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-29T19:22:06Z,2026-07-08T23:12:50Z,0,Stack from ghstack (oldest at bottom): #189316 #189315 #189314 #189190 #189188 #188739 #188112 -> #188470 #188469 #188468 Extend compressed local-reduce aux outputs beyond one 32-lane TensorSSA fragment: axis-1 groups above the fragment width (CTA-subtile groups such as 64/128) and axis-0 M-groups now lower through...
188469,flex_gemm: support compressed local reduce aux outputs,https://github.com/pytorch/pytorch/pull/188469,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"188112, 188468, 188470, 188739, 189188, 189190, 189314, 189315, 189316",,,0,19,9,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-29T19:22:02Z,2026-07-08T23:12:49Z,0,"Stack from ghstack (oldest at bottom): #189316 #189315 #189314 #189190 #189188 #188739 #188112 #188470 -> #188469 #188468 Support dense non-batched aten.mm FlexGEMM epilogues that return one compressed local-reduce aux output, e.g. acc.relu(), acc.view(M, -1, group).amax(-1): lowering classifies grouped TensorSSA re..."
188468,flex_gemm: add local reduce analysis helpers,https://github.com/pytorch/pytorch/pull/188468,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"188112, 188469, 188470, 188739, 189188, 189190, 189314, 189315, 189316",,,0,12,19,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-29T19:21:59Z,2026-07-08T23:12:48Z,0,"Stack from ghstack (oldest at bottom): #189316 #189315 #189314 #189190 #189188 #188739 #188112 #188470 #188469 -> #188468 Standalone analysis vocabulary for FlexGEMM local reductions, consumed by the compressed local-reduce aux-output support stacked on top: torch/_inductor/codegen/cutedsl/cutedsl_op_overrides.py: T..."
180826,"Remove deprecated `set_group_info`, `get_group_info` and `enable_symm_mem_for_group`",https://github.com/pytorch/pytorch/pull/180826,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/b200-symm-mem, ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, Stale, module: inductor, ciflow/inductor, release notes: distributed (symm_mem), ciflow/h100-symm-mem, ciflow/b200-symm-mem, module: symm_mem, ciflow/torchtitan",2026-04-20T09:01:57Z,2026-07-08T22:50:12Z,0,"The methods set_group_info, get_group_info and enable_symm_mem_for_group have been deprecated (#177700 (comment)). This PR removes these methods and calls to them. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjame..."
186358,Fix output mismtach for F.softshrink with bfloat16 and float scalar,https://github.com/pytorch/pytorch/pull/186358,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185484,185484,185484,,,,0,2,2,,,"module: cpu, triaged, open source, ciflow/trunk, release notes: nn, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-05T10:36:03Z,2026-07-08T22:48:00Z,0,"Fixes #185484 The softshrink decomposition and CPU eager kernel did not cast the scalar lambd to the input tensor's dtype before arithmetic. For reduced floating types like bfloat16, this caused lambd to retain float32 precision, producing results inconsistent with CUDA eager which correctly casts lambd to the tenso..."
185359,Add CE multicast low-contention all-gather,https://github.com/pytorch/pytorch/pull/185359,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"triaged, open source, ciflow/trunk, release notes: distributed (c10d), module: inductor, ciflow/inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-05-27T12:29:13Z,2026-07-08T22:47:54Z,0,"Summary Adds a copy-engine multicast low-contention all-gather implementation for CUDA symmetric memory and wires it through eager APIs, Inductor lowering, and the low-contention collective rewrite pass. The new variant is exposed as: torch.ops.symm_mem._low_contention_all_gather_ce_multicast torch.ops.symm_mem._low..."
187458,[MPS] Fuse cross_entropy into a native Metal kernel,https://github.com/pytorch/pytorch/pull/187458,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,187457,,,0,4,1,,,"triaged, open source, release notes: mps, ciflow/mps, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-16T16:48:24Z,2026-07-08T22:47:52Z,0,"[MPS] Fuse cross_entropy into a native Metal kernel Reviewing just this PR. Stacked leaf (cumulative-leaf fork on main), so the GitHub diff shows the whole softmax stack + this one. To see only what this PR adds on top of #187457, use this fork compare (renders as a normal diff of just this delta): anagnorisis2perip..."
186809,[cudaMallocAsync] Count graph-mem pool in memory_reserved(),https://github.com/pytorch/pytorch/pull/186809,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, release notes: cuda, module: CUDACachingAllocator",2026-06-09T17:33:12Z,2026-07-08T22:47:34Z,0,"Under PYTORCH_CUDA_ALLOC_CONF=backend:cudaMallocAsync, CUDA-graph captures reserve their backing in the device's graph-memory pool, which getDeviceStats never queried — so memory_reserved() / memory_stats() silently undercounted every graph-capturing workload. This adds the cudaDeviceGetGraphMemAttribute (current +..."
189295,Pin flash-attention checkout in FA3 ABI-stable H100 smoke test,https://github.com/pytorch/pytorch/pull/189295,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,189294,,,0,1,1,,,"topic: not user facing, ciflow/h100",2026-07-08T19:30:42Z,2026-07-08T22:36:47Z,0,Stack from ghstack (oldest at bottom): -> #189295 The Limited CI on H100 / ...sm90-FA3-ABI-stable-test / test job has been failing every run since 2026-07-03 at pytest collection with: ModuleNotFoundError: No module named 'flash_attn_3._C' Root cause: this is not a PyTorch regression but an unpinned external depende...
186965,[RFC][dynamo] Add ContextVar.get() tracing support (Phase 1),https://github.com/pytorch/pytorch/pull/186965,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,186449,,,,0,1,2,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-06-10T20:05:50Z,2026-07-08T22:27:06Z,0,"Adds read-only contextvars.ContextVar.get() support inside torch.compile. Values are resolved at trace time with a lambda_manager guard that re-checks at cache time, triggering recompilation on value change. .set() and .reset() graph-break with a SUPPORTABLE hint. This is a draft PR for the Phase - 1 implementation..."
189257,[rocm] Re-enable conv2d backward parametrized test on ROCm by using fp16 on NAVI,https://github.com/pytorch/pytorch/pull/189257,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-navi31, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355",2026-07-08T13:36:15Z,2026-07-08T22:19:46Z,0,"Summary This PR re-enables test_conv2d_backward_parametrized on ROCm by replacing the broad ROCm skip with a targeted NAVI path: Remove the skipIfRocmArch(...) decorator for this test. On NAVI (isRocmArchAnyOf(NAVI_ARCH)), run the test inputs in fp16. Use NAVI-specific tolerances and reference_in_float=False for the..."
189080,[inductor] Generalize reinplacing to functional ATen ops with an in-place variant,https://github.com/pytorch/pytorch/pull/189080,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: performance, topic: performance, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-07-06T19:49:11Z,2026-07-08T22:03:33Z,0,"Stack from ghstack (oldest at bottom): -> #189080 Broader fix for #188495 Under torch.compile, functionalization rewrites in-place ops (e.g. uniform_) into functional ones that clone self before writing. For write-only ops, that clone is wasted and survives as an Inductor fallback, doubling peak memory. Inductor's r..."
189310,Generalize GeneratorState input device type,https://github.com/pytorch/pytorch/pull/189310,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,3902,3902,,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-08T21:47:03Z,2026-07-08T22:01:33Z,0,Accept xpu device type. This is to fix intel/torch-xpu-ops#3902. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
180020,Add a debug tool to warn on NULL stream use,https://github.com/pytorch/pytorch/pull/180020,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,167885,,,,0,2,1,,,"triaged, open source",2026-04-10T21:30:17Z,2026-07-08T21:58:38Z,0,#167885 Implements a context manager that warns when any work is enqueued on the NULL stream. This can help detect unintended use of the NULL stream and also debug CUDA graph capture errors failing because of the NULL stream dependency. It is pretty straight-forward to also detect use of non-capture stream under CUD...
187450,[ROCm] periodic-rocm-mi350: run only the 4-GPU distributed tests,https://github.com/pytorch/pytorch/pull/187450,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/periodic-rocm-mi350, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,14,0,,,"module: rocm, triaged, open source, ciflow/trunk, topic: not user facing, keep-going, ci-no-td, ciflow/periodic-rocm-mi350",2026-06-16T16:07:01Z,2026-07-08T21:52:15Z,0,"Summary Reenable the periodic-rocm-mi350 workflow on the linux.rocm.gpu.gfx950.4 (4-GPU) runner labels, using a dedicated distributed_4gpu test config (now 5 shards) so the scarce 4-GPU runners run only the distributed tests that need >= 4 GPUs. Register ciflow/periodic-rocm-mi350 in ciflow_push_tags in .github/pyto..."
189129,inductor: Fix non-deterministic gradients in slice_scatter backward fusion,https://github.com/pytorch/pytorch/pull/189129,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188890,188890,188890,,,,0,4,2,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T05:31:18Z,2026-07-08T21:51:00Z,0,"Fixes #188890 Description When torch.use_deterministic_algorithms(True) is enabled, the AOTAutograd decomposition of repeat_interleave combined with slicing generates slice_scatter followed by sum. Inductor's scheduler aggressively fuses these, leading to complex strided layouts where the overlap analysis fails. Thi..."
186790,Remove c10/util/Array.h,https://github.com/pytorch/pytorch/pull/186790,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,2,,,"module: cpu, triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td, ciflow/vllm",2026-06-09T14:24:33Z,2026-07-08T21:50:46Z,0,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
189277,Skip redundant contiguity rescan in empty_tensor_restride (robust) (#189277),https://github.com/pytorch/pytorch/pull/189277,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, topic: not user facing, meta-exported",2026-07-08T17:02:58Z,2026-07-08T21:50:21Z,0,"Summary: TensorImpl::empty_tensor_restride(MemoryFormat::Contiguous) writes canonical row-major strides and then calls refresh_contiguous(), which re-scans sizes/strides via the out-of-line compute_contiguous() to recompute the six contiguity flags. Once row-major strides have just been written, that scan is redunda..."
175900,[Inductor] Fix Enum serialization in user-defined Triton kernel codegen (#175900),https://github.com/pytorch/pytorch/pull/175900,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, fb-exported, ciflow/trunk, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-02-26T22:11:10Z,2026-07-08T21:50:16Z,0,"Summary: User-defined Triton kernels can have triton_meta dicts containing Enum values (e.g., RoundingMode.even from fbgemm). When these are serialized with repr(), they produce invalid Python syntax like <RoundingMode.even: 2>. This adds a _sanitize_for_repr helper that recursively converts Enum values to their und..."
189290,Harden Inductor SubprocPool against compile-worker death and teardown hangs,https://github.com/pytorch/pytorch/pull/189290,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-08T18:57:17Z,2026-07-08T21:49:50Z,0,"Stack from ghstack (oldest at bottom): -> #189290 Root cause: in fork mode the compile-worker sidecar's ProcessPoolExecutor workers are created by a bare os.fork() with no exec, so each worker inherits the sidecar's entire fd table -- including the sidecar<->parent result pipe. The parent also never closed its own c..."
185617,[Profiler] Refactor test_trace_validator.py,https://github.com/pytorch/pytorch/pull/185617,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,2,1,,,"triaged, open source, topic: not user facing, module: tests/core",2026-05-29T09:30:02Z,2026-07-08T21:49:37Z,0,"Summary Refactors test/profiler/test_trace_validator.py to break the tight coupling between test cases and CUDA, enabling new hardware backends to reuse the test suite. Tests are categorized into three classes: TestTraceValidatorRules (accelerator-unrelated): 15 synthetic unit tests that validate all 6 trace validat..."
189256,[BE][MPS] Factor host-buffer wrapping into buffer_with_offset_from_tensor,https://github.com/pytorch/pytorch/pull/189256,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/mps,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,181720,,,0,1,1,,,"open source, topic: not user facing, release notes: mps, ciflow/mps",2026-07-08T13:29:03Z,2026-07-08T21:46:11Z,0,"Stack from ghstack (oldest at bottom): #181720 -> #189256 copy_from_mps_ and copy_to_mps_stride_contig both wrapped the host side of a CPU<->MPS copy identically: page-align the storage pointer, wrap it in a newBufferWithBytesNoCopy MTLBuffer, and (for non-blocking Jcopies) retain the storage in the buffer's dealloc..."
188901,Implement Double Backward for `cdist` and `pdist`,https://github.com/pytorch/pytorch/pull/188901,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,14,0,,,"triaged, open source, ciflow/inductor, release notes: inductor (aoti)",2026-07-03T15:14:16Z,2026-07-08T21:37:19Z,0,"Adds second-order backward for torch.cdist and F.pdist, which previously only supported first-order gradients. All create_graph=True calls use raised ""the derivative for '_cdist_backward' is not implemented"". The enabling change renames the first argument of _cdist_backward/_pdist_backward from grad to grad_output,..."
189155,Add CUDA Support for Fused `_add_relu` With Half and BFloat16,https://github.com/pytorch/pytorch/pull/189155,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,16,0,,,"module: cpu, triaged, open source, release notes: nn",2026-07-07T14:50:20Z,2026-07-08T21:37:10Z,0,"_add_relu (the fused relu(self + alpha*other) residual primitive) was CPU-only. On CUDA the pattern ran as separate add then relu, costing an extra kernel launch and a full round-trip of the intermediate. This adds a CUDA dispatch that reuses the existing device-generic front-end, and extends both backends from the..."
189220,Strip the arc experiment from workflow check_experiments,https://github.com/pytorch/pytorch/pull/189220,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor-pallas,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"topic: not user facing, ciflow/inductor-pallas",2026-07-08T00:54:25Z,2026-07-08T21:36:08Z,0,"Stack from ghstack (oldest at bottom): -> #189220 #189219 Follow-up to removing arc from the runner determinator: drop the now-inert arc token from every workflow's check_experiments (arc,lf -> lf, etc.; the line is removed where arc was the only value). No routing change -- the determinator already treats arc as ig..."
189230,Support multiple pools in cudagraph_trees,https://github.com/pytorch/pytorch/pull/189230,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,"185057, 188755, 189123, 189173, 189175",,,0,20,5,,,"open source, release notes: distributed (c10d), module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-08T03:34:12Z,2026-07-08T21:35:35Z,0,"This is a WIP approach to supporting multiple MemPools in cudagraph_trees and thus mode=""reduce-overhead"". It's built on top of #185057. Putting this up in draft to get some eyes on it, but it should get rebased and will be easier to review once #185057, #188755, #189123, #189173, and #189175 go in. cc @voznesenskym..."
189304,Use freeze_rng_state in distribution rsample gradient tests,https://github.com/pytorch/pytorch/pull/189304,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-07-08T21:19:58Z,2026-07-08T21:31:26Z,0,Fixes: intel/torch-xpu-ops#4020 This PR replaces manual torch.get_rng_state() / torch.set_rng_state() patterns with the freeze_rng_state() context manager in distribution regression tests. freeze_rng_state() is already available in torch.testing._internal.common_utils and properly handles CUDA/XPU backends. Using it...
189274,[dynamo] Route set/dict-view/list dunder methods through tp_slots,https://github.com/pytorch/pytorch/pull/189274,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-08T16:11:17Z,2026-07-08T21:27:00Z,0,"Stack from ghstack (oldest at bottom): -> #189274 Several SetVariable, dict-view, and list VariableTrackers reimplemented the number/sequence-protocol dunders (and, or, xor, sub and their reflected/in-place variants; add/iadd) directly inside call_method. VariableTracker.call_method already dispatches these dunders..."
188802,[MPS] Migrate conv3d to metal kernels,https://github.com/pytorch/pytorch/pull/188802,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,7,0,,,"module: cpu, triaged, open source, topic: improvements, release notes: mps, ciflow/mps",2026-07-02T13:35:27Z,2026-07-08T21:26:44Z,0,"Migrate conv3d forward from MPSGraph to Metal kernels (MPP on macOS 26+, simdgroup otherwise). Adds: int64-indexed shapes, previously unsupported Metal kernels for conv3d which are generally faster than MPS Graph ones. Perf benchs incoming below: bfloat16 config N MPSGraph ms Metal ms TFLOPS speedup bottleneck_1x1x1..."
189202,[CuteDSL][RMSNorm] use const_data_ptr for alignment check,https://github.com/pytorch/pytorch/pull/189202,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: cuda, open source, module: norms and normalization, ciflow/trunk, topic: not user facing, module: dsl-native-ops",2026-07-07T21:18:05Z,2026-07-08T21:23:13Z,0,"otherwise fails COW test checks when a mutable data_ptr is used e.g., AssertionError: False is not true : Argument 0 during forward call unexpectedly materializes. Either set `supports_cow_input_no_materialize_forward=False` in this operation's OpInfo, add the arg to the OpInfo's `allow_cow_input_materialize_forward..."
186082,Improve SymIntArrayRef concrete-int diagnostics,https://github.com/pytorch/pytorch/pull/186082,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,140960,140960,,,,0,10,1,,,"ciflow/inductor, release notes: dynamo, agentic",2026-06-03T13:56:56Z,2026-07-08T21:23:11Z,0,"Stack from ghstack (oldest at bottom): -> #186082 SymIntArrayRef to IntArrayRef conversion fails in asIntArrayRefSlow when any SymInt is heap allocated, but the previous error only said that concrete integers were expected. For symbolic sizes this hid the actual bad element and gave no guidance about the common dyna..."
189294,[CI] Realign scipy after nvmath install to fix H100 smoke NumPy ABI crash,https://github.com/pytorch/pytorch/pull/189294,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/h100,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,189034,189034,,,,0,2,2,,,"topic: not user facing, ciflow/h100",2026-07-08T19:23:55Z,2026-07-08T21:23:06Z,0,"Stack from ghstack (oldest at bottom): -> #189294 install_nvmath does an unpinned pip install nvmath-python, which pulls numpy>=2 into the H100 smoke venv, overriding the CI image's numpy 1.x pin. The image's scipy (1.10.1) is built against numpy 1.x, so once numpy is 2.x any scipy import raises a binary-incompatibi..."
189288,[inductor] Fix compute_ancestors KeyError from self-referential mutation copy,https://github.com/pytorch/pytorch/pull/189288,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,CLEAN,False,0,0,,,,189271,189271,,,,0,1,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-08T18:23:03Z,2026-07-08T21:22:14Z,0,"Stack from ghstack (oldest at bottom): -> #189288 Fixes #189271. Root cause: GraphLowering.propagate_mutation emits copy_(old_arg, new_arg) to reflect an in-place op's mutation back onto its original argument, guarding the ""no copy needed"" case with Python object identity (old_arg is new_arg). When an op mutates a b..."
189187,Retune against triton 3.8,https://github.com/pytorch/pytorch/pull/189187,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,4,1,,,"module: performance, topic: not user facing, module: inductor, ciflow/inductor, module: flex attention, ciflow/torchtitan",2026-07-07T19:19:30Z,2026-07-08T21:22:11Z,0,Stack from ghstack (oldest at bottom): -> #189187
189021,[dynamo] Support empty tuple-subclass construction (MyTuple()),https://github.com/pytorch/pytorch/pull/189021,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"185424, 189022, 189051, 189052, 189053",,,0,2,12,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-06T13:39:49Z,2026-07-08T21:21:53Z,0,Stack from ghstack (oldest at bottom): #185424 #189053 #189052 #189022 #189051 -> #189021 Constructing a tuple subclass with no arguments -- e.g. class MyTuple(tuple): pass; MyTuple() -- crashed under Dynamo with AssertionError: init_args must be provided when tuple_vt is None in UserDefinedTupleVariable.init. tuple...
189199,[DTensor] Fix spurious cross-device FakeTensor mismatch in multi-threaded tests,https://github.com/pytorch/pytorch/pull/189199,ready for maintainer decision,45,"CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-07-07T20:49:51Z,2026-07-08T21:18:46Z,0,"Problem Multi-threaded DTensor tests (those built on FSDPTestMultiThread) can fail with a spurious device-propagation error such as: RuntimeError: Unhandled FakeTensor Device Propagation for aten.where.self, found two different devices xpu:0, xpu:2 Fix common_fsdp.py: FSDPTestMultiThread.setUp swaps ShardingPropagat..."
188931,Fix matrix-vector linear fallback crashes with 1D weight and bias,https://github.com/pytorch/pytorch/pull/188931,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188891,188891,188891,,,,0,1,2,,,"triaged, open source, release notes: nn",2026-07-03T20:55:05Z,2026-07-08T21:16:55Z,0,"When F.linear (or linear_out) is called with a 2D input, 1D weight, and a defined bias, PyTorch incorrectly routes the execution to fused addmm paths. Since addmm and addmm_out expect a 2D weight matrix as their second input, invoking them with a 1D weight tensor results in a crash (e.g., RuntimeError: mat2 must be..."
189303,"Reapply ""Dispatch index_put_ to fill_ for scalar values (#188267 / #189105)""",https://github.com/pytorch/pytorch/pull/189303,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"open source, release notes: cuda",2026-07-08T21:04:52Z,2026-07-08T21:16:25Z,0,"This reapplies #188267 / #189105 which has been reverted due to Meta internal CI failures. I also addressed #189105 (comment) I will not merge it myself this time but leave it to @ngimel, @Skylion007, @izaitsevfb, or @atalman so the Meta internal CI failures can be dealt with."
188991,[NVGEMM] Support `cutlass.operators`,https://github.com/pytorch/pytorch/pull/188991,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-06T03:05:16Z,2026-07-08T21:16:04Z,0,This PR adds an adapter so NVGEMM prefers cutlass.operators and falls back to cutlass_api for the sake of backward compatibility. Authored with Claude Opus 4.8 through Cursor. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchu...
188913,[inductor] Add triton.cudagraph_max_recorded_sizes to cap dynamic shape re-recording,https://github.com/pytorch/pytorch/pull/188913,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188150,188150,188150,,,,0,2,2,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-07-03T17:40:10Z,2026-07-08T21:14:32Z,0,"Fixes #188150 Problem With mode=""reduce-overhead"", cudagraph trees record one graph per distinct dynamic-shape key: deferred_cudagraphify keys fn_cache on the concrete SymInt input values, with no upper bound. Workloads whose input sizes change almost every call — e.g. HuggingFace/TRL training loops that pad each ba..."
188998,[inductor] Fix bf16 matmul accumulation precision in Triton templates (#188492),https://github.com/pytorch/pytorch/pull/188998,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,DIRTY,False,1,1,statusCheckRollup,statusCheckRollup,188492,188492,188492,,,,0,1,2,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-06T04:15:48Z,2026-07-08T21:14:23Z,0,"Fixes #188492 Under Triton 3.8 on Hopper (SM90), bf16 Swin-Transformer and similar bf16 workloads compiled with torch.compile(mode=""max-autotune"") showed significant accuracy degradation. Root cause: Multiple Triton MM/conv/BMM templates called tl.dot(a, b, allow_tf32=ALLOW_TF32) without an explicit out_dtype. When..."
189119,Fix test-jsons artifact upload path,https://github.com/pytorch/pytorch/pull/189119,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,189118,189118,189118,,,,0,1,2,,,"triaged, open source, topic: not user facing",2026-07-07T02:15:59Z,2026-07-08T21:13:59Z,0,"Issue Fixes #189118 Summary What Problem This Solves upload_to_s3_artifacts() creates three separate intermediate artifacts for CI test output: test-reports-*.zip for XML/CSV reports, logs-*.zip for log files, and test-jsons-*.zip for JSON reports. Before this change, the third upload wrote to the test-jsons-* S3 ob..."
189102,[TEST][Inductor][CUDA] Bump tolerances for `test_comprehensive_combinations_cuda_float16`,https://github.com/pytorch/pytorch/pull/189102,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cuda, triaged, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-07-06T22:21:35Z,2026-07-08T21:13:45Z,0,"combinations uses index_put underneath which appears to do fp16 accum in inductor authored with codex example failure _assert_equal_with_significant_stride_check assert_equal_fn( File ""/usr/local/lib/python3.12/dist-packages/torch/_dynamo/test_case.py"", line 129, in assertEqual return super().assertEqual(x, y, *args..."
188996,[inductor] Fix FP8 blockwise scaling shape validation in Triton (#188711),https://github.com/pytorch/pytorch/pull/188996,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,188711,188711,"188711, 188721",,,,0,1,3,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-07-06T04:15:43Z,2026-07-08T21:13:10Z,0,Fixes #188711 The test_main_loop_scaling FP8 test was failing because Inductor's scale-shape validator (get_scaling_options/is_desired_scaling) rejected valid cuBLAS blockwise scale tensors: _is_blockwise1xTILESIZE_scaling had inverted transpose logic. The Triton kernel indexes every scale with the output dimension...
189037,Preserve strides for CuTeDSL grouped MM fake tensors,https://github.com/pytorch/pytorch/pull/189037,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-07-06T16:00:55Z,2026-07-08T21:13:02Z,0,"CuTeDSL grouped MM compiles kernels with fake tensor descriptors. The template used make_fake_compact_tensor with a stride order derived from get_stride_order. That is not enough for non-compact inputs: a padded A tensor with logical K=64 and row stride 72 has the same stride order as a compact row stride 64, so run..."
187887,[Inductor] Add decomposition for prims.uniform,https://github.com/pytorch/pytorch/pull/187887,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-22T23:54:41Z,2026-07-08T21:12:15Z,0,Currently inductor uses an eager fallback for aten.uniform and aten.uniform_. This PR adds a straight forward decomposition of prims.uniform so it can be natively lowered to prims.rand. /cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @ka...
189010,[Dynamo] Support relax frontend in tvm backend,https://github.com/pytorch/pytorch/pull/189010,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,188866,188866,188866,,,,0,1,2,,,"triaged, open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-06T06:51:16Z,2026-07-08T21:12:14Z,0,"Related Isue closes #188866 Why The tvm backend depends on tvm.relay and tvm.contrib.graph_executor, removed upstream in TVM 0.20, so every pip-installable TVM (apache-tvm==0.25.*) fails at import with ImportError: cannot import name 'relay' from 'tvm'. How Dispatch on the installed TVM API: use the legacy relay pat..."
189100,[ROCm] Fix test_is_pinned_no_context for Python 3.14 forkserver default,https://github.com/pytorch/pytorch/pull/189100,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, triaged, open source, ciflow/trunk, topic: not user facing",2026-07-06T21:54:17Z,2026-07-08T21:11:41Z,0,"test_is_pinned_no_context spawns a worker via a bare multiprocessing.Process(), implicitly relying on the ambient default start method. Python 3.14 changed that default on non-macOS POSIX platforms from ""fork"" to ""forkserver"", and this test breaks under the new default for a mechanical reason: Under ""fork"", Process...."
189153,Bump onnx from 1.21.0 to 1.22.0 in /.ci/docker,https://github.com/pytorch/pytorch/pull/189153,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, python, dependencies, ciflow/docker",2026-07-07T13:59:24Z,2026-07-08T21:09:36Z,0,Bumps onnx from 1.21.0 to 1.22.0. Release notes Sourced from onnx's releases. v1.22.0 ONNX v1.22.0 is now available with exciting new features! We would like to thank everyone who contributed to this release! Please visit onnx.ai to learn more about ONNX and associated projects. What's Changed Breaking Changes and D...
189158,Skip GHA stats artifacts from other workflow attempts,https://github.com/pytorch/pytorch/pull/189158,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,189157,189157,189157,,,,0,1,2,,,"triaged, open source, topic: not user facing",2026-07-07T14:57:39Z,2026-07-08T21:09:22Z,0,"Fixing an Issue Issue Fixes #189157 Summary What Problem This Solves _download_artifact() already detected when an artifact name's runattemptN did not match the requested workflow_run_attempt, but it only printed Skipping ... and then kept going. That meant a previous-attempt GitHub fallback artifact could still be..."
189154,Bundle unversioned libnuma for ROCm wheels,https://github.com/pytorch/pytorch/pull/189154,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,189110,189110,,,,,0,1,0,,,"module: rocm, triaged, open source, release notes: rocm",2026-07-07T14:04:38Z,2026-07-08T21:08:02Z,0,"Fixes #189110 Summary: bundle an additional torch/lib/libnuma.so copy for ROCm wheels when libnuma.so.1 is present keep the existing libnuma.so.1 bundle unchanged avoid relying on wheel installers to preserve symlinks Root cause: libtorch_rocshmem.so can call dlopen(""libnuma.so"") during import torch. The ROCm wheel..."
177757,[MPS] support for SparseCSR tensor and its unary operators,https://github.com/pytorch/pytorch/pull/177757,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,28,0,,,"triaged, open source, Stale, release notes: sparse, topic: improvements, release notes: mps, ciflow/mps",2026-03-18T19:52:39Z,2026-07-08T21:06:41Z,0,This PR adds MPS support for sparse CSR tensor format and wires sparse CSR unary operator coverage. add MPS implementations for CSR/COO index conversion. COO and CSR tensors could convert between one and another. register MPS sparse CSR dispatch for the unary operators in native_functions.yaml add MPS test coverage...
181895,[CUDA] Fix linspace integer dtype precision mismatch with CPU,https://github.com/pytorch/pytorch/pull/181895,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/h100-symm-mem",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,181807,181807,181807,,,,0,1,3,,,"module: cpu, triaged, module: mkldnn, open source, module: amp (automated mixed precision), NNC, release notes: quantization, release notes: cuda, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), release notes: distributed (checkpoint), module: compiled autograd, no-runner-experiments, ciflow/rocm-mi300, release notes: inductor (aoti), ciflow/h100, ciflow/h100-symm-mem, ciflow/vllm, ciflow/b200, ciflow/inductor-pallas, ci-refresh-hf-cache, ciflow/torchtitan, ciflow/dtensor",2026-04-29T18:33:37Z,2026-07-08T21:06:36Z,0,"The CUDA linspace kernel used float for the step when the output dtype is integral, while the CPU kernel uses double. That can make CPU and CUDA disagree on int64 linspace for the same start, end, and steps. Use double for the integral CUDA path to match RangeFactoriesKernel.cpp. Fixes #181807"
188825,"[Dynamo] Implement subgenerator support in "".throw"" / "".close"" methods",https://github.com/pytorch/pytorch/pull/188825,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"157149, 187690, 187744, 188004, 188638, 188639, 188824, 188834, 189024",,,0,13,9,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-02T16:43:23Z,2026-07-08T21:06:07Z,0,Stack from ghstack (oldest at bottom): #189024 -> #188825 #188834 #188824 #157149 #188639 #188638 #188004 #187744 #187690
188834,[Dynamo] Track generator attribute mutations for proper closure handling,https://github.com/pytorch/pytorch/pull/188834,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"157149, 187690, 187744, 188004, 188638, 188639, 188824, 188825, 189024",,,0,11,20,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-02T18:03:40Z,2026-07-08T21:06:06Z,0,Stack from ghstack (oldest at bottom): #189024 #188825 -> #188834 #188824 #157149 #188639 #188638 #188004 #187744 #187690
188824,Port doctest tests from cpython test_generators.py,https://github.com/pytorch/pytorch/pull/188824,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"157149, 187690, 187744, 188004, 188638, 188639, 188825, 188834, 189024",,,0,10,19,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-07-02T16:43:16Z,2026-07-08T21:06:06Z,0,Stack from ghstack (oldest at bottom): #189024 #188825 #188834 -> #188824 #157149 #188639 #188638 #188004 #187744 #187690
157149,[generator] Close all open generators in compile_subgraph,https://github.com/pytorch/pytorch/pull/157149,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187690, 187744, 188004, 188638, 188639, 188824, 188825, 188834, 189024",,,0,44,9,,,"open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, no-stale, ci-no-td, ciflow/torchtitan",2025-06-27T19:45:44Z,2026-07-08T21:06:05Z,0,Stack from ghstack (oldest at bottom): #189024 #188825 #188834 #188824 -> #157149 #188639 #188638 #188004 #187744 #187690 Motivation and Example Explictly close all open generators in compile_subgraph to ensure that all remaining finally blocks are executed. In CPython this is done by invoking the tp_finalize functi...
188639,[Dynamo] Corretly implement SEND opcode,https://github.com/pytorch/pytorch/pull/188639,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"157149, 187690, 187744, 188004, 188638, 188824, 188825, 188834, 189024",,,0,13,9,,,"open source, Merged, Reverted, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, ci-no-td, ciflow/torchtitan",2026-07-01T02:53:28Z,2026-07-08T21:06:04Z,0,Stack from ghstack (oldest at bottom): #189024 #188825 #188834 #188824 #157149 -> #188639 #188638 #188004 #187744 #187690
188638,[Dynamo] Fix CLEANUP_THROW,https://github.com/pytorch/pytorch/pull/188638,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"157149, 187690, 187744, 188004, 188639, 188824, 188825, 188834, 189024",,,0,13,19,,,"open source, Merged, Reverted, module: dynamo, ciflow/inductor, release notes: dynamo, ci-no-td, ciflow/torchtitan",2026-07-01T02:53:22Z,2026-07-08T21:06:03Z,0,Stack from ghstack (oldest at bottom): #189024 #188825 #188834 #188824 #157149 #188639 -> #188638 #188004 #187744 #187690
187744,[Dynamo] Add a per-generator exception stack,https://github.com/pytorch/pytorch/pull/187744,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"157149, 187690, 188004, 188638, 188639, 188824, 188825, 188834, 189024",,,0,13,10,,,"open source, Merged, Reverted, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, ci-no-td, ciflow/torchtitan",2026-06-19T21:27:04Z,2026-07-08T21:06:02Z,0,"Stack from ghstack (oldest at bottom): #189024 #188825 #188834 #188824 #157149 #188639 #188638 #188004 -> #187744 #187690 This mirrors CPython behavior where each generator has its own exception stack, which is linked for the duration the generator is active Ref: https://github.com/python/cpython/blob/e76aa128fea869..."
187690,[Dynamo] Refactor generator methods,https://github.com/pytorch/pytorch/pull/187690,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"157149, 187744, 188004, 188638, 188639, 188824, 188825, 188834, 189024",,,0,14,9,,,"open source, Merged, Reverted, topic: not user facing, module: dynamo, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-06-18T19:12:37Z,2026-07-08T21:06:01Z,0,"Stack from ghstack (oldest at bottom): #189024 #188825 #188834 #188824 #157149 #188639 #188638 #188004 #187744 -> #187690 Refactor .close, .send and .throw to follow CPython semantics."
189148,Add maybe_unused label to strides when strides is unused,https://github.com/pytorch/pytorch/pull/189148,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-07-07T12:51:27Z,2026-07-08T21:00:41Z,0,"Added [[maybe_unused]] label to strides parameter to cases where strides is present in the signature but not semantically used (i.e. when schema kind is SchemaKind.inplace, generating set_output_raw_strided, not set_output_strided and there is no structured parent super call). This avoids the issue where structured..."
189141,"[test] Add opt-in sign-sensitive special-value parity check (copysign, atan2, reciprocal)",https://github.com/pytorch/pytorch/pull/189141,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,188680,188680,188680,,,,0,1,2,,,"triaged, open source, topic: not user facing",2026-07-07T09:59:25Z,2026-07-08T20:56:01Z,0,"Fixes #188680 torch.allclose treats +0.0 and -0.0 as equal (IEEE 754 defines them as equal under ==), so the default compile-vs-eager correctness check in test_ops.py cannot detect a compiler pass that silently normalizes signed zero (e.g. -0.0 -> +0.0). This adds an opt-in check, off by default, for ops where that..."
189182,make the return more specified,https://github.com/pytorch/pytorch/pull/189182,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-07-07T18:43:36Z,2026-07-08T20:55:38Z,0,Replaced the Any return type with torch.Tensor since _weight_norm always returns a Tensor.
185424,Add CPython Dynamo agentic loop plan,https://github.com/pytorch/pytorch/pull/185424,draft/noise,13,PR is draft,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"189021, 189022, 189051, 189052, 189053",,,0,40,5,,,"open source, topic: not user facing, skip-pr-sanity-checks",2026-05-27T20:20:31Z,2026-07-08T20:53:26Z,0,"Stack from ghstack (oldest at bottom): -> #185424 #189053 #189052 #189022 #189051 #189021 Adds the CPython Dynamo expected-failure relevance rankings together with the agentic-loop docs used to work the highest-value CPython coverage gates. The docs include the fresh-context bootstrap, manager and review-agent workf..."
189053,[dynamo] Support integer arithmetic on a sourceless id()/hash() FakeIdVariable,https://github.com/pytorch/pytorch/pull/189053,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"185424, 189021, 189022, 189051, 189052",,,0,3,5,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-06T17:56:29Z,2026-07-08T20:53:25Z,0,Stack from ghstack (oldest at bottom): #185424 -> #189053 #189052 #189022 #189051 #189021 A custom hash that does integer arithmetic on id(self) -- e.g. class H(set): __hash__ = lambda self: int(id(self) & 0x7fffffff) -- crashed under Dynamo with a spurious TypeError: unsupported operand type(s) for &. id() of an ob...
189052,[dynamo] Model deque iterators with mutation-during-iteration detection,https://github.com/pytorch/pytorch/pull/189052,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"185424, 189021, 189022, 189051, 189053",,,0,3,5,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-06T17:56:23Z,2026-07-08T20:53:24Z,0,"Stack from ghstack (oldest at bottom): #185424 #189053 -> #189052 #189022 #189051 #189021 Iterating a deque under Dynamo did not detect mutation during iteration, and reversed(deque) produced the wrong iterator type. DequeVariable.tp_iter_impl returned a plain ListIteratorVariable sharing self.items (with a TODO), a..."
188814,"Fix typos in comments and docstrings across ao, distributed, masked, nn, and sparse modules",https://github.com/pytorch/pytorch/pull/188814,ready for maintainer decision,45,"CI lane labels: ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, release notes: quantization, ciflow/torchtitan",2026-07-02T15:36:55Z,2026-07-08T20:53:13Z,0,"Stack from ghstack (oldest at bottom): -> #188814 No functional changes; corrects misspellings such as ""sparsier"", ""filtters"", ""inconsistences"", ""attributesn"", ""insersection"", ""uniformaly"", ""hifen"", ""doctring"", ""spatia"", and ""scalig"" in documentation and inline comments. Authored with the assistance of an AI agent."
189131,Skip building unused XNNPACK microkernels-all; guard GCC14 workaround,https://github.com/pytorch/pytorch/pull/189131,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,188867,,,0,1,2,,,"triaged, open source",2026-07-07T05:45:49Z,2026-07-08T20:53:11Z,0,"Summary Split out of #188867 per review feedback (the KleidiAI SME feature and this change are logically independent, and bundling them made #188867 harder to review/revert as a unit): Disables XNNPACK_BUILD_ALL_MICROKERNELS, since microkernels-all is only consumed by XNNPACK's own tests/benchmarks (both already dis..."
188870,"Fix typos in comments, docstrings, and error messages",https://github.com/pytorch/pytorch/pull/188870,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,3,0,,,"ciflow/trunk, release notes: fx, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-03T02:02:20Z,2026-07-08T20:52:54Z,0,"Stack from ghstack (oldest at bottom): -> #188870 Corrects spelling and grammar in various docstrings, comments, and error strings across custom ops, inductor, distributions, fx, and sparse modules. No functional changes. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng..."
187989,[MPS] Fix linear backward SIGABRT with a 1-D weight,https://github.com/pytorch/pytorch/pull/187989,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/docker, ciflow/dtensor, ciflow/h100","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,187988,187988,"187988, 188721",,,,0,1,3,,,"module: cpu, triaged, open source, release notes: mps, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), ciflow/rocm-mi300, ciflow/h100, ciflow/h100-symm-mem, ciflow/vllm, ciflow/b200, ci-refresh-hf-cache, ciflow/torchtitan, ciflow/dtensor, ciflow/docker",2026-06-23T21:50:56Z,2026-07-08T20:52:32Z,0,"Fixes #187988 What linear() accepts a 1-D weight (shape [in_features]): it contracts and drops the last input dimension. mps_linear (forward) already handles this by unsqueezing the weight to [1, in_features]. The backward pass did not mirror that, so it lowered grad_input and grad_weight to invalid rank-1 × rank-1..."
185309,Add sparse COO dispatch for linalg.vector_norm,https://github.com/pytorch/pytorch/pull/185309,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,1,,,"open source, ciflow/trunk, release notes: sparse",2026-05-27T02:15:56Z,2026-07-08T20:52:04Z,0,"torch.linalg.vector_norm previously lacked SparseCPU/SparseCUDA/SparseMPS dispatch, so substituting it for the deprecated torch.norm raised NotImplementedError on sparse COO tensors. Adds a thin linalg_vector_norm_sparse wrapper that delegates to the existing norm_sparse kernel via aten::native_norm."
188954,[MPS] add matrix exp on MPS,https://github.com/pytorch/pytorch/pull/188954,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,4,1,,,"triaged, open source, release notes: linalg_frontend, topic: improvements, release notes: mps, ciflow/mps",2026-07-04T19:19:19Z,2026-07-08T20:50:52Z,0,Add matrix_exp on MPS. Perf measured on few shapes since larger means cpu gets much slower: float32 Matrix CPU (us) MPS (us) Speedup 256x256 632.8 1464.7 0.43x 512x512 2237.5 1967.5 1.14x 1024x1024 13315.9 5675.8 2.35x complex64 Matrix CPU (us) MPS (us) Speedup 256x256 1882.4 1711.1 1.10x 512x512 8821.6 3379.3 2.61x...
189207,Scope cuDNN Engine 5 errata to SM 8.6 GPUs only,https://github.com/pytorch/pytorch/pull/189207,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,188318,,,0,1,1,,,"triaged, open source, topic: not user facing",2026-07-07T21:49:37Z,2026-07-08T20:50:47Z,0,"[cuDNN][Convolution] Scope Engine 5 errata to SM 8.6 GPUs only Summary PR #188318 disabled cuDNN forward convolution Engine 5 unconditionally to fix nondeterminism on SM 8.6 GPUs (A10G, A2). Blocking it globally caused a severe regression on H100/H200 for depthwise-separable convolution models: 5 of 6 affected model..."
189125,[Dynamo] Raise clear error for dynamic shapes in tvm relay path,https://github.com/pytorch/pytorch/pull/189125,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,169111,169111,169111,189010,,,0,1,3,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-07T05:10:59Z,2026-07-08T20:50:41Z,0,"Related Issue closes #169111 Why With dynamic=True, Dynamo passes SymInt scalars as graph inputs, and the tvm backend feeds them straight into torch.jit.trace, which fails with a cryptic Tracer cannot infer type ... got value of type SymInt error. The relay path is trace-based and cannot support dynamic shapes. How..."
184447,[inductor][autotuning] add incremental autotune foundation,https://github.com/pytorch/pytorch/pull/184447,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"triaged, open source, fb-exported, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-19T23:13:02Z,2026-07-08T20:50:06Z,0,"Summary: Adds the standalone subsystems that the incremental autotune plugin builds on, with no plugin engagement yet: incremental/_resolver.py: background daemon that waits on CUDA timing events and dispatches per-launcher callbacks. incremental/_launcher.py: Launcher (timed-dispatch wrapper with weak-ref'd raw + s..."
183684,[inductor] extract _bench_launchers and _finalize_autotune_winner helpers,https://github.com/pytorch/pytorch/pull/183684,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, fb-exported, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-14T08:44:32Z,2026-07-08T20:49:59Z,0,"Summary: Two extractions out of the autotune flow: _bench_launchers(launchers, *args, **kwargs) -> (timings, bench_ns) -- bench-each-launcher + coordesc_tuner.cache_benchmark_result loop. Lands here unused; the streaming-compile diff wires it into the pipelined plugin. _finalize_autotune_winner(timings, autotune_tim..."
183683,[inductor] extract _iter_compile_results generator with optional thread pool,https://github.com/pytorch/pytorch/pull/183683,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, fb-exported, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-14T08:44:31Z,2026-07-08T20:49:54Z,0,"Summary: Replace the inline per-config compile loop in _precompile_worker with self.compile_results = list(self._iter_compile_results()). The generator catches OOM / PTXAS / IntelGPU errors per config (drops the result, records the last exception on self._last_compile_exception) so callers can choose between yieldin..."
183682,"[inductor] extract _bundle_compile_result, _all_failed_fallback, _maybe_put_static_autotuner",https://github.com/pytorch/pytorch/pull/183682,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, fb-exported, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-14T08:44:29Z,2026-07-08T20:49:49Z,0,"Summary: Three small extractions out of _precompile_worker / _make_launchers / precompile. Each is a literal cut-paste with no logic change. _bundle_compile_result(result) -- the TritonBundler.put(triton_hash_to_path_key(...), device) call. _all_failed_fallback(exc) -- the OOM-with-pipelining fallback that turns the..."
183663,[inductor] _make_launchers: skip rebuilding existing launchers,https://github.com/pytorch/pytorch/pull/183663,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,1,,,"triaged, open source, fb-exported, topic: not user facing, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-14T05:29:25Z,2026-07-08T20:49:42Z,0,"Summary: Modifies CachingAutotuner._make_launchers to take a list of compile results for which we make launchers, rather than making launchers for all compile results on that instance. Previously, this would have duplicated calls to CachingAutotuner._make_launcher for the initial set of compile results when dynamic..."
181826,[inductor][autotuning] extract _should_coordesc_tune cached property (#181826),https://github.com/pytorch/pytorch/pull/181826,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,1,,,"triaged, open source, Merged, Reverted, Stale, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-04-29T02:50:09Z,2026-07-08T20:49:37Z,0,Summary: Pure refactor. The two-stage gate inside coordinate_descent_tuning (skip TEMPLATE/USER_AUTOTUNE/FIXED heuristic types + skip deterministic-mode reduction kernels) is hoisted into a standalone _should_coordesc_tune cached property. coordinate_descent_tuning becomes if not self._should_coordesc_tune: return l...
181074,[inductor][coordesc] extract get_neighbor_configs and get_all_direction_configs (#181074),https://github.com/pytorch/pytorch/pull/181074,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,181074,,,,,0,1,0,,,"triaged, open source, fb-exported, Stale, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-22T05:49:33Z,2026-07-08T20:49:30Z,0,Summary: Pull Request resolved: #181074 Differential Revision: D101933332 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
188012,Add sparse triangular_solve CPU fallback,https://github.com/pytorch/pytorch/pull/188012,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,187389,187389,187389,,,,0,2,2,,,"triaged, open source, release notes: sparse",2026-06-24T01:57:06Z,2026-07-08T20:49:23Z,0,"Issue Fixes #187389 Summary Adds a non-MKL CPU reference implementation for CSR sparse triangular_solve. Supports float, double, complex float, complex double, int32/int64 CSR indices, upper/lower, transpose, unitriangular, and strided dense outputs. Keeps BSR on the existing MKL-only path so this stays scoped to th..."
189289,[TEST] Refactor test_c10d_pypg.py with hw_classification,https://github.com/pytorch/pytorch/pull/189289,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,186918,,,0,1,1,,,"triaged, open source, topic: not user facing",2026-07-08T18:46:37Z,2026-07-08T20:49:21Z,0,"Annotate all test classes in test/distributed/test_c10d_pypg.py with hw_classification as part of the ongoing test refactoring initiative. Requires #186918 to be merged first. Changes TestDDPWithWorkSubclass → GENERIC: single-rank fake Python PG on CPU, tests DDP Work subclass dispatch TestDDPWithWorkWrapper → GENER..."
188784,Fix `torch.asarray` for buffer-protocol objects with no dtype,https://github.com/pytorch/pytorch/pull/188784,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188128,188128,188128,,,,0,8,3,,,"triaged, open source, release notes: python_frontend",2026-07-02T07:25:45Z,2026-07-08T20:48:55Z,0,"Fixes #188128 The fix has two parts: When no dtype is given, read the buffer's declared format (via PyBUF_FORMAT | PyBUF_STRIDES) and map it to the matching ScalarType instead of guessing. The mapping lives in a new helper, scalar_type_from_buffer_format. It's a small, self-contained parser, but hand-rolled so the b..."
189263,Remove the Miniconda base env from the CD builder images,https://github.com/pytorch/pytorch/pull/189263,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries, ciflow/docker","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,"180247, 180250",,,0,1,2,,,"triaged, open source, ciflow/binaries, topic: not user facing, ciflow/docker",2026-07-08T14:38:24Z,2026-07-08T20:48:54Z,0,"First Docker-image increment of the conda-removal effort: removes the Miniconda3 base env from the two Linux CD builder images (manylinux2_28-builder and almalinux-builder). Nothing in the Linux CD build path actually uses that conda: The wheel build runs against the /opt/python manylinux interpreters, which set_des..."
182906,Port 11 tests under test/distributed folder to XPU,https://github.com/pytorch/pytorch/pull/182906,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/torchtitan,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,20,0,,,"triaged, open source, topic: not user facing, ciflow/torchtitan",2026-05-07T22:28:57Z,2026-07-08T20:48:29Z,0,"For #114850, we will port distributed tests to Intel GPU. We will enable Intel GPU with following methods and keep the original code styles: Example: ""torch.accelerator.current_accelerator()"" to determine the accelerator backend enabled XPU for the following files: test/distributed/test_backends.py test/distributed/..."
180230,[TEST][CUDA][TF32] Update TF32 test decorators to use new fp32 precision path,https://github.com/pytorch/pytorch/pull/180230,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"module: cuda, module: tests, triaged, open source, module: tf32, Stale, topic: not user facing",2026-04-13T17:21:24Z,2026-07-08T20:48:25Z,0,Warnings were already removed in #166956 but should help abate warning spam if they are added back... authored with claude cc @ptrblck @msaroufim @jerryzh168 @tinglvv @nWEIdia @mruberry @zasdfgbnm
181926,Enable XPU for sparse_sampled_addmm,https://github.com/pytorch/pytorch/pull/181926,draft/noise,31,PR is draft; review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,0,0,,,,,,3175,,,0,1,1,,,"open source, Stale, release notes: sparse",2026-04-29T21:07:51Z,2026-07-08T20:47:00Z,0,Wire SparseCsrXPU into sparse_sampled_addmm and sparse_sampled_addmm.out in native_functions.yaml so CSR sampled addmm can execute on XPU. Fix for: #3175 Requires PR#3534
179425,[inductor][autotuning] add logging and config for incremental autotuning (#179425),https://github.com/pytorch/pytorch/pull/179425,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,1,,,"triaged, open source, fb-exported, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-05T19:58:02Z,2026-07-08T20:41:11Z,0,Summary: This commit introduces the logging and top-level switch for Incremental Autotuning. Authored with Claude. Test Plan: CI Differential Revision: D99545697 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames...
189201,[TEST][Inductor] Fix some `test_buffer_mutation*` tests,https://github.com/pytorch/pytorch/pull/189201,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cuda, module: tests, triaged, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-07-07T21:03:18Z,2026-07-08T20:40:29Z,0,skip onednn test on ARM and defensively do not assume current device is device 0 on CUDA in case another test did not clean up the current device setting authored with codex cc @ptrblck @msaroufim @tinglvv @nWEIdia @mruberry @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng...
189132,Register sspaddmm.out for XPU backend,https://github.com/pytorch/pytorch/pull/189132,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,4112,4112,,,,,0,2,0,,,"triaged, open source, release notes: sparse",2026-07-07T06:12:27Z,2026-07-08T20:39:59Z,0,intel/torch-xpu-ops#4112 Register and implement sspaddmm and its out-variant on XPU.
189160,Use SafePyObject in PyAnomalyMetadata and PySavedVariableHooks,https://github.com/pytorch/pytorch/pull/189160,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,1,,,"triaged, open source, topic: not user facing",2026-07-07T15:04:17Z,2026-07-08T20:39:09Z,0,"Replace the raw PyObject* members (PyAnomalyMetadata::dict_ and PySavedVariableHooks::{pack_hook_,unpack_hook_,data_}) with c10::SafePyObject, and delete the hand-written GIL-acquiring destructors. Cleanup now routes through the shared PyInterpreter::decref path instead of each class reimplementing its own. Also cha..."
189127,[inductor] Decompose broadcast-bias baddbmm before pointwise,https://github.com/pytorch/pytorch/pull/189127,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,187093,187093,187093,,,,0,2,3,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T05:16:58Z,2026-07-08T20:38:12Z,0,"Fixes #187093. Summary Extend the existing addmm bias-unfuse post-grad pattern to broadcast-bias baddbmm. Only rewrite on GPU when the baddbmm input is bias-like or broadcast and all users are pointwise. Preserve standalone baddbmm, explicit stride-0 expanded bias, and alpha/beta semantics. Let the decomposed bmm +..."
189149,"[inductor] Emit pointwise kernel for small mm (K<8, N<8) instead of cuBLAS",https://github.com/pytorch/pytorch/pull/189149,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,186348,186348,"186348, 188721",,,,0,1,3,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-07-07T12:55:36Z,2026-07-08T20:37:31Z,0,"Fixes #186348 Summary When both inner dimensions of aten.mm are small (K < 8 and N < 8), the cuBLAS/Triton GEMM kernel launch overhead dominates execution time. This PR extends tuned_mm in the lowering to emit a fused pointwise Triton kernel (unsqueeze + mul + sum) for these shapes, following the same IR emission pa..."
182278,[CI] Add native build image for linux-riscv64,https://github.com/pytorch/pytorch/pull/182278,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/docker,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,19,1,,,"triaged, open source, release notes: releng, module: risc-v, no-runner-experiments, ciflow/docker",2026-05-04T10:53:18Z,2026-07-08T20:36:32Z,0,"This is used as part of the riseproject-dev out-of-tree CI [1][2], to get us closer to having native builds of PyTorch on RISC-V. [1] https://github.com/risepytorch-dev/pytorch [2] pytorch/.github/allowlist.yml Line 40 in 01bcbb4 - riseproject-dev/pytorch"
189123,cuda graphs: accept None && float graph-partition inputs,https://github.com/pytorch/pytorch/pull/189123,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-07-07T04:42:33Z,2026-07-08T20:36:26Z,0,"With capture_scalar_outputs set, a graph that materializes a scalar via .item() is partitioned for cudagraphs, and inductor threads a None or float buffer across the partition boundary as a graph input (e.g. partition0_args = [x, y, buf1=None, u0]). The record path in _allocate_and_copy_recording_inputs only accepte..."
189124,[inductor] Cache DSR-expanded single-config reductions,https://github.com/pytorch/pytorch/pull/189124,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,189121,189121,189121,,,,0,3,3,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T05:10:08Z,2026-07-08T20:36:12Z,0,"Fixes #189121. Summary Extract the existing dynamic-rblock eligibility predicate so the cache gate can see when a single original reduction config can expand later. Let DSR-eligible single-config reductions create/load an autotune cache before autotuning, while preserving the ordinary single-config no-cache fast pat..."
189206,[Inductor] register `target_code` in `input_codes` tracker,https://github.com/pytorch/pytorch/pull/189206,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-07T21:46:59Z,2026-07-08T20:35:27Z,0,"Otherwise reset appears to leave precompile entries which breaks e.g., Traceback (most recent call last): File ""/usr/lib/python3.12/unittest/case.py"", line 58, in testPartExecutor yield File ""/usr/lib/python3.12/unittest/case.py"", line 634, in run self._callTestMethod(testMethod) File ""/usr/lib/python3.12/unittest/c..."
189285,Fix MIOpen RNN dropout: gradient stall and missing per-step mask variation,https://github.com/pytorch/pytorch/pull/189285,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,6339,6339,,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-07-08T18:04:15Z,2026-07-08T20:34:34Z,0,"[ROCm] Fix MIOpen RNN dropout: gradient stall and missing per-step mask variation Fixes ROCm/ROCm#6339. Summary Setting dropout via nn.GRU(..., dropout=p) with num_layers>1 on ROCm causes training to stall, while the equivalent model using an explicit nn.Dropout between single-layer GRUs trains normally (and the fus..."
185794,Fix torch.kron RuntimeError on non-contiguous tensors - noah,https://github.com/pytorch/pytorch/pull/185794,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,,185794,,,,,0,2,1,,,"triaged, open source, release notes: linalg_frontend",2026-06-01T06:05:53Z,2026-07-08T20:31:48Z,0,Resolves #185794 This commit fixes an issue where torch.kron would crash when given non-contiguous input tensors (e.g. from a .t() operation). The bug was caused by calling at::_unsafe_view without first ensuring the tensors were contiguous in memory. A regression test has been added to test_linalg.py.
189145,[Dynamo] Preserve Stream/Event subclass types in graph reconstruction,https://github.com/pytorch/pytorch/pull/189145,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,"188325, 188405","188325, 188405","188325, 188405",,,,0,1,4,,,"triaged, open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-07T11:44:16Z,2026-07-08T20:27:40Z,0,"Related Issue Fixes #188325 Fixes #188405 Why EventVariable.python_type() returns a hardcoded torch.Event, and graph-break reconstruction rebuilds graph-created streams/events via build_stream/build_event, which hardcode torch._C.Stream/Event. Both silently drop user subclass types (the device-agnostic torch.Stream/..."
189186,"Extend ""Fix positional arg overflow in TritonBenchmarkRequest when template inputs are deduplicated"" to CUTLASS backend",https://github.com/pytorch/pytorch/pull/189186,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,188721,188268,,,0,8,2,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-07-07T19:18:55Z,2026-07-08T20:27:22Z,0,"after #188268 is merged, this extends that fix to cutlass -- do not merge before #188268 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo"
189247,Register _foreach_clone XPU dispatch,https://github.com/pytorch/pytorch/pull/189247,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,1,,,"triaged, open source",2026-07-08T08:20:11Z,2026-07-08T20:26:40Z,0,XPU: foreach_tensor_clone_xpu
186056,Support input mutations in CUDA graph control flow.,https://github.com/pytorch/pytorch/pull/186056,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186055,,,0,3,1,,,open source,2026-06-03T09:06:25Z,2026-07-08T20:26:01Z,0,"Stack from ghstack (oldest at bottom): -> #186056 #186055 This allows us to conditionally run an optimizer. This enables us to do a gradient nan check before running the optimizer without having to fuse the gradient nan check into the optimizer kernel, which is what NVIDIA has traditionally done for mlperf training..."
189044,[inductor] Handle symbolic float scalars in FX wrapper metadata hook and reference analysis (#189044),https://github.com/pytorch/pytorch/pull/189044,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, release notes: fx, module: inductor, ciflow/inductor, meta-exported",2026-07-06T16:55:08Z,2026-07-08T20:25:58Z,0,"Summary: When the Inductor FX wrapper backend (fx_wrapper=True, used by MTIA) lowers an op whose size depends on a symbolic float -- e.g. torch.arange(start, end=3.1, dtype=torch.float32) with a dynamic integer start, whose length is ceil((end - start) / step) -- it materializes the size computation as __sym_float__..."
106149,Automated submodule update: kineto,https://github.com/pytorch/pytorch/pull/106149,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2023-07-27T17:42:39Z,2026-07-08T20:25:54Z,0,This is an automated pull request to update the first-party submodule for pytorch/kineto. New submodule commit: pytorch/kineto@465ff4c Test Plan: Ensure that CI jobs succeed on GitHub before landing.
189197,make post-processing in codegen mandatory,https://github.com/pytorch/pytorch/pull/189197,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,1,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T20:45:35Z,2026-07-08T20:25:29Z,0,"We hit this in HeavyBall when running our tests. Some nodes are sometimes left dirty which can cause tests to fail when invoked often enough. I did not investigate whether this same pattern should be applied to other functions as well. Repro: """""" Minimal reproduction of the invariant violation in torch/_inductor/sch..."
189246,[DO NOT MERGE] Test L3 CRCR,https://github.com/pytorch/pytorch/pull/189246,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/crcr/crcr-test,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/crcr/crcr-test",2026-07-08T07:57:32Z,2026-07-08T20:23:09Z,0,
189146,[MPS] softmax: non-last-dim and huge-row native Metal kernels (perf specializations),https://github.com/pytorch/pytorch/pull/189146,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,187455,187456,,,0,2,2,,,"triaged, open source, release notes: mps, ciflow/mps",2026-07-07T12:06:08Z,2026-07-08T20:22:40Z,0,"[MPS] softmax: non-last-dim and huge-row native Metal kernels (perf specializations) Reviewing just this PR. This is a stacked leaf (cumulative-leaf fork on main — external contributors can't ghstack), so the GitHub diff shows the softmax core PR + this one. To see only what this PR adds on top of #187456, use this..."
188973,"Reapply ""Avoid redundant scalar-to-tensor conversion in `real`/`imag`setters on complex tensors (#188603)",https://github.com/pytorch/pytorch/pull/188973,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-07-05T14:14:04Z,2026-07-08T20:20:39Z,0,This reapplies #188603 which has been reverted due to Meta internal CI failures. @ngimel let me know which CI failure caused the revert and I can work on a fix. This reverts commit bfd1651.
189172,Fix protobuf resource-compiler build with clang-cl (RC1106 /bigobj),https://github.com/pytorch/pytorch/pull/189172,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"open source, topic: not user facing",2026-07-07T17:17:20Z,2026-07-08T20:20:05Z,0,Summary Building PyTorch on Windows with clang-cl (e.g. CC=CXX=clang-cl) and the default BUILD_CUSTOM_PROTOBUF=ON fails while compiling protobuf's version.rc:fatal error RC1106: invalid option: /bigobj /bigobj is a C/C++ compiler flag that the Windows resource compiler (rc.exe) does not understand. Root cause protob...
189162,[inductor] Use non-NaN-propagating fmax in softmax persistent reduction,https://github.com/pytorch/pytorch/pull/189162,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,189161,,,,0,2,1,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T15:28:20Z,2026-07-08T20:19:52Z,0,"Issue: #189161 What: Add an ""fmax"" reduction type (non-NaN-propagating max) and use it in the softmax two-pass fallback path (prepare_softmax_twopass_fallback). Why: The softmax-internal max only stabilizes the computation; its result is never user-visible. NaN still propagates through exp(x - xmax) into the final o..."
188681,"[Inductor] Use CUDA toolkit's libdevice by default, fallback to Triton's bundled if not found",https://github.com/pytorch/pytorch/pull/188681,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,188672,,,,0,1,1,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-01T15:47:33Z,2026-07-08T20:19:31Z,0,"For #188672 The main motivation is to use the most current compatible CUDA libdevice.10.bc available on the system. Triton bundles its own copy of libdevice.10.bc, but that bundled copy can lag the installed CUDA toolkit. When it lags, Inductor-generated Triton kernels may miss fixes, numerical updates, or target-sp..."
189174,Fix dcp privateuse1 stream support,https://github.com/pytorch/pytorch/pull/189174,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/torchtitan,review when ready,,UNKNOWN,False,0,0,,,165097,165097,165097,,,,0,2,2,,,"triaged, open source, release notes: distributed (checkpoint), ciflow/torchtitan",2026-07-07T17:21:21Z,2026-07-08T20:19:18Z,0,"Fixes: #165097 Summary: Replace hardcoded CUDA/PrivateUse1 availability checks in _write_files_from_queue with generic _get_available_device_type() and a hasattr(device_mod, ""current_stream"") capability."
189251,[Dynamo] Store backend tags in the registry,https://github.com/pytorch/pytorch/pull/189251,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-08T09:28:12Z,2026-07-08T20:18:59Z,0,"Why register_backend guards the _BACKENDS insert with compiler_fn not in _BACKENDS, but _BACKENDS is keyed by name, so the guard can never be false. Tags are stashed on the compiler function via setattr, which needs a type: ignore[attr-defined] at both the write and the read site, and leaves entry-point backends wit..."
189042,[dynamo] Add specialize_args option to assume_constant_result,https://github.com/pytorch/pytorch/pull/189042,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,14,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-06T16:36:39Z,2026-07-08T20:16:02Z,0,"Summary assume_constant_result guards a user-defined object argument only by GuardBuilder.ID_MATCH (object identity) and adds no guard for other argument types, relying on the argument already being a specialized/guarded constant. Consequently, passing a freshly allocated dataclass on every call makes ID_MATCH fail..."
189234,"Fix 1+erf catastrophic cancellation in gelu, Normal.cdf, and ndtr",https://github.com/pytorch/pytorch/pull/189234,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,187806,187806,187806,189291,,,0,1,3,,,"module: cpu, triaged, open source, release notes: mps, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-08T05:41:12Z,2026-07-08T20:12:59Z,0,"Fixes #187806. Stacked on #189291 (accurate Metal erfc): the shared erfc files sit in both PRs until it lands, then this one rebases and drops them. Root cause Exact gelu computed 0.5 * x * (1 + erf(x / sqrt(2))). For negative x, erf(...) approaches -1 and the addition cancels: fp32 is ~57% off at x = -5.5 and exact..."
189163,Enable expandable_segments on Windows (port NVML driver API as prerequisite),https://github.com/pytorch/pytorch/pull/189163,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,5,0,,,"triaged, open source, ciflow/trunk",2026-07-07T15:32:45Z,2026-07-08T20:01:21Z,0,"Enables expandable_segments:True for single-process allocation on Windows. As prerequisite, it ports the c10::cuda driver API to Windows so PYTORCH_C10_DRIVER_API_SUPPORTED can be defined."
186540,[MPS] Add segment_reduce support with native Metal kernels,https://github.com/pytorch/pytorch/pull/186540,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/mps",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,141287,,,,0,3,1,,,"triaged, open source, module: mps, release notes: mps, ciflow/mps, ciflow/inductor, release notes: inductor (aoti)",2026-06-07T12:05:15Z,2026-07-08T19:52:04Z,0,"Summary Native Metal kernel implementation of torch.segment_reduce for the MPS backend. Covers forward + backward, both lengths and offsets variants, and three dtypes (float32 / float16 / bfloat16). Up to 137× speedup over CPU on Apple Silicon. Implements both forward and backward of torch.segment_reduce on the MPS..."
180247,Migrate build system from setuptools to scikit-build-core,https://github.com/pytorch/pytorch/pull/180247,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/binaries, ciflow/docker, ciflow/dynamo, ciflow/inductor",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,"119526, 188898","180248, 180250",,,0,69,4,,,"open source, ciflow/binaries, ciflow/trunk, release notes: releng, ciflow/periodic, ciflow/inductor, ciflow/xpu, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/dynamo, ciflow/docker, ciflow/inductor-rocm-mi350",2026-04-13T18:02:09Z,2026-07-08T19:51:02Z,0,"Stack from ghstack (oldest at bottom): #180250 #180248 -> #180247 Switch pyproject.toml's build backend from setuptools to scikit-build-core (>= 1.0), which invokes CMake directly instead of having setup.py orchestrate it. This PR makes the backend swap plus the 1.0-native configuration it enables; the now-unused se..."
183910,[Test] Make test/nn/test_embedding.py device-agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/183910,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-15T15:35:25Z,2026-07-08T19:50:46Z,0,Summary Refactors test/nn/test_embedding.py to support PrivateUse1/OpenReg backends by removing hardcoded device lists and using device-agnostic test infrastructure. Changes Moved tests from TestEmbeddingNN to TestEmbeddingNNDeviceType: test_embedding_max_norm_unsorted_repeating_indices — was guarded by TEST_CUDA/TE...
185251,[Test] Move device-agnostic grouped_gemm tests to new test_matmul.py,https://github.com/pytorch/pytorch/pull/185251,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,1,,,"triaged, open source, topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-26T17:39:34Z,2026-07-08T19:50:28Z,0,"Summary Move device-agnostic grouped_gemm correctness tests from test_matmul_cuda.py to a new test_matmul.py file, making them available to out-of-tree backends. What moved test_grouped_gemm_2d_2d, test_grouped_gemm_2d_3d, test_grouped_gemm_3d_3d, test_grouped_gemm_3d_2d — correctness tests with striding/layout vari..."
184279,Avoid MultiOutput for single-output fallback kernels,https://github.com/pytorch/pytorch/pull/184279,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,136012,,136012,,,0,13,2,,,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, ci-no-td, release notes: inductor (aoti), ciflow/torchtitan, agentic",2026-05-18T21:47:49Z,2026-07-08T19:50:16Z,0,"Stack from ghstack (oldest at bottom): -> #184279 Represent single tensor FallbackKernel outputs directly instead of wrapping them in MultiOutputLayout, while preserving unbacked symbol handling and runtime-dispatch output naming. Fixes #136012 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Gu..."
186349,Improve Dynamo logger method graph-break guidance,https://github.com/pytorch/pytorch/pull/186349,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,132635,132635,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T08:24:12Z,2026-07-08T19:47:13Z,0,"Stack from ghstack (oldest at bottom): -> #186349 Dynamo already supports skipping logger methods when the exact bound method or class method is added to ignore_logging_functions / ignore_logger_methods. However, the graph-break message only said to add the logging method and did not explain that the call is skipped..."
186355,Fix AOTAutograd unused output backward pruning,https://github.com/pytorch/pytorch/pull/186355,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,131794,131794,,,,0,4,1,40760c6063f7a5cf0350e5671240a351f76b1549005e806920041390988a2ce6,4bca97412aed58761e7714f0a80a5e7fef19445b40456584e75b349bcbfb1c7e,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, module: compiled autograd, ciflow/torchtitan, agentic, needs-review",2026-06-05T09:57:06Z,2026-07-08T19:43:09Z,0,Stack from ghstack (oldest at bottom): -> #186355 AOTAutograd's custom autograd.Function previously allowed autograd to materialize undefined grad outputs as zero tensors before the compiled backward saw them. That meant a backward through only one differentiable forward output still fed zero tangents for every othe...
184121,Improve persistent tiled reduction defaults,https://github.com/pytorch/pytorch/pull/184121,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-perf-compare, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,149982,149982,,,,0,3,11,af2f2a0f0b113151559cb11597e8839b3471340fd081dbdb77fe2b53ec1900e6,b71b3422c62a420c0e47858604823c71f94bbc7e9e6fda56ac39aca7072750c4,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/inductor-perf-compare, ciflow/torchtitan, agentic, needs-review",2026-05-17T21:16:23Z,2026-07-08T19:39:04Z,0,"Stack from ghstack (oldest at bottom): -> #184121 Prefer stronger coalescing scores when tiled reduction block growth ties, and include a one-warp default config for small inner persistent tiled reductions so the default heuristic reaches the coordinate-descent winner for the dim1 block reduction case. Root cause: p..."
183884,Fix compile worker quiesce draining,https://github.com/pytorch/pytorch/pull/183884,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,7,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,"176968, 186136","186031, 186136",176968,,,2,10,11,2227b09d38ac086a99a9050d31db6198e7d6a8c2a84e0abcd2f8d8d9242d0af6,aca5df09b405904c1d353881f11c00bd9076129401c91da0e0d0743265228238,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, agentic, needs-review",2026-05-15T12:39:34Z,2026-07-08T19:35:57Z,0,"Stack from ghstack (oldest at bottom): -> #183884 Make SubprocMain quiesce drain the active inner executor before accepting later wakeup and job messages, preventing old executor threads from lingering across repeated quiesce cycles and blocking shutdown. Fixes #176968 Fixes #186136 Generated by my agent cc @voznese..."
188795,Fix `torch.asarray` to preserve an input's device when a default device is set,https://github.com/pytorch/pytorch/pull/188795,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,150199,150199,150199,,,,0,3,2,,,"triaged, open source, release notes: python_frontend, topic: bc breaking",2026-07-02T12:08:43Z,2026-07-08T19:33:54Z,0,Fixes #150199 torch.asarray preserves input's device when a default device is set except for Python scalars and sequences. _asarray_input_has_device helper checks whether the input already carries its own device and if true then DeviceContext does not override it.
186364,Prevent guard validation from mutating ShapeEnv,https://github.com/pytorch/pytorch/pull/186364,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,129936,129936,,,,0,1,1,,,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T12:07:02Z,2026-07-08T19:32:00Z,0,Stack from ghstack (oldest at bottom): -> #186364 Dynamo validates a freshly-built guard manager before the compile has fully committed. For SymBool inputs this validation can execute a guard expression that reaches back into the caller-owned outer ShapeEnv through the frame locals. That evaluation may install speci...
186372,Fix torch.Tensor construction from traced NumPy arrays,https://github.com/pytorch/pytorch/pull/186372,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,129648,129648,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T14:07:02Z,2026-07-08T19:28:15Z,0,"Stack from ghstack (oldest at bottom): -> #186372 Dynamo handles torch.Tensor as an in-graph class constructor. When nested NumPy ndarray results are traced, those ndarray leaves become fake tensor-backed variables, and fake propagation used to call torch.Tensor directly on the nested list. That path attempts Python..."
186406,Fix invalid source in AOT repro generation,https://github.com/pytorch/pytorch/pull/186406,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,128830,128830,,,,0,1,1,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T20:00:46Z,2026-07-08T19:24:07Z,0,"Stack from ghstack (oldest at bottom): -> #186406 AOT repro generation could emit text that is not valid Python source. Config codegen treated any callable with __module__ and __name__ as sourceable, so lambdas and other non-importable callables could become expressions like __main__.<lambda>. Separately, nested FX..."
186854,Simplify endianness check using std::endian,https://github.com/pytorch/pytorch/pull/186854,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/trunk, ciflow/vllm",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,176662,,,,0,3,1,,,"open source, ciflow/trunk, topic: not user facing, ciflow/vllm",2026-06-10T00:25:40Z,2026-07-08T19:23:37Z,0,std::endian is available since C++20 #176662
188440,Simplify atomic_add_float using atomic_ref,https://github.com/pytorch/pytorch/pull/188440,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/android, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,176662,,,,0,3,2,,,"module: cpu, triaged, open source, ciflow/trunk, topic: not user facing, ciflow/android, ciflow/inductor",2026-06-29T16:36:25Z,2026-07-08T19:22:53Z,0,#176662 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
186415,Fix Dynamo Parameter construction in new modules,https://github.com/pytorch/pytorch/pull/186415,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,128556,128556,,,,0,3,3,0273f69529d5fd213a53e03eb96e6164775c0fca5fd12a03ed22871352382679,2acdada02c09c190ced9ffd46f152d794cfcb3d7ce47e0063f249a460f6785ea,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-05T20:44:21Z,2026-07-08T19:21:06Z,0,"Stack from ghstack (oldest at bottom): -> #186415 Dynamo disables the traceable nn.Parameter constructor path while it instantiates general user-defined classes. That is still the right default for ordinary sourceless Parameter construction because the traceable path is behind graph_break_on_nn_param_ctor, but it is..."
186418,Fix AOT export GraphModule serialization,https://github.com/pytorch/pytorch/pull/186418,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,128554,128554,,,,0,1,1,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-06-05T21:06:18Z,2026-07-08T19:16:31Z,0,Stack from ghstack (oldest at bottom): -> #186418 GraphModules produced by aot_export_module can record _ModuleStackTracer as their graph tracer. FX GraphModule pickle loading reconstructs the graph by instantiating the saved tracer class with no arguments and retracing the generated forward source. That failed for...
187307,Fix the potential global range overflow issue,https://github.com/pytorch/pytorch/pull/187307,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, ciflow/trunk, ciflow/inductor, ciflow/xpu, release notes: inductor (aoti)",2026-06-15T06:25:32Z,2026-07-08T19:13:28Z,0,Stack from ghstack (oldest at bottom): -> #187307 Motivation globalRangeX will be truncated if gridX * threadsPerWarp * numWarps overflows. This bug is found in https://github.com/pytorch/pytorch/pull/187137/changes#diff-6b55ea9919f49e90d28a14bbacc3793c4b3ba932603c141b9452d333c4585642R6640-R6643 The same issue needs...
186403,Avoid CUDA init during CPU torch.compile,https://github.com/pytorch/pytorch/pull/186403,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,129131,129131,,,,0,9,1,,,"Merged, Reverted, ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ci-no-td, ciflow/torchtitan, agentic",2026-06-05T19:40:54Z,2026-07-08T19:13:22Z,0,"Stack from ghstack (oldest at bottom): -> #186403 CPU-only torch.compile could initialize CUDA before any CUDA tensor was used. The root cause was that several Dynamo and Inductor compile-time probes used CUDA availability as a proxy for whether it was safe to query CUDA state: RNG preservation read CUDA RNG state,..."
189076,[inductor] null-guard size/alignment asserts for optional outputs,https://github.com/pytorch/pytorch/pull/189076,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,184114,,,0,3,1,,,"ciflow/trunk, topic: bug fixes, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-07-06T19:30:20Z,2026-07-08T19:13:05Z,0,"Stack from ghstack (oldest at bottom): -> #189076 A fallback op's optional (Tensor?) output can be null at runtime even when the fake kernel returned a tensor. cpp_wrapper emitted an unconditional assert_size_stride on that null handle -> SIGSEGV, and the Python wrapper's assert_size_stride / assert_alignment would..."
189273,Fix nn.functional.threshold OpInfo reference for strict NumPy uint8 casting,https://github.com/pytorch/pytorch/pull/189273,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,189267,189267,189267,,,,0,1,3,,,"open source, ciflow/trunk, release notes: python_frontend, keep-going",2026-07-08T16:02:52Z,2026-07-08T19:12:52Z,0,Fixes #189267
180250,Add setup_.py shim for legacy build commands,https://github.com/pytorch/pytorch/pull/180250,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"180247, 180248",,,0,68,2,,,"open source, ciflow/binaries, ciflow/trunk, release notes: releng",2026-04-13T18:02:32Z,2026-07-08T19:12:48Z,0,"Stack from ghstack (oldest at bottom): -> #180250 #180248 #180247 With setup.py removed (#180248), workflows and habits that still invoke python setup.py {install,develop,bdist_wheel,clean} would simply fail. Provide a shim that forwards those to their modern equivalents (pip install, python -m build, spin clean) an..."
180248,Remove setup.py and setuptools build helpers,https://github.com/pytorch/pytorch/pull/180248,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/binaries,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"180247, 180250",,,0,71,2,,,"open source, ciflow/binaries, release notes: releng, topic: not user facing, suppress-bc-linter",2026-04-13T18:02:16Z,2026-07-08T19:12:47Z,0,"Stack from ghstack (oldest at bottom): #180250 -> #180248 #180247 With the build fully driven by scikit-build-core (#180247), the setuptools path is dead code. Remove setup.py and its helpers (build_pytorch_libs, setup_helpers.{cmake,cmake_utils,env}, test_cmake). Retain a minimal build_libtorch.py -- it calls cmake..."
186849,Fix CuTeDSL grouped_mm descriptor init shape,https://github.com/pytorch/pytorch/pull/186849,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,176178,176178,,,,0,3,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-09T22:30:36Z,2026-07-08T19:12:36Z,0,"Stack from ghstack (oldest at bottom): -> #186849 GroupedGemmKernel initializes TMA descriptors from the tensors passed to cute.compile before updating them to each runtime group. Passing the full packed A/C tensors can create descriptors whose initial shape differs from the per-group shapes, which hits illegal memo..."
189258,[inductor] Replace ExternKernel/ComputedBuffer plumbing tuples with dataclasses + Protocol,https://github.com/pytorch/pytorch/pull/189258,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,2,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-08T14:08:58Z,2026-07-08T19:12:16Z,0,"Stack from ghstack (oldest at bottom): -> #189258 This replaces three loosely-typed plumbing values in the Inductor IR layer with precise types, so the shapes are documented at the definition site and checked by the type checker rather than reconstructed by hand at each call site. ProcessKernelResult replaces the po..."
184261,[Test] Make test_serialization.py device-agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/184261,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,6,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-05-18T19:38:14Z,2026-07-08T19:12:10Z,0,"Summary Refactors test/test_serialization.py to separate device-dependent tests from CPU-only tests, following the same pattern used in test_dropout.py and test_convolution.py. Creates a new TestSerializationDeviceType(TestCase) class registered with instantiate_device_type_tests, enabling automatic test generation..."
185150,Preserve OrderedDict outputs in FX tracing,https://github.com/pytorch/pytorch/pull/185150,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,161807,161807,,,,0,6,5,,,"release notes: fx, agentic",2026-05-25T18:15:25Z,2026-07-08T19:12:06Z,0,"Stack from ghstack (oldest at bottom): -> #185150 FX TracerBase.create_arg previously handled OrderedDict through the generic dict path. Output nodes normalize dict aggregates to immutable_dict, so symbolic tracing a module that returns an OrderedDict produced a GraphModule that returned a plain dict. optimize_for_i..."
186425,Skip Dynamo inside non-traceable subclass dispatch hooks,https://github.com/pytorch/pytorch/pull/186425,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,128160,128160,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T22:34:01Z,2026-07-08T19:08:07Z,0,"Stack from ghstack (oldest at bottom): -> #186425 Dynamo already graph breaks when an operator sees a tensor subclass with a custom torch_dispatch that has not opted into the traceable wrapper subclass protocol. The eager fallback still ran with Dynamo's frame callback active, so the hook body itself could be picked..."
184319,[inductor] Align split reductions for tiled pointwise fusion,https://github.com/pytorch/pytorch/pull/184319,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,128063,128063,,,,0,4,8,6b4796a783c1e71cc32b7892af1f434e78b4fd4a6f865903e71bbedb186ff65b,6a8b63412d9d2e342909f768ec9ec0324311fc0b04f71f8b94b9ad635423303c,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T06:29:48Z,2026-07-08T19:02:41Z,0,Stack from ghstack (oldest at bottom): -> #184319 Prefer an original reduction-dimension prefix as the split count when a full reduction has a sibling pointwise chain that can expose tiled fusion. This lets the first split-reduction level share the pointwise tiling instead of blocking fusion on incompatible flat spl...
186429,Fix ONNX export of pad_sequence with symbolic split lengths,https://github.com/pytorch/pytorch/pull/186429,ready for maintainer decision,112,"needs-review label present; approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,127153,127153,,,,0,1,1,,,"ciflow/trunk, release notes: onnx, ciflow/inductor, agentic, needs-review",2026-06-05T23:16:33Z,2026-07-08T18:59:33Z,0,"Stack from ghstack (oldest at bottom): -> #186429 The pad_sequence decomposition computed the output length with Python max() over the input sequence lengths. When the inputs come from torch.tensor_split(text, batch_lengths), those lengths are unbacked symbolic values derived from data. Python max() compares them by..."
186431,Fix ModuleTracker with Dynamo runtime modules,https://github.com/pytorch/pytorch/pull/186431,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,127112,127112,,,,0,1,1,,,"topic: not user facing, agentic, needs-review",2026-06-05T23:42:43Z,2026-07-08T18:56:47Z,0,"Stack from ghstack (oldest at bottom): -> #186431 ModuleTracker uses global module hooks to maintain the active module hierarchy. When tests run under Dynamo, compiler-owned modules also receive those global hooks. Dynamo runtime OptimizedModule and GraphModule wrappers were therefore recorded as if they were user m..."
186443,Validate random_ and uniform_ meta arguments,https://github.com/pytorch/pytorch/pull/186443,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,126834,126834,,,,0,3,1,,,"topic: not user facing, agentic, needs-review",2026-06-06T03:14:44Z,2026-07-08T18:53:27Z,0,Stack from ghstack (oldest at bottom): -> #186443 Invalid random_ bounds were accepted by the Meta implementation because the from/to overloads returned self without running the native random_from_to_impl argument checks. This let fake/meta tracing record programs that eager execution would reject. Route random_ met...
188301,[inductor] Fix inductor dropping ordering dep between effectful ops with different kernel types,https://github.com/pytorch/pytorch/pull/188301,PR blocked,108,"failed checks or blocked/unstable merge state; approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,BLOCKED,False,0,0,,,,,,,,,0,3,2,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-26T21:55:08Z,2026-07-08T18:49:35Z,0,"with_effects records the ordering dependency between consecutive ORDERED-effectful ops in V.graph.additional_star_deps so the scheduler preserves their program order. The dep was keyed by new_op.get_name() (the buffer name), but the scheduler reads additional_star_deps back by operation name. When two effectful ops..."
184124,Fix Inductor multi-output replacement insertion order,https://github.com/pytorch/pytorch/pull/184124,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,148751,148751,,,,0,5,5,c13a807f7e90b40545910923ef2c0730af9e03f11e1ecaa3bb1e5af5104f29d6,4a4bde0545b1656204e3a5b1cc9c6630b54b169cfcfec3a834ef3198b94642d6,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T21:49:04Z,2026-07-08T18:48:16Z,0,Stack from ghstack (oldest at bottom): -> #184124 Inductor pattern replacement now inserts multi-output replacements at the earliest valid location after all replacement inputs and skips matches that would require reordering across external users. Fixes #148751 Generated by my agent cc @voznesenskym @penguinwu @Eika...
174699,Add CuTeDSL-based scaled grouped MM kernel,https://github.com/pytorch/pytorch/pull/174699,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/docker, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,33,0,,,"open source, topic: not user facing, matrix multiplication, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/docker",2026-02-10T19:03:11Z,2026-07-08T18:45:25Z,0,Stack from ghstack (oldest at bottom): -> #174699
184343,Prune slow Hopper WGMMA autotune configs,https://github.com/pytorch/pytorch/pull/184343,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,125277,125277,,,,0,3,1,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T09:44:11Z,2026-07-08T18:45:11Z,0,"Stack from ghstack (oldest at bottom): -> #184343 Reduce the default CUDA Triton GEMM search space on Hopper to avoid WGMMA configurations that trigger pathological ptxas compile times during max-autotune, while preserving non-Hopper behavior and exhaustive search. Add tests for target-device capability handling. Fi..."
188435,[dynamo] Model uninitialized closure cells in __build_class__,https://github.com/pytorch/pytorch/pull/188435,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-29T15:47:27Z,2026-07-08T18:44:50Z,0,"Stack from ghstack (oldest at bottom): -> #188435 A class body whose methods close over a free variable bound AFTER the class statement (or over the class's own name) carries that name in the class body's co_freevars, with a legitimately-empty closure cell at build_class time -- CPython only reads it later, when the..."
186969,Bump onnx to 1.22.0 (stable) and unify typing-extensions to 4.15.0,https://github.com/pytorch/pytorch/pull/186969,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/docker, ciflow/trunk",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,185955,185955,185955,185989,185989,,0,5,2,,,"module: onnx, module: ci, open source, ciflow/trunk, topic: not user facing, ciflow/docker",2026-06-10T20:19:56Z,2026-07-08T18:44:13Z,0,"Summary Integrates the ONNX 1.22.0 release by pinning onnx==1.22.0 (stable GA, now released on PyPI) in .ci/docker/requirements-ci.txt, and resolves the pip dependency conflict that the onnx bump exposed. This supersedes the earlier release-candidate pins (1.22.0rc1/1.22.0rc2). Root cause onnx 1.22.0 raised its depe..."
187171,[inductor] Fix test_comprehensive_addbmm_cpu_float16 precision mismatch (#187170),https://github.com/pytorch/pytorch/pull/187171,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,187170,187170,187170,,,,0,1,3,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-06-12T11:36:23Z,2026-07-08T18:44:12Z,0,"Summary Fix test_comprehensive_addbmm_cpu_float16 by comparing compiled f16 output against eager f16 output instead of an idealized f32 reference. Fixes #187170 Problem addbmm uses make_fallback in Inductor, meaning both compiled and eager paths call the same ATen kernel. The test was failing because it compared com..."
188105,Add support for custom attributes on Exceptions,https://github.com/pytorch/pytorch/pull/188105,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,,,188181,,,0,7,3,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-06-24T22:03:10Z,2026-07-08T18:44:03Z,0,Stack from ghstack (oldest at bottom): -> #188105 #188181 Authored with assistance from Claude (an AI assistant). Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang...
184347,[inductor] Add third reduction tile for strided block pointers,https://github.com/pytorch/pytorch/pull/184347,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,125077,125077,,,,0,1,3,,,"module: cpu, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-05-19T10:41:04Z,2026-07-08T18:41:19Z,0,Stack from ghstack (oldest at bottom): -> #184347 Allow reduction-only ND tiling to keep a third reduction dimension so odd-shaped discontiguous reductions can use Triton block pointers instead of collapsing logical dimensions into unsupported 1D-to-ND slices. Extend lazy compile metadata to carry R1/R2 block sizes...
186454,Warn on unsupported custom op schema types,https://github.com/pytorch/pytorch/pull/186454,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,124863,124863,,,,0,1,1,,,"release notes: python_frontend, agentic, needs-review",2026-06-06T07:36:31Z,2026-07-08T18:37:58Z,0,"Stack from ghstack (oldest at bottom): -> #186454 Custom operators could register schemas with tensor-bearing containers that are accepted by the schema parser but are not considered dispatch inputs by DispatchKeyExtractor, such as Tensor[][] or Tensor[]?. Those arguments can silently be ignored when computing the d..."
186455,Fix AOTAutograd Dynamo memory leak repro,https://github.com/pytorch/pytorch/pull/186455,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,124747,124747,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T08:11:50Z,2026-07-08T18:34:47Z,0,"Stack from ghstack (oldest at bottom): -> #186455 Issue #124747 failed when running the AOTAutograd memory leak test with PYTORCH_TEST_WITH_DYNAMO=1. There were two root causes. First, direct functorch aot_function capture did not disable compiled autograd, unlike the other AOTAutograd entry points, so compiled auto..."
186457,Fix torch.ops float parsing for scalar tensor args,https://github.com/pytorch/pytorch/pull/186457,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,124344,124344,,,,0,2,1,,,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T09:32:04Z,2026-07-08T18:30:28Z,0,"Stack from ghstack (oldest at bottom): -> #186457 Dynamo records torch.ops calls and fake-propagates them before runtime. For torch.ops bindings whose schema expects a float, eager accepts one-element tensor arguments by parsing them through Tensor.item() semantics. During Dynamo tracing, the same position contains..."
186459,Fix fake view strides for AOT group norm tracing,https://github.com/pytorch/pytorch/pull/186459,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,124181,124181,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-06T09:58:42Z,2026-07-08T18:25:42Z,0,"Stack from ghstack (oldest at bottom): -> #186459 AOTAutograd was tracing a group_norm graph that called native_group_norm without the contiguous conversion that eager group_norm performs first. The problem was not in group_norm itself: FakeTensor metadata for aten.view used _refs._reshape_view_helper, which preserv..."
186460,Support dynamic mode in Dynamo explain,https://github.com/pytorch/pytorch/pull/186460,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,124163,124163,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-06T10:10:31Z,2026-07-08T18:21:50Z,0,"Stack from ghstack (oldest at bottom): -> #186460 torch._dynamo.explain only forwarded outer kwargs to the target function. That made explain(fn, dynamic=True)(*args) pass dynamic to fn instead of to the internal optimize call, so explain could not inspect the same dynamic-shape graph that torch.compile(..., dynamic..."
186461,Fix SymBool equality handling in symbolic shapes,https://github.com/pytorch/pytorch/pull/186461,ready for maintainer decision,112,needs-review label present; approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,124110,124110,,,,0,2,1,,,"module: cpu, release notes: fx, ciflow/inductor, agentic, needs-review",2026-06-06T10:23:48Z,2026-07-08T18:17:51Z,0,"Stack from ghstack (oldest at bottom): -> #186461 SymBool equality can produce SymPy relations whose operands are themselves boolean relationals, such as Eq(Eq(u0, 1), Eq(u1, 1)). Several symbolic-shape helper paths assumed every relation had arithmetic operands and tried to subtract or order those operands, which c..."
185433,Fix FakeTensorMode handling for Inductor dynamic scalars,https://github.com/pytorch/pytorch/pull/185433,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,7,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,158088,"158088, 186031",,,,0,6,6,d0a896c43390c4e5f13452faebac0ffc6667919d8771d2e42f7eb91d827cf4e4,f5bc000b28a52ad6d91a2e550026c97902029f9e0908bc9a2d69062b76f39bd7,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-27T20:28:38Z,2026-07-08T18:14:18Z,0,"Stack from ghstack (oldest at bottom): -> #185433 Inductor lowers captured scalar reads such as Tensor.item() to DynamicScalar bindings in the generated Python wrapper. Those bindings read from real tensors created by the compiler, for example a buffer holding a lifted torch.tensor(...) value. When user code runs th..."
185776,Propagate fake RuntimeErrors to Dynamo handlers,https://github.com/pytorch/pytorch/pull/185776,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,148475,148475,,,,0,10,1,10043c8dec0c2d7a8fc60833844fa73c89569c8aaeb7e6628706833bd15aeef9,0db38197087a16ba90f20f0fa3890c78f4a23d40577a1f639d142ac2b408adad,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T00:16:47Z,2026-07-08T18:10:20Z,0,"Stack from ghstack (oldest at bottom): -> #185776 Dynamo computes example values for newly traced FX nodes by running them under fake tensor mode. If that fake execution raised a RuntimeError, get_fake_value always wrapped it as a hard TorchRuntimeError. That meant user code inside try/except RuntimeError never got..."
189276,Update third_party/kineto submodule to cb6e5f7,https://github.com/pytorch/pytorch/pull/189276,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,topic: not user facing,2026-07-08T16:40:19Z,2026-07-08T18:10:06Z,0,Includes the following commits: Rename vendored dynolog dir to prevent name collision (pytorch/kineto#1471) cb6e5f7 Fix: JSON-escape double quotes in chrome-trace event names (pytorch/kineto#1459) eba157a Pull shared test helpers into the test-utils namespace (pytorch/kineto#1470) 332e2ed Fix CQS signal modernize-us...
189265,test/inductor: relax FlexAttention invalid-block-size regex,https://github.com/pytorch/pytorch/pull/189265,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,186876,,,,,0,1,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor",2026-07-08T14:51:23Z,2026-07-08T18:09:53Z,0,"This updates test/inductor/test_flex_attention.py::test_invalid_block_size to avoid a brittle assertion on exact tuned tile sizes. The test previously expected: BLOCK_M=128 BLOCK_N=32 On ROCm/RDNA, the runtime may choose different forward tiles (e.g. 64x64) while still raising the same intended error: Q/KV block siz..."
185206,Make minifier_tests more device agnostic,https://github.com/pytorch/pytorch/pull/185206,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,11,0,,,"triaged, open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-26T09:39:30Z,2026-07-08T18:09:41Z,0,"Tried to decouple the always CPU <=> C++, CUDA <=> Triton assumption. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98"
186463,Split foreach bool inplace error samples,https://github.com/pytorch/pytorch/pull/186463,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,123663,123663,,,,0,1,1,,,"release notes: foreach_frontend, agentic, needs-review",2026-06-06T11:17:17Z,2026-07-08T18:05:31Z,0,"Stack from ghstack (oldest at bottom): -> #186463 Foreach OpInfo sample generation mixed valid bool inplace samples with scalar and scalar-list samples that correctly fail dtype promotion. test_dispatch_meta_inplace treated those real-op errors as broken OpInfo samples, so _foreach_add and _foreach_mul had to expect..."
184365,Memory aware partition selection for partitioned scatter fx pass,https://github.com/pytorch/pytorch/pull/184365,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-perf-test-nightly, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/inductor-rocm-mi355","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,168073,,,0,6,1,,,"triaged, open source, ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, ciflow/inductor-perf-test-nightly, ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/inductor-rocm-mi355, ciflow/torchtitan",2026-05-19T12:22:22Z,2026-07-08T18:04:47Z,0,Builds on the partitioned scatter FX pass (introduced in #168073) with a memory-aware partition selection strategy and improved contention gating. Replaces the previous static memory budget heuristic with per-node memory accounting modelled on overlap_scheduling.py The contention ratio is now computed as index_size...
186464,Fix singleton quotient divisibility replacement,https://github.com/pytorch/pytorch/pull/186464,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,123651,123651,,,,0,2,1,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-06-06T11:40:04Z,2026-07-08T18:02:24Z,0,"Stack from ghstack (oldest at bottom): -> #186464 Divisibility replacement can refine an unbacked size-like quotient down to a singleton such as 1. The previous range logic kept size-like symbols artificially widened when they were proven below 2, so the replacement target was still evaluated as if the quotient coul..."
178737,"Enable hipsparselt version, initCusparseltBindings and related UTs",https://github.com/pytorch/pytorch/pull/178737,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/rocm-nightly, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,0,0,,,,,,,,,0,14,2,,,"module: rocm, triaged, open source, ciflow/trunk, topic: not user facing, keep-going, ci-no-td, ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/rocm-nightly",2026-03-30T06:47:05Z,2026-07-08T17:58:34Z,0,This PR enables the following on ROCm >= 7.12 torch.backends.cusparselt.version() Enable initCusparseltBindings to avoid AttributeError: module 'torch._C' has no attribute '_cusparselt' Enable the below 31 UTs # hipSPARSELt allows INT8 sparse × INT8 dense matmul TestSparseSemiStructuredCUDA::test_mm_sparse_first_NN_...
186466,Fix scalar Tensor args for custom op float schemas in Dynamo,https://github.com/pytorch/pytorch/pull/186466,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,123470,123470,185132,,,0,2,2,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T12:37:25Z,2026-07-08T17:57:44Z,0,"Stack from ghstack (oldest at bottom): -> #186466 Dynamo represented a scalar Tensor passed to a custom operator's float schema argument as a FakeTensor during fake metadata execution. Eager dispatch accepts this pattern by converting a single-element Tensor to a Python float before calling the implementation, but D..."
186467,Preserve non-persistent buffers in compiled modules,https://github.com/pytorch/pytorch/pull/186467,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,123411,123411,,,,0,2,1,,,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T13:47:36Z,2026-07-08T17:54:05Z,0,"Stack from ghstack (oldest at bottom): -> #186467 torch.compile wraps nn.Modules in OptimizedModule, but the wrapper previously kept its own empty buffer metadata. Top-level buffers registered on the original module with persistent=False therefore disappeared from the compiled module's direct _buffers and _non_persi..."
186557,Resolve scalar overloads before AOT compilation,https://github.com/pytorch/pytorch/pull/186557,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,5,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,90923,90923,,,,0,8,1,6df3af6f3303f2f01fa13d2252acc8aeae9d87388acf1030238fae25220ff7c9,fd86ce38f42565d1ef4ed4e558caa69cfff41ae8a267098565287194ffe45468,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-07T18:10:17Z,2026-07-08T17:50:03Z,0,"Stack from ghstack (oldest at bottom): -> #186557 AOTAutograd can hand backend compilers FX graphs whose node target is a Tensor overload even though the graph still contains a Python scalar literal, for example aten.add.Tensor(x, 2). That mirrors what the Python arg parser sent to the dispatcher for operators that..."
186302,Install Rust toolchain in PyTorch CI/CD build images (#186302),https://github.com/pytorch/pytorch/pull/186302,draft/noise,13,"PR is draft; CI lane labels: ciflow/docker, ciflow/inductor",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"topic: not user facing, ciflow/inductor, meta-exported, ciflow/docker",2026-06-05T01:16:55Z,2026-07-08T17:49:26Z,0,"Summary: PyTorch's build now compiles the torch._rust extension via setup.py (cargo) and Buck, so cargo/rustc must be present in the build images. This adds a shared common/install_rust.sh (distro-agnostic: apt for the Ubuntu CI images, yum/dnf for the manylinux CD images) plus a ci_commit_pins/rust.txt version pin,..."
186470,Raise user error for joint parameter mutation export,https://github.com/pytorch/pytorch/pull/186470,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,122860,122860,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: export, agentic, needs-review",2026-06-06T14:41:14Z,2026-07-08T17:47:10Z,0,"Stack from ghstack (oldest at bottom): -> #186470 Joint forward/backward export does not support mutating module parameters. Only buffers and user inputs can be represented as mutable inputs in this path. When a parameter was mutated, the metadata still flowed into internal signature construction/remapping paths tha..."
186471,Handle symbolic tensors in overlap status,https://github.com/pytorch/pytorch/pull/186471,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,122773,122773,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-06T15:03:19Z,2026-07-08T17:42:15Z,0,"Stack from ghstack (oldest at bottom): -> #186471 get_overlap_status checked TensorImpl::numel() before considering whether either input had symbolic sizes or strides. FakeTensor uses symbolic metadata when Dynamo recompiles dynamic shapes, so gather.out meta reached assert_no_overlap(result, self) and failed while..."
185926,Preserve tuple arguments in export serde,https://github.com/pytorch/pytorch/pull/185926,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,145785,145785,,,,0,2,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-06-02T12:50:42Z,2026-07-08T17:39:36Z,0,"Stack from ghstack (oldest at bottom): -> #185926 Export serialization represented Python tuples with the same list-backed Argument variants used for Python lists. Deserialization therefore had no schema information telling it to rebuild a tuple, so higher-order operators without normal op schemas could load with di..."
189260,PERF: improve cuda/xpu Stream.__hash__,https://github.com/pytorch/pytorch/pull/189260,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: performance",2026-07-08T14:24:07Z,2026-07-08T17:38:31Z,0,"Summary torch.cuda.Stream.__hash__ and torch.xpu.Stream.__hash__ are defined as hash((self.cuda_stream, self.device)) / hash((self.sycl_queue, self.device)). self.device builds a fresh torch.device object on every call (plus a native-handle property read and a tuple allocation), so hashing a stream seems to have som..."
185929,Fix Dynamo benchmark model state isolation,https://github.com/pytorch/pytorch/pull/185929,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,145773,145773,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-02T13:23:35Z,2026-07-08T17:33:56Z,0,"Stack from ghstack (oldest at bottom): -> #185929 Dynamo performance benchmarks were intended to compare eager and compiled runs without letting one mode mutate the other mode's model state. The performance runner deep-copied the loaded model once, but then reused that single copy for eager warmup, compiled warmup,..."
185917,[Test] Refactor test_autocast.py: make behavior/cache tests device-agnostic,https://github.com/pytorch/pytorch/pull/185917,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,2,0,,,"open source, module: amp (automated mixed precision), topic: not user facing, module: tests/core",2026-06-02T11:06:53Z,2026-07-08T17:33:19Z,0,Stack from ghstack (oldest at bottom): -> #185917 Summary Create TestAutocastDeviceType(DeviceTypeTestBase) with device-generic autocast tests that automatically run on all available backends via instantiate_device_type_tests Merge TestAutocastGPU cache tests into the new class with @onlyAccelerator + @skipMPS Conso...
189233,[MPS] Fix avg_pool2d/adaptive_avg_pool2d backward on channels_last inputs,https://github.com/pytorch/pytorch/pull/189233,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,175190,175190,175190,,,,0,1,2,,,"triaged, open source, release notes: mps, ciflow/mps",2026-07-08T05:07:12Z,2026-07-08T17:32:04Z,0,Fixes #175190 Summary AvgPool2d / AdaptiveAvgPool2d backward on MPS either aborts the process (SIGABRT — buffer is not large enough from MPSNDArray) or silently produces wrong gradients when the input is channels_last. Forward works fine; only backward fails. Root cause pool2d_template takes a channels_last fast pat...
189045,Fix NaN gradients in nn.MultiheadAttention for fully-masked rows (#41…,https://github.com/pytorch/pytorch/pull/189045,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,41508,41508,41508,,,,0,2,2,,,"triaged, open source, release notes: nn",2026-07-06T17:02:57Z,2026-07-08T17:31:38Z,0,"---Fixes #41508 Root cause nn.MultiheadAttention produces NaN gradients when a key_padding_mask (or attn_mask) fully masks out every key for a given query row. Attention scores for that row become all -inf, and softmax(-inf, ..., -inf) is 0/0 NaN. That NaN propagates through backward, corrupting gradients for the wh..."
184858,Support runtime ConstDict lookups for dynamic keys,https://github.com/pytorch/pytorch/pull/184858,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,174541,174541,,,,0,6,6,015ebf21ba0d02d8af581033873f275567a6c556e7ef86f5fe2c3a1e3e67c8e3,c5816a9b0aa26be2e05e102b4f04f1d4ab350b8e1d71d77307a476dd94ad92c8,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-22T06:56:38Z,2026-07-08T17:31:22Z,0,"Stack from ghstack (oldest at bottom): -> #184858 ConstDictVariable previously resolved every read through HashableTracker. For source-backed scalar keys such as cache[self.key], that forced Dynamo to compare the concrete key value at trace time and emit value guards, so each module instance with a different float k..."
189215,[MPS] Add `largeTensorTest` to `test_group_norm_backward_large_input`,https://github.com/pytorch/pytorch/pull/189215,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"186604, 186605, 186866, 186867","186605, 186867",,,,0,3,2,,,"open source, release notes: mps, ciflow/mps",2026-07-08T00:20:05Z,2026-07-08T17:28:51Z,0,"Stack from ghstack (oldest at bottom): -> #189215 Fixes #186867 Fixes #186866 Fixes #186605 Fixes #186604 Follow-up from #188855, which did not fix the failures Updates largeTensorTest to synchronize and collect any unused MPS buffers and then check remaining MPS memory usage against the absolute maximum memory that..."
188682,Limit ROCm build targets to architectures covered by testing,https://github.com/pytorch/pytorch/pull/188682,draft/noise,13,"PR is draft; CI lane labels: ciflow/rocm-mi300, ciflow/rocm-mi350, ciflow/trunk",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,5,0,,,"module: rocm, open source, ciflow/trunk, topic: not user facing, ciflow/rocm-mi300, ciflow/rocm-mi350",2026-07-01T15:56:40Z,2026-07-08T17:25:27Z,0,"This PR makes the MI350 ROCm build compile only for gfx950 instead of inheriting the shared Docker image’s broader ROCm arch list. It adds a rocm-arch-list input to the reusable Linux build workflow, passes that value through as PYTORCH_ROCM_ARCH, and sets rocm-arch-list: gfx950 for the MI350 build paths. This lets..."
185947,Fix Dynamo logit with tensor eps,https://github.com/pytorch/pytorch/pull/185947,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,145596,145596,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-02T15:29:46Z,2026-07-08T17:24:18Z,0,"Stack from ghstack (oldest at bottom): -> #185947 Dynamo traced torch.logit(input, eps_tensor) directly even though the operator schema takes eps as a float? scalar. Eager accepts a 0-dim tensor there by scalarizing it. For float64 fake tensor propagation can use an item_memo-backed hint and specialize, but float32..."
185952,Support compiled async all_reduce wait,https://github.com/pytorch/pytorch/pull/185952,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,145574,145574,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T15:57:37Z,2026-07-08T17:20:37Z,0,"Stack from ghstack (oldest at bottom): -> #185952 Dynamo already rewrites synchronous dist.all_reduce calls to traceable functional collectives, but it rejected async_op=True before modeling the returned Work handle. This caused torch.compile(fullgraph=True) and strict export to fail for the common pattern dist.all_..."
185971,Fix Dynamo random.Random replay on cache hits,https://github.com/pytorch/pytorch/pull/185971,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,145445,145445,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T18:10:38Z,2026-07-08T17:07:28Z,0,"Stack from ghstack (oldest at bottom): -> #185971 Source-backed random.Random objects were handled like in-frame RNGs: Dynamo recorded random method calls as closures over the trace-time RNG state, then side-effect replay reset the live RNG object to the trace-time final state on every compiled invocation. The first..."
185972,Handle malformed MSVC help output,https://github.com/pytorch/pytorch/pull/185972,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,145383,145383,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T18:20:28Z,2026-07-08T17:03:22Z,0,"Stack from ghstack (oldest at bottom): -> #185972 _is_msvc_cl only needs to identify whether the compiler is Microsoft's cl.exe, but it decoded the full /help output before checking the first line. Some MSVC versions and Windows locale settings can emit bytes that are not valid for the selected Python decode codec,..."
189216,[Dynamo] Implement DELETE_DEREF,https://github.com/pytorch/pytorch/pull/189216,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-08T00:41:10Z,2026-07-08T17:02:56Z,0,"Stack from ghstack (oldest at bottom): -> #189216 Dynamo previously had no handler for the DELETE_DEREF bytecode, which is emitted for del x when x is a cell variable (captured by a nested function) or a free variable. This models cell deletion the same way CPython does: CPython's DELETE_DEREF does PyCell_SET(cell,..."
185984,Fix Dynamo out= resize for internal tensors,https://github.com/pytorch/pytorch/pull/185984,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,145220,145220,,,,0,3,10,aa7b347902e11f79b6c91cefa5876f63b76fa3893093f56ff7662751b928d283,f011c0dbdd4c0c9dcda220b2b1a880de894dee91a7f1e4075c5812f0f8727026,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T19:43:01Z,2026-07-08T17:00:33Z,0,"Stack from ghstack (oldest at bottom): -> #185984 Dynamo saved the pre-fake-propagation shape of out= tensors and unconditionally graph-broke if fake propagation resized them. That is necessary for graph-input out= tensors, because resizing them mutates tracked fake metadata used for guards and backend signatures. I..."
187509,test: add backend-agnostic c10d collective tests,https://github.com/pytorch/pytorch/pull/187509,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,187508,,,0,1,3,,,topic: not user facing,2026-06-16T22:23:07Z,2026-07-08T17:00:29Z,0,"Stack from ghstack (oldest at bottom): -> #187509 #187508 Add a MultiProcessTestCase suite that exercises the core distributed collective and point-to-point API across an explicit backend allowlist. The suite covers Gloo on CPU and NCCL plus nccl2 on CUDA, with operation coverage derived from the existing NCCL c10d..."
189142,fx: Fix incorrect return_annotation for tuple types in operator schemas,https://github.com/pytorch/pytorch/pull/189142,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,189106,189106,189106,,,,0,3,2,,,"open source, release notes: fx",2026-07-07T11:02:17Z,2026-07-08T17:00:08Z,0,"Fixes #189106 Description Currently, get_signature_for_torch_op returns incorrect annotations for operations that return multiple tensors (e.g., aten.var_mean.default). Instead of returning standard tuple[torch.Tensor, torch.Tensor], it returns a raw Python tuple of types (<class 'torch.Tensor'>, <class 'torch.Tenso..."
187508,c10d: register nccl2 backend,https://github.com/pytorch/pytorch/pull/187508,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,187509,,,0,1,3,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-06-16T22:23:04Z,2026-07-08T17:00:06Z,0,"Stack from ghstack (oldest at bottom): #187509 -> #187508 Add ProcessGroupNCCL2 under c10d/nccl2 as a Backend implementation that delegates the core collective, point-to-point, barrier, split, allocator, and lifecycle calls to ProcessGroupNCCL while reporting a distinct nccl2 backend name. Register nccl2 through tor..."
189165,[aot_compile] Add torch.compiler.precompile,https://github.com/pytorch/pytorch/pull/189165,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187859, 188366, 188367, 188376",,,0,4,4,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T16:13:12Z,2026-07-08T16:58:43Z,0,"Stack from ghstack (oldest at bottom): -> #189165 #187859 #188376 #188367 #188366 Adds torch.compiler.precompile, an ahead-of-time precompile API that captures a whole computation with make_fx and lowers it to a self-contained, runnable Python source string plus an acceleration cache: python_code, cache = torch.comp..."
188366,[functorch] Inline codegen'd epilogue fns directly into the runtime wrapper,https://github.com/pytorch/pytorch/pull/188366,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"187859, 188367, 188376, 189165",,,0,8,4,,,"topic: not user facing, ciflow/inductor",2026-06-28T15:06:29Z,2026-07-08T16:57:44Z,0,"Stack from ghstack (oldest at bottom): #189165 #187859 #188376 #188367 -> #188366 _create_runtime_wrapper codegens the output-alias (_alias_fn) and input-mutation (_apply_mutations) epilogue helpers as standalone functions, then took a needless detour: it wrapped each in a dynamically-created bound-method shim (type..."
188367,[functorch] Extract _compile_and_exec_source into a generic codegen.py,https://github.com/pytorch/pytorch/pull/188367,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"187859, 188366, 188376, 189165",,,0,8,4,,,ciflow/inductor,2026-06-28T15:39:06Z,2026-07-08T16:57:43Z,0,"Stack from ghstack (oldest at bottom): #189165 #187859 #188376 -> #188367 #188366 _compile_and_exec_source -- the chokepoint that compiles a generated wrapper source string into a live function -- lived in subclass_codegen.py, but it is not subclass-specific: runtime_wrappers.py reaches into the ""subclass"" module 13..."
188376,[functorch] Add source_emit: reconstruct live values as standalone source,https://github.com/pytorch/pytorch/pull/188376,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"187859, 188366, 188367, 189165",,,0,10,4,,,ciflow/inductor,2026-06-28T20:57:49Z,2026-07-08T16:57:40Z,0,"Stack from ghstack (oldest at bottom): #189165 #187859 -> #188376 #188367 #188366 source_emit.py is a small leaf module with one job: given a live Python object, return a Python expression string that, when exec'd, reconstructs an EQUAL value -- emitted as readable, auditable source (never a pickle / base64 blob) --..."
187859,[functorch] Add torch._functorch.aot_autograd.compile_to_python,https://github.com/pytorch/pytorch/pull/187859,draft/noise,13,PR is draft; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"188366, 188367, 188376, 189165",,,0,36,4,,,ciflow/inductor,2026-06-22T17:58:41Z,2026-07-08T16:57:39Z,0,"Stack from ghstack (oldest at bottom): #189165 -> #187859 #188376 #188367 #188366 Add torch._functorch.aot_autograd.compile_to_python(gm, example_inputs) -> (python, cache), the outer half of the backend contract behind torch.compiler.precompile. It composes AOTAutograd's own codegen'd runtime wrappers (the prelude/..."
186007,Preserve eager post accumulate hook ordering in AOTAutograd,https://github.com/pytorch/pytorch/pull/186007,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,144376,144376,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, module: compiled autograd, agentic, needs-review",2026-06-02T23:09:49Z,2026-07-08T16:56:34Z,0,"Stack from ghstack (oldest at bottom): -> #186007 AOTAutograd returns backward gradients in forward input order, which is required to map each gradient to its input. The autograd engine also used that output order to schedule next edges, so a collapsed Python autograd.Function could fire post_accumulate_grad_hook ca..."
184094,Clone aliased mm/addmm inputs on CPU to avoid in-place gemm,https://github.com/pytorch/pytorch/pull/184094,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,85852,85852,85852,,,,0,1,2,,,"triaged, open source, release notes: linalg_frontend",2026-05-17T13:27:28Z,2026-07-08T16:56:12Z,0,"Issue Fixes #85852 Summary addmm_impl_cpu_ delegates to BLAS gemm, which cannot operate in-place on its input matrices. When a caller passes the same tensor for both an input and the output (e.g. torch.mm(C, C, out=C)), gemm reads from and writes to the same buffer, silently producing corrupt results. The existing a..."
186011,[dynamo] skip empty graph fallback recursively,https://github.com/pytorch/pytorch/pull/186011,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,144360,"144360, 186031",,,,0,4,2,7e243494cb93f44dfca0bcb3fec30c27a2b052af58acfcf84bca5c8d3ea8c518,39fd3ce3cb89c2b84c754a99328a23df6222eebe298aab6f65aad0189f01252b,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T23:32:23Z,2026-07-08T16:52:11Z,0,"Stack from ghstack (oldest at bottom): -> #186011 When Dynamo skips a frame because tracing produced an empty FX graph, it falls back to eager execution of that frame. Previously this used the generic SkipFrame path, whose execution strategy skipped only the current code object and left recursive frames at DEFAULT...."
188472,[inductor] Bucket all_reduce in the manual overlap scheduler,https://github.com/pytorch/pytorch/pull/188472,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-29T19:31:53Z,2026-07-08T16:51:13Z,0,The manual overlap bucketer (ManualOverlapPreservingBucketer) only handled all_gather and reduce_scatter and silently skipped all_reduce. As a result the HSDP replicate-axis / DDP gradient all_reduce was left unbucketed (one launch per parameter) and never scheduled to overlap compute. Teach the manual scheduler to...
186016,Reject invalid bitwise dtypes in Inductor lowering,https://github.com/pytorch/pytorch/pull/186016,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,144247,144247,,,,0,2,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T00:02:59Z,2026-07-08T16:48:18Z,0,"Stack from ghstack (oldest at bottom): -> #186016 CPU Inductor lowered bitwise pointwise ops without checking that tensor inputs had integer or Boolean dtypes. Invalid floating-point inputs could then reach backend code generation; on the current checkout this produces invalid C++ such as float & float, and the orig..."
184963,Fix FakeTensor shallow copy state,https://github.com/pytorch/pytorch/pull/184963,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,171977,171977,,,,0,8,5,9ecda0f70b34c45c037549593d2ba21136947e4b6a797ee03f5316eb2cdd83fc,7c4a0621eb4d8888c7f46177dcb2651f98520bf20e31bde5e1eb1d207d9b616f,"module: dynamo, ciflow/inductor, release notes: export, ciflow/torchtitan, agentic, needs-review",2026-05-23T04:57:06Z,2026-07-08T16:46:07Z,0,"Stack from ghstack (oldest at bottom): -> #184963 copy.copy(tensor) is a Python operation, not an ATen op. In non-strict torch.export, inputs are already FakeTensors when user Python runs under make_fx, so stdlib copy.copy fell through Tensor serialization/rebuild machinery. That rebuilt the fake tensor wrapper with..."
186021,[Inductor][CPU] Guard compiled integer div by zero,https://github.com/pytorch/pytorch/pull/186021,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,143649,143649,,,,0,2,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T01:14:46Z,2026-07-08T16:39:16Z,0,"Stack from ghstack (oldest at bottom): -> #186021 CPU Inductor lowered integer floor division and trunc division to raw C++ / and % expressions. When a compiled kernel saw a zero divisor, the hardware raised SIGFPE before PyTorch could turn the condition into a Python exception. torch.fmod and torch.remainder alread..."
186833,Remove stale export OpInfo xfails,https://github.com/pytorch/pytorch/pull/186833,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,170539,170539,,,,0,7,1,,,"topic: not user facing, agentic, needs-review",2026-06-09T20:01:23Z,2026-07-08T16:34:31Z,0,"Stack from ghstack (oldest at bottom): -> #186833 The affected OpInfo export tests now pass, but test_export_opinfo still marked them as expected failures. That made the test harness report successful runs as unexpected successes instead of letting the cases run normally. Remove only the stale xfail markers for the..."
186472,Graph break on dynamic unspecialized ModuleList indexing,https://github.com/pytorch/pytorch/pull/186472,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,122578,122578,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T15:30:03Z,2026-07-08T16:31:06Z,0,"Stack from ghstack (oldest at bottom): -> #186472 Fixes #122578 When Dynamo sees mutation on an nn.Module, it restarts analysis and tracks the module as an UnspecializedNNModuleVariable. For ModuleList-like containers, self[self._pos] then fell through to the generic user-object subscript path. That path inlined the..."
186478,Fix AOTAutograd resized out tensor mutation replay,https://github.com/pytorch/pytorch/pull/186478,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,122386,122386,,,,0,3,1,,,"topic: not user facing, module: inductor, ciflow/inductor, agentic, needs-review",2026-06-06T17:37:50Z,2026-07-08T16:27:56Z,0,"Stack from ghstack (oldest at bottom): -> #186478 Auto-functionalized out= operators can replace a functional wrapper with a result that has different size and stride metadata. AOTAutograd's metadata mutation analysis trusted the functionalization metadata-mutation flag, but this flag is not set for the replace_ pat..."
186487,Run Dynamo compilation with empty functorch stack,https://github.com/pytorch/pytorch/pull/186487,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,122200,122200,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T19:41:18Z,2026-07-08T16:24:08Z,0,"Stack from ghstack (oldest at bottom): -> #186487 Dynamo already disables Python modes while compiling, but it did not clear functorch dynamic layers. Calling torch.compile from an ambient functorch context, such as torch.func.functionalize, therefore left the functorch interpreter stack active during Dynamo tracing..."
186488,Improve schema errors for symbolic int list mismatches,https://github.com/pytorch/pytorch/pull/186488,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,122129,122129,,,,0,1,1,,,"release notes: jit, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-06T20:14:25Z,2026-07-08T16:20:31Z,0,"Stack from ghstack (oldest at bottom): -> #186488 When Dynamo routes a Python list through FX, the list can become an immutable_list. If a custom operator schema expects int[] but receives symbolic integer elements from capture_scalar_outputs, schema conversion fails before the user can see that the element type is..."
186248,Simplify Min/Max using ShapeEnv value ranges,https://github.com/pytorch/pytorch/pull/186248,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,137096,137096,,,,0,5,1,,,"ciflow/trunk, release notes: fx, module: inductor, module: dynamo, ciflow/inductor, agentic",2026-06-04T18:13:57Z,2026-07-08T16:17:54Z,0,"Stack from ghstack (oldest at bottom): -> #186248 ShapeEnv.simplify previously only had narrow handling for Max(0, x) and Max(1, x), so value-range facts such as torch._check(u0 <= 5) were not used to simplify Min(u0, 5). That left guards like Eq(Min(5, u0), u0) data-dependent in Dynamo even though the ordering was..."
186918,[Test] Add test hardware classification for test classes,https://github.com/pytorch/pytorch/pull/186918,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,9,0,,,"open source, module: testing, ciflow/trunk, topic: not user facing",2026-06-10T08:45:30Z,2026-07-08T16:17:52Z,0,"Summary This PR introduces hardware classification for PyTorch test classes. Tests can now explicitly declare their hardware classification through a hw_classification class attribute, and users can filter which categories of tests to run via --hw-classification. When --hw-classification is not specified, all existi..."
186189,Report dynamo expected-failure XPASSes,https://github.com/pytorch/pytorch/pull/186189,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,139080,139080,,,,0,16,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic",2026-06-04T05:52:23Z,2026-07-08T16:17:51Z,0,Stack from ghstack (oldest at bottom): -> #186189 Direct PYTORCH_TEST_WITH_DYNAMO=1 runs of tests under test/dynamo were returning before invoking the common unittest runner because torch._dynamo.test_case.run_tests() treated Dynamo mode the same as crossref mode. That meant a test listed in test/dynamo_expected_fai...
184048,Enable Inductor OpInfo CPU coverage,https://github.com/pytorch/pytorch/pull/184048,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,161132,161132,,,,0,15,1,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, agentic",2026-05-16T21:00:51Z,2026-07-08T16:17:39Z,0,"Stack from ghstack (oldest at bottom): -> #184048 Enable CPU-generated Inductor OpInfo tests by replacing the global GPU/Triton gate with device-specific skips, and keep stale skipped timing data from preventing pytest sharding of the large OpInfo file. Fixes #161132 Generated by my agent cc @voznesenskym @penguinwu..."
188181,Move __dict__ view (get_dict_vt) onto base VariableTracker,https://github.com/pytorch/pytorch/pull/188181,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,188105,,,0,5,1,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-06-25T19:36:10Z,2026-07-08T16:17:37Z,0,"Stack from ghstack (oldest at bottom): #188105 -> #188181 Deduplicates the identical dict_vt field + get_dict_vt accessor that lived on UserDefinedObjectVariable, the user-function variables, and the unspecialized nn.Module variable. The lazily-created DunderDictVariable cache now lives on VariableTracker; per-VT __..."
189147,Fix unload utility for newer Triton pyds,https://github.com/pytorch/pytorch/pull/189147,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,"4037, 4038","4037, 4038",,,,,0,4,0,,,"open source, module: inductor, ciflow/inductor, ciflow/xpu, ciflow/torchtitan",2026-07-07T12:28:21Z,2026-07-08T16:17:35Z,0,unload_xpu_triton_pyds() hardcodes two intel-xpu-backend-for-triton internals that no longer exist in current backends: XPULauncher.mod was removed when the per-kernel launcher was replaced with a shared driver-level launcher (intel/intel-xpu-backend-for-triton#6650) XPUUtils's singleton attribute was renamed instan...
184223,Run Inductor fallbacks below autograd,https://github.com/pytorch/pytorch/pull/184223,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,139629,139629,,,,0,13,1,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic",2026-05-18T16:25:49Z,2026-07-08T16:17:34Z,0,"Stack from ghstack (oldest at bottom): -> #184223 Guard Inductor-generated fallback dispatcher calls with AutoDispatchBelowADInplaceOrView so fallback execution skips Autograd and ADInplaceOrView dispatch keys that AOTDispatcher already handled. Covers Python and C++ wrapper fallback paths, fallback out variants, an..."
186493,Optimize AOTAutograd compiled forward wrapper,https://github.com/pytorch/pytorch/pull/186493,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,122029,122029,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-06T21:27:08Z,2026-07-08T16:17:07Z,0,"Stack from ghstack (oldest at bottom): -> #186493 AOTAutograd already codegen'd the CompiledFunction.forward entry point, but the generated function still delegated every call to generic save/finalize helpers. For tiny compiled functions those helpers did mostly static or empty work on each invocation: slicing saved..."
189178,Add PyTorch AI policy,https://github.com/pytorch/pytorch/pull/189178,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,topic: not user facing,2026-07-07T17:47:25Z,2026-07-08T16:15:18Z,0,Stack from ghstack (oldest at bottom): -> #189178 Move the AI contribution guidance into a dedicated AI_POLICY.md file and leave CONTRIBUTING.md as the entry point for the broader contribution guide. Update issue and pull request templates so their AI-assisted development policy links go directly to the new policy d...
189038,[MPS] Fix uint8 casting in MPS `copy_` kernels,https://github.com/pytorch/pytorch/pull/189038,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,160744,160744,160744,,,,0,2,3,,,"triaged, open source, ciflow/mps",2026-07-06T16:04:28Z,2026-07-08T16:12:55Z,0,"Fixes #160744 On MPS, copy_ from a negative floating point source into a torch.uint8 destination can produce 0 for broadcasted shapes like [3] and [2, 2]. CPU and CUDA convert float -> uint8 through a signed integer intermediate, -1.0 becomes 255 MPS Metal helper was casting directly to uchar, which is platform dive..."
186494,[Dynamo] Constant-fold CUDA Tensor.is_pinned(),https://github.com/pytorch/pytorch/pull/186494,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,121526,121526,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T21:44:36Z,2026-07-08T16:12:01Z,0,"Stack from ghstack (oldest at bottom): -> #186494 Dynamo traced Tensor.is_pinned() as a generic tensor method call. In user Triton paths such as Triton's pinned-memory helper, this meant a CUDA fake tensor produced an aten.is_pinned bool result while tracing. Older builds failed while fake-running aten.is_pinned; cu..."
186495,Initialize compile context for non-strict export tracing,https://github.com/pytorch/pytorch/pull/186495,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,121318,121318,,,,0,1,1,,,"release notes: export, agentic, needs-review",2026-06-06T22:03:08Z,2026-07-08T16:09:24Z,0,Stack from ghstack (oldest at bottom): -> #186495 Non-strict export goes directly through the make_fx/AOT export path and does not enter Dynamo's compile-frame wrapper. Structured logging therefore sees no CompileContext for TORCH_TRACE records and falls back to emitting stack metadata without frame_id values. tlpar...
188047,fix(dlpack): raise Python RuntimeError instead of aborting for negative DLPack strides,https://github.com/pytorch/pytorch/pull/188047,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188023,188023,188023,,,,0,10,3,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-06-24T09:26:39Z,2026-07-08T16:06:27Z,0,"Fixes #188023 Summary torch.from_dlpack() aborted the entire Python process (SIGABRT / libc++abi: terminating due to uncaught exception) when given any array with negative strides (e.g. numpy_array[::-1]). The failure was silent at the library boundary—no Python RuntimeError was raised, so callers such as scikit-lea..."
184910,Reject direct aten.set in Dynamo,https://github.com/pytorch/pytorch/pull/184910,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), linux-jammy-py3.10-clang18 / test-osdc (crossref, 1, 2, mt-l-x86iavx512-8-64)",statusCheckRollup,,174176,174176,,,,8,5,1,4bdc884203e04b5c74f297d5afc71c2d78fe7ae684c70686eacdc10388754841,2ba1c80e37653dd4b5fc46e959aa47ae0a4f3287583a145ae7dfce8b5df06067,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-22T17:26:07Z,2026-07-08T16:05:11Z,0,"Stack from ghstack (oldest at bottom): -> #184910 Direct torch.ops.aten.set calls bypassed the Tensor.set_ method handling and entered the generic TorchInGraphFunctionVariable proxy path. That let Dynamo produce an FX graph containing aten.set followed by aten.lift, even though the out-of-place aten.set storage alia..."
185069,Fix AOTI Windows cross-compile package loading,https://github.com/pytorch/pytorch/pull/185069,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,166093,166093,,,,4,3,5,24e26cafc90177db3dffd0d19efccbbb81d452fe56f07b9b62adfda9e5e03ac0,741e3698b1cf52d17d6ab60c81a206112ca6a75683fcb236a4a9f99fb5a03336,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-05-24T13:08:20Z,2026-07-08T16:01:31Z,0,"Stack from ghstack (oldest at bottom): -> #185069 Linux-to-Windows AOTI cross-compilation was sending MinGW link-only runtime flags through every C++ builder path. That included precompiled-header generation, where the command is a header compile, not a link. With MinGW this can drive the toolchain into linker start..."
185701,[Test] Make test_packed_sequence.py device agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/185701,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,1,,,"triaged, open source, topic: not user facing, module: tests/core",2026-05-30T09:51:18Z,2026-07-08T15:58:34Z,0,"Summary Refactors test/nn/test_packed_sequence.py to break the tight coupling between tests and specific hardware (CUDA), enabling new accelerator backends to seamlessly reuse these tests. Split the monolithic PackedSequenceTest class into two: TestPackedSequenceCPU (6 tests): Pure validation/error-path tests with n..."
186498,Add dynamic shape test assertions for input dims,https://github.com/pytorch/pytorch/pull/186498,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,120911,120911,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, agentic, needs-review",2026-06-06T22:59:38Z,2026-07-08T15:58:12Z,0,"Stack from ghstack (oldest at bottom): -> #186498 The dynamic-shape Inductor harness only verified that generated code had some dynamic-shape evidence. It could not assert that specific user input dimensions stayed symbolic, so partial specialization such as one static input dim plus another symbolic dim could still..."
185918,[Test] Make autocast op-list tests device-agnostic via registry,https://github.com/pytorch/pytorch/pull/185918,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,185917,,,0,1,1,,,"open source, module: amp (automated mixed precision), module: tests/core",2026-06-02T11:06:59Z,2026-07-08T15:57:15Z,0,"Stack from ghstack (oldest at bottom): -> #185918 #185917 Summary Add test list registry to torch/testing/_internal/autocast_test_lists.py: register_autocast_test_lists(device_type, cls) for new backends get_autocast_test_lists(device_type, dev) to retrieve list by device ""cpu"" pre-registered with AutocastCPUTestLis..."
186897,Warn when custom op int schema specializes SymInts,https://github.com/pytorch/pytorch/pull/186897,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,154851,154851,,,,0,6,5,697fe96a0bd2c2a4bb3b408433135423afc3f28dddcc463801e1e4749aee81c0,31b535a37b7caa1e4c815994a23d517706c207a01d23a112a6c13c9fc907e350,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-06-10T05:44:20Z,2026-07-08T15:53:41Z,0,"Stack from ghstack (oldest at bottom): -> #186897 Fixes #154851 Generated by my agent Custom operators defined through torch.library preserve an important user-facing distinction in their schema text: int arguments require a concrete Python integer, while SymInt arguments can carry dynamic shape values. After the di..."
184380,Avoid fp32 materialization in checkpoint recompute,https://github.com/pytorch/pytorch/pull/184380,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,120381,120381,,,,0,1,5,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T14:49:07Z,2026-07-08T15:49:47Z,0,Stack from ghstack (oldest at bottom): -> #184380 Realize checkpoint recompute producers before they feed larger pointwise broadcasts or dtype upcasts. This keeps Inductor from materializing full-size fp32 intermediates in checkpoint backward graphs while leaving normal fusion behavior unchanged. Fixes #120381 Gener...
186500,Fix Dynamo resume after f-string logging graph breaks,https://github.com/pytorch/pytorch/pull/186500,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,120375,120375,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T23:18:21Z,2026-07-08T15:45:31Z,0,"Stack from ghstack (oldest at bottom): -> #186500 TorchDynamo translated FORMAT_VALUE conversion flags by immediately tracing str, repr, or ascii. For debug f-strings such as f""{a=}"" on tensors, that forced repr(tensor) during tracing. Tensor repr is intentionally unsupported because it depends on runtime values, so..."
184392,Fix ModularIndexing printer semantics,https://github.com/pytorch/pytorch/pull/184392,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"119883, 187027","119883, 187027",,,,0,2,2,,,"module: cpu, ciflow/mps, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T15:53:40Z,2026-07-08T15:41:30Z,0,"Stack from ghstack (oldest at bottom): -> #184392 Lower ModularIndexing through FloorDiv and PythonMod so Python, C++, Triton, and MPS printers share Python floor-division and modulo semantics, while preserving the MPS safe_mod workaround for the known Metal compiler case. Fixes #119883 Fixes #187027 Generated by my..."
186373,Disable duck sizing by default,https://github.com/pytorch/pytorch/pull/186373,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,129456,129456,,,,0,15,1,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, ciflow/dtensor, agentic",2026-06-05T14:24:45Z,2026-07-08T15:35:48Z,0,"Stack from ghstack (oldest at bottom): -> #186373 Duck sizing assigned the same symbolic variable to unrelated input dimensions when their first-run concrete values happened to match. With dynamic=True this emitted false equality guards, so later calls with the unrelated dimensions diverging recompiled even though t..."
189103,Support dist.split_group on CPU-only (gloo) process groups,https://github.com/pytorch/pytorch/pull/189103,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"Merged, Reverted, ciflow/trunk, release notes: distributed (c10d), ci-no-td, ciflow/torchtitan",2026-07-06T22:26:44Z,2026-07-08T15:35:21Z,0,"Stack from ghstack (oldest at bottom): -> #189103 split_group unconditionally required the default pg to have a bound device id and resolved the parent backend via torch.accelerator.current_accelerator(), so a cpu:gloo-only process group (or a bare ""gloo"" group on a machine without an eagerly-bound accelerator) coul..."
186283,Trace data_ptr through tensor constructors,https://github.com/pytorch/pytorch/pull/186283,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,136271,136271,,,,0,14,1,,,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic",2026-06-04T22:52:45Z,2026-07-08T15:35:09Z,0,"Stack from ghstack (oldest at bottom): -> #186283 Dynamo already represents tensor data pointer accesses as DataPtrVariable for identity comparisons and TMA descriptor plumbing, but that variable had no FX representation. As a result, code such as torch.tensor([x.data_ptr(), y.data_ptr()]) either graph-broke at the..."
181720,[MPS] Make pin_memory return CPU-aliased storage backed by a unified MTLBuffer,https://github.com/pytorch/pytorch/pull/181720,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,CLEAN,False,0,0,,,,"181374, 188970","181374, 188970",189256,,,0,4,7,,,"open source, topic: bug fixes, release notes: mps, ciflow/mps",2026-04-28T05:37:09Z,2026-07-08T15:34:47Z,0,"Stack from ghstack (oldest at bottom): -> #181720 #189256 torch.empty(..., device=""cpu"", pin_memory=True) and tensor.pin_memory() previously returned tensors whose .device was mps:0, because the MPS pinned memory allocator handed back DataPtrs tagged with the MPS device. This adds a thin pinned allocator that alloca..."
186502,Make context decorator frames identify the context manager,https://github.com/pytorch/pytorch/pull/186502,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,118747,118747,,,,0,2,1,,,"topic: not user facing, agentic",2026-06-07T00:13:38Z,2026-07-08T15:34:38Z,0,"Stack from ghstack (oldest at bottom): -> #186502 Python tracebacks use a frame's code object name, not the wrapper function's name metadata. torch.utils._contextlib.context_decorator used nested wrappers named decorate_context and generator_context for every context-manager decorator, so tracebacks could not distin..."
186503,Preserve input aliases for multi-output view returns,https://github.com/pytorch/pytorch/pull/186503,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,118739,118739,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-07T00:30:29Z,2026-07-08T15:31:42Z,0,"Stack from ghstack (oldest at bottom): -> #186503 AOTAutograd classified differentiable outputs from multi-output view ops as non-aliases when all aliases came from the same multi-output view. That is fine for intermediate tensors, but it is wrong when the outputs are views of graph inputs. Pure view functions like..."
186504,Allow HOP speculation through torch autograd.Function skipfiles,https://github.com/pytorch/pytorch/pull/186504,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,118334,118334,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T00:56:08Z,2026-07-08T15:28:59Z,0,"Stack from ghstack (oldest at bottom): -> #186504 Dynamo normally treats most framework files under torch/ as skipfiles. That is correct for ordinary tracing, but it also applied while higher-order operators speculatively trace an autograd.Function forward or backward to prove that the function can be represented as..."
181308,[CPU] Fix std_mean/var_mean all-reduce ~4x slower than std + mean (#122191),https://github.com/pytorch/pytorch/pull/181308,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,122191,122191,122191,43858,,,0,1,4,,,"triaged, open source, Stale, no-stale",2026-04-24T00:06:08Z,2026-07-08T15:27:39Z,0,"Fixes #122191. Summary torch.set_num_threads(1); x = torch.rand(128, 128, 128) torch.std_mean(x) # ~7 ms on M5 torch.std(x); torch.mean(x) # ~1.8 ms combined std_var_out has a CPU all-reduce fast path (added in #43858) that routes fp32/fp64 to a parallel two-pass (x - mean)^2 in std_var_all_cpu. std_var_mean_out doe..."
188650,[SymmMem] Allocate symmetric memory through the CUDACachingAllocator v2,https://github.com/pytorch/pytorch/pull/188650,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200-symm-mem, ciflow/h100-symm-mem, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: distributed (c10d), ciflow/h100-symm-mem, ciflow/b200-symm-mem, ciflow/torchtitan",2026-07-01T07:56:58Z,2026-07-08T15:26:33Z,0,"This is a follow up to my original PR #187054. The original descrioption: This PR makes symm_mem.empty() allocate through the CUDACachingAllocator when using the implicit symmetric memory pool. This lets symmetric-memory allocations participate in CUDAGraph private-pool reuse, enabling symm_mem.empty() inside CUDAGr..."
187236,Make Overlap Runtime Estimation Device Aware,https://github.com/pytorch/pytorch/pull/187236,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,4,1,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-13T04:54:11Z,2026-07-08T15:25:16Z,0,"This PR improves inductor’s overlap scheduling path for xpu and other noncuda accelerators . When analytical estimates are requested, the scheduler now stays on the analytical estimation path instead of falling back to device-specific runtime estimation, allowing overlap scheduling to run on xpu while keeping the ex..."
186506,Register missing CIA decomposition py_impls,https://github.com/pytorch/pytorch/pull/186506,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,117555,117555,,,,0,2,1,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-06-07T02:00:29Z,2026-07-08T15:23:46Z,0,"Stack from ghstack (oldest at bottom): -> #186506 Some Python decompositions were registered to DispatchKey.Autograd only. After AOTAutograd inference-mode tracing changes, those registrations can be skipped, so tracing may bypass the decomposition and hit lowerings that expect the decomp to have run. Mirror the saf..."
186507,Invalidate Dynamo caches when module hooks change,https://github.com/pytorch/pytorch/pull/186507,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,117265,117265,,,,0,2,7,0b6a76447346767a73987cf42235e5ff866840a3fa8e0f35f8f8aef7d4625d5e,2cb993b4047ca01c40b37c9f6fec8cf5e6456df9e12d769a7488795b5718b2fb,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T02:52:30Z,2026-07-08T15:20:55Z,0,"Stack from ghstack (oldest at bottom): -> #186507 Dynamo skips detailed guards for empty nn.Module hook dictionaries, and global hook dictionaries can be covered only by type guards. Public hook registration and RemovableHandle.remove() mutate those OrderedDicts in place, so the module dict version and existing guar..."
185976,Fix Dynamo autograd.Function backward grad mode tracing,https://github.com/pytorch/pytorch/pull/185976,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,145231,"145231, 186031",,,,0,4,10,698a4877df1604853f2e4117301f48073d80a63f2b212943312848b4a48f3b40,514e6b1020bbb2a3d843634d20ec0db6fe7b22e7d41d48dfa13f9783e64c8753,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T18:50:05Z,2026-07-08T15:18:05Z,0,"Stack from ghstack (oldest at bottom): -> #185976 TestAutograd.test_gradcheck_nondeterministic reaches this bug through gradgradcheck, which invokes torch.autograd.grad(..., create_graph=True) on NonDetFunc. Dynamo traces custom autograd.Function backward bodies while ambient grad mode is disabled. Before this chang..."
186510,Fix Dynamo low-precision addc opmath,https://github.com/pytorch/pytorch/pull/186510,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,4,1,"statusCheckRollup, linux-jammy-py3.14t-clang18 / test-osdc (default, 2, 5, mt-l-x86iavx512-16-128), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,116202,116202,,,,2,3,3,f70a17fca74d17def4ad8283b88e853f7da9b2450e3d274bddaa86e86ace7b39,064cefce309439317b23f7b1772d0cb72bac603ad36bae0ace6106dc4b802044,"module: dynamo, ciflow/inductor, release notes: optim, ciflow/torchtitan, agentic, needs-review",2026-06-07T03:26:27Z,2026-07-08T15:12:30Z,0,"Stack from ghstack (oldest at bottom): -> #186510 Dynamo decomposes Tensor.addcmul_ and Tensor.addcdiv_ when scalar tensor values would otherwise force a graph break. The decomposition used low precision mul/div inputs directly before feeding the result through prims.fma. On CUDA, ATen's addcmul/addcdiv kernels use..."
186512,[dynamo] Graph break on shape-changing Parameter.data assignment,https://github.com/pytorch/pytorch/pull/186512,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,115484,115484,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T03:57:26Z,2026-07-08T15:09:42Z,0,"Stack from ghstack (oldest at bottom): -> #186512 DeepSpeed ZeRO-3 style forward hooks gather partitioned parameters by assigning a full-sized tensor to Parameter.data before a submodule runs, then repartition the same parameter back to an empty tensor in a forward hook. Dynamo traced those hook-side Parameter.data..."
186517,Detect aliased subclass inputs after desugaring,https://github.com/pytorch/pytorch/pull/186517,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,114415,114415,,,,0,1,1,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-06-07T05:26:58Z,2026-07-08T15:05:12Z,0,Stack from ghstack (oldest at bottom): -> #186517 AOTAutograd detects duplicate and aliased input mutations before tensor subclasses are unwrapped. That misses a case where two distinct wrapper subclass inputs do not alias at the outer level but desugar to dense tensor inputs that share storage. If one of those inne...
185793,Add ONNX lowering for prims.remainder,https://github.com/pytorch/pytorch/pull/185793,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,147973,,,,,0,2,0,,,"release notes: onnx, agentic, needs-review",2026-06-01T05:44:38Z,2026-07-08T15:02:11Z,0,"Stack from ghstack (oldest at bottom): -> #185793 The Dynamo ONNX exporter can surface torch.ops.prims.remainder.default in exported programs, but torchlib did not register a lowering for that prim. Translation therefore failed with a DispatchError whenever a graph contained prims.remainder, including scalar/tensor..."
186520,Preserve SymInt bindings in dynamic-shape minifier repros,https://github.com/pytorch/pytorch/pull/186520,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,114296,114296,,,,0,1,1,,,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T05:51:09Z,2026-07-08T15:01:48Z,0,"Stack from ghstack (oldest at bottom): -> #186520 Dynamic-shape FX graphs can have SymInt placeholders whose only purpose is to bind a symbol used by FakeTensor metadata. The minifier's unused-input cleanup only looked at graph users, so it could erase one of these otherwise unused SymInt placeholders while leaving..."
189237,Add c-shims for aten.unsqueeze and aten.expand,https://github.com/pytorch/pytorch/pull/189237,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,,,,,,,0,1,0,,,"open source, ciflow/inductor, release notes: inductor (aoti)",2026-07-08T06:11:48Z,2026-07-08T14:58:22Z,0,"AOTInductor was missing c-shim entries for aten.unsqueeze.default and aten.expand.default, so these fallback ops could not use direct shim calls. Register both ops in inductor_fallback_ops and regenerate the exported shim headers and version map for CPU, CUDA, MPS, and XPU."
184432,Serialize cudagraph tree runtime,https://github.com/pytorch/pytorch/pull/184432,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,113809,113809,,,,0,2,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T20:52:45Z,2026-07-08T14:57:12Z,0,"Stack from ghstack (oldest at bottom): -> #184432 CUDAGraphTreeManager state and its CUDA graph private pool are shared with autograd worker threads, so serialize manager runtime entry to avoid concurrent pool recording from compiled backward. Add a regression test for overlapping cudagraph tree runs. Fixes #113809..."
184436,Fix Triton async compile cache keys,https://github.com/pytorch/pytorch/pull/184436,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,112788,112788,,,,0,3,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T21:24:52Z,2026-07-08T14:52:50Z,0,"Stack from ghstack (oldest at bottom): -> #184436 Avoid embedding graph-specific Triton kernel names in the cached source passed to async_compile.triton, while preserving the descriptive name for profiling and runtime metadata. Fixes #112788 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobi..."
186529,Preserve grad state in Dynamo minifier repros,https://github.com/pytorch/pytorch/pull/186529,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,111414,111414,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T07:58:46Z,2026-07-08T14:45:26Z,0,Stack from ghstack (oldest at bottom): -> #186529 Dynamo after-dynamo repro generation serialized autocast state but sampled grad and inference state from the environment where the generated repro was run. That meant minifier launchers produced under torch.no_grad() or torch.inference_mode() could run again under de...
186530,[dynamo] Restore autocast dtype after trace-time updates,https://github.com/pytorch/pytorch/pull/186530,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,111385,111385,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-07T08:19:15Z,2026-07-08T14:41:41Z,0,"Stack from ghstack (oldest at bottom): -> #186530 Dynamo already has special handlers for autocast enabled state, nesting, and cache state because generated autocast helper calls mutate global state while tracing. torch.set_autocast_dtype was the missing piece: it used the generic in-graph function path, so tracing..."
186531,Remove stale Dynamo skip for max_pool1d pooling test,https://github.com/pytorch/pytorch/pull/186531,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,111320,111320,,,,0,1,1,,,"topic: not user facing, agentic, needs-review",2026-06-07T08:33:47Z,2026-07-08T14:39:01Z,0,"Stack from ghstack (oldest at bottom): -> #186531 The skip on test_max_pool1d was added when running the CPU pooling suite under Dynamo reliably hit an LLVM out-of-memory failure. On the current tree, the max_pool1d Dynamo cases and the broader CPU pooling class no longer reproduce the OOM, but the decorator still p..."
186533,Fix SymInt vector dispatcher signature detection,https://github.com/pytorch/pytorch/pull/186533,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,106135,106135,,,,0,1,1,,,"topic: not user facing, agentic, needs-review",2026-06-07T10:27:07Z,2026-07-08T14:33:01Z,0,"Stack from ghstack (oldest at bottom): -> #186533 Registering a C++ kernel that takes const std::vector<c10::SymInt>& against a SymInt[] schema failed schema checking with a confusing int[] vs SymInt[] mismatch. The dispatcher only classified a small set of exact SymInt-related parameter types as SymInt-bearing, so..."
186534,Account for fanout in Inductor reuse realization,https://github.com/pytorch/pytorch/pull/186534,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,105768,105768,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T10:47:10Z,2026-07-08T14:28:35Z,0,"Stack from ghstack (oldest at bottom): -> #186534 Inductor decides whether to materialize a reused pointwise expression using a read-count heuristic. The old check only considered reads for one expansion of the expression. For residual blocks like the repro in #105768, a four-read residual add with two downstream us..."
187601,[functorch] Fix dynamic layers under inference_mode,https://github.com/pytorch/pytorch/pull/187601,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,177750,177750,,,,0,2,1,,,"release notes: torch.func, agentic, needs-review",2026-06-17T20:04:02Z,2026-07-08T14:25:35Z,0,"Stack from ghstack (oldest at bottom): -> #187601 Under torch.inference_mode(), c10::InferenceMode changes both AutogradState and dispatch-key TLS: it sets the inference-mode bit, removes ADInplaceOrView from included keys, and excludes the autograd dispatch keyset. The existing functorch handling only saved and cle..."
186538,Fix CUDA driver linking for Inductor CUTLASS builds,https://github.com/pytorch/pytorch/pull/186538,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,103417,103417,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T11:40:26Z,2026-07-08T14:17:49Z,0,"Stack from ghstack (oldest at bottom): -> #186538 Inductor's CUDA/CUTLASS compile flags linked the CUDA driver with -lcuda. That requires an unversioned libcuda.so linker symlink, but common NVIDIA driver installs and runtime container setups expose only the runtime soname libcuda.so.1. In those environments torch.c..."
182024,[docs] Fix typo in data page,https://github.com/pytorch/pytorch/pull/182024,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, release notes: dataloader",2026-04-30T13:39:35Z,2026-07-08T14:17:29Z,0,Description: Corrected a minor typo in the [utils.data] docs (see multiprocessing_context) to improve readability. No functional changes included.
185220,[Test] Make test_wrap_inductor_compiled_regions.py device-agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/185220,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-05-26T13:34:34Z,2026-07-08T14:08:40Z,0,"Summary Make test_wrap_inductor_compiled_regions.py device-agnostic by using instantiate_device_type_tests to generate device-specific test subclasses. Changes Use instantiate_device_type_tests with only_for=(""cuda"", ""xpu"") to generate device-specific test subclasses. Add device parameter to 20 accelerator tests and..."
188016,Update torch-xpu-ops commit pin,https://github.com/pytorch/pytorch/pull/188016,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing, ciflow/binaries_wheel, ciflow/xpu",2026-06-24T02:12:22Z,2026-07-08T13:52:31Z,0,"Update the torch-xpu-ops commit to intel/torch-xpu-ops@d87e68, includes: Build system: Replace enable_if SFINAE with C++20 requires-clauses Kernel optimization/enabling/fixes: Replace shift_group_left with reduce_over_group in Reduce Sum for better portability Make deterministic tie-breaking stable in XPU topk kerne..."
189224,[xpu][skip] Skip test_combo_kernel_dynamic_scale_rblock on XPU,https://github.com/pytorch/pytorch/pull/189224,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,"187810, 187811","187810, 187811","187810, 187811",,,,0,1,4,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/xpu, module: xpu",2026-07-08T01:35:37Z,2026-07-08T13:52:22Z,0,Fixes #187810 Fixes #187811 Summary Skip test_combo_kernel_dynamic_scale_rblock on XPU for both ComboKernelTests and ComboKernelTestsPerSubkernelBlocks (inherits the skip from parent class). Root Cause The test was added in #186957 to validate occupancy-driven rblock halving for combo kernel reductions. This optimiz...
187354,Match Inductor i0/i1 infinities to eager CUDA,https://github.com/pytorch/pytorch/pull/187354,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,187332,187332,,,,0,2,1,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic",2026-06-15T16:10:11Z,2026-07-08T13:52:08Z,0,"Stack from ghstack (oldest at bottom): -> #187354 CUDA Inductor lowers unscaled modified Bessel I functions to libdevice.cyl_bessel_i0/i1. libdevice returns infinities for infinite inputs, while PyTorch's eager Cephes-derived CPU/CUDA kernels and SciPy return NaN for these unscaled functions. That made torch.compile..."
186539,[Dynamo] Limit FX node growth from loop unrolling,https://github.com/pytorch/pytorch/pull/186539,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,102839,102839,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-07T11:59:58Z,2026-07-08T13:38:21Z,0,"Stack from ghstack (oldest at bottom): -> #186539 Dynamo symbolically follows Python loop backedges and emits FX nodes for tensor operations in each unrolled iteration. There was no guardrail on the amount of graph growth caused by repeatedly taking the same backedge, so a static Python loop could build arbitrarily..."
186361,Improve autograd.Function backward graph grad names,https://github.com/pytorch/pytorch/pull/186361,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,130182,"130182, 186031",,,,0,5,2,6476cc7537ceb091912f468f4262bf3260f89ad6e427378d8da9010f9dbed436,bca421559946a054b05c091eba777ab173153e74bc9ed76e847feebdcfb9cb20,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T11:23:47Z,2026-07-08T13:36:08Z,0,"Stack from ghstack (oldest at bottom): -> #186361 The autograd.Function HOP rebuilds the backward FX graph after tracing so its inputs line up with the rewritten forward graph outputs and saved tensors. During that rewrite, gradient placeholders reused the raw forward-output placeholder names, so a backward graph co..."
184345,Use freezing compiler for traced inference graphs,https://github.com/pytorch/pytorch/pull/184345,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,125236,125236,,,,2,6,18,33e55aed4395203663485aa89c96853f0d3231ce28628acea33e3c7ae5d6266a,d4f5f34f2bfbc9e0e224bbca39d374a8876417bc2e699e3fb441fc202d7b735f,"module: inductor, ciflow/inductor, release notes: inductor, agentic, needs-review",2026-05-19T10:01:14Z,2026-07-08T13:19:57Z,0,Stack from ghstack (oldest at bottom): -> #184345 Let AOTAutograd decide whether the graph is inference before applying Inductor's freezing compiler. This lets no_grad entered inside a compiled region still trigger freezing optimizations while preserving autograd compilation for training graphs. Fixes #125236 Genera...
187592,Support registered HOPs in FlopCounterMode,https://github.com/pytorch/pytorch/pull/187592,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,134385,134385,,,,0,3,1,,,"release notes: python_frontend, agentic, needs-review",2026-06-17T18:45:57Z,2026-07-08T13:16:23Z,0,"Stack from ghstack (oldest at bottom): -> #187592 FlopCounterMode already had flop formulas registered for flex_attention and flex_attention_backward, but its higher-order-operator dispatch path only handled Triton wrapper HOPs and cond. Every other HOP returned NotImplemented before the registry lookup could run, s..."
185304,Fix while_loop stride metadata preservation,https://github.com/pytorch/pytorch/pull/185304,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,4,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,160939,160939,,,,0,13,1,f3346b747e3c8a9b89cd1cbe0aad89cc05c88a1d120cfdb1e0bcc35376408d89,a6d35544b0eabf035da27f05ebbe041583106c9cba264abd898c72798f6fe2cd,"module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan, ciflow/dtensor, agentic, needs-review",2026-05-27T01:11:10Z,2026-07-08T13:12:21Z,0,"Stack from ghstack (oldest at bottom): -> #185304 while_loop requires body outputs to be metadata-compatible with the carried inputs, because each body output becomes the next iteration's input. Two stride-specific issues could violate that invariant for otherwise valid programs. First, the refs implementation of em..."
189198,[DO NOT MERGE][ROCm] Validate 7.14 skip set on ROCm preview lane with TheRock ROCm deps preloading,https://github.com/pytorch/pytorch/pull/189198,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/dtensor, ciflow/inductor, ciflow/rocm-nightly","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,17,0,,,"module: rocm, open source, release notes: releng, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), keep-going, ci-no-td, ciflow/rocm-nightly, ciflow/dtensor, ciflow/docker",2026-07-07T20:46:40Z,2026-07-08T13:01:49Z,0,cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coc...
189218,[DO NOT MERGE][ROCm] Validate 7.14 skip set on TheRock preview lane,https://github.com/pytorch/pytorch/pull/189218,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/dtensor, ciflow/inductor, ciflow/rocm-nightly","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,15,0,,,"module: rocm, open source, release notes: releng, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), keep-going, ci-no-td, ciflow/rocm-nightly, ciflow/dtensor, ciflow/docker",2026-07-08T00:45:25Z,2026-07-08T13:01:47Z,0,cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coc...
188720,[ROCm] Enable cholesky_ex via hipSOLVER xgeev (ROCm >= 7.14),https://github.com/pytorch/pytorch/pull/188720,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/rocm-mi200, ciflow/rocm-mi300, ciflow/rocm-nightly","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,"181647, 181648, 181649",,,,0,6,3,,,"module: rocm, open source, ciflow/trunk, release notes: linalg_frontend, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ci-no-td, ciflow/rocm-mi300, ciflow/rocm-mi200, ciflow/rocm-nightly",2026-07-01T19:17:19Z,2026-07-08T13:01:38Z,0,"hipSOLVER >= 3.5.0 (ROCm >= 7.14) introduced hipsolverDnXgeev, the ROCm equivalent of cuSOLVER's cusolverDnXgeev. This PR leverages that support to enable non-Hermitian eigendecomposition (linalg.eig) and cholesky_ex on ROCm without requiring MAGMA. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport..."
189250,[Test][Refactor] Decouple test_dataloader and test_comparison_utils according to the accelerator relevance.,https://github.com/pytorch/pytorch/pull/189250,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-07-08T09:13:01Z,2026-07-08T13:01:10Z,0,Summary Refactored test/test_dataloader.py and test/test_comparison_utils.py following the S1/S2/S3 decoupling strategy: test_dataloader.py S1 (accelerator-unrelated): TestDataLoader — 66 tests using no device APIs. S2 (accelerator-agnostic): TestDataLoaderDevice (renamed from TestDataLoaderDeviceType) — pin_memory...
186091,Preserve CUDA LSTM in Inductor,https://github.com/pytorch/pytorch/pull/186091,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,140845,140845,,,,0,9,1,a983f5a89f30a47847009cbf46aca36477acccfc971daf33c3ad89d8d818350c,bc06e0ba3be24d76de06191633b3accfe00cbdffe1b2e9b020783a15e2aa8d9c,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T15:21:41Z,2026-07-08T12:47:20Z,0,"Stack from ghstack (oldest at bottom): -> #186091 CUDA exported LSTM currently reaches Inductor through AOTAutograd as a CompositeImplicitAutograd op. Because Inductor's default decomposition table includes the generic LSTM decomposition, AOT tracing expands aten.lstm.input into the Python cell loop before runtime d..."
185280,Fix Dynamo automatic dynamic int specialization,https://github.com/pytorch/pytorch/pull/185280,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,161119,161119,,,,0,4,1,,,"ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic",2026-05-26T20:57:44Z,2026-07-08T12:27:31Z,0,"Stack from ghstack (oldest at bottom): -> #185280 Dynamo's automatic dynamic PGO correctly learned that the vLLM Mamba hidden_states size was varying, but a related Python int source could still stay on the integer-specialization path when it came from a global or other normally-specialized source. In the reported p..."
188049,Fix Inductor div floor nonfinite CUDA results,https://github.com/pytorch/pytorch/pull/188049,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188048,188048,188048,,,,0,1,3,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-24T09:57:44Z,2026-07-08T12:27:26Z,0,"Inductor lowered floating torch.div with rounding_mode=""floor"" as floor(div_rn(a, b)). That preserves +/-inf when the dividend is infinite and leaves signed-zero quotients unchanged when a finite nonzero dividend is divided by an infinite divisor, but eager floor division uses a Python-style fmod-based implementatio..."
188330,Dispatch `index_put_` to `masked_fill_.Scalar` when possible,https://github.com/pytorch/pytorch/pull/188330,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"open source, ciflow/trunk, release notes: cuda",2026-06-27T10:09:39Z,2026-07-08T12:14:25Z,0,"This is a followup on #188267 and #61612 and tries to dispatch index_put_ to masked_fill_.Scalar where possible - e.g. gpu_tensor[mask] = 1.0 - instead of converting the scalar to a CPU tensor first. While this prevents an unnecessary CPU tensor, it has the main advantage that we don't need any special handling for..."
186546,Preserve original Dynamo minifier backend failures,https://github.com/pytorch/pytorch/pull/186546,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,97750,97750,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T14:28:43Z,2026-07-08T12:09:12Z,0,"Stack from ghstack (oldest at bottom): -> #186546 The Dynamo minifier backend used one try/except block for both the real compiler invocation and the internal ""No issue was detected"" sentinel. When the compiler path succeeded, that sentinel was caught as if it were a minifiable backend failure. When the compiler did..."
186549,Fix fake deepcopy for custom module mappings,https://github.com/pytorch/pytorch/pull/186549,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,97078,97078,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-07T15:25:23Z,2026-07-08T12:06:30Z,0,"Stack from ghstack (oldest at bottom): -> #186549 Dynamo fake tensor propagation deepcopies modules before running FX call_module nodes. The previous implementation delegated directly to copy.deepcopy under FakeCopyMode, which reconstructs dict and OrderedDict subclasses through their reduce constructor. DeepSpeed-s..."
186553,[dynamo] Graph break on packed sequence lower ops,https://github.com/pytorch/pytorch/pull/186553,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,93501,93501,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T16:28:10Z,2026-07-08T12:03:34Z,0,"Stack from ghstack (oldest at bottom): -> #186553 _pack_padded_sequence and _pad_packed_sequence both produce output shapes from the runtime values of their length/batch-size tensors. Dynamo previously avoided the public pack_padded_sequence wrapper with a skip rule, but the lower aten/_VF entry points could still b..."
189253,[TEST] Refactor test_c10d_spawn_ucc.py with hw_classification,https://github.com/pytorch/pytorch/pull/189253,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,186918,,,0,1,1,,,"open source, release notes: distributed (c10d), topic: not user facing",2026-07-08T11:48:58Z,2026-07-08T12:01:34Z,0,"Add hw_classification to test/distributed/test_c10d_spawn_ucc.py. Changes: Add hw_classification = MULTI_DEVICE_SPECIFIC to TestDistributedNNFunctionsUcc Tests UCC collective ops (broadcast, reduce, allreduce, all_gather, all_to_all) requiring the UCC backend and >=2 GPUs. Backend-specific, not device-generic. Test..."
186551,Preserve symbolic tensor metadata in after-AOT repros,https://github.com/pytorch/pytorch/pull/186551,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,93367,93367,,,,0,1,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-07T15:55:45Z,2026-07-08T12:01:01Z,0,"Stack from ghstack (oldest at bottom): -> #186551 AOT accuracy minifier repros were still losing dynamic-shape context for tensor-only symbolic inputs. The repro generator only chose symbolic tracing and symbolic compile args when the concrete input args had free symbols, or when placeholder metadata contained scala..."
186561,Preserve autograd metadata for functionalized inplace views,https://github.com/pytorch/pytorch/pull/186561,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,90507,90507,,,,0,2,1,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-06-07T18:49:40Z,2026-07-08T11:56:59Z,0,Stack from ghstack (oldest at bottom): -> #186561 AOTAutograd used functionalization to trace inplace view ops on intermediate tensors as out-of-place view ops. For a function like: out = x * 2 out.unsqueeze_(0) return out that made the compiled autograd.Function return a differentiable view even though eager return...
187057,Fix value-opaque graph output replay,https://github.com/pytorch/pytorch/pull/187057,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,187041,187041,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-11T15:12:35Z,2026-07-08T11:53:46Z,0,"Stack from ghstack (oldest at bottom): -> #187057 Value-opaque objects can be represented by FakeScriptObject while tracing. When such an object is emitted as a graph output and then consumed by generated Python bytecode, Dynamo replayed the FakeScriptObject directly. Tensor subclass metadata reconstruction already..."
186435,Fix mkldnn FakeTensor error handling,https://github.com/pytorch/pytorch/pull/186435,draft/noise,23,PR is draft; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,127111,127111,,,,0,15,1,,,"module: cpu, module: mkldnn, topic: not user facing, module: dynamo, ciflow/inductor, agentic",2026-06-06T01:41:49Z,2026-07-08T11:53:23Z,0,"Stack from ghstack (oldest at bottom): -> #186435 FakeTensor converted real mkldnn tensors into ordinary strided meta tensors, so Dynamo fake propagation lost mkldnn layout information. As a result, unsupported mkldnn view and invalid mkldnn max-pool cases could be traced successfully and then fail later when the ge..."
186397,Keep constant-false torch._check in compiled graph,https://github.com/pytorch/pytorch/pull/186397,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,129327,129327,,,,0,2,5,,,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T18:47:33Z,2026-07-08T11:51:17Z,0,"Stack from ghstack (oldest at bottom): -> #186397 Dynamo's torch._check handling returned immediately for Python-constant true predicates, but for Python-constant false predicates it evaluated the message eagerly and raised an observed exception while tracing. In fullgraph mode this turned a user RuntimeError into a..."
186395,Preserve overlap deps without blocking fusion,https://github.com/pytorch/pytorch/pull/186395,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,186084,186084,,,,0,1,3,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T18:26:23Z,2026-07-08T11:46:42Z,0,Stack from ghstack (oldest at bottom): -> #186395 Overlap scheduling used preserve_node_ordering() to encode extra ordering constraints when insert_overlap_deps=True. That rewrote the dependent FX nodes as higher_order.control_deps subgraphs. The HOP is opaque to Inductor fusion and its lowering realizes dependency...
186409,Fix subclass view replay with dynamic metadata,https://github.com/pytorch/pytorch/pull/186409,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,128649,128649,,,,0,4,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-05T20:21:19Z,2026-07-08T11:44:39Z,0,"Stack from ghstack (oldest at bottom): -> #186409 Subclass view replay can allocate ephemeral symbols for saved view state while reconstructing a fake view from its hidden base. For traceable wrapper subclass views with symbolic output metadata, the hidden base was fakeified with static metadata, and replay-internal..."
185861,Fix ViewMeta serialization for special views,https://github.com/pytorch/pytorch/pull/185861,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,185630,185630,,,,0,2,1,,,"topic: not user facing, agentic, needs-review",2026-06-01T22:21:56Z,2026-07-08T11:37:57Z,0,"Stack from ghstack (oldest at bottom): -> #185861 The hand-written resize_ and _unsafe_view ViewMeta classes declared their SerializableTuple shape fields as const vector references. When those tuples were exposed through the pybind pickle/as_tuple path, the vector references could point at temporaries, corrupting s..."
188053,test: add svdvals NaN propagation test (fixes #187759),https://github.com/pytorch/pytorch/pull/188053,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187759,187759,187759,,,,0,1,2,,,"triaged, open source, module: linear algebra, topic: not user facing",2026-06-24T12:00:46Z,2026-07-08T11:37:03Z,0,"Summary Adds a test verifying that torch.linalg.svdvals() correctly handles NaN inputs, fixing the inconsistency with torch.linalg.svd() documented in #187759. Problem svdvals() silently swallows NaN in some backend configs, returning finite singular values for a NaN matrix. This is a silent correctness bug. svd() c..."
186420,Run torchrec tests in Dynamo CI,https://github.com/pytorch/pytorch/pull/186420,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,128531,128531,,,,0,3,5,,,"release notes: releng, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-05T21:23:13Z,2026-07-08T11:35:42Z,0,"Stack from ghstack (oldest at bottom): -> #186420 The Dynamo core suite already selected test/dynamo/test_torchrec.py, but OSS Dynamo CI did not build or install torchrec/fbgemm for that CPU job. As a result, the file imported no torchrec dependency, replaced its TestCase with NoTest, and reported zero executed test..."
185871,Fix symbolic stride-order sorting without explicit ShapeEnv,https://github.com/pytorch/pytorch/pull/185871,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,185510,185510,,,,0,2,5,5ac1dca51ecba38cb601e1a446b083d549d712ba3ec3e002cb5a06a5b515e480,ead11f5be07dacf9c5aaae6246d2db6c2c02db19bb2cbb76fb029999f61721e9,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T01:07:50Z,2026-07-08T11:27:18Z,0,Stack from ghstack (oldest at bottom): -> #185871 Inductor stride-order computation was only symbolic-aware when callers explicitly passed a ShapeEnv into get_stride_order/get_fill_order. Some lowering paths call get_stride_order with symbolic strides while an Inductor graph is active but without passing shape_env....
185873,Preserve mutation renames after scheduler recompute,https://github.com/pytorch/pytorch/pull/185873,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185509,185509,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T01:31:28Z,2026-07-08T11:24:49Z,0,"Stack from ghstack (oldest at bottom): -> #185873 Inductor may recompute a scheduler node's loop body while preparing fusion. That recomputation re-extracts read/write dependencies from the new body. The old code only reapplied mutation_renames on the fake-dependency path, so nodes with mutation renames but no fake..."
185324,Reject mixed-device fake convolution inputs,https://github.com/pytorch/pytorch/pull/185324,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,160399,160399,,,,0,6,3,,,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ci-no-td, agentic",2026-05-27T05:10:32Z,2026-07-08T11:15:02Z,0,"Stack from ghstack (oldest at bottom): -> #185324 FakeTensor's convolution implementation validated dtype mismatches but did not validate that the convolution input, weight, and bias were on the same device. If a module was compiled on CPU and its parameters were later moved to CUDA in place, Dynamo could fake-propa..."
188965,Fix non-deterministic node ordering in get_source_partitions(),https://github.com/pytorch/pytorch/pull/188965,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,147170,147170,147170,,,,0,2,3,,,"triaged, open source, release notes: fx",2026-07-05T07:06:47Z,2026-07-08T20:33:53Z,0,"Fixes #147170 Summary get_source_partitions() in torch/fx/passes/utils/source_matcher_utils.py could return input_nodes, output_nodes, and params in non-deterministic order across runs due to the use of Python set() which has non-deterministic iteration order due to hash randomization (PYTHONHASHSEED). Root Cause In..."
186476,Fix Dynamo resume tensor lifetimes across graph breaks,https://github.com/pytorch/pytorch/pull/186476,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,122512,"122512, 174929",,,,0,13,2,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic",2026-06-06T17:10:46Z,2026-07-08T11:14:32Z,0,"Stack from ghstack (oldest at bottom): -> #186476 Dynamo resume functions used to pass saved stack/local state by expanding a list with CALL_FUNCTION_EX. That kept the list, and any tensor locals inside it, alive while the resumed frame ran. If resumed execution crossed another compiled graph or immediately ran eage..."
185876,Prevent saving mutated grad inputs in AOT min-cut,https://github.com/pytorch/pytorch/pull/185876,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185497,185497,,,,0,2,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T02:10:50Z,2026-07-08T11:05:12Z,0,Stack from ghstack (oldest at bottom): -> #185876 AOTAutograd clones data-mutated grad inputs in the joint trace so the backward graph can read their pre-mutation values. The Inductor min-cut partitioner could still choose the original primal input placeholder as a saved activation. At runtime that placeholder alias...
186246,Fix stale scalar value in addcdiv graph-break decomposition,https://github.com/pytorch/pytorch/pull/186246,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185382,185382,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T17:57:05Z,2026-07-08T11:00:22Z,0,"Stack from ghstack (oldest at bottom): -> #186246 Dynamo decomposes in-place addcmul_ and addcdiv_ into product/division plus inductor_prims.fma plus copy_ so tensor-valued value arguments avoid item graph breaks and keep CUDA FMA parity. When a .item() graph break appears inside a helper called from a loop, the res..."
188653,Add test decorator for device-type-specific RNG seed setting.,https://github.com/pytorch/pytorch/pull/188653,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,3,1,,,"open source, topic: not user facing",2026-07-01T09:06:40Z,2026-07-08T10:51:18Z,0,"For tests that have randomized inputs, i.e. statistical tests of distributions, and need to have RNG seeds hardcoded in their bodies, it could get tricky to manage all device types with single seed value. See Note: [Randomized statistical tests] in test_distributions.py. This PR provides a new decorator, that allows..."
188736,[xpu][inductor] update Intel Triton commit pin,https://github.com/pytorch/pytorch/pull/188736,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing, ciflow/inductor, keep-going, ciflow/xpu, ciflow/docker",2026-07-01T21:11:23Z,2026-07-08T10:34:30Z,0,
184508,Fix export with unused DTensor param attrs,https://github.com/pytorch/pytorch/pull/184508,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,184229,184229,,,,0,2,1,,,"release notes: export, ciflow/dtensor, agentic, needs-review",2026-05-20T07:56:58Z,2026-07-08T10:33:24Z,0,"Stack from ghstack (oldest at bottom): -> #184508 Drop unused non-tensor subclass parameter and buffer placeholders before export builds user input metadata, and fill missing nn_module_stack metadata for DTensor dispatch nodes. Fixes #184229 Generated by my agent"
185897,Support concrete enum types in torch.library.define schemas,https://github.com/pytorch/pytorch/pull/185897,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,183957,183957,,,,0,3,1,,,"topic: not user facing, agentic, needs-review",2026-06-02T06:03:36Z,2026-07-08T10:30:52Z,0,"Stack from ghstack (oldest at bottom): -> #185897 Fixes #183957 torch.library.define accepts string schemas and passes them to the dispatcher schema parser. Generic enum.Enum worked because it is registered as an opaque type name, but concrete enum subclasses such as Mode or __main__.Mode were not registered before..."
188236,[inductor]: Include convolution_backward in decide_layout_opt conv node detection,https://github.com/pytorch/pytorch/pull/188236,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, ciflow/xpu, ciflow/torchtitan",2026-06-26T01:58:47Z,2026-07-08T10:30:31Z,0,"Motivation When AOTAutograd splits a model into separate forward and backward graphs, the backward graph may contain only convolution_backward without a corresponding forward convolution. Previously, decide_layout_opt pytorch/torch/_inductor/graph.py Lines 757 to 779 in 994f78b def decide_layout_opt(gm: GraphModule,..."
186564,Skip uninitialized outputs during accuracy minification,https://github.com/pytorch/pytorch/pull/186564,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,93437,93437,,,,0,2,1,,,"release notes: fx, module: dynamo, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-06-07T19:18:22Z,2026-07-08T10:26:53Z,0,"Stack from ghstack (oldest at bottom): -> #186564 Accuracy minification can truncate an FX graph suffix and use an intermediate node as the candidate graph output. When that intermediate node is an uninitialized tensor factory such as aten.new_empty or prims.empty_strided, the candidate compares undefined allocation..."
185939,Fix int8 WoQ GEMM residual epilogue fusion,https://github.com/pytorch/pytorch/pull/185939,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,145687,145687,,,,0,4,3,f76e3977c2afa5a91555378c86b632ee2c819ac80fdf484a034552c0cfc1df5f,d6b17f02a01661cb94786903487067f2894bb975f1c52d67219f4effb3a7a764,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T14:27:20Z,2026-07-08T10:22:03Z,0,Stack from ghstack (oldest at bottom): -> #185939 Fixes #145687 Generated by my agent The LLaMA-style int8 weight-only quantized MLP tail has a down-projection GEMM followed by residual add and RMSNorm. The residual add has multiple consumers: RMSNorm's reduction and the final pointwise normalization. Because that a...
172074,Fixes grad_dtype property being lost in fsdp,https://github.com/pytorch/pytorch/pull/172074,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,170648,170648,170648,,,,0,2,2,,,"triaged, open source, ciflow/inductor, release notes: distributed (fsdp2), ciflow/torchtitan",2026-01-09T13:18:57Z,2026-07-08T10:18:22Z,0,"Fixes #170648 Problem When fully_shard() is applied, it creates new nn.Parameter objects to hold the sharded data. The original parameter's grad_dtype property was not being copied to these new parameters, so any user-set grad_dtype was lost. Fix _fsdp_param.py: Before creating the sharded parameter, save the origin..."
178336,[Distributed] Make DDP tests and tensor parallel dependencies backend agnostic,https://github.com/pytorch/pytorch/pull/178336,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, Stale, release notes: distributed (ddp), module: PrivateUse1, ciflow/torchtitan",2026-03-24T22:20:35Z,2026-07-08T10:17:37Z,0,This PR is part of a series to refactor distributed tests to be completely device-agnostic. Changes Source: torch/nn/parallel/distributed.py — BF16 comm hook check: Query torch.accelerator + torch.get_device_module for is_bf16_supported() so non-CUDA accelerators can use BF16 compress hooks torch/distributed/tensor/...
189240,[TEST] Refactor test_c10d_logger.py with hw_classification,https://github.com/pytorch/pytorch/pull/189240,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,,,,186918,,,0,1,1,,,"open source, topic: not user facing",2026-07-08T06:42:24Z,2026-07-08T10:13:21Z,0,"Split mixed test class in test/distributed/test_c10d_logger.py by hardware scope. Changes: Extract test_get_or_create_logger into C10dLoggerGenericTest (hw_classification = GENERIC) Pure CPU-side logger singleton check, no distributed or device dependency. Keep test_exception_logger in C10dErrorLoggerTest (hw_classi..."
188539,Add support for win_arm64 CUDA builds,https://github.com/pytorch/pytorch/pull/188539,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,,,,,,,0,6,0,,,"open source, release notes: releng",2026-06-30T08:02:13Z,2026-07-08T10:06:30Z,0,WIP changes: MAGMA build for win_arm64
186380,Run python decompositions under functionalization,https://github.com/pytorch/pytorch/pull/186380,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,5,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,129418,129418,,,,0,8,2,2baf26b8ac7c120dd4e6bae9158ee6664d0f2877c78dc0a4e63b9e7d1bbfc72b,b6fb1313e4aeb8c20ff888b260c30d0d5cd4272adc5930b25ba826972cb13262,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-06-05T15:00:48Z,2026-07-08T10:01:47Z,0,"Stack from ghstack (oldest at bottom): -> #186380 Functionalization currently runs before proxy tracing, but AOTAutograd and make_fx decomposition tables lived on ProxyTorchDispatchMode. As a result, FunctionalTensorMode saw the original functional operator, then proxy tracing ran the Python decomposition afterward...."
186446,Fix fake overlap handling for Dynamo try/except,https://github.com/pytorch/pytorch/pull/186446,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,5,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,126474,126474,,,,0,8,1,2da5f3170d79cb53d0c315a9f8c91a007c17c407cd6604023e57c8e9713f03b0,7b613d531e87c4c9d05e8a0a1684e380a10f9586ee67b82df4543da9589c77c4,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T05:11:40Z,2026-07-08T09:56:39Z,0,"Stack from ghstack (oldest at bottom): -> #186446 FakeTensor did not model eager memory-overlap checks for mutable ops during tracing, so Dynamo could trace through code that eager would raise from. In try/except this skipped the user's exception handler, and inductor could compile an invalid mutation that silently..."
186505,Simplify mod-one symbolic shape guards,https://github.com/pytorch/pytorch/pull/186505,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,118332,118332,,,,0,2,1,,,"module: cpu, release notes: fx, ciflow/inductor, agentic, needs-review",2026-06-07T01:25:40Z,2026-07-08T09:54:20Z,0,"Stack from ghstack (oldest at bottom): -> #186505 The remaining actionable part of #118332 is that guards like 1 % divisor != 0 stay as modular guards even when range reasoning already knows the divisor is positive. For an integer divisor with lower bound at least one, 1 % divisor != 0 is equivalent to divisor >= 2;..."
187457,[MPS] Migrate log_softmax to native Metal kernels,https://github.com/pytorch/pytorch/pull/187457,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,189146,,,0,3,1,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-16T16:47:47Z,2026-07-08T09:53:32Z,0,"[MPS] Migrate log_softmax to native Metal kernels Reviewing just this PR. Stacked leaf (cumulative-leaf fork on main), so the GitHub diff shows the softmax core+perf PRs + this one. To see only what this PR adds on top of #189146, use this fork compare (renders as a normal diff of just this delta): anagnorisis2perip..."
186462,Allow cond branch mutation during export,https://github.com/pytorch/pytorch/pull/186462,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,123972,123972,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T10:43:08Z,2026-07-08T09:50:39Z,0,"Stack from ghstack (oldest at bottom): -> #186462 Torch cond tracing allowed branch input mutation and aliasing only when grad mode was disabled. torch.export runs with grad enabled by default, including when cond internally uses torch.compile to capture the higher order op, so export rejected cond branches that it..."
186923,Graph break on custom op fake tensor failures,https://github.com/pytorch/pytorch/pull/186923,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185141,185141,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-10T10:01:59Z,2026-07-08T09:45:51Z,0,"Stack from ghstack (oldest at bottom): -> #186923 Dynamo fake propagation uses fake tensors to infer FX node metadata. When a non-ATen custom op has no fake/meta support, or its fake path reaches an opaque kernel that needs real tensor data, fake propagation can raise a generic RuntimeError. Dynamo was treating thos..."
186485,Fix export dynamic dims with size-one examples,https://github.com/pytorch/pytorch/pull/186485,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,122294,122294,,,,0,7,3,92c868cc1ad73e7aab31e2199bc5d1d82df8f29d9d2a7c9fe3bde21ca8a4e6c2,92b9c6f074a097a46c338dcf62d369ae1d1973f5b481482664ab214945723b12,"module: inductor, ciflow/inductor, release notes: export, ciflow/torchtitan, agentic, needs-review",2026-06-06T19:08:21Z,2026-07-08T09:42:45Z,0,"Stack from ghstack (oldest at bottom): -> #186485 Export currently turns user-declared dynamic dimensions with example size 0 or 1 into ordinary backed symbols. ShapeEnv's 0/1 specialization then collapses those symbols to constants, so a valid dynamic_shapes spec such as two inputs sharing Dim(""n_labels"", min=1, ma..."
187493,Fix Inductor nextafter for low precision CUDA tensors,https://github.com/pytorch/pytorch/pull/187493,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,187340,187340,,,,0,2,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-16T19:58:40Z,2026-07-08T09:39:13Z,0,"Stack from ghstack (oldest at bottom): -> #187493 Inductor lowered torch.nextafter through the generic numeric pointwise path, which let Triton represent fp16 and bf16 loads as fp32 compute values before calling libdevice.nextafter. That computes the next fp32 value and then rounds back to the low precision output d..."
184694,Fix AOTAutograd input overlap guards,https://github.com/pytorch/pytorch/pull/184694,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,178680,178680,,,184481,2,11,6,9a7f1e5419b5a8060063a6efa126f1aa63a8bb0ed93657850e3e8c18abfe9149,1111e06a78692f781caeab20ed2230e249b049d759bd768afc49b80863491cf5,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-21T10:17:39Z,2026-07-08T09:35:56Z,0,"Stack from ghstack (oldest at bottom): -> #184694 Preserve input storage-overlap relationships for AOTAutograd mutation handling, including static traces and AOTAutograd cache hits, so alias-sensitive synthetic-base graphs are not reused for incompatible inputs. Fixes #178680 Generated by my agent cc @voznesenskym @..."
187311,Preserve forward AD tangents for Inductor dual inputs,https://github.com/pytorch/pytorch/pull/187311,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor_cpp_wrapper, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,187284,187284,,,,0,6,1,7cd5d7f02ec29fc8512489c9c3851fda37072374ca4a75838027eecb6d502eb4,eddb8fad349def970e7ebc4850b456afd0d61871a66af87089a1f6f4b5cfa5a4,"module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-15T07:27:27Z,2026-07-08T09:33:04Z,0,"Stack from ghstack (oldest at bottom): -> #187311 Inductor generated callables compute primal tensors only. When a compiled callable is invoked with dual tensor inputs, the generated runtime bypasses normal dispatcher JVP propagation and silently returns an output without a forward AD tangent. Dynamo guards the acti..."
184355,[3/N][Test] Migrate TestBwdGradients device-agnostic skips to `@skipOps`,https://github.com/pytorch/pytorch/pull/184355,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,177259,"184356, 184678",,,0,16,7,,,"module: tests, open source, ciflow/trunk, topic: not user facing",2026-05-19T11:47:25Z,2026-07-08T09:32:56Z,0,Stack from ghstack (oldest at bottom): #184678 #184356 -> #184355 Fixes parts of #177259 cc @mruberry
184323,Share foreach combo kernel bodies,https://github.com/pytorch/pytorch/pull/184323,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,127863,127863,,,,0,5,1,,,"ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic",2026-05-19T06:57:46Z,2026-07-08T09:32:51Z,0,"Stack from ghstack (oldest at bottom): -> #184323 Detect matching pointwise foreach subkernel bodies and emit a single shared Triton body with per-branch setup, reducing repeated code that slows Triton compilation. Fixes #127863 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @Xiao..."
184573,Fix compile support for CTC loss tensor lengths,https://github.com/pytorch/pytorch/pull/184573,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,183396,183396,,,,2,4,5,178771eaef9819c970e292cd120104ff70d7a451cd43b1036215802d5e9d8ebc,7486acdf5d9bcb382a860e9e71043de67009c6e58f2fabeb6c7e84fa067a55a3,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-20T17:22:19Z,2026-07-08T09:32:28Z,0,"Stack from ghstack (oldest at bottom): -> #184573 Teach FakeTensor and meta handling for CTC loss selector/internal paths while capturing the public tensor-length CTC op for Dynamo/Inductor fallback, so torch.compile preserves runtime backend selection.\n\nFixes #183396\nGenerated by my agent cc @voznesenskym @pengu..."
184678,[5/N][Test] Migrate TestMeta device-agnostic skips to `@skipOps`,https://github.com/pytorch/pytorch/pull/184678,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,177259,"184355, 184356",,,0,18,3,,,"module: tests, open source, ciflow/trunk, topic: not user facing",2026-05-21T07:47:08Z,2026-07-08T09:31:55Z,0,Stack from ghstack (oldest at bottom): -> #184678 #184356 #184355 Fixes parts of #177259 cc @mruberry
183628,Fix fractional max pool RNG in Inductor,https://github.com/pytorch/pytorch/pull/183628,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,183249,183249,,,,0,3,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-13T22:38:56Z,2026-07-08T09:29:00Z,0,Stack from ghstack (oldest at bottom): -> #183628 Preserve eager CUDA RNG semantics for fractional max pool random samples and keep fallback random ordering when those samples appear in mixed-random Inductor graphs. Also fix fractional_max_pool3d sample dimension ordering to match native kernels. Fixes #183249 Gener...
183654,Fix Inductor CPU group norm affine precision,https://github.com/pytorch/pytorch/pull/183654,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"181696, 183120","181696, 183120",,,,0,2,5,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-14T03:26:24Z,2026-07-08T09:26:14Z,0,Stack from ghstack (oldest at bottom): -> #183654 Use FMA in Inductor's CPU native_group_norm affine decomposition to match eager's folded affine computation and avoid near-cancellation differences before clamp/log. Covers eval-mode BatchNorm/GroupNorm chains feeding clamp/log that diverged from eager under torch.co...
184595,Allow functorch grad fake propagation in Dynamo,https://github.com/pytorch/pytorch/pull/184595,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,183015,183015,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-20T19:25:44Z,2026-07-08T09:20:54Z,0,Stack from ghstack (oldest at bottom): -> #184595 Temporarily allow non-fake functorch wrapper tensors while fake-propagating _autograd_grad so compiled vmap(hessian) works across CPU and CUDA. Fixes #183015 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe...
187486,Fix unbacked memo reuse during fake replay,https://github.com/pytorch/pytorch/pull/187486,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,187415,187415,,,,0,2,1,,,"release notes: fx, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-16T19:20:12Z,2026-07-08T09:18:17Z,0,Stack from ghstack (oldest at bottom): -> #187486 Fake replay of an FX graph that already has unbacked bindings must allocate a fresh symbol at each old binding site so rebind_unbacked can rename that fresh symbol back to the original binding. Boolean mask indexing violates that invariant when two aten.index.Tensor...
187456,[MPS] Migrate softmax and _softmax_backward to native Metal kernels (core),https://github.com/pytorch/pytorch/pull/187456,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,187455,,,,0,1,1,,,"triaged, open source, topic: improvements, release notes: mps, ciflow/mps",2026-06-16T16:46:45Z,2026-07-08T09:10:26Z,0,[MPS] Migrate softmax and _softmax_backward to native Metal kernels (core) What / why First of a small series (Part of #187455) migrating MPS softmax off MPSGraph. This core PR lands the correctness foundation: native Metal kernels for _softmax / _softmax_backward_data over the last dimension — a single-row kernel (...
184602,Preserve frozen ShapeEnv during fake tensor guards,https://github.com/pytorch/pytorch/pull/184602,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,182940,182940,,,,0,5,3,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-05-20T20:08:23Z,2026-07-08T09:05:31Z,0,Stack from ghstack (oldest at bottom): -> #184602 Ignore guard-installing side effects when a ShapeEnv is frozen and keep fake tensor cache entries distinct across guard-mutation state so eager fake execution preserves symbolic metadata. Fixes #182940 Generated by my agent
184632,Fix CUDA fake strides for mixed dtype pointwise ops,https://github.com/pytorch/pytorch/pull/184632,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,"182200, 184101","182200, 184101",,,,0,5,5,,,"topic: not user facing, module: inductor, ciflow/inductor, agentic, needs-review",2026-05-21T01:19:34Z,2026-07-08T09:01:26Z,0,"Stack from ghstack (oldest at bottom): -> #184632 CUDA TensorIterator chooses the output layout for these binary pointwise ops from the original broadcasted operands, then keeps dtype promotion inside the generated kernel. FakeTensor was instead trusting the normal meta result after promotion, which can report a con..."
184639,Support None returns from torch.cond,https://github.com/pytorch/pytorch/pull/184639,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_huggingface, 1, 1, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,181891,181891,,,,8,3,3,2af1701e66c9b1b7f3e466351ead4023e43c8fb361ecce23313ec4984f4f333c,34b1f59efcc2628c9c75833ae175c869c22e1559d9b0a0c017470ecccc57d376,"module: inductor, module: dynamo, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-05-21T02:06:01Z,2026-07-08T08:58:55Z,0,"Stack from ghstack (oldest at bottom): -> #184639 Allow torch.cond branches that only mutate inputs to return None by accepting None constants through Dynamo, CUDA graph conditional capture, and Inductor while guarding unsupported differing constants. Fixes #181891 Generated by my agent cc @voznesenskym @penguinwu @..."
185469,Fix stable TMA descriptor reconstruction in Dynamo,https://github.com/pytorch/pytorch/pull/185469,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), linux-docs / build-docs-cpp-false, linux-docs / build-docs-python-false, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,155856,155856,,,,0,2,1,0e882540551a5d5cb264d8d37b4a06d0d6234a205b29f9932af279b5905f0f02,78836fc9b2013e6ef78a333cf90c13b07d26e935105a869ea3ab61d43354f9da,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-28T05:10:08Z,2026-07-08T08:54:06Z,0,"Stack from ghstack (oldest at bottom): -> #185469 Dynamo's stable Triton TMA descriptor variable had a reconstruction method, but it invoked TensorVariable.reconstruct directly. TensorVariables are reconstructed through PyCodegen.call, which turns them into graph outputs for resume bytecode; direct reconstruction is..."
183701,Fix CPU vectorized reduction tail stores,https://github.com/pytorch/pytorch/pull/183701,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,181690,181690,,,,0,2,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-14T10:57:10Z,2026-07-08T08:51:21Z,0,"Stack from ghstack (oldest at bottom): -> #183701 When a vertical reduction is vectorized over a non-reduction dimension with a smaller tail, Inductor reused the main vector kernel reduction suffix and emitted full-width non-contiguous stores. Reuse the tail vector kernel suffix so tail_size bounds the final store...."
185995,Fix stale pending-backward state for dead CUDA graph paths,https://github.com/pytorch/pytorch/pull/185995,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,144961,,,,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T20:56:50Z,2026-07-08T08:46:38Z,0,"Stack from ghstack (oldest at bottom): -> #185995 A compiled forward that produced no live outputs could leave running_forwards_with_pending_backwards set on the CUDA graph tree manager. That stale guard prevented a later unrelated inference graph from starting a new generation, so the later graph stayed in warmup i..."
185496,Fix export autograd.grad saved tensor tracing,https://github.com/pytorch/pytorch/pull/185496,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,"146719, 155044","146719, 155044",,,,0,6,14,450a27209c01eb5cf30cfa65807365994324df5d6364c2cc33de380a5e12d382,0787be33801bd43ae2bf91e3e0166a757da4909bd1afa2a092d155ed8d509f56,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-05-28T10:31:31Z,2026-07-08T08:44:24Z,0,"Stack from ghstack (oldest at bottom): -> #185496 torch.export non-strict tracing can see autograd saved tensors as fresh FakeTensor wrappers when tracing backward formulas through torch.autograd.grad. Proxy lookup only knew the original forward FakeTensor by object identity, so the unpacked saved result from sqrt w..."
184356,[4/N][Test] Migrate TestFwdGradients device-agnostic skips to `@skipOps`,https://github.com/pytorch/pytorch/pull/184356,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,177259,"184355, 184678",,,0,19,11,,,"module: tests, open source, ciflow/trunk, topic: not user facing",2026-05-19T11:47:31Z,2026-07-08T08:41:27Z,0,Stack from ghstack (oldest at bottom): #184678 -> #184356 #184355 Fixes parts of #177259 cc @mruberry
185804,Fix non-strict export tensor len handling,https://github.com/pytorch/pytorch/pull/185804,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,147326,147326,,,,0,4,1,,,"Merged, Reverted, ciflow/trunk, ciflow/inductor, release notes: export, ci-no-td, agentic",2026-06-01T09:49:24Z,2026-07-08T08:41:16Z,0,"Stack from ghstack (oldest at bottom): -> #185804 Non-strict export patches a small set of Python builtins while tracing fake tensors so Python operators preserve symbolic shape expressions. It handled max, min, and math.pow, but left builtins.len untouched. CPython len(tensor) therefore asked the fake tensor's symb..."
183511,Fix Inductor addmm beta-zero input handling,https://github.com/pytorch/pytorch/pull/183511,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,178040,"178040, 188721",,,,0,5,2,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, agentic",2026-05-13T05:12:42Z,2026-07-08T08:41:01Z,0,"Stack from ghstack (oldest at bottom): -> #183511 Route CUDA addmm beta=0 through lowering that validates dtype/device and matmul metadata before dropping the ignored input, including zero-output and out_dtype cases. Fixes #178040 Generated with AI cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @Xiaobi..."
186388,Avoid native eager router graph breaks under Dynamo,https://github.com/pytorch/pytorch/pull/186388,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,186354,186354,,,,0,11,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic",2026-06-05T17:04:31Z,2026-07-08T08:41:00Z,0,"Stack from ghstack (oldest at bottom): -> #186388 Fixes #186354 The native override eager router is installed as the real backend kernel for aten overrides. Normally that path should evaluate eager-only predicates and fall back through the captured boxed kernel, which avoids recursively redispatching to the same ove..."
188990,[FSDP] Fix to_cpu when flat_param is already on CPU,https://github.com/pytorch/pytorch/pull/188990,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,188989,188989,188989,,,,0,2,2,,,"triaged, open source, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-07-06T02:44:05Z,2026-07-08T08:40:41Z,0,"Fix #188989 If flat_param is already on CPU, summon_full_params(offload_to_cpu=True) should not attempt to move it to CPU or free the unsharded parameter. Doing so would leave flat_param and its views aliasing freed storage, leading to reading freed memory. This change adds a check in FlatParamHandle.to_cpu to exit..."
186656,Autotune XBLOCK for mix-order reductions,https://github.com/pytorch/pytorch/pull/186656,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185814,185814,,,,0,3,5,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-08T21:35:52Z,2026-07-08T08:40:37Z,0,"Stack from ghstack (oldest at bottom): -> #186656 Mix-order persistent reductions derived XBLOCK only from RSPLIT_SIZE. For large RMSNorm backward rows, RSPLIT_SIZE grows with M and pushes XBLOCK to 2 or 4 even when N is large enough that the XBLOCK * N footprint hurts bandwidth. Keep the existing heuristic candidat..."
189227,[xpu][feature]Extend SYCL device-specific math functions to NumericUtils,https://github.com/pytorch/pytorch/pull/189227,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,,,,,,,0,2,0,,,"open source, module: xpu",2026-07-08T02:59:00Z,2026-07-08T08:40:33Z,0,"This pull request updates the aten/src/ATen/NumericUtils.h header to improve support for SYCL devices by ensuring correct usage of SYCL math functions and handling type conversions where necessary. The changes primarily add SYCL-specific branches to several mathematical utility functions, ensuring compatibility and..."
186345,[Test] Make test_comparison_utils.py device-generic,https://github.com/pytorch/pytorch/pull/186345,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-05T08:01:16Z,2026-07-08T08:40:27Z,0,"Refactor test/test_comparison_utils.py so torch._assert_tensor_metadata tests run on every available backend, not only implicit CPU. Summary TestComparisonUtils tests metadata assertion for export/compile guards. The op compares sizes, strides, dtype, device, and layout -- no kernel numerics -- so it is safe to run..."
184656,Fix CPU SDPA higher-order grad compile,https://github.com/pytorch/pytorch/pull/184656,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,14,1,"statusCheckRollup, linux-jammy-py3.10-clang18 / test-osdc (einops, 1, 1, mt-l-x86iavx512-8-64), linux-jammy-py3.10-clang18 / test-osdc (default, 1, 5, mt-l-x86iavx512-16-128), linux-noble-xpu-n-py3.10 / build-osdc, linux-jammy-py3.10-clang18-asan / build-osdc, linux-jammy-py3.10-gcc11 / build-osdc, linux-jammy-py3.10-gcc11-no-ops / build-osdc, linux-jammy-py3.14-clang18 / build-osdc, linux-jammy-py3.14t-clang18 / build-osdc, linux-jammy-aarch64-py3.10 / build, before-test / llm-retrieval, unit-test / inductor-cpu-build / build, unit-test / inductor-build / build-osdc, unit-test / inductor-triton-cpu-build / build-osdc",statusCheckRollup,,181177,181177,,,,0,3,2,a4d9933d1648863ea4df1bb1da64d98db558953b8598af4e529c9402ec4af515,afefe953ea6fe2c777e45f3b713e2571a3b5921213bea05792a980207576fd7c,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-05-21T05:02:38Z,2026-07-08T08:38:02Z,0,Stack from ghstack (oldest at bottom): -> #184656 Register a differentiable Autograd-key path for CPU flash SDPA that preserves the fused op aux-output contract while avoiding the unsupported fused backward derivative in compiled torch.func.grad. Add regression coverage for the compile path and direct op mask/logsum...
186857,Use real flex block counts for autotune inputs,https://github.com/pytorch/pytorch/pull/186857,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,180651,180651,181096,,,0,2,2,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-10T00:41:36Z,2026-07-08T08:33:36Z,0,"Stack from ghstack (oldest at bottom): -> #186857 Flex attention autotuning used create_num_blocks_fake_generator to fill num_blocks benchmark tensors with the sparse index tensor width. That made both the partial and full block count inputs look maximally populated during autotune, even when the real BlockMask had..."
188388,[WIP][XPU][Test]Port 12 UT test suites to Intel GPU,https://github.com/pytorch/pytorch/pull/188388,draft/noise,13,PR is draft; CI lane labels: ciflow/xpu,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,7,0,,,"open source, topic: not user facing, ciflow/xpu",2026-06-29T05:36:19Z,2026-07-08T08:28:23Z,0,"For #114850, we will port 12 UT test suites to Intel GPU. We could enable Intel GPU with following methods and try the best to keep the original code styles: Instantiate_device_type_tests() Use ""torch.accelerator.current_accelerator()"" to determine the accelerator backend Enabled XPU for some test path Added allow_x..."
185878,Fix Inductor expm1 on float subnormal inputs,https://github.com/pytorch/pytorch/pull/185878,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185480,185480,,,,0,3,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T02:34:22Z,2026-07-08T08:28:20Z,0,"Stack from ghstack (oldest at bottom): -> #185878 CELU and ELU lower through expm1. On CUDA Inductor, Triton emitted libdevice.expm1 directly, and the libdevice path flushes float32 subnormal results to positive zero under the default FTZ behavior. For negative bf16 subnormal inputs this changed CELU/ELU from eager'..."
183744,Update ACL dependencies for BF16 inner product,https://github.com/pytorch/pytorch/pull/183744,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/docker, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,9,1,"statusCheckRollup, linux-jammy-py3.14t-clang18 / test (crossref, 2, 2, linux.2xlarge), linux-jammy-py3.10-clang18 / test (crossref, 2, 2, linux.2xlarge), linux-jammy-py3.10-clang18 / test (dynamo_wrapped, 3, 3, linux.2xlarge), linux-jammy-py3.10-clang18 / test (onnx, 1, 1, linux.4xlarge), linux-jammy-py3.14-clang18 / test (crossref, 2, 2, linux.2xlarge), linux-jammy-py3.14-clang18 / test (dynamo_wrapped, 3, 3, linux.2xlarge), manylinux2_28-builder:rocm7.2, manylinux2_28-builder:rocm7.1",statusCheckRollup,,180447,180447,,,,2,6,7,31bf0ed5ec63ef72ce1288c250f9188e58649a8ad09ef2016c2492d306c1838a,7aee1a1c9a7522f1d9180baaad5a183ccb54fe861ed6350dcbd8e6b73542f9df,"module: mkldnn, ciflow/trunk, topic: not user facing, ci-no-td, no-runner-experiments, ciflow/docker, agentic, needs-review",2026-05-14T17:51:19Z,2026-07-08T08:23:33Z,0,"Stack from ghstack (oldest at bottom): -> #183744 Bump IDEEP to the oneDNN v3.12 integration and update the ACL installer default to the merged ComputeLibrary commit that accepts BF16 CpuFullyConnected validation, so AArch64 BF16 linear can use the optimized ACL inner-product path under freezing. Partially fixes #18..."
184667,Fix cond branch tensor constants in run_decompositions,https://github.com/pytorch/pytorch/pull/184667,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,180354,180354,,,,0,5,5,,,"release notes: export, agentic, needs-review",2026-05-21T06:44:47Z,2026-07-08T08:20:27Z,0,Stack from ghstack (oldest at bottom): -> #184667 Unwrap functional tensor buffers on HOP subgraphs while retracing and keep branch-local tensor constants out of the top-level decomposed export signature/state_dict. Fixes #180354 Generated by my agent
182987,[Inductor] Parallel batched precompile in combo sequential autotune,https://github.com/pytorch/pytorch/pull/182987,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,183028,,,0,2,3,,,"Stale, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-08T20:14:18Z,2026-07-08T08:18:47Z,0,"Stack from ghstack (oldest at bottom): #183028 -> #182987 CachingAutotuner._combo_sequential_autotune runs a search over Phase 1 (block sizes per sub-kernel group) and Phase 2 (warps/stages). Each trial does a triton.compile(). Trials within a Phase 1 group, and all Phase 2 trials, are independent, they all build on..."
183041,[FSDP1] Test forward overlap with use_orig_params,https://github.com/pytorch/pytorch/pull/183041,draft/noise,13,PR is draft,wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,1,0,,,"Stale, topic: not user facing",2026-05-09T07:09:13Z,2026-07-08T08:18:44Z,0,Stack from ghstack (oldest at bottom): -> #183041 Add a forward overlap regression test for the use_orig_params=True path with mixed precision. This exercises pre_unshard writeback behavior that can serialize all-gather against compute. Test Plan: lintrunner test/distributed/fsdp/test_fsdp_overlap.py
180209,XPU oneDNN: thread-local LRU cache for FP matmul primitives + post_ops deduplication,https://github.com/pytorch/pytorch/pull/180209,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,24,0,,,"module: cpu, triaged, open source, Stale, topic: not user facing",2026-04-13T08:26:32Z,2026-07-08T08:18:40Z,0,"Summary Introduces thread-local LRU caching for oneDNN FP matmul primitives on XPU (keyed by shapes, strides, dtypes, bias, flags, and fused post-op fingerprint) and a second thread-local LRU that reuses identical dnnl::post_ops built from Attr. Refactors Attr so post-op emission and cache keys share one code path v..."
183854,Filter C++-wrapper-incompatible extern autotune choices,https://github.com/pytorch/pytorch/pull/183854,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,179777,,179777,,,0,7,4,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, agentic, needs-review",2026-05-15T05:57:10Z,2026-07-08T08:17:45Z,0,"Stack from ghstack (oldest at bottom): -> #183854 Filter autotune choices that cannot produce C++ wrapper code when Inductor is compiling with the C++ wrapper. The original failure happened when an empty 1x1 convolution selected conv1x1_via_mm, a Python-only extern choice with no C++ wrapper kernel name, and wrapper..."
176574,Increase dynamo coverage of the `operator` module by adding polyfills,https://github.com/pytorch/pytorch/pull/176574,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: cpu, triaged, open source, Stale, module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-03-05T08:05:57Z,2026-07-08T08:08:34Z,0,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98 @jerryzh168 @Lucaskabela @mlazos @guilhermeleobas
187609,Match CUDA threshold scalar dtype in decomp,https://github.com/pytorch/pytorch/pull/187609,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,185470,185470,185846,,,0,5,2,,,"module: inductor, ciflow/inductor, release notes: inductor, agentic",2026-06-17T21:26:55Z,2026-07-08T07:46:01Z,0,"Stack from ghstack (oldest at bottom): -> #187609 The threshold decomposition used the generic DEFAULT elementwise promotion wrapper, so low-precision inputs compared against Python scalar thresholds in fp32. That matches the CPU threshold kernel, but not CUDA: CUDA casts the scalar threshold to the input dtype befo..."
185501,Fix Issue #185396 inductor complex64 triton signature,https://github.com/pytorch/pytorch/pull/185501,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185396,185396,185396,,,,0,8,2,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-28T11:27:25Z,2026-07-08T07:45:41Z,0,"Fixes #185396 This updates the PR to fix the underlying issue rather than only improve the diagnostic. Root cause: Inductor was generating a Triton pointwise exact-stride copy for a complex64 view_as_complex output. Triton signature generation does not support complex pointer dtypes, which caused KeyError: 'complex6..."
186915,[WIP][XPU]Enable inductor cases when issues are fixed,https://github.com/pytorch/pytorch/pull/186915,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,19,0,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/xpu, ci-no-td, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, ciflow/dtensor",2026-06-10T08:14:49Z,2026-07-08T07:45:36Z,0,Enable inductor test cases on XPU when the blocking issues are fixed. Some of the skip decorators and DISABLED conditions have been updated to enable running more tests on XPU now that the blocking issues are resolved. Changes test_combo_kernels.py: Remove @skipIfXpu for issue #181863 test_compile_subprocess.py: Upd...
186444,Fix Adam capturable zero-lr NaNs,https://github.com/pytorch/pytorch/pull/186444,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,126514,126514,,,,0,5,1,,,"release notes: optim, agentic",2026-06-06T04:11:24Z,2026-07-08T07:45:11Z,0,Stack from ghstack (oldest at bottom): -> #186444 The capturable single-tensor Adam path folds tensor learning-rate math into the denominator because addcdiv_ only accepts a numeric value. The old formula divided the square-rooted second moment by step_size_neg before adding eps. When lr was zero and a row had zero...
185082,Preserve lifted SymInts during strict export capture,https://github.com/pytorch/pytorch/pull/185082,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,165073,165073,,,,0,17,1,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: export, ciflow/torchtitan, agentic",2026-05-24T20:53:56Z,2026-07-08T07:45:02Z,0,Stack from ghstack (oldest at bottom): -> #185082 Strict export skipped Dynamo's basic-symbol lifting while creating graph inputs. That made the export torch-level graph diverge from Dynamo: tensor shape symbols were immediately recomputed with aten.sym_size nodes during later tracing instead of being consumed as li...
184695,Fix all-masked TransformerEncoder compile consistency,https://github.com/pytorch/pytorch/pull/184695,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,178677,178677,,,,0,1,1,,,"release notes: nn, agentic, needs-review",2026-05-21T10:40:39Z,2026-07-08T07:35:30Z,0,Stack from ghstack (oldest at bottom): -> #184695 Skip the nested-tensor sparsity conversion for the all-masked src_key_padding_mask case so eager follows the dense path that torch.compile already uses. Fixes #178677 Generated by my agent
186150,Fix compile storage semantics for as_strided scatter,https://github.com/pytorch/pytorch/pull/186150,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,140171,140171,,,,0,8,4,59bc95fc2baa6d89b675cacd9d39e5fcd0d8878791a383ca39d63bd96d980841,37c37a8d25cd379394e0ae346d4e081a655675f38bfcca6881e4a0f3eac32622,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T23:27:10Z,2026-07-08T07:32:08Z,0,"Stack from ghstack (oldest at bottom): -> #186150 The original failure came from functionalized view mutation lowering using as_strided_scatter on a view input. The prim meta check validated the requested scatter view against input.numel(), but eager as_strided_scatter clones with clone_preserve_strides semantics: n..."
184709,Fix Dynamo tensor subclass constructor side effects,https://github.com/pytorch/pytorch/pull/184709,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,178388,178388,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-21T13:17:06Z,2026-07-08T07:25:01Z,0,"Stack from ghstack (oldest at bottom): -> #184709 Treat tensor subclass construction from an existing tensor as a fresh Python wrapper, track it before tracing custom init, and forward constructor arguments so init side effects are replayed correctly. Add coverage for module attribute assignment and preserved non-co..."
183866,Preserve native matmul lowp output precision,https://github.com/pytorch/pytorch/pull/183866,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,7,1,"statusCheckRollup, linux-jammy-cuda13.0-py3.10-gcc11-sm90 / test-osdc (smoke, 1, 1, mt-l-x86iamx-22-225-h100), linux-jammy-cuda13.0-py3.10-gcc11-sm100 / test-osdc (smoke_b200, 1, 1, mt-l-x86iamx-22-225-b200), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), linux-jammy-cuda13.0-py3.10-gcc11-sm100 / test (smoke_b200, 1, 1, linux.dgx.b200), macos-py3-arm64 / test (mps, 1, 1, macos-m2-15)",statusCheckRollup,,178317,178317,,,,0,6,2,bedd0f1fee1dafa372f7b61b3f430fcd9f092d7f99d9c06e70cc1cfe0fb7825e,cf860930ab37115bb6a4603a2694b93e69cdfa4a4c4e2b8b42bd419d70486046,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, agentic, needs-review",2026-05-15T08:21:33Z,2026-07-08T07:22:01Z,0,"Stack from ghstack (oldest at bottom): -> #183866 Native tl.dot accumulates fp16/bf16 matmuls in fp32, but aten.mm/bmm expose fp16/bf16 outputs before fused consumers. Add an explicit low-precision dtype boundary for native matmul reductions and cover remove_no_ops/native bmm precision. Fixes #178317 Generated by my..."
184713,Support Dynamo class setattr for hook handles,https://github.com/pytorch/pytorch/pull/184713,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,178248,178248,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-21T14:17:59Z,2026-07-08T07:19:25Z,0,Stack from ghstack (oldest at bottom): -> #184713 Teach Dynamo to replay standard class attribute mutations and mutable nn.Module hook dictionaries so forward-hook registration inside compiled forward no longer fails on RemovableHandle.next_id. Fixes #178248 Generated by my agent cc @voznesenskym @penguinwu @EikanWa...
184716,Fix conv_transpose2d meta output padding validation,https://github.com/pytorch/pytorch/pull/184716,ready for maintainer decision,112,needs-review label present; approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,178127,178127,,,,0,1,1,,,"release notes: nn, agentic, needs-review",2026-05-21T14:27:39Z,2026-07-08T07:16:54Z,0,Stack from ghstack (oldest at bottom): -> #184716 Mirror native transposed convolution validation in the meta shape path so invalid output_padding combinations raise instead of producing a shape. Fixes #178127 Generated by my agent
185585,Fix mark_unbacked stride recompiles,https://github.com/pytorch/pytorch/pull/185585,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,153204,153204,,,,0,9,5,ec3246a4d8de92974d84621379f1d846705cb062bfa64b4fcb1d82f211d1ac76,91dd968e6e9fbcda60ef153fe51c7b6e3908682dd6bc8002446b82badfbd48c7,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-29T02:48:59Z,2026-07-08T07:13:24Z,0,"Stack from ghstack (oldest at bottom): -> #185585 mark_unbacked made selected sizes unbacked, but left the corresponding strides on INFER_STRIDE. ShapeEnv then applied the hard-coded non-contiguous stride 0/1 specialization before stride dynamism was consulted, so switching between non-contiguous and contiguous mark..."
183871,[inductor] Match eval BatchNorm CUDA numerics,https://github.com/pytorch/pytorch/pull/183871,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,178096,178096,,,,12,3,8,b4c3885906c5e02ec9bcf6c459fb17845f60cd8df15da578443bb753dec9e895,207171a344498437b80fc4b8f2f684604fe701ea0b57fc4352414974a58fe1fe,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-15T09:11:55Z,2026-07-08T07:10:13Z,0,"Stack from ghstack (oldest at bottom): -> #183871 Override Inductor's eval batch norm decomposition to use CUDA eager's rsqrt and FMA-friendly affine ordering, while preserving Conv-BN freezing by exposing only single-use foldable addcmul chains to binary folding. Adds CUDA regressions for the dynamic BatchNorm+Conv..."
188967,Update logging docs to match current API,https://github.com/pytorch/pytorch/pull/188967,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,137285,137285,137285,,,,0,2,4,,,"open source, topic: not user facing",2026-07-05T07:37:30Z,2026-07-08T07:05:30Z,0,"Fixes #137285 Summary Updates docs/source/logging.md to match the current PyTorch logging API: Added native_dsl component Added autotuning_inputs, caching, and compute_dependencies artifacts"
184793,Fix no-tensor-aliasing recompile reason for stale sources,https://github.com/pytorch/pytorch/pull/184793,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,176211,176211,,,,0,4,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-21T20:26:47Z,2026-07-08T07:05:29Z,0,"Stack from ghstack (oldest at bottom): -> #184793 NO_TENSOR_ALIASING guard failure reporting re-evaluates the tensor source expressions recorded by the first compile. When a later invocation changes a container that held tensors into a value such as None, sources like L['data'][0] no longer evaluate. The guard had a..."
184917,Make FX traceback preserve state thread-local,https://github.com/pytorch/pytorch/pull/184917,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,173816,173816,,,,0,1,1,,,"release notes: fx, agentic, needs-review",2026-05-22T18:20:29Z,2026-07-08T07:03:00Z,0,"Stack from ghstack (oldest at bottom): -> #184917 FX traceback stored preserve-node metadata state in module globals. A preserve_node_meta or _preserve_node_seq_nr context in one thread therefore changed has_preserved_node_meta(), current_meta, and replay-node state observed by other threads. This was the same class..."
184928,Fix linalg_pinv tensor tolerance broadcasting,https://github.com/pytorch/pytorch/pull/184928,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,173052,173052,,,,0,1,1,,,"release notes: linalg_frontend, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-22T21:33:31Z,2026-07-08T06:57:51Z,0,"Stack from ghstack (oldest at bottom): -> #184928 aten.linalg_pinv.atol_rtol_tensor used to unsqueeze tensor tolerances before broadcasting them with the singular values. For a 2D input, singular values have shape [k], so a tensor rcond/rtol with shape [1] became [1, 1], broadcasted the singular values to [1, k], an..."
189203,[DO NOT MERGE] This is a reference branch to test 7.14,https://github.com/pytorch/pytorch/pull/189203,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi350","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, topic: not user facing, keep-going, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/docker",2026-07-07T21:19:24Z,2026-07-08T06:51:35Z,0,
181000,[inductor] Dump Python stacks on CI test subprocess timeout,https://github.com/pytorch/pytorch/pull/181000,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/rocm-mi200, ciflow/rocm-mi300, ciflow/rocm-mi355","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ci-no-td, ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/rocm-mi200, ciflow/torchtitan",2026-04-21T17:32:39Z,2026-07-08T06:51:27Z,0,"Flaky CUDA Inductor tests occasionally hang past 30 minutes and trigger the CI outer timeout. The existing termination path (SIGINT then SIGKILL) yields no diagnostics because processes blocked in native CUDA/Triton code don't act on SIGINT — by the time SIGKILL fires, in-process state is gone. This PR adds an opt-i..."
184009,Handle SymPy boolean Inductor graph inputs,https://github.com/pytorch/pytorch/pull/184009,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,167070,167070,,,,0,2,3,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-05-16T09:34:04Z,2026-07-08T06:43:00Z,0,"Stack from ghstack (oldest at bottom): -> #184009 Treat SymPy boolean relations as symbolic scalar graph inputs across Inductor wrapper dtype checks, input unpacking, and generated scalar expressions. Add focused regression coverage for AOT C++ wrapper handling so SymBool inputs use their runtime graph input name in..."
183862,[Inductor] Preserve view_as_complex input layout before fake prop,https://github.com/pytorch/pytorch/pull/183862,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,179368,179368,,,,0,5,5,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-15T07:15:17Z,2026-07-08T06:38:21Z,0,"Stack from ghstack (oldest at bottom): -> #183862 Insert a contiguous clone for single-use computed inputs to view_as_complex in compiled backward graphs before Inductor's additional fake propagation, so layout-optimized SDPA/conv backward graphs do not fail before lowering can apply fallback layout constraints. Fix..."
185059,Fix non-strict export tensor subclass attr tracing,https://github.com/pytorch/pytorch/pull/185059,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,167007,167007,,,,0,1,1,,,"release notes: export, agentic, needs-review",2026-05-24T09:06:01Z,2026-07-08T06:35:29Z,0,"Stack from ghstack (oldest at bottom): -> #185059 Non-strict export proxies tensor subclass attribute reads by temporarily installing a custom getattribute on wrapper subclasses. The attribute set only used set(dir(instance)) - set(dir(torch.Tensor)), so wrapper inner tensors whose names overlap Tensor attributes, s..."
184015,[inductor] Handle scalar workaround args in pattern checks,https://github.com/pytorch/pytorch/pull/184015,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,166604,166604,,,,0,1,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-16T12:05:45Z,2026-07-08T06:31:50Z,0,Stack from ghstack (oldest at bottom): -> #184015 Preserve matched scalar workaround values when rechecking replacement patterns and fall back to registered constants only when the scalar was optimized out. This avoids crashing the randperm indexing pattern when slice_shape is absent from the initial match. Fixes #1...
189238,[XPU] Register linalg._powsum and _foreach_powsum.Scalar for XPU,https://github.com/pytorch/pytorch/pull/189238,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, release notes: xpu",2026-07-08T06:28:58Z,2026-07-08T06:31:23Z,0,Registers the XPU dispatch keys for linalg._powsum and _foreach_powsum.Scalar so the kernels added in torch-xpu-ops can be reached. Related XPU PR Companion to intel/torch-xpu-ops#4195 IMPORTANT: DO NOT MERGE THIS PR UNTIL THE ABOVE XPU PR IS MERGED Changes No new ops or schemas are introduced; this only wires exist...
189183,[inductor][mtia] Fix fp16 reduction accumulator dtype mismatch,https://github.com/pytorch/pytorch/pull/189183,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,1,1,,,"topic: not user facing, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-07-07T18:51:27Z,2026-07-08T06:21:15Z,0,"Summary: On MTIA the Inductor Triton backend sets codegen_upcast_to_fp32=False, so a low-precision reduction keeps its source dtype for the loop-carried accumulator (upcast_acc_dtype(torch.float16) returns torch.float16). But an op inside the reduction body can still upcast the reduced value to fp32 unconditionally..."
185063,Handle scalar carries in Inductor while_loop,https://github.com/pytorch/pytorch/pull/185063,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,166525,166525,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-05-24T10:30:47Z,2026-07-08T06:19:11Z,0,"Stack from ghstack (oldest at bottom): -> #185063 Fixes #166525 WhileLoop.create assumed every carried and additional input from the outer higher order op node was an FX node with meta[""val""]. A Python scalar loop carry is represented there as a literal, while the traced cond/body subgraphs use an unbacked SymInt pl..."
188966,Add operator.concat/iconcat support to dynamo builtins,https://github.com/pytorch/pytorch/pull/188966,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,116396,116396,116396,,,,0,2,4,,,"triaged, open source, release notes: releng, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-05T07:30:23Z,2026-07-08T06:17:57Z,0,"Fixes #116396 Summary operator.concat and operator.iconcat were missing from Dynamo's builtin operator registries, causing graph breaks when these operators were used. This PR adds them to the appropriate operator lists. Changes Added operator.concat and operator.iconcat to torch/_dynamo/variables/builtin.py in: IN_..."
185064,Add unroll option to scan,https://github.com/pytorch/pytorch/pull/185064,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,166516,166516,,,,12,4,6,901c36c896e41fe419b1c59a4121847ca1d0aae64d463a2ee2428b8c4b0aef5d,3f9beb936f9fdb775817e11f4e12258489378e7536b9cf1fb1a94739172336ef,"release notes: onnx, module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, release notes: inductor, module: higher order operators, ciflow/torchtitan, agentic, needs-review",2026-05-24T11:26:52Z,2026-07-08T06:15:05Z,0,"Stack from ghstack (oldest at bottom): -> #185064 Fixes #166516 The root cause of the reported scan performance gap was that scan had no way to request unrolled code generation. Dynamo always captured scan as a higher order op without an unroll parameter, and Inductor decomposed it into a while_loop whose body perfo..."
185800,Fix TORCH_LOGS stderr handling at process exit,https://github.com/pytorch/pytorch/pull/185800,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,147486,147486,,,,0,2,1,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, agentic",2026-06-01T08:28:53Z,2026-07-08T06:14:19Z,0,"Stack from ghstack (oldest at bottom): -> #185800 PT2 logging installed normal StreamHandler instances for TORCH_LOGS console output. When pytest capture was active during log initialization, those handlers retained pytest's temporary sys.stderr object. Pytest restores and closes that capture stream before Python at..."
185908,[Dynamo] Preserve live dict view semantics around mappingproxy mutation,https://github.com/pytorch/pytorch/pull/185908,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,149583,149583,,,,0,11,1,97a0d1fefb664e1999358c91b91023ad485c9ec7ebe56030b0d971d774ec16bc,8c5fb842608903e89b577401a1df11a52cd015b860f212fea81a138f85ae3874,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T09:00:46Z,2026-07-08T06:11:11Z,0,"Stack from ghstack (oldest at bottom): -> #185908 Fixes #149583 EasyDict constructs objects by iterating self.__class__.__dict__.keys(). Dynamo previously either failed to trace the mappingproxy keys call, or after that support was added could model dict view objects as stale snapshots. That is incorrect because CPy..."
185085,[dynamo] Support datetime.now attrs as runtime inputs,https://github.com/pytorch/pytorch/pull/185085,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,164922,164922,,,,0,4,1,c71d3befe7187374a5ecd62a230a77783b300721caf1a49c249283ec3133a57f,f54835df730fdf953180c872071a427e87cb900aaf9fac2d2b34fab66035084f,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-24T22:44:53Z,2026-07-08T06:07:41Z,0,"Stack from ghstack (oldest at bottom): -> #185085 Dynamo handled datetime.datetime as a generic UserDefinedClassVariable, so a zero-argument datetime.datetime.now() call fell through to unsupported class-method handling. In fullgraph mode this raised an unsupported-method error, and in non-fullgraph mode it forced a..."
185088,Avoid graphsafe RNG inputs in descriptor compile,https://github.com/pytorch/pytorch/pull/185088,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,164559,164559,,,,0,1,1,,,"ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-05-25T00:04:24Z,2026-07-08T06:04:40Z,0,"Stack from ghstack (oldest at bottom): -> #185088 aot_compile_joint_with_descriptors lets callers inspect and transform the partitioned forward and backward graphs after aot_export_joint_with_descriptors. The descriptor export has a public calling convention for params, buffers, user inputs, tangents, tokens, and fu..."
186801,Allow swap_tensors with internal TensorWeakRef,https://github.com/pytorch/pytorch/pull/186801,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/slow, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,186796,186796,,,,0,5,1,,,"Merged, Reverted, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/slow, ci-no-td, agentic",2026-06-09T16:06:00Z,2026-07-08T06:02:14Z,0,"Stack from ghstack (oldest at bottom): -> #186801 torch.utils.swap_tensors rejected any Python weakref on either tensor before swapping. Dynamo guards use torch.utils.weak.TensorWeakRef, but that class wrapped a plain weakref.ref, so weakref.getweakrefs() could not distinguish Dynamo's internal tensor weakrefs from..."
160585,Add dim=None support to torch.logsumexp,https://github.com/pytorch/pytorch/pull/160585,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,156075,156075,"156075, 174929",,,,0,4,4,,,"oncall: distributed, open source, Stale, release notes: jit, module: inductor, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2025-08-13T22:45:58Z,2026-07-08T05:58:24Z,0,"Fixes #156075 Update: Rebased and brought to date. In order to support dim=None for torch.logsumexp, I followed the pattern used by torch.sum and torch.mean, as far as possible. Changed the function signature in native_functions.yaml to make dimension argument optional Changes in ReduceOps.cpp to modify (1) meta fun..."
189084,[DO NOT MERGE] Testing Rocm 7.14,https://github.com/pytorch/pytorch/pull/189084,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/inductor-rocm-mi200","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"module: rocm, open source, ciflow/trunk, topic: not user facing, keep-going, ciflow/rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/docker",2026-07-06T20:38:08Z,2026-07-08T05:57:52Z,0,This PR is not meant to be merged but only serves to test Inductor on ROCm 7.14 cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
188480,[Dynamo] Add UNSPECIFIED_BACKEND lint rule requiring explicit backend=,https://github.com/pytorch/pytorch/pull/188480,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,8,2,,,"open source, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/docker",2026-06-29T21:01:30Z,2026-07-08T05:57:36Z,0,"Stack from ghstack (oldest at bottom): -> #188480 Follow-up to the test_misc.py eager-backend cleanup, addressing @williamwen42's review ask for a lint rule on test/dynamo tests. torch.compile's default backend is ""inductor"", which pulls in the full codegen stack. Most Dynamo tests only exercise tracing/correctness..."
183994,Specialize numpy scalar module attrs,https://github.com/pytorch/pytorch/pull/183994,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,172942,172942,,,,0,6,5,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-16T03:07:47Z,2026-07-08T05:50:01Z,0,Stack from ghstack (oldest at bottom): -> #183994 Treat NumPy floating module attributes as guarded scalar constants so they do not enter compiled graphs as CPU tensor inputs. Preserve NumPy dtype semantics for tensor factories and keep CUDA graphs free of scalar DeviceCopy partitions.\n\nFixes #172942\nGenerated by...
186764,fix: Use typing_extensions.TypeAliasType for better reexport of `__module__`,https://github.com/pytorch/pytorch/pull/186764,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,171905,171905,171905,,,,0,3,3,,,"triaged, open source, release notes: quantization",2026-06-09T04:35:04Z,2026-07-08T05:48:41Z,0,"Summary PyTorch uses a pattern of setting module on re-exported symbols to fix their apparent origin module. For plain TypeAlias annotations (e.g., MyAlias: TypeAlias = int | str), the underlying object is a raw Python type like types.UnionType or typing.Any, which does not support module mutation. Root cause PyTorc..."
186913,Fix int64 value_expr arithmetic in Triton codegen,https://github.com/pytorch/pytorch/pull/186913,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,183901,183901,,,,0,4,3,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-10T08:06:48Z,2026-07-08T05:47:06Z,0,"Stack from ghstack (oldest at bottom): -> #186913 Fixes #183901 Inductor already distinguishes value-producing symbolic expressions from pure indexing expressions through value_expr, but Triton still printed the symbolic expression using the original range symbols. For small kernels those range symbols are int32, so..."
188327,[test] Add error_inputs for nn.Conv2d module,https://github.com/pytorch/pytorch/pull/188327,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,174183,174183,174183,,,,0,1,2,,,"triaged, open source, topic: not user facing",2026-06-27T08:29:59Z,2026-07-08T05:39:05Z,0,"Summary nn.Conv2d had no module_error_inputs_func defined in common_modules.py, meaning error message regressions were not tested. This adds module_error_inputs_torch_nn_Conv2d with three cases covering the main error paths: 2D input (expects 3D/4D) -> RuntimeError: expected 3D (unbatched) or 4D (batched) input to c..."
182376,[CUDA][CPU] Unify CPU/GPU relu/max behavior,https://github.com/pytorch/pytorch/pull/182376,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,181801,,,,0,3,2,,,"module: cuda, module: cpu, triaged, open source, Stale, release notes: nn",2026-05-04T22:13:54Z,2026-07-08T05:34:02Z,0,For #181801 authored with codex cc @ptrblck @msaroufim @jerryzh168 @tinglvv @nWEIdia @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
185751,"[nn] Raise ValueError early for invalid (ndim, pad_size) in non-constant F.pad modes",https://github.com/pytorch/pytorch/pull/185751,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/trunk,review when ready,,UNKNOWN,False,0,0,,,184845,184845,184845,,,,0,1,2,,,"triaged, open source, ciflow/trunk, release notes: nn, topic: not user facing",2026-05-31T09:39:23Z,2026-07-08T05:20:35Z,0,"Summary torch.nn.functional.pad with mode='reflect', 'replicate', or 'circular' silently accepted invalid (input_ndim, pad_size) combinations and raised NotImplementedError from deep inside the C++ kernel dispatch table (PadNd.cpp _pad_enum_symint). The error arrived at the wrong level with the wrong exception type,..."
187908,[fix] torch.where silently overflows fp16 scalars (issue #187429),https://github.com/pytorch/pytorch/pull/187908,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,187429,187429,187429,,,,0,1,3,,,"triaged, open source, ciflow/trunk, release notes: python_frontend",2026-06-23T03:48:25Z,2026-07-08T05:20:33Z,0,"Fixes #187429 Root cause Commit 0eae6b68f42 (""Unify torch.tensor and torch.ops.aten.scalar_tensor behavior"") intentionally relaxed the cast in fill_inplace<Half> for numel==1 tensors: it switched from checked_convert to static_cast so scalar_tensor matches torch.tensor() semantics. That change had an unintended side..."
185756,[clamp] Fix float16 scalar overflow check inconsistency between CPU and GPU,https://github.com/pytorch/pytorch/pull/185756,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/trunk,review when ready,,UNKNOWN,False,0,0,,,171356,171356,171356,187908,,,0,1,3,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-05-31T12:04:31Z,2026-07-08T05:20:15Z,0,"Summary torch.clamp / torch.clip on float16 tensors had inconsistent validation between CPU and GPU. When a scalar bound exceeds the float16 range (~65504), CPU correctly raises RuntimeError but GPU silently succeeded and returned incorrect results - the out-of-range bound saturates to inf when stored back as float1..."
187614,Use sym_numel in linalg_cond string overload,https://github.com/pytorch/pytorch/pull/187614,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,186028,186028,,,,0,4,1,,,"release notes: linalg_frontend, module: inductor, ciflow/inductor, agentic, needs-review",2026-06-17T21:52:30Z,2026-07-08T04:57:51Z,0,"Stack from ghstack (oldest at bottom): -> #187614 torch.linalg.cond has separate overloads for scalar/None norms and string norms. The scalar/None overload already checks empty inputs with sym_numel(), but the string overload used by p=""fro"" and p=""nuc"" still queried numel(). Fake tensors with symbolic sizes cannot..."
173948,Fixes LazyConvTranspose modules when used with torch.compile(dynamic=True),https://github.com/pytorch/pytorch/pull/173948,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,173252,173252,173252,,,,0,4,2,,,"triaged, open source, Stale, release notes: nn, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-01-30T19:28:17Z,2026-07-08T04:51:20Z,0,Fixes #173252 The issue was that lazy modules tried to use SymInt dimensions directly for materialization which requires concrete integers. Now converts SymInt to concrete int using optimization_hint( ). cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiay...
189205,[DO NOT MERGE] Harmless comment just to enable CI,https://github.com/pytorch/pytorch/pull/189205,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi350","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing, keep-going, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi350",2026-07-07T21:33:53Z,2026-07-08T04:51:10Z,0,
176072,Fix Triton launcher arg mismatch silent failure,https://github.com/pytorch/pytorch/pull/176072,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,146018,146018,,"146018, 174534, 175228",,,0,7,6,,,"triaged, open source, Stale, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-02-28T14:04:23Z,2026-07-08T04:50:40Z,0,"Fixes #146018. I saw the reviewer feedback from @jansel on #175228 and decided to implement the requested changes. Added test/inductor/test_launcher_validation.py Ensured _validate_launcher_args(None, ...) handles the None case safely This supersedes #175228 and #174534 by fully addressing the requested reviews. cc..."
183652,Fix dynamic slice lowering for generic views,https://github.com/pytorch/pytorch/pull/183652,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,183259,183259,,,,0,19,3,,,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ci-no-td, autorevert: disable, ciflow/torchtitan, agentic",2026-05-14T02:43:44Z,2026-07-08T04:50:00Z,0,"Stack from ghstack (oldest at bottom): -> #183652 Handle dynamic slice bounds on generic Inductor views without requiring stride metadata, and fall back for unresolved slice_scatter bounds so eager validation is preserved. Fixes #183259 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Ch..."
181109,tools/stats: add script to generate test-times.json from local JUnit XML reports,https://github.com/pytorch/pytorch/pull/181109,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,"141550, 171659, 180975",,,,0,3,9,,,"module: tests, triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-22T14:22:35Z,2026-07-08T04:49:48Z,0,"Summary External architecture CI pipelines cannot reach PyTorch's internal ClickHouse/S3 infrastructure to pull the test timing data that run_test.py uses for time-based sharding. Without it, the sharding falls back to round-robin, producing unbalanced shards and inflated wall time. This PR adds tools/stats/generate..."
182377,[DTensor] SymmetricMemory-backed DTensors,https://github.com/pytorch/pytorch/pull/182377,ready for maintainer decision,45,"CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,182382,,,,0,11,1,,,"triaged, open source, ciflow/inductor, release notes: distributed (dtensor), module: symm_mem, ciflow/torchtitan, ciflow/dtensor",2026-05-04T22:17:37Z,2026-07-08T04:44:20Z,0,"This PR adds an opt-in path for DTensor to allocate local tensor shards using SymmetricMemory. This is part of the one-sided DTensor work described in #182382. Allocating local tensor shards using SymmetricMemory makes them remotely addressable, enabling one-sided communication operations. Later PRs can introduce al..."
184934,Validate packed RNN batch_sizes before fallback,https://github.com/pytorch/pytorch/pull/184934,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,173046,173046,,,,0,1,1,,,"topic: not user facing, agentic, needs-review",2026-05-22T21:49:11Z,2026-07-08T04:29:01Z,0,"Stack from ghstack (oldest at bottom): -> #184934 Packed RNN fallback code reads batch_sizes through const_data_ptr<int64_t>() and indexes it on the host. If a CUDA batch_sizes tensor reaches this path, for example through torch.compile(..., backend=""eager"") on aten.rnn_tanh, the fallback can dereference device memo..."
184938,Fix AOTAutograd forward output overwritten by backward mutation,https://github.com/pytorch/pytorch/pull/184938,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,172956,172956,,,,0,2,4,,,"module: dynamo, ciflow/inductor, release notes: aot autograd, agentic, needs-review",2026-05-22T22:17:45Z,2026-07-08T04:26:00Z,0,"Stack from ghstack (oldest at bottom): -> #184938 AOTAutograd traces the joint forward/backward graph under functionalization. If a custom autograd.Function saves a forward-created tensor, returns the same tensor from forward, and mutates it in backward, functionalization updates the wrapper for that returned tensor..."
184950,Fix mkldnn transposed conv pointwise shape inference,https://github.com/pytorch/pytorch/pull/184950,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,4,1,"statusCheckRollup, linux-jammy-cuda13.0-py3.10-gcc11 / test (default, 2, 5, linux.g6.4xlarge.experimental.nvidia.gpu), linux-jammy-cuda13.0-py3.10-gcc11 / test-osdc (default, 1, 5, mt-l-x86aavx2-29-113-l4), linux-jammy-rocm-py3.10-mi355 / test (default, 1, 10, linux.rocm.gpu.gfx950.1)",statusCheckRollup,,172711,172711,,,,0,1,2,eb0a7159b1cb775ff9b8cfd3380ceb29a14782fbaf7946987ce87b812abad623,7ea5cd17b32d6c470464abac0abb4600fbe113fec6f9e3828df61ec210c15b2b,"module: cpu, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-23T00:54:55Z,2026-07-08T04:19:40Z,0,Stack from ghstack (oldest at bottom): -> #184950 The mkldnn transposed convolution pointwise meta and Inductor lowering paths treated every transposed-convolution weight as if it were a reordered oneDNN weight. That is only valid for prepacked MKLDNN weights. Regular PyTorch transposed-convolution weights are laid...
184747,Remove no-op SDPA masks before backend dispatch,https://github.com/pytorch/pytorch/pull/184747,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,177712,177712,,,,0,1,1,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-21T16:23:09Z,2026-07-08T04:16:39Z,0,"Stack from ghstack (oldest at bottom): -> #184747 Dynamo can trace library masking helpers that materialize full-attention masks to avoid data-dependent checks, but a non-null mask prevents FlashAttention dispatch. Canonicalize structurally proven all-true boolean masks and zero additive masks to None before eager/I..."
183879,[inductor][cpu] Enable int8 WOQ GEMM template on Arm,https://github.com/pytorch/pytorch/pull/183879,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,177327,177327,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-15T10:38:21Z,2026-07-08T04:13:41Z,0,"Stack from ghstack (oldest at bottom): -> #183879 Register Arm vector micro-GEMM configs for bf16 activation and int8 WOQ weights, and use portable generated int8 load/prefetch helpers so the CPU GEMM template can be selected on AArch64. Fixes #177327 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgo..."
183881,Fix online softmax epilogue fusion for T5,https://github.com/pytorch/pytorch/pull/183881,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,177118,177118,,,,6,4,3,c6fda8258e58750535f9c31283904e5c3b1900ae09b0f017a0b172b4ed74b021,32dca12c83a725b66782cb6ca3828a0a363702f0dec6965f9fae5b846edc3298,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-15T11:25:03Z,2026-07-08T04:09:02Z,0,"Stack from ghstack (oldest at bottom): -> #183881 Keep T5-style online softmax epilogues fused when their non-contiguous pointwise reads are compatible recomputations of reduction inputs, while preserving fully non-contiguous softmax behavior. Fixes #177118 Generated by my agent cc @voznesenskym @penguinwu @EikanWan..."
183908,Fix deferred runtime assert symbol availability,https://github.com/pytorch/pytorch/pull/183908,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,"154146, 176770","154146, 176770",,,,0,1,2,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-15T15:17:13Z,2026-07-08T04:05:11Z,0,"Stack from ghstack (oldest at bottom): -> #183908 Avoid emitting deferred Inductor runtime asserts until every symbol in the assertion is available to the generated wrapper. Stale backward assertions that reference unavailable backed symbols are skipped instead of producing Python with undefined names, while missing..."
183944,Fix compiled autograd clear saved tensors semantics,https://github.com/pytorch/pytorch/pull/183944,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,176112,,176112,,,0,3,2,,,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, module: compiled autograd, ciflow/torchtitan, agentic, needs-review",2026-05-15T19:25:18Z,2026-07-08T04:01:54Z,0,Stack from ghstack (oldest at bottom): -> #183944 Compiled autograd now preserves clear_saved_tensors_on_access behavior for Python custom Functions by avoiding graph locals that keep saved tensors alive and by enforcing one-shot fake ctx access. Fixes #176112 Generated by my agent cc @voznesenskym @penguinwu @Eikan...
188043,[Test] Make test/test_sort_and_select.py device-agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/188043,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing, module: tests/core",2026-06-24T08:27:19Z,2026-07-08T03:59:54Z,0,"Refactors test/test_sort_and_select.py to decouple tests from specific device types, enabling out-of-tree backends to run the test suite without CUDA-only restrictions. Changes Split into four test classes: TestSortAndSelect — Device-independent tests (no device parametrization). Contains tests that do not require a..."
183954,Disallow mid-trace reference opaque creation,https://github.com/pytorch/pytorch/pull/183954,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,175968,175968,,,,8,9,16,8293951c7c7b60c01423bad4eabec894bd175cffbfbc402cb34d3f300cdd83ad,778a1f478262a5a875c8710c0fb4c168d2020fb37a576e22905040e6b7b1aed8,"ciflow/trunk, release notes: fx, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-15T20:02:10Z,2026-07-08T03:56:03Z,0,"Stack from ghstack (oldest at bottom): -> #183954 Reference-type opaque objects created during tracing can reach FX argument creation as raw Python objects. If FX interns such an untracked object as an _opaque_obj* get_attr constant, the compiled graph has no input guard or reconstruction path for it, and Inductor c..."
184803,[dynamo] Avoid key-set guard for dict setitem,https://github.com/pytorch/pytorch/pull/184803,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175858,175858,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-21T21:52:19Z,2026-07-08T03:53:31Z,0,Stack from ghstack (oldest at bottom): -> #184803 Dynamo installed DICT_KEYS_MATCH before tracing ConstDictVariable.setitem. That full key-set guard is not needed for setitem side-effect replay: replay updates the assigned key and does not depend on unrelated keys. The guard made compiled functions recompile once af...
183958,Enable SDPA constraint cache by default,https://github.com/pytorch/pytorch/pull/183958,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,175833,175833,,,,0,1,2,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-15T20:17:46Z,2026-07-08T03:48:03Z,0,Stack from ghstack (oldest at bottom): -> #183958 The existing SDPA constraint cache deduplicates copies when shared K/V tensors hit the same layout requirement. Enable it by default and add a regression test for the issue repro. Fixes #175833 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guo...
186366,Support vmapped compiled autograd.Functions,https://github.com/pytorch/pytorch/pull/186366,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,129845,129845,,,,10,4,8,d3c6240c88571090fdfd3a93e6a718976282bdef42b9e76d5e49196fedfc099b,a4bd5d25cd7adbf3a1637e2ab53990a2535961d3aa8862b6feb27b21538e47c2,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-05T12:25:53Z,2026-07-08T03:44:03Z,1,"Stack from ghstack (oldest at bottom): -> #186366 Dynamo lowers captured autograd.Function calls through an internal ApplyTemplate autograd.Function. Two pieces were missing for the issue's composition of vmap, autograd.Function with generate_vmap_rule=True, and compile. First, ctx.mark_non_differentiable handling a..."
180186,Try to enable LTO in OSS,https://github.com/pytorch/pytorch/pull/180186,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,15,0,,,"module: cpu, open source, Stale, ciflow/binaries, release notes: releng",2026-04-12T19:27:53Z,2026-07-08T03:41:17Z,1,Modern revival of #137866 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
183681,Fix Triton compile-time autotune fallback,https://github.com/pytorch/pytorch/pull/183681,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,182093,,182093,,,0,22,2,,,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ci-no-td, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, agentic",2026-05-14T08:39:44Z,2026-07-08T03:40:48Z,1,Stack from ghstack (oldest at bottom): -> #183681 Recover Triton launchers by disabling pipelining for each recoverable shared-memory/OOM failure during compile-time autotune. This keeps compile_results aligned with the launchers that survived filtering and avoids hardware-specific skips. The separate free-threaded...
186845,[make_fx] Decompose detach by default for higher-order grads,https://github.com/pytorch/pytorch/pull/186845,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175477,175477,,,,0,2,1,,,"release notes: fx, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-09T21:07:14Z,2026-07-08T03:40:38Z,1,"Stack from ghstack (oldest at bottom): -> #186845 make_fx's post-dispatch tracing can record detach calls emitted by autograd saved-variable handling while tracing a backward with create_graph=True. When the resulting graph is replayed and users take a second derivative through it, those detach nodes cut gradient fl..."
186273,[TEST][CUDA][TF32] Re-enable TF32 setter check test,https://github.com/pytorch/pytorch/pull/186273,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cuda, module: tests, triaged, open source, module: tf32, ciflow/trunk, topic: not user facing",2026-06-04T21:50:10Z,2026-07-08T03:39:16Z,1,Also check if cuBLAS TF32 override was set authored with codex cc @ptrblck @msaroufim @tinglvv @nWEIdia @mruberry @zasdfgbnm
184838,Fix Dynamo raw descriptor building,https://github.com/pytorch/pytorch/pull/184838,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175264,175264,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-22T03:09:07Z,2026-07-08T03:35:58Z,1,Stack from ghstack (oldest at bottom): -> #184838 Dynamo had descriptor-specific VariableTracker classes and used them when UserDefinedClassVariable performed normal class attribute lookup. Raw descriptor objects reached through class dictionaries or mappingproxy values used the generic sourceful and sourceless buil...
187104,Fix embedding_bag CPU segfault when offsets is empty and indices is not,https://github.com/pytorch/pytorch/pull/187104,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,175370,175370,175370,,,,0,1,2,,,"triaged, open source, release notes: nn, topic: bug fixes",2026-06-11T23:21:32Z,2026-07-08T03:34:58Z,1,"Fixes #175370 F.embedding_bag on CPU segfaults when offsets is empty and indices is not. Root cause With zero bags, make_offset2bag fills offset2bag with -1: the index_add_ over empty offsets is a no-op, so the initial offset2bag[0] -= 1 propagates through the cumsum. The slow-path kernels (index_select_add, index_s..."
184848,Fix nearest upsample scale shortcuts in compile,https://github.com/pytorch/pytorch/pull/184848,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175154,175154,,,,0,1,1,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-22T03:56:14Z,2026-07-08T03:32:45Z,1,"Stack from ghstack (oldest at bottom): -> #184848 torch.compile used the Python decomposition and Inductor lowering for nearest upsampling, and both always computed source indices from the explicit inverse scale factor when scales were supplied. Native eager has backend-specific legacy paths that do not always do th..."
189181,[CI] Advisor: split garbage verdict into infra_issue + garbage (minimal),https://github.com/pytorch/pytorch/pull/189181,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,"8213, 188042",,,0,2,4,,,topic: not user facing,2026-07-07T18:36:43Z,2026-07-08T03:15:15Z,1,"Adds an infra_issue verdict to the Claude CI Advisor (claude-autorevert-advisor.yml) and cleanly separates the three ""no-revert, not-the-suspect"" verdicts. This is a minimal alternative to the prompt change in #188042 — no new principle paragraph and no enumerated error-string signatures; it mostly moves infra out o..."
189011,include activation epilogue in GemmAndBiasParams::Signature(),https://github.com/pytorch/pytorch/pull/189011,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, release notes: linalg_frontend",2026-07-06T07:08:02Z,2026-07-08T02:58:16Z,1,Issue Description This issue can be triggered by torch._addmm_activation() when called with different use_gelu values on the same tensor shape. GemmAndBiasParams defines two signature methods: BLASSignature() — includes activation Signature() — omits activation Because GemmAndBiasParams::Signature() omits activation...
185764,Recommend nonstrict_trace for skipped C extensions,https://github.com/pytorch/pytorch/pull/185764,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,149565,149565,,,,3,6,3,2493a4f731c75be4f1e2aad9d687aa16482ab3f95e6ceede1a9edfaa26130f60,ba0a3fe7339e9758345eaea96b7536669e17a002cd998d2e24e09f75118de37a,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-05-31T17:21:23Z,2026-07-08T02:39:09Z,1,Stack from ghstack (oldest at bottom): -> #185764 The Dynamo skipped builtin/C-extension hint still pointed traceable functions at torch.compiler.allow_in_graph. That guidance is now stale for this path: nonstrict_trace is the better escape hatch because it traces the wrapped callable without forcing users to allow...
183969,Fix CSE across unrelated mutation regions,https://github.com/pytorch/pytorch/pull/183969,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,174472,174472,,,,0,2,1,,,"release notes: inductor, agentic, needs-review",2026-05-15T22:24:59Z,2026-07-08T02:36:28Z,1,"Stack from ghstack (oldest at bottom): -> #183969 Allow fx_graph_cse to reuse duplicate pure aten ops across mutation region boundaries when all intervening mutable aten writes are known not to alias the candidate inputs or the previous duplicate result. Add regression coverage for unrelated writes, input aliases, u..."
184864,Error on overlapping as_strided mutations in compile,https://github.com/pytorch/pytorch/pull/184864,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,174371,174371,,,,0,1,1,,,"release notes: aotdispatcher, agentic, needs-review",2026-05-22T08:12:47Z,2026-07-08T02:34:02Z,1,"Stack from ghstack (oldest at bottom): -> #184864 Functionalization rewrites in-place mutations on as_strided views into scatter-style updates. For views with internal overlap, eager mutation has undefined write ordering and may update the same storage location more than once, while the functionalized as_strided_sca..."
185108,Prevent non-strict export from leaking fake buffers,https://github.com/pytorch/pytorch/pull/185108,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,5,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,163189,163189,,,,0,5,1,4df3357c06d6f3bc8599c62ee848adfdc124f2df31eaa95f6efadbbe1a00e914,85672602720e42f3ce3667b9bef4d5136496bc544ed4b0463c8014d211b95513,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-25T05:52:58Z,2026-07-08T02:29:12Z,1,"Stack from ghstack (oldest at bottom): -> #185108 Non-strict export runs the module under FakeTensorMode. A buffer registered with a None value is not returned by named_buffers(), so it is not reparameterized with the rest of the module state. If forward assigns a tensor to that buffer, the fake tracing value is wri..."
184948,Fix dynamo export for tensor closures in output pytrees,https://github.com/pytorch/pytorch/pull/184948,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,172839,172839,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-23T00:22:50Z,2026-07-08T02:25:10Z,1,"Stack from ghstack (oldest at bottom): -> #184948 The private _dynamo_graph_capture_for_export path traces ModuleToTrace.forward, which flattens the user output and returns an ExportTracerOutput containing both flat outputs and the output TreeSpec. OutputGraph then tried to convert that TreeSpec to a Python constant..."
189195,Cuda Graph Default Enable (#189195),https://github.com/pytorch/pytorch/pull/189195,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-07-07T20:27:25Z,2026-07-08T02:23:24Z,1,Summary: Differential Revision: D110936261
189209,[inductor] Fix with_effects star deps keyed by buffer name instead of operation name,https://github.com/pytorch/pytorch/pull/189209,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T23:11:30Z,2026-07-08T02:23:04Z,1,"Stack from ghstack (oldest at bottom): -> #189209 The with_effects lowering records effect-chain ordering into V.graph.additional_star_deps keyed by new_op.get_name(). For FallbackKernel (an OperationBuffer), get_name() resolves to Buffer.get_name() and returns the buffer name (bufN), but Scheduler.compute_dependenc..."
183731,[inductor] Validate fallback output dtype metadata,https://github.com/pytorch/pytorch/pull/183731,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,180807,180807,,,,0,17,1,,,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, ci-no-td, release notes: inductor (aoti), ciflow/torchtitan, agentic",2026-05-14T15:35:13Z,2026-07-08T02:23:00Z,1,"Stack from ghstack (oldest at bottom): -> #183731 Extend Inductor's fallback output metadata checks to validate dtype alongside size and stride, so bad custom-op fake/meta kernels fail fast before downstream static Triton kernels can consume mismatched tensor dtypes and silently compute incorrect results. Fixes #180..."
188747,Python c10d Backend trampoline cleanup,https://github.com/pytorch/pytorch/pull/188747,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,187479,,,0,11,1,,,"triaged, open source, release notes: distributed (c10d)",2026-07-01T22:21:57Z,2026-07-08T02:21:56Z,1,"Built on top of #187479 Summary Cleans up dead code in the override dispatch. PyBackend objects are created via c10::make_intrusive in wrap(), which bypasses pybind11's type system. pybind11::get_override can never find overrides on these objects, so all PYBIND11_OVERRIDE fallbacks were dead code. Replaced them with..."
184958,Prevent escaped vjp TensorWrappers from reaching AOTAutograd,https://github.com/pytorch/pytorch/pull/184958,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,24,1,"statusCheckRollup, linux-jammy-py3.10-clang18-asan / test (default, 4, 7, linux.4xlarge), linux-jammy-py3.10-clang18-asan / test-osdc (default, 5, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test-osdc (default, 4, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test-osdc (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test (default, 4, 5, linux.4xlarge), linux-jammy-py3.10-gcc11 / test (default, 1, 5, linux.4xlarge), linux-jammy-py3.14t-clang18 / test (crossref, 2, 2, linux.2xlarge), linux-jammy-py3.14t-clang18 / test-osdc (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test (default, 1, 5, linux.4xlarge), linux-jammy-py3.14t-clang18 / test (default, 4, 5, linux.4xlarge), linux-jammy-py3.10-clang18 / test (crossref, 1, 2, linux.2xlarge), linux-jammy-py3.10-clang18 / test-osdc (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test (dynamo_wrapped, 3, 3, linux.2xlarge), linux-jammy-py3.10-clang18 / test (default, 1, 5, linux.4xlarge), linux-jammy-py3.10-clang18 / test (default, 4, 5, linux.4xlarge), linux-jammy-py3.14-clang18 / test (default, 4, 5, linux.4xlarge), linux-jammy-py3.14-clang18 / test (dynamo_wrapped, 3, 3, linux.2xlarge), linux-jammy-py3.14-clang18 / test (crossref, 2, 2, linux.2xlarge), linux-jammy-py3.14-clang18 / test-osdc (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test (default, 1, 5, linux.4xlarge), linux-jammy-aarch64-py3.10 / test-osdc (default, 4, 5, mt-l-arm64g4-16-62), linux-jammy-aarch64-py3.10 / test (default, 3, 5, linux.arm64.m8g.4xlarge), linux-jammy-aarch64-py3.10 / test (default, 1, 5, linux.arm64.m8g.4xlarge)",statusCheckRollup,,172026,172026,,,,6,2,5,70a71962285f7a0246512e42ead921591c8fd68636528fc0fa47821e438b7b31,e67b6fa4c919693fc4a31ece4c96d45281662fe7164790b99394f78f3284eda7,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-23T03:02:08Z,2026-07-08T02:20:24Z,1,"Stack from ghstack (oldest at bottom): -> #184958 When Dynamo traces through torch.func.vjp, the returned vjp_fn can capture functorch TensorWrappers from the temporary grad level. This is fine when the vjp_fn is consumed inside the compiled region, but if the callable escapes as a graph output those wrapped tensors..."
181357,[dynamo] Add polyfill for operator.indexOf,https://github.com/pytorch/pytorch/pull/181357,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,116396,,,,0,2,2,,,"triaged, open source, Stale, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-24T09:31:14Z,2026-07-08T02:18:56Z,1,"operator.indexOf currently causes a graph break in Dynamo because it is a C builtin with no traceable Python implementation: Failed to trace builtin operator indexOf with argument types ['list', 'int'] This PR adds a substitute_in_graph polyfill following the same pattern used for operator.countOf, operator.concat,..."
179376,fix(export): add SDPA decomposition for contiguous output,https://github.com/pytorch/pytorch/pull/179376,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,179287,179287,,,,,0,4,0,,,"triaged, open source, Stale, topic: not user facing, module: dynamo",2026-04-04T09:54:22Z,2026-07-08T02:18:50Z,1,"Summary When exporting nn.MultiheadAttention with need_weights=False and running on CUDA, the exported graph contains permute->view without contiguous() in between, causing a view error. Problem The exported graph shows: scaled_dot_product_attention = ... permute = torch.ops.aten.permute.default(...) view_8 = torch...."
180689,[ROCm] Handle profiler trace format differences in combo kernel tests,https://github.com/pytorch/pytorch/pull/180689,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/rocm-mi300",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,0,0,,,,"180017, 180018, 180019",,"180017, 180018, 180019",,,0,1,3,,,"module: rocm, open source, Stale, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300",2026-04-17T17:18:12Z,2026-07-08T02:18:49Z,1,"Fixes #180017. Fixes #180018. Fixes #180019. Three combo kernel tests access event[""args""][""grid""] from profiler chrome traces. The grid key structure differs between CUPTI and ROCTracer trace formats, so this throws KeyError on ROCm. Use .get() and only assert on grid values when present. All other validation (nume..."
176876,[inductor] Add device-type benchmarking dispatch registry,https://github.com/pytorch/pytorch/pull/176876,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, Stale, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-03-09T09:44:54Z,2026-07-08T02:18:46Z,1,"Introduce a pluggable device-type → function registry for benchmarking (torch._inductor.runtime.benchmarking): • Add _BENCHMARK_DISPATCH and register_benchmarker(device_type, fn, *, override=False) • Have Benchmarker.benchmark consult the registry first; retain the existing CPU/GPU fallback for backward compatibilit..."
184961,Fix packed RNN export fake metadata,https://github.com/pytorch/pytorch/pull/184961,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,171978,171978,,,,0,4,1,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-05-23T04:12:02Z,2026-07-08T02:15:57Z,1,"Stack from ghstack (oldest at bottom): -> #184961 torch.export could fail on models using pack_padded_sequence followed by an RNN module because packed RNN module paths used batch_sizes[0] directly when constructing default hidden states. During export, that value can be represented by a FakeTensor scalar, which the..."
184964,Handle tensor scalar slice bounds in export setitem,https://github.com/pytorch/pytorch/pull/184964,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,171974,171974,,,,0,1,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-23T05:29:13Z,2026-07-08T02:13:43Z,1,"Stack from ghstack (oldest at bottom): -> #184964 Non-strict export already rewrites Tensor.getitem with SymInt indices and slices into traceable select/slice calls. Tensor.setitem still fell through to Python indexing, and tensor scalar slice bounds inside slice objects reached that path as Tensor objects. Python i..."
184966,Preserve unflattened module forward signatures,https://github.com/pytorch/pytorch/pull/184966,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,171964,171964,,,,0,1,1,,,"release notes: export, agentic, needs-review",2026-05-23T05:47:24Z,2026-07-08T02:11:28Z,1,"Stack from ghstack (oldest at bottom): -> #184966 torch.export.unflatten produces an UnflattenedModule whose runtime forward implementation accepts *args, **kwargs. Re-exporting that module with dynamic shapes asks export to bind example inputs through inspect.signature(module.forward), so a call like module(x) was..."
184967,Fix export CIA decomps with backend kernels,https://github.com/pytorch/pytorch/pull/184967,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,171897,171897,,,,0,1,1,,,"release notes: export, agentic, needs-review",2026-05-23T06:33:16Z,2026-07-08T02:08:48Z,1,"Stack from ghstack (oldest at bottom): -> #184967 ExportedProgram.run_decompositions splits CompositeImplicitAutograd (CIA) decompositions into a CIA override table and a regular ProxyTensor decomposition table. For CIA ops with runtime backend kernels, ProxyTensor preserves the backend kernel for eager fidelity ins..."
184970,Preserve pin_memory in generated view copy kernels,https://github.com/pytorch/pytorch/pull/184970,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,171894,171894,173517,,,0,2,4,,,"topic: not user facing, agentic, needs-review",2026-05-23T06:56:02Z,2026-07-08T02:05:15Z,1,"Stack from ghstack (oldest at bottom): -> #184970 Generated view-copy kernels clone their view outputs to implement copy semantics. That includes aten.lift_fresh_copy, which torch.compile uses for lifted tensor constants. Tensor.clone intentionally does not preserve CPU pinned-memory allocation, so a user tensor cre..."
185815,Fix Dynamo NumPy scalar identity handling,https://github.com/pytorch/pytorch/pull/185815,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,146951,146951,,,,0,13,1,71f50c3ec1781fa4589635a6db0c69327adf8e6eeaebee9b6bf186de6b9f8aea,be89140e88bb10601c9bfdbb3ebd4e3307f31bb51630e1bfb194c34598530624,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T13:27:48Z,2026-07-08T02:02:30Z,1,"Stack from ghstack (oldest at bottom): -> #185815 Dynamo represents NumPy scalars as 0-d NumpyNdarrayVariable values, which inherits from TensorVariable. The existing identity comparison path could then fall through to tensor fake-value identity, losing Python object identity for NumPy scalar singletons. This made c..."
184974,Fix register_autograd wrappers under torch.func.grad,https://github.com/pytorch/pytorch/pull/184974,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,170834,170834,,,,0,2,5,,,"module: custom-operators, release notes: torch.func, agentic, needs-review",2026-05-23T08:48:22Z,2026-07-08T01:57:49Z,1,"Stack from ghstack (oldest at bottom): -> #184974 torch.library.register_autograd generates an autograd wrapper for the registered operator formula. That wrapper used to inherit from autograd.Function while keeping the legacy forward(ctx, ...) calling convention and not defining setup_context. When torch.func.grad w..."
184975,Support FakeTensor module materialization via to_empty,https://github.com/pytorch/pytorch/pull/184975,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,170781,170781,,,,0,2,1,,,"topic: not user facing, agentic, needs-review",2026-05-23T09:04:32Z,2026-07-08T01:53:49Z,1,"Stack from ghstack (oldest at bottom): -> #184975 Module._apply forced the swap_tensors path whenever the existing parameter was a FakeTensor. FakeTensorMode's converter keeps weak references to converted tensors, so swapping fake parameters can fail with the weakref guard instead of using the normal module conversi..."
184976,Fix FakeTensor Module._apply conversion with weakrefs,https://github.com/pytorch/pytorch/pull/184976,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,"153056, 170770","153056, 170770",,,,0,1,2,,,"topic: not user facing, agentic, needs-review",2026-05-23T09:21:41Z,2026-07-08T01:51:21Z,1,"Stack from ghstack (oldest at bottom): -> #184976 FakeTensorMode keeps converter memo entries using weak references. Module._apply was forcing every existing FakeTensor parameter through torch.utils.swap_tensors, but swap_tensors rejects tensors that have weakrefs associated with them. This made even a no-op dtype c..."
184977,Fix Dynamo wrapper method sources for nested compile,https://github.com/pytorch/pytorch/pull/184977,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, unit-test / inductor-cpu-core-test (3.12) / test (inductor_core, 2, 2, linux.c7i.2xlarge)",statusCheckRollup,,170696,170696,,,,0,2,1,84326d921c85728553c27c2d7d228a0be7e9c4e17c19e2712c14450767f68d7e,1eadfbd26843d1eea236a0e8038408e9a985484f66e697f22ee6826792eaef37,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-23T09:41:47Z,2026-07-08T01:48:44Z,1,Stack from ghstack (oldest at bottom): -> #184977 A method decorated with torch.compile keeps the callable users invoke as an outer compile wrapper while the function Dynamo should inline lives on _torchdynamo_inline. Method resolution for user-defined objects and nn.Module class attributes was unwrapping _torchdyna...
188641,Mark linalg.eigh as cudagraph-unsafe in Inductor,https://github.com/pytorch/pytorch/pull/188641,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,188355,188355,,,,0,1,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-07-01T03:22:14Z,2026-07-08T01:43:28Z,1,"Stack from ghstack (oldest at bottom): -> #188641 torch.linalg.eigh lowers through Inductor as the fallback op aten._linalg_eigh.default. On CUDA this path is backed by cuSOLVER and is not safe to execute while a CUDA graph is being captured. Inductor's cudagraph partitioning did not classify the op as unsafe, so ma..."
184051,Honor custom op layout constraints in fallbacks,https://github.com/pytorch/pytorch/pull/184051,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,161113,161113,,,,2,8,7,5cc4e7db2c018d5de68a1defe7ea24bca18dff74ae334132b8c141c6269ed50b,7f00ac0848182919b078e2f2a23dbdc89f31544fb5b5fafbe8adbf3d161bf4cd,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ci-no-td, ciflow/torchtitan, agentic, needs-review",2026-05-16T22:41:31Z,2026-07-08T01:40:54Z,1,"Stack from ghstack (oldest at bottom): -> #184051 Inductor has several paths that can run an eager fallback for an op. The normal lowering path already had layout-constraint handling, but the direct fallback paths used by fallback-by-default and unsupported-type fallback called the fallback handler directly. That me..."
185953,Clear Dynamo weakrefs before graph-break resumes,https://github.com/pytorch/pytorch/pull/185953,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,145529,145529,,,,0,12,4,74abd4137a679d2704f71ebdb6c468a9940e664726e48cf7ef471210410930f0,1926cc5393cc76e133a5fbf7fc0f3866f81d55b7a7c3a35b9596a380ec525269,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T16:14:45Z,2026-07-08T01:37:41Z,1,"Stack from ghstack (oldest at bottom): -> #185953 Dynamo compile contexts can keep WeakIdRef entries in tracing and FakeTensor metadata after a subgraph is compiled. Registered backends already clear that state by default, but custom backends skipped the cleanup to preserve full-graph standalone/custom backend use c..."
185288,Support symbolic list lengths in Triton grids,https://github.com/pytorch/pytorch/pull/185288,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,161111,161111,,,,0,12,1,37523a36e82ab8896f88591639dd5aa3d785fbdc8be2d66c9b7b251e2058f362,c69bc96738ab44510817dc8c1fdb86873c9b73f786448ed1d21414c8a1f8afa3,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-26T22:21:23Z,2026-07-08T01:35:05Z,1,"Stack from ghstack (oldest at bottom): -> #185288 Dynamo failed to compile Triton grid functions that derive launch sizes from host-side tensor data. In the issue path, the grid computes sequence lengths with np.diff(lengths.cpu()), iterates NumPy scalar arrays through range(num), extends a Python list with those ra..."
184978,Improve export error for untraceable lifted constants,https://github.com/pytorch/pytorch/pull/184978,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,170672,170672,,,,0,1,1,,,"release notes: export, agentic, needs-review",2026-05-23T09:58:14Z,2026-07-08T01:31:12Z,1,"Stack from ghstack (oldest at bottom): -> #184978 When non-strict export sees a bogus lifted fake tensor constant, it used to raise an internal diagnostic ending with ""Please file an issue on github."" Issue #170672 hits this path by reading parameters through .data before copying back into the parameter under no_gra..."
187969,Enable traceable DTensor RNG ops on XPU,https://github.com/pytorch/pytorch/pull/187969,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,3010,3010,,,,,0,3,0,,,"open source, ciflow/trunk, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-23T19:30:45Z,2026-07-08T01:30:36Z,1,"Fixes intel/torch-xpu-ops#3010 Summary DTensor random ops (bernoulli, dropout, normal_, rand_like, etc.) failed under torch.compile on XPU because the traceable, graph-safe RNG path was CUDA-only. XPU fell back to the eager _distribute_region path, which materializes a real size-16 uint8 RNG-state tensor — leaking a..."
184981,Fix torch._numpy bool numeric subtract,https://github.com/pytorch/pytorch/pull/184981,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,170509,170509,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-23T10:38:55Z,2026-07-08T01:27:26Z,1,"Stack from ghstack (oldest at bottom): -> #184981 Dynamo traces NumPy and SciPy code through torch._numpy. scipy.stats geom.pmf computes expressions like (1 - cond) + np.isnan(k), where cond is a bool ndarray. NumPy rejects bool-bool subtraction, but mixed numeric/bool subtraction is valid and promotes the bool oper..."
189017,Respect the blocking parameter in torch.Event,https://github.com/pytorch/pytorch/pull/189017,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/mps, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,4,0,,,"open source, ciflow/trunk, release notes: mps, ciflow/mps",2026-07-06T09:20:32Z,2026-07-08T01:25:26Z,1,"Stack from ghstack (oldest at bottom): -> #189017 Motivation This issue was discovered in vllm-project/vllm#47081. Previously, torch.Event accepted the blocking parameter but ignored it. This PR makes torch.Event honor the blocking parameter, reducing the behavioral gap between torch.Event and torch.cuda.Event."
184983,Make SDPA backend checks safe for unbacked head dims,https://github.com/pytorch/pytorch/pull/184983,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,170127,170127,,,,0,2,1,,,"release notes: export, agentic, needs-review",2026-05-23T11:11:36Z,2026-07-08T01:23:08Z,1,"Stack from ghstack (oldest at bottom): -> #184983 torch.export can trace models that build repeat counts as tensors and pass them to repeat_interleave. Without an explicit output size, repeat_interleave creates unbacked symbolic sizes for the repeated K/V head dimension. The fused SDPA backend selectors then used no..."
184987,Fix index_select refs bounds semantics,https://github.com/pytorch/pytorch/pull/184987,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,169779,"169779, 187463",,,,0,2,4,,,"module: inductor, ciflow/inductor, release notes: inductor, agentic, needs-review",2026-05-23T12:56:00Z,2026-07-08T01:18:24Z,1,"Stack from ghstack (oldest at bottom): -> #184987 torch.index_select rejects negative and out-of-range indices, but the reference decomposition used advanced indexing directly. That gave Inductor Python indexing semantics, including negative-index wrapping, so compiled index_select could silently return data for [-1..."
186027,Fix symbolic broadcast inference for unbacked sizes,https://github.com/pytorch/pytorch/pull/186027,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,143495,143495,,,,0,12,1,af7870cc125f31998c5c865f62cc407effd64b241b26b463138a815faa4a03ae,6a608a4197c514d9e54ccd0750a3deb5678b6a07f08aa9bb77db652904bbb604,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor, agentic, needs-review",2026-06-03T01:48:08Z,2026-07-08T01:14:58Z,1,"Stack from ghstack (oldest at bottom): -> #186027 Symbolic broadcasting currently collapses ambiguous singleton/equality cases into an equality-only assumption in a few fake/ref/meta paths. For unbacked sizes produced by operators such as nonzero, this rejects valid runtime inputs where one side is actually a single..."
188636,Handle reciprocal SymPy powers in shape range eval,https://github.com/pytorch/pytorch/pull/188636,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"188422, 188423","188422, 188423",,,,0,1,2,,,"module: cpu, topic: not user facing, module: inductor, ciflow/inductor, agentic, needs-review",2026-07-01T02:49:14Z,2026-07-08T01:11:14Z,1,"Stack from ghstack (oldest at bottom): -> #188636 Inductor stride-order checks ask the symbolic shape environment to evaluate modulo divisibility guards. For the Sapiens2 dynamic SDPA case in transformers, SymPy simplifies part of one of those guards into a plain sympy.Pow with exponent -1. The SymPy interpreter pre..."
189004,Add priority&priority_range to torch.Stream,https://github.com/pytorch/pytorch/pull/189004,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,6,0,,,"open source, ciflow/trunk, release notes: python_frontend",2026-07-06T05:45:02Z,2026-07-08T01:09:05Z,1,"Stack from ghstack (oldest at bottom): -> #189004 Motivation This PR adds priority and priority_range support to torch.Stream. Some applications, such as vLLM (see https://github.com/vllm-project/vllm/pull/37160/changes#diff-c9d929754a80ae9da82b43467ecc3103586054754ce6943b94334de41118fcaaR171), use low-priority stre..."
186893,Fix dynamic size-one tensor recompiles,https://github.com/pytorch/pytorch/pull/186893,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,4,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), pallas-tpu-py3.12-inductor / test (inductor-pallas-tpu, 1, 1, linux.google.tpuv7x.1)",statusCheckRollup,,160271,160271,,,,0,8,5,1e5826747f840775fb4749f6c06089edd144f57fe60f58d8db3258d122ef8a47,689d55800b2b2fda23e9e8666fff5d46aee65a77d50dade7fa70fe7f58b86fcd,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan, agentic, needs-review",2026-06-10T04:58:03Z,2026-07-08T01:08:57Z,1,"Stack from ghstack (oldest at bottom): -> #186893 torch.compile(dynamic=True) disabled static-by-default shape assumptions, but tensor sizes with concrete value 1 still used the general 0/1 specialization path. A first call with batch size 1 therefore installed tensor guards with size=[1], so a later call with batch..."
184988,Allow HOPs to pass through unchanged input tensors,https://github.com/pytorch/pytorch/pull/184988,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,169769,169769,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-23T13:23:12Z,2026-07-08T01:06:12Z,1,"Stack from ghstack (oldest at bottom): -> #184988 Dynamo and HOP functionalization rejected every input-output storage alias for control-flow higher order ops. That was too strict for branches and loop bodies that return an exact input placeholder unchanged: there is no new view relationship or hidden mutation, but..."
184990,[dynamo] Route fake RuntimeError through user exception handlers,https://github.com/pytorch/pytorch/pull/184990,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,5,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,"157217, 169538, 183887","157217, 169538, 183887",,,,0,6,4,56d2e234188d5d73a53da5b44b39a287bfe0c8666262f9fdac2a3502e020ff32,e5568ee2ae3a76951d69e7515269432b430c76c988106cbc6cd7a2296c4319bb,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-23T15:03:13Z,2026-07-08T01:03:39Z,1,"Stack from ghstack (oldest at bottom): -> #184990 When Dynamo evaluates an FX node with fake tensors, run_node wraps operator failures in a fake-node RuntimeError so unhandled failures produce a useful TorchRuntimeError diagnostic. That same wrapping was also used when the user operation was inside a try/except bloc..."
182334,[inductor][autotuning] pipeline AsyncCompile + CachingAutotuner (#182334),https://github.com/pytorch/pytorch/pull/182334,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"open source, fb-exported, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-04T19:37:11Z,2026-07-08T01:01:49Z,1,"Summary: Adds a 3-stage pipelined autotuner that overlaps Triton compile, parent-side launcher creation, and benchmarking. The worker subprocess streams each `CompileResult` back to the parent over a per-kernel AF_UNIX socket as it finishes; a per-kernel daemon drains those into launchers in parallel with inductor c..."
184995,Fix AOT getitem replay under inference_mode,https://github.com/pytorch/pytorch/pull/184995,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,169477,169477,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-23T16:46:34Z,2026-07-08T01:01:23Z,1,Stack from ghstack (oldest at bottom): -> #184995 Dynamo captures the inference_mode decorator around a module forward as enter/exit calls in the FX graph. During AOT metadata collection we replay that graph with functionalized inputs via PropagateUnbackedSymInts. If a basic tensor slice/getitem is replayed while in...
186032,Avoid reshape metadata guards on symbolic size-one checks,https://github.com/pytorch/pytorch/pull/186032,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,5,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,143463,143463,,,,0,9,1,3a05399bf7eb6b53886e752016bf0f144635ab1b0e766f164c39801375d0027f,394398424a05e1f601d2b6886e0148a0f7fd291b1f384c5b440854aa7b8d5229,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-06-03T03:32:52Z,2026-07-08T00:58:50Z,1,"Stack from ghstack (oldest at bottom): -> #186032 Export can fail dynamic shape validation when metadata-only decisions guard on whether a symbolic extent is equal to one. In the issue repro, temporal slicing feeds GroupNormSpatial through flatten/reshape and GroupNorm decompositions. Several reshape, contiguity, an..."
188276,[FSDP] Add MORI SDMA all-gather backend and zero-copy output capability,https://github.com/pytorch/pytorch/pull/188276,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-06-26T15:56:21Z,2026-07-08T00:54:36Z,1,cc: @weifengpy Summary This follows the design discussed in pytorch/pytorch#186601. Add an opt-in MORI SDMA all-gather backend for FSDP2. Add a conservative param-contiguous zero-copy output path for eligible FSDP parameter groups. Keep default FSDP2 behavior unchanged unless users explicitly install the custom all-...
185008,Add DebugMode tensor hash export,https://github.com/pytorch/pytorch/pull/185008,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,168975,168975,,,,0,3,1,,,"topic: not user facing, ciflow/dtensor, agentic, needs-review",2026-05-23T20:08:56Z,2026-07-08T00:53:37Z,1,"Stack from ghstack (oldest at bottom): -> #185008 DebugMode.log_tensor_hashes collected tensor hashes inside the live call records and rendered them into debug_string(), but there was no structured way to export those hashes for external processing or line-oriented diffs. Add DebugMode.tensor_hashes() to flatten col..."
185009,Add fuzzy DebugMode hash mismatch comparison,https://github.com/pytorch/pytorch/pull/185009,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,168974,168974,,,,0,2,1,,,"topic: not user facing, ciflow/dtensor, agentic, needs-review",2026-05-23T20:30:56Z,2026-07-08T00:50:28Z,1,"Stack from ghstack (oldest at bottom): -> #185009 DebugMode.check_hash_mismatches currently assumes the two log streams have identical positional structure before it compares hashes. That makes the tool fail at the first inserted, deleted, or shifted call with a length/operator error, which leaves accuracy debugging..."
183620,Port 2 tests under test/distributed folder to XPU,https://github.com/pytorch/pytorch/pull/183620,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,13,0,,,"triaged, open source, release notes: distributed (c10d), ciflow/torchtitan",2026-05-13T21:22:01Z,2026-07-08T00:49:47Z,1,"For #114850, we will port distributed tests to Intel GPU. We will enable Intel GPU with following methods and keep the original code styles: Example: ""torch.accelerator.current_accelerator()"" to determine the accelerator backend enabled XPU for the following files: test/distributed/test_device_mesh.py test/distribut..."
185017,Fix OpOverloadPacket tracing during initialization,https://github.com/pytorch/pytorch/pull/185017,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,168136,168136,,,,0,3,3,,,"topic: not user facing, agentic, needs-review",2026-05-23T22:41:56Z,2026-07-08T00:45:01Z,1,"Stack from ghstack (oldest at bottom): -> #185017 Tracing tools that use CPython frame hooks can inspect an OpOverloadPacket at the entry of init. Previously, the fields used by repr(self), str(self), and getattr were first assigned inside init, so a trace hook that logged function arguments could observe the packet..."
184628,Fix SingletonInt static guard evaluation,https://github.com/pytorch/pytorch/pull/184628,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"182217, 183369","182217, 183369",,,,0,4,5,,,"release notes: fx, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-21T00:13:54Z,2026-07-08T00:41:33Z,1,Stack from ghstack (oldest at bottom): -> #184628 Evaluate determinate SingletonInt-only symbolic expressions before guard emission so NestedTensor keepdim reductions do not create unguardable ephemeral guards. Leave mixed or unsupported SingletonInt expressions unknown. Fixes #182217 Fixes #183369 Generated by my a...
184481,Guard synthetic-base input alias layouts,https://github.com/pytorch/pytorch/pull/184481,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,93617,"93617, 188133","184694, 185891","184694, 185891",,16,5,9,f992c13663fbaba64c2bfbf603d6c7ae512992fd1ac8186fa4b57adca1e2a9dc,4eb6ebe1deb04d401f89acdf886e1c418da484980cf47de9c17eabe124c579d9,"module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-20T04:47:57Z,2026-07-08T00:38:07Z,1,Stack from ghstack (oldest at bottom): -> #184481 Add AOTAutograd guards and cache-key data for synthetic-base storage groups so compiled InputBuffer view reconstruction is not reused after input alias topology or storage offsets change. Fixes #93617 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgon...
189027,[dynamo] Guard os.environ reads with an ambient env-var guard,https://github.com/pytorch/pytorch/pull/189027,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-07-06T14:22:26Z,2026-07-08T00:38:03Z,1,"Problem os.getenv results are baked into compiled graphs as constants with no guard. This compiles under fullgraph=True and then never recompiles when the variable changes, silently returning stale results: os.environ[""FLAG""] = ""1"" @torch.compile(fullgraph=True) def f(x): if int(os.getenv(""FLAG"", ""0"")): return x + 1..."
189122,Fix sparse-dense mul dropping data when broadcasting a size-1 sparse dim,https://github.com/pytorch/pytorch/pull/189122,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,188900,188900,188900,,,,0,1,3,,,"open source, release notes: sparse",2026-07-07T04:23:42Z,2026-07-08T00:36:19Z,1,"Fixes #188900 Summary Multiplying a dense tensor by a sparse COO tensor whose size-1 sparse dimension must broadcast up to a larger size silently dropped data: only the index-0 slice along the broadcast dimension was populated, and the rest were left as zeros. No error or warning was raised. Root cause In intersecti..."
184079,Enable graph deduplication for Inductor compile,https://github.com/pytorch/pytorch/pull/184079,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,156103,,156103,,,0,10,4,,,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-05-17T07:25:43Z,2026-07-08T00:35:13Z,1,"Stack from ghstack (oldest at bottom): -> #184079 Use Dynamo graph deduplication for Inductor torch.compile by default so repeated full-model regions can be emitted as invoke_subgraph calls and compiled once. Keep an opt-out and skip this path for cpp_wrapper, which cannot lower invoke_subgraph yet. Fixes #156103 Ge..."
185025,[export][serde] Initialize symbol ranges before constants,https://github.com/pytorch/pytorch/pull/185025,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,167872,167872,,,,0,1,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-23T23:25:13Z,2026-07-08T00:32:29Z,1,"Stack from ghstack (oldest at bottom): -> #185025 GraphModuleDeserializer installs itself as the current deserializer before loading serialized artifacts. Constants can contain FakeTensors, and unpickling those constants reconstructs tensor metadata immediately. If that metadata has symbolic sizes, reconstruction ca..."
185033,Support Dynamo Tensor.backward by default,https://github.com/pytorch/pytorch/pull/185033,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,167732,167732,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-24T00:49:33Z,2026-07-08T00:29:13Z,1,"Stack from ghstack (oldest at bottom): -> #185033 Tensor.backward already lowers through Dynamo's autograd.grad handler plus accumulate_grad_, but method_backward rejected the call whenever trace_autograd_ops was left at its default False. That meant normal torch.compile users still hit an Unsupported graph break ev..."
185034,Enable Dynamo capture of torch.autograd.grad,https://github.com/pytorch/pytorch/pull/185034,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,167729,"167729, 174929",,,,0,1,2,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-24T01:04:50Z,2026-07-08T00:25:51Z,1,"Stack from ghstack (oldest at bottom): -> #185034 Dynamo already had a torch.autograd.grad tracing handler and graph-boundary validation, but the handler was blocked by the trace_autograd_ops config defaulting to False. As a result, the issue repro failed before Dynamo could put autograd.grad in the FX graph for AOT..."
185035,Make symbolic formatting guard data-dependent scalars,https://github.com/pytorch/pytorch/pull/185035,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,167719,167719,,,,0,3,5,fe421faa8e60a937b5efa992594cce59d4830c4f88fbf7eeba3d58ade87ca68a,c071a84f65b57f1d285bb6359d9d5774e78682081eeddd3bab6046a3bf1b20fa,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-24T01:33:36Z,2026-07-08T00:23:05Z,1,"Stack from ghstack (oldest at bottom): -> #185035 Formatting an unbacked symbolic scalar currently falls back to the symbolic expression text. In non-strict export this allowed idx.item() + 1 to become ordinary Python string data inside an f-string, so a ModuleDict lookup tried the key domainu0 + 1 and leaked a raw..."
185042,Avoid value guards for residual lazy constants,https://github.com/pytorch/pytorch/pull/185042,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,167596,167596,,,,0,4,2,36586fbef79199da63d28018f41eb667af1e5f6a83b2c10d63ea9820ac2ad1aa,c15f89adbbb24af07602e3e7cb6db91810148097617a4f86011a507c1e5575e1,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-24T03:26:03Z,2026-07-08T00:20:19Z,1,Stack from ghstack (oldest at bottom): -> #185042 Dynamo's LazyConstantVariable is meant to defer guarding on primitive constants until their value is used for graph-specializing behavior. Residual bytecode codegen was defeating that by calling is_python_constant()/as_python_constant() and emitting LOAD_CONST for so...
186953,Fix CPU amin/amax backward rematerialization mismatch,https://github.com/pytorch/pytorch/pull/186953,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,186799,186799,,,,0,4,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-10T17:20:45Z,2026-07-08T00:16:47Z,1,Stack from ghstack (oldest at bottom): -> #186953 CPU Inductor can compute value-selecting reductions such as amin and amax with a scalar math path in the forward reduction while AOTAutograd rematerializes the reduced input in backward through a different vectorized math path. The backward formula routes gradients w...
184590,Fix uncached NestedTensor seqlens under compile,https://github.com/pytorch/pytorch/pull/184590,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,183266,183266,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-20T19:00:50Z,2026-07-08T00:13:03Z,1,Stack from ghstack (oldest at bottom): -> #184590 Trace uncached NestedTensor private seqlen access as a graph computation and rewrap the NestedTensor with the same zero-width metadata cache tensor eager would create. This avoids specializing on data-dependent .item() results while preserving the returned NestedTens...
185043,Functionalize mutations in deferred local_map,https://github.com/pytorch/pytorch/pull/185043,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,167568,167568,,,,0,3,1,,,"topic: not user facing, agentic, needs-review",2026-05-24T03:49:31Z,2026-07-08T00:08:21Z,1,"Stack from ghstack (oldest at bottom): -> #185043 Deferred local_map builds a joint forward/backward graph from the already captured local body while functionalization is suspended. For Triton launches captured as triton_kernel_wrapper_mutation, this means the existing Triton functionalization rule never gets a chan..."
185050,Handle SymInt sizes in lazy module initialization,https://github.com/pytorch/pytorch/pull/185050,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,"167191, 173252","167191, 173252",,,,0,3,2,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-05-24T05:49:47Z,2026-07-08T00:03:22Z,1,"Stack from ghstack (oldest at bottom): -> #185050 Lazy modules infer their materialized parameter sizes from input.shape during their first forward. Under torch.compile(dynamic=True), those inferred entries can be torch.SymInt values. LazyLinear, lazy conv, and lazy norm then stored those SymInts and passed them to..."
185051,Fix FX GraphModule serialization of string annotations,https://github.com/pytorch/pytorch/pull/185051,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,167117,167117,,,,0,1,1,,,"release notes: fx, agentic, needs-review",2026-05-24T06:03:45Z,2026-07-07T23:59:51Z,1,"Stack from ghstack (oldest at bottom): -> #185051 FX codegen represents forward-reference annotations as generated globals whose values are the original strings. GraphModule serialization formatted every generated global as an importable Python object, so a string annotation reached Importer.get_name and crashed bec..."
185052,Fix Inductor reuse for disjoint strided slices,https://github.com/pytorch/pytorch/pull/185052,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,167098,167098,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-24T06:34:21Z,2026-07-07T23:56:48Z,1,"Stack from ghstack (oldest at bottom): -> #185052 Inductor decides whether to realize a pointwise producer based partly on how many FX users consume it. The bf16-to-float4 helper produces unpacked uint8 fp4 values and then packs them with two strided slices, uint8_data[1::2] and uint8_data[::2]. Those two users look..."
184920,[dynamo] Store FunctionIdSet entries by identity weakrefs,https://github.com/pytorch/pytorch/pull/184920,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,9,1,"statusCheckRollup, linux-jammy-py3.10-clang18 / test-osdc (dynamo_wrapped, 3, 3, mt-l-x86iavx512-8-64), linux-jammy-py3.14t-clang18 / test-osdc (dynamo_wrapped, 3, 3, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test-osdc (dynamo_wrapped, 2, 3, mt-l-x86iavx512-8-64), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,173578,173578,,,,2,3,1,f1b41992b5b8379326cdbefb6fe4b89602eef554a2d2db49ea1fac96e797b7de,6102fbab9b27bdb529508a400c911c72c6f9aeb2ff21b577071c279da5125c8a,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-22T19:34:08Z,2026-07-07T23:53:56Z,1,"Stack from ghstack (oldest at bottom): -> #184920 FunctionIdSet previously stored only id(obj) values. When a registered callable was collected, CPython could reuse that id for an unrelated object, causing the new object to inherit allow/disallow/nonstrict/leaf behavior. Several decorators worked around part of this..."
185775,Avoid storage access for no-storage tensors in MetaConverter,https://github.com/pytorch/pytorch/pull/185775,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,148695,148695,,,,0,13,2,455f1c863e1738d3cea844b67b3b25c3fb98021c7ad94aac4f268f1ef5a202ea,f8626e850232a1e23bc83a698d39e884eb01f0aaf31a5efeab435a065d254226,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-05-31T23:34:56Z,2026-07-07T23:50:14Z,1,"Stack from ghstack (oldest at bottom): -> #185775 OpaqueTensorImpl tensors can expose dense strided metadata while deliberately not owning a Storage. MetaTensorDescriber only skipped storage access for a fixed list of layouts and wrappers, so fakifying a PrivateUse1 OpaqueTensorImpl called untyped_storage() and rais..."
186879,Fix export dynamic shapes for variadic kwargs,https://github.com/pytorch/pytorch/pull/186879,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,164313,164313,,,,0,3,1,,,"release notes: export, agentic, needs-review",2026-06-10T02:55:06Z,2026-07-07T23:46:33Z,1,"Stack from ghstack (oldest at bottom): -> #186879 Dynamic shape validation binds inputs through the module signature. For modules with a **kwargs parameter, that means user keyword arguments are nested under the variadic parameter name, for example {""model_kwargs"": {""return_dict"": None}}. Non-strict range constraint..."
184830,[Dynamo] Guard source-backed identity comparisons,https://github.com/pytorch/pytorch/pull/184830,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,CLEAN,False,0,0,,,,175267,175267,,,,0,20,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic",2026-05-22T01:17:32Z,2026-07-07T23:45:32Z,1,"Stack from ghstack (oldest at bottom): -> #184830 Dynamo's identity helper only trusted VariableTracker real Python values. Source-backed values such as aliases, inherited descriptors, and class-dict descriptor entries can represent concrete Python objects while reporting NO_SUCH_SUBOBJ, so is comparisons could be f..."
186886,Fix keyed user Triton autotune in Inductor,https://github.com/pytorch/pytorch/pull/186886,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,164124,164124,,,,0,2,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-10T03:53:02Z,2026-07-07T23:43:59Z,1,"Stack from ghstack (oldest at bottom): -> #186886 Inductor preserved user Triton autotune configs but dropped Triton's key list when lowering user-defined Triton kernels. Runtime CachingAutotuner then selected one global best launcher for the generated kernel, so different runtime key values could reuse a config tun..."
185100,Keep fake tensor cache synthesis out of proxy traces,https://github.com/pytorch/pytorch/pull/185100,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,163713,163713,,,,0,1,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-25T03:09:28Z,2026-07-07T23:35:34Z,1,"Stack from ghstack (oldest at bottom): -> #185100 FakeTensor dispatch cache rebuilds cached FakeTensor outputs from stored tensor metadata. That reconstruction can call empty_strided() and, for views, set_(), both of which run symbolic shape/storage checks. When this happened while a proxy tracer was active, those c..."
185103,Fix unbacked scalar equality reasoning through div/mod,https://github.com/pytorch/pytorch/pull/185103,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,163641,163641,,,,0,1,1,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-05-25T03:57:20Z,2026-07-07T23:31:06Z,1,"Stack from ghstack (oldest at bottom): -> #185103 When torch._check records equality between data-dependent unbacked symbols from Tensor.item(), ShapeEnv intentionally does not install a global replacement. A global replacement can move a runtime data-dependent binding before the point where it is actually produced...."
185105,Fix export move_to_device_pass device rewrites,https://github.com/pytorch/pytorch/pull/185105,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,163323,163323,,,,0,1,1,,,"release notes: export, agentic, needs-review",2026-05-25T04:39:21Z,2026-07-07T23:28:41Z,1,"Stack from ghstack (oldest at bottom): -> #185105 move_to_device_pass rewrites device-bearing FX nodes and tensor metadata, but it did not recompile the affected GraphModules. The generated Python could therefore keep stale _assert_tensor_metadata device kwargs after the pass, so running the moved exported graph dir..."
185106,Fix non-strict export state aliases,https://github.com/pytorch/pytorch/pull/185106,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,163303,163303,,,,0,1,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-25T05:05:05Z,2026-07-07T23:24:33Z,1,"Stack from ghstack (oldest at bottom): -> #185106 Non-strict export reparametrizes registered parameters and buffers before tracing, but plain Python containers can still hold aliases to the original registered tensors. When forward reads such an alias, make_fx emits a get_attr tensor constant and lift_constants_pas..."
185116,Thread effect tokens through cond branches,https://github.com/pytorch/pytorch/pull/185116,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"163143, 165981","163143, 165981",,,,0,1,2,,,"ciflow/inductor, release notes: export, ciflow/torchtitan, agentic, needs-review",2026-05-25T07:40:15Z,2026-07-07T23:21:55Z,1,"Stack from ghstack (oldest at bottom): -> #185116 The failure in #163143 came from functionalization discovering an effectful op inside a torch.cond branch after the cond operands had already been fixed. That left the branch graph with an extra effect token input/output that the outer cond call did not thread, so AO..."
179550,Enable eager tests for inductor-unittest - Enable TestViewOps and TestDevicePrecision on Intel GPU,https://github.com/pytorch/pytorch/pull/179550,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,188987,"176688, 176689, 178849, 179549",,,0,15,5,,,"open source, topic: not user facing, ciflow/mps, keep-going, ciflow/xpu, ci-no-td",2026-04-07T07:05:18Z,2026-07-07T23:20:53Z,1,"In order to let Intel GPU run the inductor-unittest workflow we need to need to enable Intel GPU on the following files: test_ops.py, test_torch.py and test_ops_gradiant.py. This PR is used to port TestViewOps and TestDevicePrecision in test_torch.py, they are enabled with instantiate_device_type_tests and onlyAccel..."
184944,Make OpaqueBase usable as a pybind base,https://github.com/pytorch/pytorch/pull/184944,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,172869,172869,,,,0,17,3,,,"release notes: distributed (c10d), ciflow/inductor, ciflow/dtensor, agentic",2026-05-22T23:33:57Z,2026-07-07T23:19:14Z,1,"Stack from ghstack (oldest at bottom): -> #184944 Pybind assumes that explicit bases passed to py::class_ are pybind-registered classes. The old OpaqueBase was a pure Python marker class, so trying to pass it as the base for a pybinded opaque class left pybind without parent type_info and hit an assertion in pybind..."
184007,[inductor] Clear stale subprocess cache env vars,https://github.com/pytorch/pytorch/pull/184007,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,167725,167725,,,,0,12,1,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, agentic",2026-05-16T08:22:34Z,2026-07-07T23:19:03Z,1,"Stack from ghstack (oldest at bottom): -> #184007 Persistent Inductor worker processes now receive explicit cache environment resets, including absent parent values, so fresh_cache() cannot leave stale cache directories in child processes. This also keeps Triton's libdevice path knob synchronized with the worker env..."
185121,Raise TypeError for invalid torch.export.load inputs,https://github.com/pytorch/pytorch/pull/185121,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,163040,163040,,,,0,1,1,,,"release notes: export, agentic, needs-review",2026-05-25T08:26:36Z,2026-07-07T23:18:53Z,1,"Stack from ghstack (oldest at bottom): -> #185121 torch.export.load delegates to load_pt2, which only warned when f was neither a .pt2 path nor a readable, seekable buffer. It still passed the invalid object to PT2ArchiveReader, so calling torch.export.load(ep) with an ExportedProgram crashed later with AttributeErr..."
188004,rewrite exception handling to be more closer to CPython,https://github.com/pytorch/pytorch/pull/188004,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,,,"157149, 187690, 187744, 188638, 188639, 188824, 188825, 188834, 189024",,,0,17,19,,,"open source, Merged, Reverted, module: dynamo, ciflow/inductor, release notes: dynamo, ci-no-td, ciflow/torchtitan",2026-06-24T00:20:34Z,2026-07-08T21:06:03Z,0,Stack from ghstack (oldest at bottom): #189024 #188825 #188834 #188824 #157149 #188639 #188638 -> #188004 #187744 #187690
189184,[inductor][fb] Whole-graph batch_addmm->bmm fusion for independent same-shape addmm (#189184),https://github.com/pytorch/pytorch/pull/189184,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-07-07T19:05:36Z,2026-07-07T23:17:51Z,1,"Summary: What Adds a config-gated Inductor post-grad pass (caffe2/torch/_inductor/fx_passes/fb/batch_addmm_fusion.py, fbcode-only) that batches independent same-(M, K, N) aten.addmm ops into a single bmm (+ per-node bias add + select) across the whole graph. It targets the per-feature embedding-projection pattern in..."
188374,[HOP] Deduplicate lifted args for switch,https://github.com/pytorch/pytorch/pull/188374,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,12,0,,,"triaged, open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-28T18:18:04Z,2026-07-07T23:15:33Z,1,"This PR deduplicates lifted arguments across the branches of switch. In the example below, a and b are lifted by the first branch, but also by the second branch. Similar c is first lifted by the second and afterwards by the third: def branch0(inp_x): return inp_x + a + b def branch1(inp_x): return inp_x + a + b + c..."
185124,Fix prims.broadcast_in_dim functionalization,https://github.com/pytorch/pytorch/pull/185124,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,"157610, 162734, 163037","157610, 162734, 163037",,,,0,2,11,,,"module: inductor, ciflow/inductor, release notes: inductor, agentic, needs-review",2026-05-25T08:58:44Z,2026-07-07T23:15:13Z,1,"Stack from ghstack (oldest at bottom): -> #185124 prims::broadcast_in_dim is a view prim whose schema aliases the input, but it did not have a Functionalize kernel. When AOTAutograd functionalized a graph that called the prim directly, the generic non-ATen fallback rejected the alias annotation before later clones o..."
179549,Enable eager tests for inductor-unittest - define xfails for xpu in op_db for test_ops.py using device_type list in DecorateInfo,https://github.com/pytorch/pytorch/pull/179549,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"176688, 176689, 178849, 179550",,,0,11,4,,,"open source, release notes: python_frontend, ciflow/xpu",2026-04-07T06:30:42Z,2026-07-07T23:05:37Z,1,define xfails for xpu in op_db for test_ops.py using device_type list in DecoratorInfo as xpu has same xfail behavior as cuda. Stack from ghstack (oldest at bottom): #179550 #178849 -> #179549 #176689 #176688 #178565
182224,Enable skipped static_runtime tests on AArch64,https://github.com/pytorch/pytorch/pull/182224,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,178522,178522,,,,0,1,1,,,"open source, Stale, ciflow/trunk, topic: not user facing",2026-05-02T13:31:10Z,2026-07-07T22:49:48Z,1,Stack from ghstack (oldest at bottom): -> #182224 Fixes #178522
176830,feat: add polyfills for setitem/delitem op,https://github.com/pytorch/pytorch/pull/176830,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,116396,,,,0,1,2,,,"triaged, open source, Stale, module: dynamo, ciflow/inductor, release notes: dynamo, bot-triaged, ciflow/torchtitan",2026-03-08T15:45:18Z,2026-07-07T22:49:45Z,1,"#116396 Add support for operator.setitem and operator.delitem by handling them in BuiltinVariable(same as the existing getitem). Add tests covering operator.setitem on list, dict, and tensor, and operator.delitem on list and dict. Test plan .venv/bin/python test/dynamo/test_functions.py FunctionTests.test_operator_s..."
189191,Cuda Graph Default Enable (#189191),https://github.com/pytorch/pytorch/pull/189191,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,189191,,,,,0,4,0,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-07-07T20:17:22Z,2026-07-07T22:49:03Z,1,Summary: Pull Request resolved: #189191 Differential Revision: D110935008
185126,Optimize GB_REGISTRY linter scanning,https://github.com/pytorch/pytorch/pull/185126,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,162973,162973,,,,0,1,1,,,"topic: not user facing, agentic, needs-review",2026-05-25T09:22:46Z,2026-07-07T22:44:44Z,1,Stack from ghstack (oldest at bottom): -> #185126 The GB_REGISTRY adapter was paying for two full scans of torch/_dynamo: one AST parse/walk to find direct Unsupported raises and another parse/walk to collect unimplemented(...) call sites. Its ast.get_source_segment() optimization also cached ast._splitlines_no_ff()...
176689,Enable eager tests for inductor-unittest - enable test_ops_fwd_gradients.py and test_ops_gradients.py on Intel GPU,https://github.com/pytorch/pytorch/pull/176689,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"176688, 178849, 179549, 179550",,,0,27,4,,,"open source, topic: not user facing, ciflow/xpu",2026-03-06T07:22:58Z,2026-07-07T22:44:17Z,1,Enable xpu with instantiate_device_type_tests() allow_xpu=True augument skip cases in op db if xpu has limitations. Stack from ghstack (oldest at bottom): #179550 #178849 #179549 -> #176689 #176688 #178565
176688,Enable eager tests for inductor-unittest - enable test_dtypes and align XPU op dyptes with OpInfo,https://github.com/pytorch/pytorch/pull/176688,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"176689, 178849, 179549, 179550",,,0,20,4,,,"triaged, open source, topic: not user facing, ciflow/xpu, ci-no-td",2026-03-06T07:22:52Z,2026-07-07T22:41:50Z,1,Enable test_dtypes() of TestCommon for xpu Update the op_db to align XPU op dyptes with OpInfo Stack from ghstack (oldest at bottom): #179550 #178849 #179549 #176689 -> #176688 #178565 disable torch.half and torch.complex32 on XPU for some fft ops before they are implemented update dot and vdot dtypes according to l...
185132,Fix custom op autograd defaults with tensorlist outputs,https://github.com/pytorch/pytorch/pull/185132,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,162687,162687,,,,0,1,3,,,"release notes: autograd, agentic, needs-review",2026-05-25T12:14:04Z,2026-07-07T22:41:33Z,1,Stack from ghstack (oldest at bottom): -> #185132 Custom op autograd wraps generated autograd.Functions with supports_tensorlist when an operator has TensorList-like inputs or outputs. That wrapper records the pytree input spec from the arguments that the dispatcher actually passed to autograd. The dispatcher may el...
185138,[dynamo] Enforce one compiled frame for fullgraph skipped frames,https://github.com/pytorch/pytorch/pull/185138,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,162415,162415,,,,0,5,1,431c955bb604899dd3e01c5e85676cba7a2574167511f84545255714729e7c1b,b345f19c52e3b25fed1ceef38683cecc3e4a9d23f074f1fa95ba7de54d12f077,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-25T14:01:15Z,2026-07-07T22:37:36Z,1,"Stack from ghstack (oldest at bottom): -> #185138 torch.compile(..., fullgraph=True) promises a single compiled region or an error. When the compiled function's code object was explicitly skipped with skip_code, Dynamo ran that outer frame eagerly while leaving recursive frame tracing enabled. A call like f(x) = g(x..."
185140,Preserve reentrant checkpoint in non-strict export,https://github.com/pytorch/pytorch/pull/185140,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,162386,162386,,,,0,1,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-25T14:49:17Z,2026-07-07T22:34:40Z,1,"Stack from ghstack (oldest at bottom): -> #185140 Non-strict export traces reentrant torch.utils.checkpoint through CheckpointFunction.forward. That implementation wraps the checkpointed body in no_grad, and the export set-grad pass previously lowered the region as wrap_with_set_grad_enabled(False, body). When aot_e..."
185147,Fix Z3 translation validation for torch.sym_not,https://github.com/pytorch/pytorch/pull/185147,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,162287,162287,,,,0,1,1,,,"release notes: fx, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-25T16:13:15Z,2026-07-07T22:32:09Z,1,"Stack from ghstack (oldest at bottom): -> #185147 torch.fx.experimental.validator.z3op translated operator.not_ to z3.Not, but did not translate torch.sym_not. Shape-env graphs can record symbolic boolean negation with torch.sym_not, so the validator fell back to invoking the public helper on a z3.BoolRef. The helpe..."
185149,Reject partial-overlap copy aliases,https://github.com/pytorch/pytorch/pull/185149,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,162151,"162151, 174929",,,,0,1,2,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-05-25T17:17:07Z,2026-07-07T22:28:51Z,1,Stack from ghstack (oldest at bottom): -> #185149 Fixes #162151 The issue repro assigns x[0] = x[0].t() after a permute. That lowers to a copy where the source and destination are aliases with the same first element but different strides. Eager copy_ previously let this through because the memory-overlap classifier...
163503,[dynamo] prevent recompilation limit exceeded on external_utils.wrap_inline,https://github.com/pytorch/pytorch/pull/163503,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"171826, 185524, 186100, 188173",,,0,114,5,,,"topic: not user facing, module: dynamo, ciflow/inductor, keep-going, ciflow/torchtitan",2025-09-22T10:44:28Z,2026-07-07T22:27:47Z,1,"Stack from ghstack (oldest at bottom): #171826 -> #163503 #185524 #188173 #186100 #186657 This will prevent recompilations (i.e. logs to TORCH_LOGS=""recompiles"") due to wrap_inline's guard on fn. If we call wrap_inline on enough different functions, we would hit the recompile limit. cc @voznesenskym @penguinwu @Eika..."
186100,"[dynamo, nested graph breaks] enable nested_graph_breaks by default on OSS",https://github.com/pytorch/pytorch/pull/186100,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"163503, 171826, 185524, 188173",,,0,39,5,,,"topic: not user facing, module: dynamo, ciflow/inductor, keep-going, ciflow/dynamo, ciflow/torchtitan",2026-06-03T16:40:56Z,2026-07-07T22:27:42Z,1,Stack from ghstack (oldest at bottom): #171826 #163503 #185524 #188173 -> #186100 #186657 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
185524,"[dynamo, nested graph breaks] add strong nested graph break test infrastructure",https://github.com/pytorch/pytorch/pull/185524,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"163503, 171826, 186100, 188173",,,0,42,5,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-28T14:46:30Z,2026-07-07T22:27:31Z,1,Stack from ghstack (oldest at bottom): #171826 #163503 -> #185524 #188173 #186100 #186657 Add NestedGraphBreaksStrong test variant that forces graph breaks at every leaf function return via debug_force_graph_break_on_leaf_return. This includes a _debug_force_graph_break_on_leaf_return_disable_codes WeakSet to preven...
185155,Avoid warning spam for backend fallback graph breaks,https://github.com/pytorch/pytorch/pull/185155,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,161796,161796,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-25T18:52:48Z,2026-07-07T22:25:38Z,1,"Stack from ghstack (oldest at bottom): -> #185155 Dynamo lets a small set of backend compiler exceptions fall back to eager by turning them into graph breaks. That path still used unimplemented_with_warning, so allowed fallback exceptions were emitted as user-facing WARNING records. For scalar extraction paths such..."
185159,Inductor: gate max-autotune 1x1 conv GEMM lowering,https://github.com/pytorch/pytorch/pull/185159,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,8,1,"statusCheckRollup, linux-noble-rocm-py3.12-mi300 / test (default, 1, 6, linux.rocm.gpu.gfx942.1), linux-noble-rocm-py3.12-mi300 / test (default, 4, 6, linux.rocm.gpu.gfx942.1), linux-jammy-cuda13.0-py3.10-gcc11-sm100 / test (smoke_b200, 1, 1, linux.dgx.b200), linux-jammy-cuda13.0-py3.10-gcc11-sm90 / test-osdc (smoke, 1, 1, mt-l-x86iamx-22-225-h100), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), get-label-type / runner-determinator",statusCheckRollup,,161764,161764,,,,0,7,3,d432a6dd36d63d37d92d74ad72cbe4ce93f2df369f8fb711b5ef22f252be43b0,1c1855ad9176af4720e52947b1a4c7c68a815e2577a1dceb387c7372a1ba9301,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, agentic, needs-review",2026-05-25T20:01:50Z,2026-07-07T22:21:35Z,1,"Stack from ghstack (oldest at bottom): -> #185159 The max-autotune 1x1 convolution lowering converted channels-last convolutions to GEMM whenever max_autotune or max_autotune_gemm was enabled. On GPUs that fail the GEMM template availability gate, this could still replace a cudnn convolution with ATen mm/addmm. Dens..."
181027,Factor nanmean dtype checks into a helper + add kwarg coverage,https://github.com/pytorch/pytorch/pull/181027,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,131043,131043,,"131043, 172809",,,0,1,2,,,"triaged, open source, Stale, release notes: python_frontend, no-stale",2026-04-21T21:20:16Z,2026-07-07T22:17:01Z,1,"Summary This PR was originally a duplicate of #172809, which was merged to main on 2026-05-03 and closed #131043. Rebased and refocused as a small follow-up: Refactor. #172809 added opt_dtype validation to both nanmean and nanmean_out, leaving identical 14-line check blocks in each. Pull them into a single check_nan..."
185166,Fix export slice bounds across dynamic dim,https://github.com/pytorch/pytorch/pull/185166,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,161705,161705,,,,0,7,1,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: export, agentic, needs-review",2026-05-25T20:48:26Z,2026-07-07T22:16:53Z,1,"Stack from ghstack (oldest at bottom): -> #185166 Export's fake/decomposition slice normalization specialized static slice bounds against backed symbolic dimension sizes. For cases like x[:, :5], the sample-width branch created range guards even though eager slice clamps at runtime. Using sym_min/sym_max exposes the..."
181062,Validate grad_output shape in multi_margin_loss_backward,https://github.com/pytorch/pytorch/pull/181062,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,146790,146790,146790,,,,0,1,3,,,"triaged, open source, Stale, release notes: nn, no-stale",2026-04-22T02:33:57Z,2026-07-07T22:15:59Z,1,"Fixes #146790. Summary torch.ops.aten.multi_margin_loss_backward segfaults when reduction='none' and grad_output doesn't have shape [nframe]. Both CPU and CUDA backward kernels index grad_output per sample without validating its shape, so an empty or undersized tensor reads past storage. Mirrors the check_dim_size(g..."
185170,Fix ExportedProgram pickle after graph_module edits,https://github.com/pytorch/pytorch/pull/185170,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,161671,161671,,,,0,1,1,,,"release notes: export, agentic, needs-review",2026-05-25T21:55:14Z,2026-07-07T22:13:22Z,1,"Stack from ghstack (oldest at bottom): -> #185170 Modified ExportedPrograms can contain call_module nodes when users edit the underlying FX graph. The schematized torch.export.save path still rejects that graph form by design, but the Python torch.save path should be able to round-trip the object when weights_only=F..."
189033,Auto-triage open source PRs when a maintainer engages,https://github.com/pytorch/pytorch/pull/189033,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,topic: not user facing,2026-07-06T15:33:43Z,2026-07-07T22:10:04Z,1,"Author Notes Excluding jason for now until his bot is migrated. Agent Notes Updated: restructured into the two-workflow (capture + privileged) design per review, so the org-scoped PAT never lives in a PR-triggerable job. Adds automation that applies the triaged label to a PR carrying the open source label as soon as..."
181036,Clamp spdiags diagonal length to zero for out-of-range offsets,https://github.com/pytorch/pytorch/pull/181036,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,178089,178089,178089,,,,0,2,4,,,"open source, Stale, release notes: sparse, no-stale",2026-04-21T22:30:33Z,2026-07-07T22:08:00Z,1,"Fixes #178089. Summary torch.sparse.spdiags computes per-diagonal nnz as nnz_per_diag = at::where( offsets_1d.le(0), offsets_1d.add(shape[0]).clamp_max_(diagonals_2d.size(1)), offsets_1d.add(-std::min<int64_t>(shape[1], diagonals_2d.size(1))).neg()); For target shape (R, C) and diagonal offset k, this yields a negat..."
185171,Document activation memory budget config,https://github.com/pytorch/pytorch/pull/185171,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,161650,161650,,,,0,3,3,,,"topic: not user facing, agentic, needs-review",2026-05-25T22:12:35Z,2026-07-07T22:07:35Z,1,"Stack from ghstack (oldest at bottom): -> #185171 Older public material pointed users at torch._dynamo.config.activation_memory_budget, but the knob is defined in torch._functorch.config and consumed by AOTAutograd's partitioner after Dynamo captures the graph. Without in-tree documentation, users could not tell whe..."
188039,autograd: fix PyObject* refcount leak in setup_context path,https://github.com/pytorch/pytorch/pull/188039,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188034,188034,188034,,,,0,1,2,,,"triaged, open source",2026-06-24T07:49:27Z,2026-07-07T22:06:39Z,1,"Summary python_function.cpp: In _call_function_forward, the return value of PyObject_CallObject for the setup_context call was stored as auto result (raw PyObject*), so Py_DECREF was never called on the success path, leaking one reference per invocation. Changed declaration to THPObjectPtr result(...) for RAII clean..."
188173,"[dynamo, nested graph breaks] propagate NGB suppression through intermediate inlined functions",https://github.com/pytorch/pytorch/pull/188173,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"163503, 171826, 185524, 186100",,,0,15,5,,,"ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/dynamo, ciflow/torchtitan",2026-06-25T18:45:46Z,2026-07-07T22:00:17Z,1,"Stack from ghstack (oldest at bottom): #171826 #163503 #185524 -> #188173 #186100 #186657 When a graph break originates from a function on NGB_SUPPRESS_INLINELIST (e.g. torch.distributed), the prior fix only suppressed NGB inside the suppressed function itself. Intermediate caller functions between the model's forwa..."
189095,[inductor] Preserve stream annotations across pattern replacements,https://github.com/pytorch/pytorch/pull/189095,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-07-06T21:34:16Z,2026-07-07T21:59:07Z,1,"Stack from ghstack (oldest at bottom): (to be filled) The pattern matcher's percolate_tags propagated only ""recompute"" and ""ac_graph_id"" meta from matched nodes to their replacements, dropping the ""custom"" meta that carries user stream assignments. When a post-grad pattern (e.g. online softmax) replaced a node compu..."
185191,Support wrapper subclass storages in HOP alias checks,https://github.com/pytorch/pytorch/pull/185191,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,161456,161456,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-26T05:51:37Z,2026-07-07T21:53:38Z,1,"Stack from ghstack (oldest at bottom): -> #185191 HOP subgraph aliasing checks assumed each FX node contributed one fake tensor. For traceable wrapper subclasses, _collect_fake_inputs returns the flattened inner fake tensors for a single node, so callers that indexed [0] ignored later inner tensors. That let aliasin..."
185274,Fix ci expected accuracy updates without ClickHouse credentials,https://github.com/pytorch/pytorch/pull/185274,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,161289,161289,,,,0,4,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-26T20:01:56Z,2026-07-07T21:49:29Z,1,Stack from ghstack (oldest at bottom): -> #185274 update_expected.py used ClickHouse as the only source for mapping a commit SHA to the inductor workflow and job metadata needed to construct public S3 artifact URLs. That made the documented local update command fail immediately for contributors without CH_KEY_ID/CH_...
184047,Enable cooperative reductions by default,https://github.com/pytorch/pytorch/pull/184047,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-perf-compare, ciflow/inductor-periodic, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,161138,161138,,,,0,1,5,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/inductor-perf-compare, ciflow/inductor-periodic, ciflow/torchtitan, agentic, needs-review",2026-05-16T20:42:29Z,2026-07-07T21:46:44Z,1,Stack from ghstack (oldest at bottom): -> #184047 Use the existing cooperative reduction heuristic by default so large scalar reductions can avoid split-reduction follow-up kernels when cross-block synchronization is profitable. Fixes #161138 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guob...
148674,Set /NODEFAULTLIB:vcomp for MSVC when linking caffe2::mkl with libiomp5md.lib,https://github.com/pytorch/pytorch/pull/148674,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,,113490,113490,,,,0,1,1,,,"module: build, module: windows, triaged, open source, release notes: build, topic: bug fixes, no-stale",2025-03-06T15:08:15Z,2026-07-07T21:45:46Z,1,"Fixes: #113490 When using the Microsoft Visual C++ Compiler with Intel® OpenMP, it's needed to avoid linking the Microsoft OpenMP runtime library (vcomp) and explicitly pass the name of the Intel® OpenMP compatibility library as linker options. More details: https://www.intel.com/content/www/us/en/docs/cpp-compiler/..."
187131,[muon] add multi-tensor (foreach) standard Muon path (#187131),https://github.com/pytorch/pytorch/pull/187131,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, release notes: optim, meta-exported",2026-06-12T02:12:25Z,2026-07-07T21:29:05Z,1,"Summary: Adds an opt-in foreach implementation for torch.optim.Muon. This change introduces _multi_tensor_muon and wires the foreach argument through Muon, the functional API, and state-dict BC. The foreach path uses foreach ops for momentum, weight decay, and grouped parameter updates while preserving the existing..."
185293,Require bool conditions for torch.where,https://github.com/pytorch/pytorch/pull/185293,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/mps,review diff and decide acceptance path,,CLEAN,False,0,0,,,,161088,161088,,,,0,3,1,,,"module: bc-breaking, release notes: python_frontend, topic: bc breaking, release notes: mps, ciflow/mps, agentic, needs-review",2026-05-26T22:46:16Z,2026-07-07T21:18:40Z,1,"Stack from ghstack (oldest at bottom): -> #185293 where_self_out still accepted Byte conditions by warning once and casting them to bool before TensorIterator. That kept eager ternary torch.where inconsistent with export, which already rejects uint8 predicates. Remove the Byte compatibility path from the shared nati..."
186859,Recover fake/proxy views from functional tensor bases,https://github.com/pytorch/pytorch/pull/186859,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,181304,181304,,,,0,4,1,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-06-10T01:17:18Z,2026-07-07T21:17:17Z,1,"Stack from ghstack (oldest at bottom): -> #186859 Fixes #181304 Generated by my agent When torch.compile traces torch.func.grad over a model with an N-D linear input, AOTAutograd's joint tracing can run C++ backward code that creates a plain Tensor view whose base is a FunctionalTensor. The FunctionalTensor is backe..."
184779,"[MPS] Rewrite _bincount_mps with native Metal atomic-add (~18 s → 2.1 ms, 8,500× speedup)",https://github.com/pytorch/pytorch/pull/184779,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,"77764, 149325",,,,0,1,4,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-21T18:45:01Z,2026-07-07T21:14:06Z,1,"Title: [MPS] Rewrite _bincount_mps with native Metal atomic-add (~18 s → 2.1 ms, 8,500× speedup) TL;DR Unweighted bincount on MPS is now faster than CPU. On the workload-shape input (2M int64 elements, mostly zeros with one 57.6k run), runtime drops from ~18,000 ms (legacy MPSGraph scatter-add) to 2.1 ms — a ~8,500×..."
184780,[MPS] Fast atomic-free path for flat torch.unique (fixes #97310 perf + #111173 correctness),https://github.com/pytorch/pytorch/pull/184780,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,97310,97310,"77764, 97310, 111173, 149325",,,,0,1,5,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-21T18:45:10Z,2026-07-07T21:14:05Z,1,"Title: [MPS] Fast atomic-free path for flat torch.unique (~19 s → 1.7 ms, fixes #97310) TL;DR Closes #97310 (perf, open since March 2023) and #111173 (silent correctness bug, open since October 2023). On the workload-shape input (2M int64 elements, single long duplicate run), torch.unique(..., return_counts=True) on..."
184787,[export] Deserialize PT2 schema in C++ during load,https://github.com/pytorch/pytorch/pull/184787,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,176428,176428,,,,0,3,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-21T19:46:05Z,2026-07-07T21:11:39Z,1,"Stack from ghstack (oldest at bottom): -> #184787 Root cause: torch.export.load() deserialized PT2 schema JSON through Python-side recursive dataclass construction. Large schemas spent most load time in _dict_to_dataclass, _bytes_to_dataclass, and typing.get_type_hints even though C++ schema parsing/construction alr..."
185054,[autograd] Validate forward-mode jacobian/hessian require vectorize=T…,https://github.com/pytorch/pytorch/pull/185054,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184842,184842,184842,,,,0,1,2,,,"triaged, open source, release notes: autograd, topic: improvements",2026-05-24T07:41:30Z,2026-07-07T21:10:35Z,1,"Fixes #184842 torch.autograd.functional.jacobian(strategy=""forward-mode"", vectorize=False) is an unsupported combination, but the API used to accept it at call time and then raise NotImplementedError from deep inside _jacfwd's fallback branch. The same shape failure applied to torch.autograd.functional.hessian(outer..."
189177,[inductor] Drop unused cache_info from CacheCompiledArtifact._validate_and_unpack,https://github.com/pytorch/pytorch/pull/189177,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-07T17:37:05Z,2026-07-07T21:08:24Z,1,"Addresses review feedback on #188377. CacheCompiledArtifact._validate_and_unpack returned (artifact_bytes, cache_info, key), but the middle cache_info element was discarded at both call sites: _to_binary_bytes needs only key, and save's unpacked branch needs only artifact_bytes. This narrows the return to (artifact_..."
187161,[FSDP] Cache post-forward DeviceMesh to avoid duplicate NCCL communicators per layer,https://github.com/pytorch/pytorch/pull/187161,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,187155,187155,187155,,,,0,2,4,,,"triaged, open source, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-06-12T08:33:55Z,2026-07-07T20:54:49Z,1,"Summary Fixes #187155 When fully_shard is called with reshard_after_forward as an int (e.g., 2), each layer creates its own DeviceMesh and NCCL communicator in _get_post_forward_mesh_info, even though the mesh topology is identical across layers. On NVSwitch-connected systems (H200) with NCCL 2.29.7, this exhausts t..."
171826,"[dynamo] add break_graph_if_unsupported to COMPARE_OP, BINARY_*, and INPLACE_* instructions",https://github.com/pytorch/pytorch/pull/171826,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"163503, 185524, 186100, 188173",,,0,77,5,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-01-06T22:00:30Z,2026-07-07T20:52:40Z,1,Stack from ghstack (oldest at bottom): -> #171826 #163503 #185524 #188173 #186100 #186657 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo
187133,[muon] add CUDA graph capture for the NS computation (#187133),https://github.com/pytorch/pytorch/pull/187133,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"release notes: optim, meta-exported",2026-06-12T02:19:36Z,2026-07-07T20:51:41Z,1,Summary: Context Diff 3 of the upstream port of the Gram Muon stack. Diff 1 added the multi-tensor scaffolding; Diff 2 added the Gram NS algorithm and batched standard NS for square groups. This diff layers CUDA graph capture on top of the multi-tensor NS dispatch -- collapsing the ~ns_steps * 3 per-param kernel lau...
187132,[muon] add Gram Newton-Schulz algorithm + batched standard NS for square groups (#187132),https://github.com/pytorch/pytorch/pull/187132,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"release notes: optim, meta-exported",2026-06-12T02:13:09Z,2026-07-07T20:51:40Z,1,Summary: Context Diff 2 of the upstream port of the Gram Muon stack. Diff 1 added the multi-tensor (foreach) scaffolding for the standard NS path with no algorithmic changes. This diff introduces the alternate orthogonalization algorithm itself plus the bmm-batched standard NS fast path for same-shape square groups...
188979,remaining HOPs register_fake,https://github.com/pytorch/pytorch/pull/188979,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,188978,,,0,9,3,,,topic: not user facing,2026-07-05T22:10:59Z,2026-07-07T20:51:23Z,1,"good way to review this is to expand the code above my changes to see original python impl, added 2 tests to test the register_fake mechanism Stack from ghstack (oldest at bottom): -> #188979 #188978"
188999,Add RISC-V CPU pause hint to cpu_atomic_add_float spin loop,https://github.com/pytorch/pytorch/pull/188999,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,180975,,,,0,1,2,,,"module: cpu, triaged, open source, topic: not user facing, module: risc-v",2026-07-06T04:36:22Z,2026-07-07T20:50:48Z,1,"Summary cpu_atomic_add_float retries a compare_exchange_weak in a busy loop and emits an architecture-specific CPU relax hint between attempts — yield on aarch64, _mm_pause() on x86. On RISC-V neither branch applied: __aarch64__ is undefined, so the loop fell through to _mm_pause(), which this header defines as an e..."
188042,[CI] Split AI advisor garbage verdict into infra_issue + garbage,https://github.com/pytorch/pytorch/pull/188042,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,0,0,,,,,,,,,0,1,1,,,topic: not user facing,2026-06-24T07:58:38Z,2026-07-07T20:50:42Z,1,"Adds an infra_issue verdict to the Claude CI Advisor (claude-autorevert-advisor.yml) — JSON-schema enum + prompt. Companion consumer changes (lambda, HUD, ClickHouse schema): pytorch/test-infra#8213. Why: a 7-day sample of advisor verdicts showed nearly all garbage verdicts were actually CI infrastructure failures,..."
185073,[export] Treat free new-tracer state as constants,https://github.com/pytorch/pytorch/pull/185073,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,165641,165641,,,,0,2,1,,,"release notes: export, agentic, needs-review",2026-05-24T15:52:34Z,2026-07-07T20:37:31Z,1,Stack from ghstack (oldest at bottom): -> #185073 Strict export with the new Dynamo tracer can install tensors captured from free/global scope onto the traced GraphModule as parameters or buffers. The state-dict cleanup code then builds a remapping table only for traced state that corresponds to parameters or buffer...
185075,Improve FakeTensorMode deepcopy error for real tensors,https://github.com/pytorch/pytorch/pull/185075,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,165547,165547,,,,0,2,3,,,"topic: not user facing, agentic, needs-review",2026-05-24T17:26:16Z,2026-07-07T20:32:17Z,1,"Stack from ghstack (oldest at bottom): -> #185075 Tensor.deepcopy is a Python entry point, so FakeTensorMode only saw the later storage-copy internals when a real CPU/CUDA tensor was copied under the mode. That fell through to set_ on a fake/meta tensor with real storage and produced a low-level meta-vs-device stora..."
182785,[Inductor] Fix FX wrapper codegen for symbolic floor/ceil offsets in ReinterpretView (#182785),https://github.com/pytorch/pytorch/pull/182785,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cpu, fb-exported, Stale, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-07T06:50:53Z,2026-07-07T20:12:15Z,1,Summary: The FX wrapper codegen (wrapper_fxir.py) crashes when generating as_strided nodes for ReinterpretView with symbolic offsets containing sympy.floor. This surfaces for ops like median/mode whose sort decomposition creates views with offset floor(rows * cols / 2 - 0.5). Two issues fixed: OptimizedPythonReferen...
182145,[ROCm] Reland #181955: Redirect index_add_ to scatter_add_ for large tensors,https://github.com/pytorch/pytorch/pull/182145,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/rocm-mi300,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,"180430, 181955",,,0,1,2,,,"module: rocm, open source, Stale, release notes: rocm, ci-no-td, ciflow/rocm-mi300",2026-05-01T17:26:05Z,2026-07-07T20:12:10Z,1,Reland of #181955 (and #180430). Both reverts came from public CI signal: the second was test_consistency_SparseCSR_mul_cuda_complex32 on linux.rocm.gpu.gfx950.1. Root cause: the original redirect's gate did not check dtype. index_add_cuda_impl dispatches via AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND4 (Indexing.cu:1275)...
176678,[CI][Windows] Add NVIDIA RTX workflow,https://github.com/pytorch/pytorch/pull/176678,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/win-arm64",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,170753,,,,0,8,3,,,"module: windows, triaged, open source, Stale, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/win-arm64",2026-03-06T03:06:09Z,2026-07-07T20:11:55Z,1,Summary: This PR adds a new experimental GitHub Actions workflow (.github/workflows/win-rtx.yml) to enable CI testing on Windows NVIDIA RTX GPUs. This workflow targets self-hosted runners to validate PyTorch functionality on consumer-grade RTX hardware. Motivation: To improve test coverage for Windows users running...
175566,Fix test_extension_backend race condition,https://github.com/pytorch/pytorch/pull/175566,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,137027,137027,,,,,0,3,0,,,"open source, Stale, topic: not user facing, module: inductor, ciflow/inductor",2026-02-23T20:42:48Z,2026-07-07T20:11:47Z,1,"Fixes #137027 Removing the build tree removes a lock file that we also are trying to separately remove. Instead, perform each build in its own directory. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhan..."
187481,Add overflow check to FBGEMM version of quantize_val_arm,https://github.com/pytorch/pytorch/pull/187481,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,187480,187480,187480,,,,0,1,2,,,"triaged, open source, release notes: quantization",2026-06-16T19:00:24Z,2026-07-07T20:09:25Z,1,"AffineQuantizerBase.cpp has two copies of quantize_val_arm, one inside #ifdef USE_FBGEMM and one on the #else branch. The latter copy calls __builtin_add_overflow, but the FBGEMM version does not. This PR adds the overflow check to the FBGEMM copy, fixing some quantization test failures on GB200. Fixes #187480 Autho..."
188250,Remove dead USE_ROCM block in CUDADeviceAssertionHost,https://github.com/pytorch/pytorch/pull/188250,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, triaged, open source, topic: not user facing",2026-06-26T06:41:34Z,2026-07-07T20:00:42Z,1,The inner #ifdef USE_ROCM is unreachable: its enclosing guard requires !defined(USE_ROCM). Also condense the stale v1/v2 comment. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
186245,[1/N] Optimize PyTorch CUDA build with LLVM BOLT,https://github.com/pytorch/pytorch/pull/186245,ready for maintainer decision,45,"CI lane labels: ciflow/binaries, ciflow/docker, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,188602,,,,0,6,2,,,"triaged, open source, ciflow/binaries, ciflow/trunk, release notes: cuda, topic: performance, release notes: cpu (aarch64), ciflow/docker",2026-06-04T17:46:45Z,2026-07-07T19:56:04Z,1,"LLVM BOLT is a post-link optimizer that optimize's application's code layout based on execution profile gathered by perf. It has been used to optimize NVIDIA's PyTorch containers since 26.04. This PR brings that optimization into the PyTorch repo. LLVM BOLT works in two phases, first a profile is collected using an..."
182972,added graph break on exception based on if user code would catch it,https://github.com/pytorch/pytorch/pull/182972,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,175608,175608,175608,,,,0,10,3,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/dynamo, ciflow/torchtitan",2026-05-08T17:27:29Z,2026-07-07T19:53:45Z,1,"fixes #175608 I realize this does not fix every try/except behavior error, but it does fix the repro with a pattern that can be resued. I can extend it other instances if that is the agreed approach. I also want to point out that try/except needs to be handled differently based on what triggers it. There are some ex..."
188605,Fix infinite recursion with pybind11 enum conversions,https://github.com/pytorch/pytorch/pull/188605,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,188547,188547,188547,,,,0,4,3,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-06-30T21:43:10Z,2026-07-07T19:50:37Z,1,"Fixes #188547 Fixes infinite recursion when calling int( ), float( ), or indexing pybind11 enums in compiled functions. Basically detects GetAttrVariable from pybind11 instancemethods and directly evaluates conversions on constants instead of recursing through call_method which previously led to infinite recursion c..."
184276,[inductor] fix torch.randperm for slice_shape node in fx_passes,https://github.com/pytorch/pytorch/pull/184276,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,183988,183988,"183988, 188721",,,,0,1,4,,,"module: cpu, triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/rocm-mi300, ciflow/h100, ciflow/vllm, ciflow/b200, ciflow/inductor-pallas, ci-refresh-hf-cache, ciflow/torchtitan, ciflow/dtensor",2026-05-18T21:26:22Z,2026-07-07T19:39:05Z,1,"Fixes: #183988 In [1]: import torch ...: ...: def fn(x): ...: idx = torch.randperm(x.shape[0], device=x.device) ...: return x[idx] ...: ...: x = torch.randn(8, 64, device=""cuda"") ...: ...: # Eager: OK ...: print(fn(x).shape) ...: ...: # aot_eager: OK ...: torch._dynamo.reset() ...: print(torch.compile(fn, backend=""a..."
185084,Avoid contiguity guards on maybe-singleton symbolic dims,https://github.com/pytorch/pytorch/pull/185084,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,165032,165032,,,,0,1,1,,,"topic: not user facing, agentic, needs-review",2026-05-24T21:54:06Z,2026-07-07T19:23:43Z,1,"Stack from ghstack (oldest at bottom): -> #185084 check_contiguous_sizes_strides(..., false_if_dde=True) is used by metadata and fake-tensor paths that should return a conservative boolean instead of specializing input shapes. Its singleton-dimension skip used guard_or_false(x == 1), which recorded a guard when the..."
185460,Fix CPU perf smoketest baseline drift checks,https://github.com/pytorch/pytorch/pull/185460,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,156511,156511,,,,0,16,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic",2026-05-28T03:05:41Z,2026-07-07T19:21:55Z,1,"Stack from ghstack (oldest at bottom): -> #185460 The CPU TorchBench smoketest compared measured speedup only against a lower bound: speedup < target * scale. When a PR made a model materially faster, the check still passed, so the expected baseline could remain stale. A later slowdown could then evade CI as long as..."
185946,[CUDA] Warn instead of assert in `reportProcessMemoryInfo` to prevent crash on Tegra OOM,https://github.com/pytorch/pytorch/pull/185946,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185240,185240,185240,,,186375,0,3,3,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-06-02T15:27:46Z,2026-07-07T19:21:47Z,1,"This PR replaces assertions for successful NVML calls with softer checks and warnings to prevent hard crashes on Tegra, which does not have full NVML support. Fixes #185240"
185299,Run CSE in AOT inference graphs,https://github.com/pytorch/pytorch/pull/185299,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,161067,161067,,,,0,3,5,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-05-26T23:59:12Z,2026-07-07T19:20:47Z,1,"Stack from ghstack (oldest at bottom): -> #185299 AOTAutograd's training path runs fx_graph_cse in the partitioner, but the forward-only inference path bypasses that partitioner and never performs this CSE. Quantization-like code that computes the same activation transform for several consumers therefore reaches Ind..."
185300,Fix AttrProxy module reuse in Dynamo,https://github.com/pytorch/pytorch/pull/185300,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,161053,161053,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-27T00:17:17Z,2026-07-07T19:17:11Z,1,"Stack from ghstack (oldest at bottom): -> #185300 Dynamo checks side-effect tracking before VariableBuilder wraps a value, but proxy_tensor _AttrProxy modules are unwrapped inside wrap_module before they are tracked. When the same proxy-backed module is first seen as a local key and later through WeakKeyDictionary i..."
165766,Enable kineto XPU scope profiler extending existing XPU profiler plugin,https://github.com/pytorch/pytorch/pull/165766,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,56,0,,,"triaged, open source, topic: not user facing, keep-going, ciflow/xpu",2025-10-17T16:29:12Z,2026-07-07T19:13:14Z,1,"Summary: As XPU became a PyTorch built-in device, the profiler support is indispensable part of functionality completeness. In this PR, the XPU scope profiler is introduced by extending existing XPU profiler plugin. It allows to gather XPU hardware metrics. The LIBKINETO_NOXPUPTI option is used to enable or disable..."
181860,[c10] Add InlineEventBase to unify event APIs in DeviceGuardImplInter…,https://github.com/pytorch/pytorch/pull/181860,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,163836,163836,163836,,,,0,3,2,,,"open source, Stale",2026-04-29T12:23:45Z,2026-07-07T19:12:51Z,1,"…face Event methods on DeviceGuardImplInterface have inconsistent signatures some take void* + DeviceIndex, others take just void*. This makes it impossible for backends to uniformly access event metadata like device type or flags without extra bookkeeping. Adding InlineEventBase, a minimal non-templated base class..."
185320,Improve scan HOP direct call error,https://github.com/pytorch/pytorch/pull/185320,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,160544,160544,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-27T04:13:25Z,2026-07-07T19:11:00Z,1,"Stack from ghstack (oldest at bottom): -> #185320 Direct calls to torch.ops.higher_order.scan are not the supported user-facing API. Dynamo's scan HOP path still accepted any functools.partial as if it came from the torch._higher_order_ops.scan frontend, then unconditionally read combine_fn.keywords[""combine_fn""]. W..."
184052,Preserve CIA ops in AOT joint export,https://github.com/pytorch/pytorch/pull/184052,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,160388,160388,,,,0,2,5,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-16T23:17:15Z,2026-07-07T19:06:21Z,1,Stack from ghstack (oldest at bottom): -> #184052 Apply export's CompositeImplicitAutograd decomposition table handling to aot_export_joint_simple so backend AOT graphs agree with export on preserved functional CIA ops. Keep direct aot_export_module decomposition behavior unchanged. Fixes #160388 Generated by my agent
184059,Fix split reduction kernel metadata,https://github.com/pytorch/pytorch/pull/184059,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,160124,160124,,,,0,3,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T01:18:39Z,2026-07-07T19:02:38Z,1,"Stack from ghstack (oldest at bottom): -> #184059 Build split-reduction graph fragments from scheduler-local IO so generated metadata reflects each kernel's actual inputs and outputs, including fused producers and epilogues. Fixes #160124 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-..."
187012,Support optional Tensor list schemas,https://github.com/pytorch/pytorch/pull/187012,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,160094,160094,,,,0,2,1,,,"release notes: python_frontend, agentic, needs-review",2026-06-11T04:20:19Z,2026-07-07T18:59:07Z,1,"Stack from ghstack (oldest at bottom): -> #187012 torch.library.infer_schema derives supported parameter annotations from a small table of base types and generated list/optional variants. Tensor enabled Tensor[] and Tensor?[] but explicitly disabled Tensor[]?, so annotations such as Optional[list[torch.Tensor]] were..."
185336,[dynamo] Add regression test for prepended forward pre-hooks,https://github.com/pytorch/pytorch/pull/185336,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,160083,160083,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-27T07:48:16Z,2026-07-07T18:55:57Z,1,"Stack from ghstack (oldest at bottom): -> #185336 Registering a later forward pre-hook with prepend=True mutates the module hook OrderedDict order with move_to_end(last=False). In the failure from #160083, Dynamo traced hooks in OrderedDict iteration order but used guard sources derived from dict.keys(...) positions..."
185347,Support torch.compile under user FakeTensorMode,https://github.com/pytorch/pytorch/pull/185347,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,45,1,"statusCheckRollup, linux-jammy-py3.10-clang18-asan / test-osdc (default, 7, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test-osdc (default, 5, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test-osdc (default, 6, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (default, 2, 7, linux.4xlarge), linux-jammy-py3.10-clang18-asan / test-osdc (default, 3, 7, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18-asan / test (default, 3, 7, linux.4xlarge), linux-jammy-py3.10-clang18-asan / test (default, 6, 7, linux.4xlarge), linux-jammy-py3.10-clang18-asan / test (default, 4, 7, linux.4xlarge), linux-jammy-py3.10-gcc11 / test (default, 4, 5, linux.4xlarge), linux-jammy-py3.10-gcc11 / test-osdc (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test (default, 3, 5, linux.4xlarge), linux-jammy-py3.14t-clang18 / test-osdc (default, 5, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test (distributed, 1, 3, linux.2xlarge.amx), linux-jammy-py3.10-gcc11 / test-osdc (distributed, 2, 3, mt-l-x86iamx-8-64), linux-jammy-py3.10-gcc11 / test (default, 5, 5, linux.4xlarge), linux-jammy-py3.10-gcc11 / test-osdc (default, 5, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-gcc11 / test-osdc (default, 3, 5, mt-l-x86iavx512-16-128), linux-jammy-aarch64-py3.10 / test (default, 2, 5, linux.arm64.m8g.4xlarge), linux-jammy-py3.10-clang18 / test-osdc (default, 5, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test (default, 5, 5, linux.4xlarge), linux-jammy-py3.10-clang18 / test-osdc (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test-osdc (default, 3, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.10-clang18 / test (default, 4, 5, linux.4xlarge), linux-jammy-py3.14t-clang18 / test (default, 4, 5, linux.4xlarge), linux-jammy-py3.10-clang18 / test (default, 3, 5, linux.4xlarge), linux-jammy-py3.14t-clang18 / test (default, 3, 5, linux.4xlarge), linux-jammy-py3.14t-clang18 / test (default, 2, 5, linux.4xlarge), linux-jammy-aarch64-py3.10 / test (default, 5, 5, linux.arm64.m8g.4xlarge), linux-jammy-py3.14t-clang18 / test-osdc (default, 4, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14t-clang18 / test-osdc (default, 3, 5, mt-l-x86iavx512-16-128), linux-jammy-aarch64-py3.10 / test (default, 3, 5, linux.arm64.m8g.4xlarge), linux-jammy-aarch64-py3.10 / test (default, 4, 5, linux.arm64.m8g.4xlarge), linux-jammy-aarch64-py3.10 / test-osdc (default, 4, 5, mt-l-arm64g4-16-62), linux-jammy-aarch64-py3.10 / test-osdc (default, 3, 5, mt-l-arm64g4-16-62), linux-jammy-aarch64-py3.10 / test-osdc (default, 5, 5, mt-l-arm64g4-16-62), linux-jammy-py3.14-clang18 / test (crossref, 2, 2, linux.2xlarge), linux-jammy-py3.14-clang18 / test (dynamo_wrapped, 3, 3, linux.2xlarge), linux-jammy-py3.14-clang18 / test (default, 4, 5, linux.4xlarge), linux-jammy-py3.14-clang18 / test-osdc (crossref, 1, 2, mt-l-x86iavx512-8-64), linux-jammy-py3.14-clang18 / test (default, 2, 5, linux.4xlarge), linux-jammy-py3.14-clang18 / test-osdc (default, 2, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test (default, 3, 5, linux.4xlarge), linux-jammy-py3.14-clang18 / test-osdc (default, 5, 5, mt-l-x86iavx512-16-128), linux-jammy-py3.14-clang18 / test-osdc (default, 4, 5, mt-l-x86iavx512-16-128)",statusCheckRollup,,"136586, 160057","136586, 160057",,,,0,1,8,779f6cdda464b7f91ae90189a7f7f97d054b73af4778bde66c09af8276267f4c,3bfc3f4a2f2e8def8f2be6adcd417e5af93ebd4d4fdd27068404f4d911ae3faf,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-27T09:15:19Z,2026-07-07T18:51:04Z,1,"Stack from ghstack (oldest at bottom): -> #185347 Dynamo creates its own FakeTensorMode for tracing and then switches to a fresh backend FakeTensorMode before handing graphs to AOTAutograd/Inductor. When torch.compile was invoked under a user-created FakeTensorMode, backend example inputs could still belong to the u..."
185349,Fix flex decoding BLOCK_M validation for short GQA,https://github.com/pytorch/pytorch/pull/185349,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,160018,160018,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-27T09:47:33Z,2026-07-07T18:47:13Z,1,"Stack from ghstack (oldest at bottom): -> #185349 Flex decoding packs grouped query attention heads into the M tile, so its BLOCK_M is measured across query heads while the sparse block mask's query block size is measured in per-head query rows. The previous lowering compared SPARSE_Q_BLOCK_SIZE directly against the..."
185353,Normalize export state names for subclass decomposition,https://github.com/pytorch/pytorch/pull/185353,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,159918,159918,,,,0,3,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-27T10:43:12Z,2026-07-07T18:42:42Z,1,Stack from ghstack (oldest at bottom): -> #185353 run_decompositions() unwraps tensor subclass parameters and buffers ahead of time by registering an nn.utils.parametrize parametrization. The resulting ExportedProgram was exposing parametrization implementation details such as parametrizations.p2.original0 in the gr...
185356,Fix DTensor embedding mask state under compile,https://github.com/pytorch/pytorch/pull/185356,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,159843,159843,,,,0,4,1,,,"module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor, agentic, needs-review",2026-05-27T12:08:34Z,2026-07-07T18:38:46Z,1,"Stack from ghstack (oldest at bottom): -> #185356 Rowwise DTensor embedding uses _MaskPartial placements to carry the mask created while redistributing indices. That mask is tensor state, but it lived inside placement metadata rather than DTensor.local_tensor. AOTAutograd functionalization only transformed DTensor's..."
186003,Fix Dynamo condition-dependent frame skip caching,https://github.com/pytorch/pytorch/pull/186003,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,144820,144820,,,,0,15,1,3945ee225b03e48d094049b7d7c7eca9f35bb8def955d6d84cd6afe7e9572b17,745365ea9006e21626694b1d1ad775128b7249ca0ae74c0363133f772e1cd9a9,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T22:20:13Z,2026-07-07T18:37:28Z,1,"Stack from ghstack (oldest at bottom): -> #186003 Dynamo previously returned a default ConvertFrameReturn when tracing decided to skip a frame. The default frame execution strategy applies SKIP to the code object, so a skip discovered under one set of locals became an unconditional skip for all future calls to the s..."
185357,Forward container methods on OptimizedModule,https://github.com/pytorch/pytorch/pull/185357,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,159831,159831,,,,0,4,3,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-27T12:24:19Z,2026-07-07T18:33:56Z,1,"Stack from ghstack (oldest at bottom): -> #185357 torch.compile wraps nn.Module instances in OptimizedModule and proxies normal attribute access to the original module. Python special method lookup bypasses getattr, so ModuleDict.getitem was never found on the wrapper and compiled ModuleDict instances were not subsc..."
188698,Make Fsspec filesystem checkpointing public and exposing cache_staged_state_dict to enable async save,https://github.com/pytorch/pytorch/pull/188698,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/torchtitan,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,188772,188772,188772,,,,0,2,4,,,"open source, release notes: distributed (checkpoint), ciflow/torchtitan",2026-07-01T16:42:42Z,2026-07-07T18:23:02Z,1,Fixes #188772 This PR makes FsspecReader and FsspecWriter public by renaming _fsspec_filesystem.py to fsspec_filesystem.py and exposing them in the torch.distributed.checkpoint namespace. The changes also include exposing cache_staged_state_dict in its constructor and passing it to the parent FileSystemWriter to ful...
188946,[Bugfix][FSDP] Capture out-of-place redistribute() return value in optimizer state reconstruction,https://github.com/pytorch/pytorch/pull/188946,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,111774,,,0,1,2,,,"triaged, open source, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-07-04T11:26:32Z,2026-07-07T18:20:48Z,1,"Capture out-of-place redistribute() return value in _unflatten_orig_param_states Fixes a correctness bug in TP+FSDP optimizer-state reconstruction where DTensor.redistribute() was called without capturing its return value. Root Cause In _unflatten_orig_param_states(), the code calls: value.redistribute(placements=(R..."
188816,"[Metal][ops] remove nonzero INT_MAX limit, widen to 64-bit (#149325)",https://github.com/pytorch/pytorch/pull/188816,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,149325,,,,0,2,2,,,"triaged, open source, release notes: mps, ciflow/mps",2026-07-02T15:49:04Z,2026-07-07T18:14:26Z,1,"Summary Addresses the nonzero op case of #149325 (the umbrella tracking MPS 32-bit index limits). MPS nonzero rejected any tensor with numel >= INT_MAX, and the internal flat accumulator in the scatter_nonzero_indices Metal kernel used 32-bit arithmetic so large-dim modulo/division would wrap even below the guard. R..."
177540,[inductor] Handle Triton cluster_dims metadata from both binary and metadata layouts,https://github.com/pytorch/pytorch/pull/177540,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,1,,,"triaged, open source, Stale, ciflow/trunk, module: inductor, ciflow/inductor",2026-03-16T16:19:47Z,2026-07-07T18:13:39Z,1,"Summary Handle Triton cluster_dims metadata from both binary and metadata layouts. Problem Some Triton versions expose cluster dims on the compiled binary, while others surface them on metadata. Inductor currently assumes a single layout, which can break launcher construction when num_ctas and cluster_dims come from..."
188396,[Test] Improve test reuse in test/test_tensor_iterator.py for out-of-tree backends,https://github.com/pytorch/pytorch/pull/188396,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,2,,,"triaged, open source, topic: not user facing",2026-06-29T07:12:19Z,2026-07-07T18:13:06Z,1,"Summary Improve test reuse in test/test_tensor_iterator.py for out-of-tree backends by adopting device-generic test patterns and removing hardcoded CUDA assumptions. Changes Move the cross-device tests (test_cross_device_check_raises, test_allow_cpu_scalars) from TestTensorIteratorBuild (CPU-only) to TestTensorItera..."
186622,Fix bf16 autocast precision barriers in Inductor,https://github.com/pytorch/pytorch/pull/186622,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,168126,168126,,,,0,5,1,2b2234fadb15d798e236f8c2fd0089c7f7b393b69203192820f6b44f8ffb11f2,02a891af9a885eec13472015df41bd3db679b98c7769600c43cb3487794e5b22,"release notes: fx, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-08T17:23:01Z,2026-07-07T18:09:44Z,1,"Stack from ghstack (oldest at bottom): -> #186622 The issue repro combines bf16 autocast with torch.compile. Inductor was diverging from eager in two ways: pointwise fusion elided eager bf16 round trips across linear/sigmoid/mul/cast boundaries, and the native_layer_norm decomposition used reference var_mean statist..."
185365,Fix FX subgraph rewrite of adjacent aggregate matches,https://github.com/pytorch/pytorch/pull/185365,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,159613,159613,,,,0,1,3,,,"release notes: fx, agentic, needs-review",2026-05-27T12:55:13Z,2026-07-07T18:06:57Z,1,"Stack from ghstack (oldest at bottom): -> #185365 replace_pattern computes all matches before it starts mutating the graph, so later matches may contain references to nodes that earlier replacements erase. The existing bookkeeping handled this only when a later pattern placeholder directly matched a replaced Node. F..."
188331,[TEST] Refactor test_profiler_tree.py with hw_classification,https://github.com/pytorch/pytorch/pull/188331,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,1,,,"triaged, open source, topic: not user facing",2026-06-27T11:02:27Z,2026-07-07T18:05:28Z,1,"Extract duplicated assertTreesMatch helper into a ProfilerTreeTestMixin shared by both test classes, and add hw_classification attributes: GENERIC for TestProfilerTree, CUDA for TestProfilerTreeCUDA. Test Cases: Before and after: 10 tests collected. 3 passed, 7 skipped."
188828,Skip conv_transpose3d redispatch test on Windows CI - flaky test,https://github.com/pytorch/pytorch/pull/188828,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,"182869, 188757","182869, 188757",188757,,,,0,1,2,,,"open source, ciflow/trunk, release notes: python_frontend",2026-07-02T16:53:54Z,2026-07-07T18:03:47Z,1,Summary Extend the existing conv_transpose3d OpInfo skip for TestTorchFunctionRedispatchOps.test_redispatch to include Windows CI. TestTorchFunctionRedispatchOpsCUDA.test_redispatch_nn_functional_conv_transpose3d_cuda_complex64 fails consistently on Windows CI with the same error as Linux. This issue disabled the te...
185369,Fix dynamic output sizes for adaptive avg pool lowering,https://github.com/pytorch/pytorch/pull/185369,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,"159550, 185575","159550, 185575",,,,0,1,2,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-27T13:42:54Z,2026-07-07T18:01:43Z,1,"Stack from ghstack (oldest at bottom): -> #185369 The Inductor lowering for aten._adaptive_avg_pool2d computes the maximum adaptive pooling window from the input and output sizes. When output_size is provided as symbolic Python arguments under torch.compile(dynamic=True), that maximum window can be a SymPy expressio..."
185371,Fix symbolic antialiased resize export shapes,https://github.com/pytorch/pytorch/pull/185371,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,4,1,"statusCheckRollup, win-vs2022-cpu-py3 / build, win-vs2022-cuda13.0-py3 / build, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,159468,159468,,,,0,2,5,fb1666005be245ea22f79996fa9a71f20b7ffa7a95041e66309350bc2b3b6c96,fbfd4eb39307137357509e0fef7039c26922ed9692e4869fce511752120a5818,"ciflow/trunk, release notes: onnx, ciflow/inductor, agentic, needs-review",2026-05-27T14:23:13Z,2026-07-07T17:59:04Z,1,"Stack from ghstack (oldest at bottom): -> #185371 Dynamo export could specialize dynamic interpolate scale factors for antialiased bilinear resize. In the reported model the scale is computed as 16 / image.shape[2], but the vector upsample decomposition checked int(s) == s; with a symbolic float this uses the exampl..."
184648,[Test] Add has_sufficient_memory() hook to DeviceTypeTestBase,https://github.com/pytorch/pytorch/pull/184648,PR blocked,108,failed checks or blocked/unstable merge state; approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,BLOCKED,False,0,0,,,,,,,,,0,7,1,,,"triaged, open source",2026-05-21T03:19:01Z,2026-07-07T17:58:35Z,1,Summary This PR adds a has_sufficient_memory() hook to DeviceTypeTestBase and updates largeTensorTest() to use it when available. Motivation largeTensorTest() currently relies on _has_sufficient_memory() to determine whether a device has enough available memory for a requested allocation size. _has_sufficient_memory...
188434,[dcp] Free staged dict on caller thread via weakref.finalize in async_save (#188434),https://github.com/pytorch/pytorch/pull/188434,ready for maintainer decision,45,"CI lane labels: ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, release notes: distributed (checkpoint), meta-exported, ciflow/torchtitan",2026-06-29T15:36:06Z,2026-07-07T17:56:04Z,1,"Summary: _ThreadBasedAsyncCheckpointExecutor.execute_save submits the staged state-dict to a worker thread for the upload, then returns the Future to the caller. The worker holds the only strong reference to the staged dict (via the ThreadPool WorkItem's args), so when save_wrapper returns the dict's CPU storages ge..."
185373,Fix assume_constant_result tensor constants after graph breaks,https://github.com/pytorch/pytorch/pull/185373,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,159457,159457,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-27T14:45:10Z,2026-07-07T17:55:12Z,1,"Stack from ghstack (oldest at bottom): -> #185373 When a function marked with torch._dynamo.assume_constant_result returned a Tensor, Dynamo registered that Tensor through register_attr_or_module using a ConstantSource built from the function name. Unlike the non-Tensor path, the Tensor registration path does not in..."
185892,Improve export device mismatch errors,https://github.com/pytorch/pytorch/pull/185892,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,7,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,184841,184841,,,,0,4,10,0b020a5e62f41ed6d2c782c0da9c498f1d1b385b15aabe03fba11379b5f86d60,1af81d5de74dc78d9d4166065e5149b91dfde2ee5669751a9a4c41cdcf3c42f4,"module: dynamo, ciflow/inductor, release notes: export, ciflow/torchtitan, agentic, needs-review",2026-06-02T05:12:39Z,2026-07-07T17:52:17Z,1,"Stack from ghstack (oldest at bottom): -> #185892 Non-strict torch.export.export fakifies example inputs and module state before tracing. When those tensors are on different devices, FakeTensor device propagation can raise an implementation-facing FakeTensorDeviceMismatchError before export has a chance to present a..."
184103,Fix FakeTensorUpdater HOP dtype propagation,https://github.com/pytorch/pytorch/pull/184103,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,152548,152548,,,,16,10,10,72facefc09ddd73da51a632a8f4a2f4c309409e4dddba20c138a9be3f8c12e36,750c3e295ddd983a54e5ce127d97053f3c7097441efbd40d6697ba6df5d7125e,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T16:28:43Z,2026-07-07T17:48:51Z,1,Stack from ghstack (oldest at bottom): -> #184103 Compare fake tensor dtypes when deciding whether metadata changed and retrace higher-order operators even when they are lowerable. Seed subgraph updater work from placeholder metadata updates so dtype changes propagate through nested graphs. Fixes #152548 Generated b...
187901,[TEST] Prevent PLATFORM_SUPPORTS_* from skipping tests on privateuse1 backends,https://github.com/pytorch/pytorch/pull/187901,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, module: PrivateUse1",2026-06-23T02:57:20Z,2026-07-07T17:45:17Z,1,"Summary Add TEST_PRIVATEUSE1 handling to all evaluate_platform_supports_*() functions and PLATFORM_SUPPORTS_* constants in common_cuda.py, preventing tests guarded by @unittest.skipUnless(PLATFORM_SUPPORTS_*, ...) from being skipped on PrivateUse1 backends. Motivation Many tests are guarded by PLATFORM_SUPPORTS_* fl..."
185404,[inductor] Fold small-row expanded bmm to mm in post-grad,https://github.com/pytorch/pytorch/pull/185404,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,159346,159346,,,,0,1,3,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-27T16:55:17Z,2026-07-07T17:44:24Z,1,"Stack from ghstack (oldest at bottom): -> #185404 A dynamic linear on batch-first transposed activations can reach post-grad as an expanded batched matmul: [T, B, K] @ expand([K, N], [T, K, N]) For small B this produces many small-row GEMMs. On CUDA fp16 this path can be much slower than flattening the transposed ac..."
187828,Add privateuse1 backend support to has_triton(),https://github.com/pytorch/pytorch/pull/187828,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, module: PrivateUse1",2026-06-22T11:36:25Z,2026-07-07T17:43:48Z,1,"Motivation PrivateUse1 backends may provide Triton support through their own Triton integrations. However, has_triton() currently only recognizes a fixed set of built-in device types. Many Triton-related tests rely on has_triton(). Since registered PrivateUse1 backends are not recognized by this check, existing test..."
189089,[fmha_fwd][rocm] Extend CK ck_tile FMHA seqlen-pad guard to the GQA swap path,https://github.com/pytorch/pytorch/pull/189089,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, topic: not user facing, meta-exported",2026-07-06T21:19:18Z,2026-07-07T17:42:50Z,1,"Summary: Follow-up to the gfx9 CK ck_tile FMHA tile-tail OOB guard: the guard excluded the seqlenq_ngroups_swapped (GQA decode) path, but that path is exactly where a real MoE/MTML inference model still faults during AOTT GPU lowering on AMD (MI300X gfx942): ""Memory access fault by GPU node-N ... Reason: Unknown"". R..."
189108,[Inductor] Bound exact index expressions for int32 indexing eligibility,https://github.com/pytorch/pytorch/pull/189108,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-06T23:04:06Z,2026-07-07T17:42:13Z,1,"[Inductor] Bound exact index expressions for int32 indexing eligibility Summary Decide int32-vs-int64 indexing by checking the actual addressing expressions a kernel will emit, instead of approximating them with a per-buffer storage-size or stride proxy. Background / Root cause When unrolled chunked-loop slices are..."
184061,Fix fused moving avg observer functional meta,https://github.com/pytorch/pytorch/pull/184061,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,159192,159192,,,,0,2,5,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T01:47:12Z,2026-07-07T17:40:44Z,1,Stack from ghstack (oldest at bottom): -> #184061 Model the qparam tensor resize side effects in the meta kernel so the generated functional variant returns accurate metadata for per-row fake quantization. Fixes #159192 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper...
186894,Fix non-strict export of vmap tensor indexing,https://github.com/pytorch/pytorch/pull/186894,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,158540,158540,,,,0,1,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-06-10T05:13:52Z,2026-07-07T17:37:06Z,1,"Stack from ghstack (oldest at bottom): -> #186894 Non-strict export rewrites scalar tensor indices in getitem through select/slice helpers, and the scalar-index path converted 0-d integer tensor indices with item(). Under vmap those logical scalar indices are BatchedTensor values, where item() is unsupported, so exp..."
188130,[Test] Make test_autocast device-agnostic for Out-of-Tree backends,https://github.com/pytorch/pytorch/pull/188130,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, module: amp (automated mixed precision), ciflow/trunk, topic: not user facing, module: tests/core",2026-06-25T03:31:28Z,2026-07-07T17:33:31Z,1,"Changes Refactors test/test_autocast.py to decouple tests from CUDA-specific assumptions, enabling Out-of-Tree backends to reuse accelerator-agnostic tests. Class restructuring Before After Notes TestAutocastCPU (11 tests) Unchanged CPU-only tester using AutocastCPUTestLists TestAutocastGPU (3 tests) → TestAutocastD..."
184065,Fix static launcher 16-bit float scalars,https://github.com/pytorch/pytorch/pull/184065,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,158521,158521,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T02:50:14Z,2026-07-07T17:30:43Z,1,Stack from ghstack (oldest at bottom): -> #184065 Encode fp16 and bf16 scalar arguments with distinct launcher ABI codes so static and fast CUDA launchers pack them as 16-bit values instead of 32-bit floats. Add regression coverage for fp16 and bf16 scalar launches. Fixes #158521 Generated by my agent cc @voznesensk...
185419,[dynamo] avoid stale verbose hint for backend failures,https://github.com/pytorch/pytorch/pull/185419,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,158384,158384,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-27T18:20:48Z,2026-07-07T17:26:18Z,1,"Stack from ghstack (oldest at bottom): -> #185419 Backend compilation runs under TracingContext.clear_frame(), which intentionally marks exceptions with real_stack = None so Dynamo does not attach an incorrect tracing user stack to backend failures. augment_exc_message() only checked whether the real_stack attribute..."
185436,Preserve eager exception assertions for eager_noexcept,https://github.com/pytorch/pytorch/pull/185436,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,158087,158087,,,,0,4,1,,,"module: cpu, module: mkldnn, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-27T21:05:37Z,2026-07-07T17:22:40Z,1,Stack from ghstack (oldest at bottom): -> #185436 PYTORCH_TEST_WITH_DYNAMO=1 runs tests through the eager_noexcept backend so unexpected exceptions from Dynamo-generated GraphModules remain visible as Dynamo bugs. That wrapper was too opaque for tests that explicitly assert eager exception behavior: when a generated...
185438,Scope distribution validation disabling to Dynamo tracing,https://github.com/pytorch/pytorch/pull/185438,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,157926,157926,,,,0,3,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-27T22:06:06Z,2026-07-07T17:20:01Z,1,"Stack from ghstack (oldest at bottom): -> #185438 TorchDynamo permanently disabled torch.distributions parameter validation from TorchPatcher.patch() when a user constructed a torch.compile wrapper. That changed Distribution._validate_args for the whole Python process, so unrelated eager distribution constructors co..."
184718,[Test] Make test_testing.py device-agnostic for out-of-tree accelerators,https://github.com/pytorch/pytorch/pull/184718,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,1,,,"open source, topic: not user facing, module: tests/others",2026-05-21T14:41:19Z,2026-07-07T17:17:47Z,1,"Summary This PR removes CUDA-specific assumptions from test/test_testing.py to enable out-of-tree accelerator backends to run these tests. Changes: test_get_supported_dtypes: The else branch previously hardcoded op.dtypesIfCUDA, which would return incorrect dtypes for any non-CUDA accelerator. Replaced with op.suppo..."
185450,Fix reinplace alias checks for index_put,https://github.com/pytorch/pytorch/pull/185450,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,156786,156786,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-28T00:27:35Z,2026-07-07T17:16:47Z,1,Stack from ghstack (oldest at bottom): -> #185450 Inductor's reinplace pass can rewrite functional index_put into its in-place variant when the destination has no later observable uses. That check was only reasoning about later uses of the destination tensor. It did not account for operands that index_put reads duri...
185813,[ROCm][Windows] Audit license-files: explicit enumeration + audit test (#183434),https://github.com/pytorch/pytorch/pull/185813,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/docker,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,180237,,,0,3,2,,,"module: rocm, triaged, open source, topic: not user facing, ciflow/docker",2026-06-01T12:53:35Z,2026-07-07T17:15:10Z,1,"Fix for pytorch/pytorch#183434. The PEP 639 migration (#180237) replaced the old concatenated LICENSE blob with per-file license-files in pyproject.toml. The existing recursive globs: license-files = [ ""LICENSE"", ""third_party/**/LICENSE"", ""third_party/**/LICENSE.txt"", ""third_party/**/LICENSE.rst"", ""third_party/**/CO..."
187165,[ROCm] Enable FMA inductor lowering on ROCm,https://github.com/pytorch/pytorch/pull/187165,draft/noise,31,"PR is draft; review decision requests changes; CI lane labels: ciflow/inductor, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/rocm-mi200, ciflow/rocm-mi300",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,1,,,"module: rocm, triaged, open source, release notes: foreach_frontend, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/rocm-mi200, ciflow/torchtitan",2026-06-12T10:48:03Z,2026-07-07T17:14:35Z,1,Co-authored by Claude cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @a...
185816,[WIP] [PoC] Enable CPU FlexAttention on AArch64,https://github.com/pytorch/pytorch/pull/185816,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,3,1,,,"open source, ciflow/trunk, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-01T13:46:39Z,2026-07-07T17:14:11Z,1,"This PR enables Inductor's existing CPU C++ FlexAttention template on AArch64 devices. The generated FlexAttention C++ template uses at::native::cpublas::gemm for the QK matmul on SVE256, matching the approach used by the CPU FlashAttentionKernel.cpp instead of relying on the transpose-B micro-GEMM path. Summary Thi..."
184946,[Inductor] Fix split reduction codegen and skip redundant masking for block ptr backends (#184946),https://github.com/pytorch/pytorch/pull/184946,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,1,,,"fb-exported, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-23T00:01:24Z,2026-07-07T17:14:10Z,1,"Summary: Previously when use_block_ptr=True, split reduction kernels were unable to generate block pointer loads because ops.masked() sets _load_mask which forces scalar pointer arithmetic. This is bad for performance on backends that require block pointers for efficient memory access. To address this, instead of us..."
185451,Wrap internal Dynamo assertion errors,https://github.com/pytorch/pytorch/pull/185451,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,156640,156640,,,,0,4,1,3939f5013d99d3580bdd893d30ae9f4eeedb5cb3b302f5f15849df7435c8932f,3d602b1400ce18d7277787cc31ec066d922fce9434d2325790dd5e26d622191f,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-28T01:12:18Z,2026-07-07T17:12:06Z,1,"Stack from ghstack (oldest at bottom): -> #185451 Dynamo's frame compilation error path previously treated every AssertionError as an expected pass-through exception. That kept user assertions working, but it also let internal Dynamo assertions escape as raw AssertionError, which makes unexpected compiler bugs look..."
184985,Fix nested compile handling in AOT joint export,https://github.com/pytorch/pytorch/pull/184985,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,169991,169991,,,,2,8,12,4671f53e38c8c48f2d3b1dccd7115b5bb41499603be01b2b8752be78ae9a1894,471d14edf0cc85e628b2908707267ef32c20976a6c1b18bfea2563ec9b2a5bd5,"module: dynamo, ciflow/inductor, release notes: export, ciflow/torchtitan, agentic, needs-review",2026-05-23T11:40:48Z,2026-07-07T17:10:59Z,1,"Stack from ghstack (oldest at bottom): -> #184985 Direct calls to aot_export_joint_with_descriptors did not install the export compile state that strict and non-strict torch.export paths already use. When a model contained an existing torch.compile wrapper, or called torch.compile inside forward, Dynamo still tried..."
188994,[dynamic shapes] Preserve algebraic relationships in foreign unbacked SymInt transfer (#188723),https://github.com/pytorch/pytorch/pull/188994,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,0,0,,,188723,188723,188723,,,,0,1,3,,,"triaged, open source, release notes: fx, ciflow/inductor",2026-07-06T04:15:37Z,2026-07-07T17:08:55Z,1,"Fixes #188723 ShapeEnv._transfer_foreign_expr_as_unbacked had an order-dependent bug: when a composite foreign expression (e.g. u0 + u1) was transferred before its base unbacked symbols (u0, u1) had been seen individually, the algorithm collapsed the composite to a single opaque local symbol, losing algebraic relati..."
185057,Support torch.cuda.use_mem_pool in compile,https://github.com/pytorch/pytorch/pull/185057,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,167026,167026,,,,22,6,7,eddc620384774756e102471aad24eeac95ac6724b1a9f7ec515441c95167796c,eb143c09fac2ec310913b40a34bed3a34381ad57c1a5e5184630de94a45811f4,"module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-24T08:03:10Z,2026-07-07T17:08:52Z,1,"Stack from ghstack (oldest at bottom): -> #185057 Dynamo previously treated torch.cuda.use_mem_pool as a skipped Python context manager, so fullgraph tracing failed before the context body could be compiled. Even if Dynamo emitted begin/end calls, Inductor could not rely on those side-effect nodes to order generated..."
184312,Fix efficient zero tensor handling in torch.compile,https://github.com/pytorch/pytorch/pull/184312,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,177261,177261,,,,,0,13,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ci-no-td, release notes: inductor (aoti), ciflow/torchtitan",2026-05-19T03:50:59Z,2026-07-07T17:08:27Z,1,Follow-up for #178933 Fixes #177261 PR #178933 attempted to fix the IMA by materializing the full tensor but internal tests failed due to an OOM. I analyzed the ZeroTensor eager fallback implementation more closely and it materializes the efficient zeros tensor using reshape of a 1-element 0 tensor. It should be saf...
188766,[CP] Fix FlexCP + load balancing enabled + aot_fx_trace issue,https://github.com/pytorch/pytorch/pull/188766,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186282,,,0,2,1,,,"ciflow/trunk, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-07-02T01:10:13Z,2026-07-07T17:06:29Z,1,"Stack from ghstack (oldest at bottom): #186282 -> #188766 Without the fix, TorchTitan graph_trainer + FlexCP with load balancing will trigger the following errors: Without squeeze(0) fix: RuntimeError: Attempting to use FunctionalTensor on its own. Instead, please use it with a corresponding FunctionalTensorMode() W..."
185461,[dynamo] Handle module-level logging ignores,https://github.com/pytorch/pytorch/pull/185461,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,156191,156191,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-28T03:32:55Z,2026-07-07T17:06:15Z,1,"Stack from ghstack (oldest at bottom): -> #185461 Module-level stdlib logging helpers such as logging.info first check len(logging.root.handlers) before delegating to the root logger method. Dynamo tracked logging.root as LoggingLoggerVariable, but logging.root.handlers fell back to an untyped GetAttrVariable. That..."
188948,Fix triangular_solve for sparse CPU tensors on non-MKL platforms,https://github.com/pytorch/pytorch/pull/188948,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,153410,153410,153410,,,,0,2,2,,,"triaged, open source, module: inductor, ciflow/inductor",2026-07-04T14:25:28Z,2026-07-07T17:06:02Z,1,"Fixes #153410 This change provides a reference implementation for triangular_solve on sparse CPU tensors when MKL is not available, by materializing diagonal indices and falling back to the dense implementation. Additionally, it reverts an incorrect modification in test/inductor/test_compile_worker.py that attempted..."
185462,Fix NumPy ndarray operators under default device contexts,https://github.com/pytorch/pytorch/pull/185462,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"140884, 156162","140884, 156162",,,,0,3,2,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-28T03:51:41Z,2026-07-07T17:01:40Z,1,"Stack from ghstack (oldest at bottom): -> #185462 Dynamo models NumPy arrays as NumpyNdarrayVariable, which subclasses TensorVariable for implementation sharing. Builtin operator handling checked active torch_function modes before it reached the NumPy operator wrapper. A torch.device(...) context installs a DeviceCo..."
185466,[dynamo] Show carets in graph break user stacks,https://github.com/pytorch/pytorch/pull/185466,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,156127,156127,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-28T04:14:46Z,2026-07-07T16:57:14Z,1,"Stack from ghstack (oldest at bottom): -> #185466 Dynamo captured column metadata on its synthetic FrameSummary objects, but several graph-break paths formatted those stacks with Python's traceback.format_list(). That formatter is inconsistent for synthesized frames: on Python 3.12 it can omit marker lines entirely..."
185467,Keep tensors alive for Dynamo pointer reconstruction,https://github.com/pytorch/pytorch/pull/185467,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,156059,156059,,,,0,3,1,6cc47cbe77184af05ec80e3dc42b40bbb84dceb2509d1a371370873c820fb745,2baf3caa1390ddc42afbfbc2ab3597b571bef6393483da347a4cba565e32368b,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-28T04:40:15Z,2026-07-07T16:53:45Z,1,"Stack from ghstack (oldest at bottom): -> #185467 When Dynamo graph-breaks after creating a tensor in the captured graph, the resume bytecode may need to reconstruct raw pointer-like values such as a.untyped_storage()._cdata. Before this change, _cdata was represented as a generic GetAttrVariable, so the resumed byt..."
185471,Handle Dynamo tracing of torch._C._infer_size,https://github.com/pytorch/pytorch/pull/185471,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,155800,155800,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-28T05:49:39Z,2026-07-07T16:50:04Z,1,"Stack from ghstack (oldest at bottom): -> #185471 Dynamo treats torch._C._infer_size as an in-graph callable. When the arguments are torch.Size values, Dynamo represents the SizeVariable contents as FX tuple-like aggregates, so fake execution and generated FX code end up calling the Python C binding with tuples. Tha..."
184083,Preserve user Triton allocator in Inductor,https://github.com/pytorch/pytorch/pull/184083,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,155584,155584,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T08:42:18Z,2026-07-07T16:45:45Z,1,"Stack from ghstack (oldest at bottom): -> #184083 Inductor now installs its fallback Triton allocator only when Triton is still using its null allocator or one that Inductor installed earlier, so user-provided allocators remain active during Inductor Triton launches. Add focused tests for preserving custom allocator..."
189028,[HOP] Moved branch clamping to user surface for switch,https://github.com/pytorch/pytorch/pull/189028,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-07-06T14:42:34Z,2026-07-07T16:41:56Z,1,This PR moves the clamping of the branch index into the user-facing switch(...) call. This has be advantage that it can be handled in a central place and no plumbing into the dense implementation or Inductor is required. cc @ydwu4
185474,Fix nested tensor tangent metadata coercion,https://github.com/pytorch/pytorch/pull/185474,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,155421,155421,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-28T06:45:08Z,2026-07-07T16:41:41Z,1,Stack from ghstack (oldest at bottom): -> #185474 AOTAutograd normalizes traced tensor subclass tangents with coerce_tangent_metadata before collecting tangent metadata. Runtime backward tangents were not using the same normalization before comparing against the traced metadata. For jagged NestedTensor outputs from...
185483,Add fake support for pad_packed_sequence,https://github.com/pytorch/pytorch/pull/185483,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,155238,155238,,,,0,3,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-05-28T08:13:46Z,2026-07-07T16:36:25Z,1,"Stack from ghstack (oldest at bottom): -> #185483 aten::_pad_packed_sequence was registered as CompositeImplicitAutograd even though its native kernel reads batch_sizes through data_ptr, writes the lengths output, and copies real tensor data. Fake tensor dispatch tried to run that native kernel while Dynamo was trac..."
185207,[Test] Make dynamo/test_aot_autograd_cache.py device-agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/185207,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,2,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-05-26T09:46:51Z,2026-07-07T16:34:21Z,1,"Summary Replace hardcoded CUDA references with accelerator-agnostic APIs (torch.accelerator) to enable AOT autograd cache tests to run on any accelerator backend with Triton support, rather than being restricted to CUDA/XPU only. Changes: Add ACC_TYPE and HAS_ACCELERATOR using torch.accelerator APIs Replace @require..."
184016,Fix Inductor DCE for fallback alias mutations,https://github.com/pytorch/pytorch/pull/184016,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,166009,166009,,,,6,3,4,fdbabc8a1f0b1c1e63bb2f61e4d105472d2011856f47407e9abf121b9742f47a,749ebf86ef3ff8c6fb7c825dd15191a1571bd9836ef9c848b15dd9369354d992,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-16T12:33:47Z,2026-07-07T16:32:58Z,1,Stack from ghstack (oldest at bottom): -> #184016 Preserve real mutations into fallback alias outputs through lowering and scheduler DCE so returned base buffers observe the copy. Fixes #166009 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenz...
184086,Fix Inductor scaled_mm scale_result lowering,https://github.com/pytorch/pytorch/pull/184086,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,154807,154807,,,,0,3,3,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, agentic, needs-review",2026-05-17T10:11:18Z,2026-07-07T16:30:06Z,1,Stack from ghstack (oldest at bottom): -> #184086 Fixes #154807 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @mlazos
185499,Fix unbacked memo reuse during decomposition retrace,https://github.com/pytorch/pytorch/pull/185499,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,154647,154647,,,,0,2,3,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-05-28T11:04:27Z,2026-07-07T16:25:41Z,1,"Stack from ghstack (oldest at bottom): -> #185499 When run_decompositions retraces an exported graph, PropagateUnbackedSymInts expects every old unbacked binding site to allocate a fresh symbol so rebind_unbacked can rename the fresh symbol back to the original one. For boolean indexing with the same mask twice, the..."
184090,Preserve benchmark integer inputs for Inductor debug,https://github.com/pytorch/pytorch/pull/184090,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,154592,154592,,,,0,2,3,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T11:54:06Z,2026-07-07T16:12:27Z,1,"Stack from ghstack (oldest at bottom): -> #184090 Add an opt-in benchmark harness mode that preserves non-floating input values for generated Inductor repros, while keeping default metadata-only inputs to avoid persisting user data. The mode bypasses FX graph cache, skips overlapping tensors, and compacts serialized..."
185504,Support unbacked SymInt closures in cond export,https://github.com/pytorch/pytorch/pull/185504,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,154559,154559,,,,0,5,1,9fa097d71d7588435f5e2b7025168482072d1d74e08a70a45f6221d407aa316c,21ebc40bfb3a5e522c4eabebfcd17bf2269e4570661a202457736072b03a8a80,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-28T11:44:54Z,2026-07-07T16:09:25Z,1,"Stack from ghstack (oldest at bottom): -> #185504 Dynamo already handled hinted SymInt and SymBool values captured by higher-order-op branch functions, but unhinted SymInt values still forced an explicit graph break. If the captured symbol was then used as a tensor size inside a torch.cond branch, the replacement sy..."
185508,[dynamo] Preserve observed exception stack for reraises,https://github.com/pytorch/pytorch/pull/185508,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,154454,154454,,,,0,4,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-28T12:13:31Z,2026-07-07T16:02:55Z,1,"Stack from ghstack (oldest at bottom): -> #185508 Observed exceptions created by raise_observed_exception did not remember the user stack from the first observed raise site. When that same exception variable was handled and later reraised with a bare raise, Dynamo attributed the graph break to the reraising bytecode..."
184091,Handle cudagraph liveness changes during replay,https://github.com/pytorch/pytorch/pull/184091,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,154306,154306,,,,0,3,3,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T12:18:31Z,2026-07-07T15:58:50Z,1,Stack from ghstack (oldest at bottom): -> #184091 Non-reentrant checkpoint recomputation can keep a parent cudagraph output live longer than it was during graph recording. Treat after-graph liveness mismatches as a re-recordable invariant failure and restore cleared managed inputs before recording the new path. Fixe...
184092,Skip tiny cudagraph partitions in max-autotune,https://github.com/pytorch/pytorch/pull/184092,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,154301,154301,,,,3,11,5,628424792dbb97f6b2f36a92c18198b4d753d2e4fa905f2d80ab8f54166915b4,9ca1556b2046ee8db753483f6499d6566d7d15c77daa106a88a8c4d1dc2e6fc6,"Merged, Reverted, ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ci-no-td, ciflow/torchtitan, agentic, needs-review",2026-05-17T12:34:09Z,2026-07-07T15:54:29Z,1,Stack from ghstack (oldest at bottom): -> #184092 Use the existing cudagraph partition-size threshold for max-autotune so single-kernel pointwise graphs avoid replay input-copy overhead while still enabling cudagraphs for larger partitions. Fixes #154301 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @...
185528,Fix Dynamo tracing of compile support checks,https://github.com/pytorch/pytorch/pull/185528,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,154259,154259,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-28T15:22:28Z,2026-07-07T15:50:42Z,1,"Stack from ghstack (oldest at bottom): -> #185528 Newer torchvision roi_align calls torch._dynamo.utils.is_compile_supported(input.device.type) to choose between the Python ROIAlign path and the custom op path. Since torch._dynamo is skiplisted, Dynamo treated this helper as skipped and inserted graph breaks when vi..."
185536,Teach Dynamo PGO about sparse tensor strides,https://github.com/pytorch/pytorch/pull/185536,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,154161,154161,,,,0,2,1,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-28T16:26:32Z,2026-07-07T15:46:33Z,1,"Stack from ghstack (oldest at bottom): -> #185536 Issue #154161 exposed a PGO invariant mismatch. record_automatic_dynamic stored tensor sizes and strides for automatic dynamic shape decisions, but sparse tensors have logical sizes without strided-layout per-dimension strides. The previous sparse path encoded that a..."
185545,Handle symbolic bounds in slice decomposition,https://github.com/pytorch/pytorch/pull/185545,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,154137,154137,,,,0,1,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-28T17:27:48Z,2026-07-07T15:43:08Z,1,"Stack from ghstack (oldest at bottom): -> #185545 slice_forward normalized start and end with Python comparisons against the input dimension size. When that size is an unbacked SymInt, as in Detectron2 ImageList copying into a symbolic slice during export, comparisons such as end > size require a data-dependent guar..."
185564,Handle pre-dispatch export of subclass tensor constants,https://github.com/pytorch/pytorch/pull/185564,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,153387,153387,,,,0,6,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-28T22:11:09Z,2026-07-07T15:33:38Z,1,"Stack from ghstack (oldest at bottom): -> #185564 Pre-dispatch non-strict export could fail on a traceable wrapper tensor subclass stored as a normal module attribute. ProxyTensorMode rejected the untracked wrapper subclass before FX had a chance to lift it as a constant, so the dispatch stack propagated NotImplemen..."
185568,Fix export of arange with data-dependent float step,https://github.com/pytorch/pytorch/pull/185568,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,153247,153247,,,,0,3,1,,,"module: cpu, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-28T23:40:00Z,2026-07-07T15:30:11Z,1,"Stack from ghstack (oldest at bottom): -> #185568 torch.arange can receive a step derived from a captured scalar, such as 1 / x.sum() or 1 / mask.sum().item(). During export/fake execution this creates unbacked symbolic scalar values. The reference arange implementation was using Python control flow on the step sign..."
187956,Propagate FX subgraph replacement meta values,https://github.com/pytorch/pytorch/pull/187956,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,187945,187945,,,,0,3,1,,,"release notes: fx, agentic, needs-review",2026-06-23T16:37:51Z,2026-07-07T15:27:38Z,1,"Stack from ghstack (oldest at bottom): -> #187956 replace_pattern and replace_pattern_with_filters copy replacement graph nodes into the target graph with Graph.graph_copy. Replacement graphs created by symbolic_trace do not have meta[""val""], so newly inserted nodes lost the tensor metadata that export graphs use fo..."
185569,Handle symbolic bool short-circuit with unbacked terms,https://github.com/pytorch/pytorch/pull/185569,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,153227,153227,,,,0,3,1,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-05-29T00:12:28Z,2026-07-07T15:22:19Z,1,"Stack from ghstack (oldest at bottom): -> #185569 When evaluating a boolean expression with unbacked symbols, ShapeEnv first tries static simplification and unbacked elimination before deciding whether to raise a data-dependent error or return a guard_or fallback. For a compound boolean like (s0 > 3) | (u0 > 200), t..."
185588,[inductor] Codegen backed SymInt input runtime asserts,https://github.com/pytorch/pytorch/pull/185588,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,153175,153175,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-29T03:21:28Z,2026-07-07T15:19:08Z,1,"Stack from ghstack (oldest at bottom): -> #185588 When prefer_deferred_runtime_asserts_over_guards=True, Dynamo records backed-SymInt shape checks as deferred runtime asserts instead of guards. FX materializes those backed-only checks as input-graph aten._assert_scalar nodes, but regular Inductor lowering dropped th..."
185591,Canonicalize avg_pool2d singleton lists during export,https://github.com/pytorch/pytorch/pull/185591,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,153149,153149,,,,0,1,1,,,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-05-29T04:04:37Z,2026-07-07T15:14:17Z,1,"Stack from ghstack (oldest at bottom): -> #185591 aten.avg_pool2d accepts int[2] arguments for kernel_size, stride, and padding, and eager execution permits a single-element list or tuple by expanding it internally. torch.export records the proxy node before the native/meta implementation performs that expansion, so..."
185604,Improve infer_size unbacked broadcast error,https://github.com/pytorch/pytorch/pull/185604,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,152656,152656,,,,0,4,1,,,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-29T07:04:31Z,2026-07-07T15:11:46Z,1,"Stack from ghstack (oldest at bottom): -> #185604 fake_impls.infer_size first checked singleton broadcasting with guard_or_false and then deferred sizeA == sizeB. When both singleton checks were data-dependent unbacked expressions, guard_or_false collapsed both to False and the runtime assert only said Eq(u0, u1) fa..."
185690,Optimize AOT input view mutations,https://github.com/pytorch/pytorch/pull/185690,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-pallas",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,152346,152346,,,,0,3,3,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/inductor-pallas, agentic, needs-review",2026-05-30T00:57:40Z,2026-07-07T15:06:00Z,1,"Stack from ghstack (oldest at bottom): -> #185690 Fixes #152346 AOTAutograd functionalization represents input view mutations by building an updated full input and then copying that full tensor back into the original input. For simple slice mutations like x[0] += 1.0, Python's augmented assignment adds a STORE_SUBSC..."
185624,Fix subclass static input remapping for SymInt args,https://github.com/pytorch/pytorch/pull/185624,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,152343,152343,,,,0,1,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-05-29T12:38:30Z,2026-07-07T15:02:30Z,1,"Stack from ghstack (oldest at bottom): -> #185624 AOTAutograd remapped a static wrapper subclass input by counting every unwrapped component of the subclass. With dynamic shapes, subclass unwrapping also appends symbolic size/stride placeholders, so those SymInt graph args were included in fw_metadata.static_input_i..."
185632,Make AOTAutograd metadata repr subclass-safe,https://github.com/pytorch/pytorch/pull/185632,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,152183,152183,,,,0,1,1,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-05-29T14:18:34Z,2026-07-07T14:59:34Z,1,Stack from ghstack (oldest at bottom): -> #185632 ViewAndMutationMeta and related subclass metadata are included in AOTAutograd structured logging for tlparse. These dataclasses can hold tensor subclasses in traced_tangents and SubclassCreationMeta.original_subclass. The generated dataclass repr recursively calls th...
184109,Fix Inductor replacement patterns across view-to-reshape,https://github.com/pytorch/pytorch/pull/184109,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,151649,151649,,,,0,2,5,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T18:13:50Z,2026-07-07T14:52:56Z,1,"Stack from ghstack (oldest at bottom): -> #184109 Allow register_replacement patterns traced through view to match post-grad graphs rewritten to reshape, while preserving exact matching for manual patterns. Fixes #151649 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSupe..."
185718,Fix fake functorch-wrapped tensor handling in jacfwd vmap,https://github.com/pytorch/pytorch/pull/185718,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,"151196, 172428","151196, 172428",,,,0,1,2,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-05-30T18:48:11Z,2026-07-07T14:44:05Z,1,"Stack from ghstack (oldest at bottom): -> #185718 Dynamo fake propagation for vmap(jacfwd(...)) could see GradTrackingTensor/BatchedTensor wrappers around FakeTensor inputs as non-fake. That caused FakeTensor validation and common-device logic to miss the inner fake tensor; after validation was fixed, forward AD Zer..."
188114,Fix CK SDPA flash attention on gfx1200/gfx1201 (RDNA4),https://github.com/pytorch/pytorch/pull/188114,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/docker, ciflow/h100, ciflow/inductor",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,188113,188113,188113,"144777, 155103, 187267",,,0,2,10,,,"module: rocm, module: cpu, open source, module: amp (automated mixed precision), ciflow/trunk, release notes: quantization, release notes: rocm, topic: not user facing, ciflow/mps, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/vllm, ciflow/b200, ciflow/inductor-pallas, ci-refresh-hf-cache, ciflow/torchtitan, ciflow/docker",2026-06-24T23:36:50Z,2026-07-07T14:35:08Z,1,"The CK Tile FMHA generator already compiles working kernels for gfx12 (--targets=gfx1201 passes through to the kernel template), but two guards prevented them from being used on RDNA4 hardware: launch_kernel_pt.hpp: kentry_pt kernel entry stubs added in PR #144777 ([ROCm] CK SDPA - Move arch check to CK patch) had a..."
188237,Optimize CSR SpMM CPU grain size and inner accumulation,https://github.com/pytorch/pytorch/pull/188237,PR blocked,108,failed checks or blocked/unstable merge state; approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,BLOCKED,False,0,0,,,,,,,,,0,1,1,,,"triaged, open source, release notes: sparse",2026-06-26T02:12:02Z,2026-07-07T14:34:05Z,1,CSR SpMM does substantial work per sparse row when each row has many nonzeros and the dense RHS has a large column dimension. Account for that work when choosing the row parallel grain size so rows are split more aggressively for these cases. Also inline the per-nonzero dense-row accumulation instead of issuing one...
176404,Add prioritized text linker optimization for lld.,https://github.com/pytorch/pytorch/pull/176404,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/linux-aarch64, ciflow/trunk",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, ciflow/trunk, topic: not user facing, ciflow/linux-aarch64",2026-03-04T06:57:32Z,2026-07-07T14:18:47Z,1,"During upgrade to PyTorch 2.10, we observed that on aarch64, when USE_PRIORITIZED_TEXT_FOR_LD is enabled, the build finishes successfully but importing torch (e.g. python3 -c ""import torch"" or directly dlopen-ing libtorch_global_deps.so) consistently crashes with SIGSEGV, while the same setup on x86 works fine. The..."
187164,[ROCm] unskip export tests on ROCm,https://github.com/pytorch/pytorch/pull/187164,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/rocm-mi200, ciflow/rocm-mi300, ciflow/rocm-mi355","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: rocm, triaged, open source, topic: not user facing, ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/rocm-mi200",2026-06-12T09:37:03Z,2026-07-07T14:16:38Z,1,Working locally testing CI cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
185755,Use gmem efficient attention kernel for long fp32 KV,https://github.com/pytorch/pytorch/pull/185755,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,149857,149857,,,,0,2,1,,,"topic: not user facing, agentic, needs-review",2026-05-31T11:29:21Z,2026-07-07T14:12:52Z,1,"Stack from ghstack (oldest at bottom): -> #185755 The sm80 fp32 memory-efficient attention forward registry launches the first compatible CUTLASS kernel. For long-key cross-attention it selected the RF 64x64/64x128 kernels, which avoid the output accumulator but require more QK key-tile iterations. When TF32 makes t..."
185722,[dynamo] Wrap module.compile leaf module call frames,https://github.com/pytorch/pytorch/pull/185722,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,150915,150915,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-30T19:22:44Z,2026-07-07T14:10:05Z,1,"Stack from ghstack (oldest at bottom): -> #185722 module.compile() installs torch.compile(self._call_impl). For built-in leaf modules such as Conv2d, both Module._call_impl and the module forward live in skip-rule files. The torch.compile(module) path already handles this by wrapping modules whose forward has a skip..."
184119,Avoid fp8 bf16 prologue fusion on pre-sm90,https://github.com/pytorch/pytorch/pull/184119,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,8,1,"statusCheckRollup, linux-jammy-cuda13.0-py3.10-gcc11-sm100 / test (smoke_b200, 1, 1, linux.dgx.b200), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), linux-jammy-cuda13.0-py3.10-gcc11-sm90 / test-osdc (smoke, 1, 1, mt-l-x86iamx-22-225-h100), get-label-type / runner-determinator",statusCheckRollup,,150621,150621,,,,4,6,11,abfa793255fa37377f38ba5bf0826a025c2a936b2d38e4d79d070c48823af9f7,83ddf047fecbea06944cad59e5c7cb523609cdedfd73f55246c4b9ea7ba4fe2c,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, agentic, needs-review",2026-05-17T20:38:06Z,2026-07-07T14:03:15Z,1,"Stack from ghstack (oldest at bottom): -> #184119 Block standard Triton mm prologue fusion when a pre-sm90 CUDA target would fuse a float8 read into a bf16 dot operand. The standalone fp8-to-bf16 pointwise conversion path lowers the fp8 load through fp32 and stores bf16, so it does not create the same invalid instru..."
185724,Fix scalar searchsorted decomposition for unbacked values,https://github.com/pytorch/pytorch/pull/185724,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,150613,150613,,,,0,4,1,3f8b0beb74a306f202d736f38505bf5e507c715b5cf222920eb0b28009cc266c,1c318c3cd973baac18351239cbe3d9ff43e30d6130dc39d1572702758e58292b,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-30T20:03:57Z,2026-07-07T13:58:50Z,1,"Stack from ghstack (oldest at bottom): -> #185724 The scalar Inductor decompositions for searchsorted and bucketize converted the scalar argument with torch.tensor([self], device=...) and then extracted the single result. When capture_scalar_outputs traces a value from .item(), that scalar can be an unbacked SymInt..."
184593,[Test] Make test_autograd.py device-agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/184593,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing, module: tests/core",2026-05-20T19:15:47Z,2026-07-07T13:56:39Z,1,"Replace hardcoded CUDA references with device-agnostic equivalents so out-of-tree accelerator backends can run these tests. Replace @onlyCUDA with @onlyAccelerator for tests with no CUDA-specific logic Replace TEST_CUDA / torch.cuda.is_available() with TEST_ACCELERATOR Replace device=""cuda"" / .cuda() with torch.acce..."
189097,[RFC] Expose Kineto custom logger registration,https://github.com/pytorch/pytorch/pull/189097,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,3,0,,,open source,2026-07-06T21:35:31Z,2026-07-07T13:56:30Z,1,Working draft RFC for changes made to support Pytorch users registering their own custom loggers and exposing a new public method for explicitly defining which logger kineto should utilize when exporting.
185726,Improve Dynamo stack for nonstrict_trace input errors,https://github.com/pytorch/pytorch/pull/185726,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,150465,150465,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-30T20:29:39Z,2026-07-07T13:55:55Z,1,"Stack from ghstack (oldest at bottom): -> #185726 Directly compiling a function decorated with torch._dynamo.nonstrict_trace wraps the callable with torch._dynamo.external_utils.wrap_inline so Dynamo has an inlineable Python frame to trace. When nonstrict_trace rejected an unsupported input type, the Unsupported exc..."
188951,Add ROCm Docker release images,https://github.com/pytorch/pytorch/pull/188951,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,188319,,,,0,2,1,,,"module: rocm, triaged, open source, topic: not user facing",2026-07-04T16:16:26Z,2026-07-07T13:51:06Z,1,"Implements the changes proposed in #188319. Changes generate_binary_build_matrix.py: add ROCM_ARCHES_FULL_VERSION mapping (7.1→7.1.1, 7.2→7.2.3) generate_docker_release_matrix.py: add ROCm matrix rows alongside existing CUDA rows docker-release.yml: pass ROCM_VERSION and ROCM_VERSION_SHORT env vars to build job dock..."
185731,Clear stale prefix exception table entries in Dynamo,https://github.com/pytorch/pytorch/pull/185731,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,150319,150319,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-30T21:46:54Z,2026-07-07T13:50:06Z,1,"Stack from ghstack (oldest at bottom): -> #185731 Python 3.12 coroutine bytecode can attach an exception table entry to the initial RESUME prefix instruction. Dynamo copies prefix instructions into the generated output bytecode when compiling a subgraph, but the generated output owns a new bytecode body and does not..."
185732,Fix nested compile during subclass fake tracing,https://github.com/pytorch/pytorch/pull/185732,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,150262,150262,,,,0,2,1,,,"ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic",2026-05-30T22:18:57Z,2026-07-07T13:44:45Z,1,"Stack from ghstack (oldest at bottom): -> #185732 Compiler-internal fake propagation and AOTAutograd tracing can execute tensor subclass hooks while FakeTensorMode or functionalization modes are active. If that subclass hook calls a torch.compile wrapper, Dynamo can re-enter compilation using those compiler-owned fa..."
184742,[Test] Make test_expanded_weights.py device-agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/184742,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: tests/core",2026-05-21T16:09:11Z,2026-07-07T13:41:45Z,1,"Replace hardcoded CUDA references with device-agnostic equivalents so out-of-tree accelerator backends can run these tests. test/test_expanded_weights.py Replace TEST_CUDA with TEST_ACCELERATOR in module test generation loop Replace hardcoded ""cuda"" device with torch.accelerator.current_accelerator().type Dynamic te..."
185734,Preserve torch._check messages in runtime asserts,https://github.com/pytorch/pytorch/pull/185734,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,4,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,150063,150063,,,,0,8,1,f8df2e62f6aa0b9b539cb92800dd7d8e0261b6d2ba92aca578ceddce5cc76c07,2486533cefe291d46ad79abdc565627527b81f9abacf86953daeb5927bc2f844,"release notes: jit, ciflow/inductor, agentic, needs-review",2026-05-30T23:49:22Z,2026-07-07T13:40:16Z,1,"Stack from ghstack (oldest at bottom): -> #185734 torch._check already accepts lazy message callables, but the symbolic expect_true path discarded them before creating deferred runtime asserts. As a result, export and compile graphs emitted only the generated symbolic expression text for aten._assert_scalar, which h..."
185740,Support inspect.signature on nested functions in Dynamo,https://github.com/pytorch/pytorch/pull/185740,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149968,149968,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-31T03:39:14Z,2026-07-07T13:32:17Z,1,Stack from ghstack (oldest at bottom): -> #185740 Dynamo previously fell back to tracing inspect.signature when the target was a nested function created in the frame being compiled. That path tried to execute inspect.py against partially modeled FunctionType state and could not construct signatures for nested closur...
185742,Support None leaves in nonstrict_trace,https://github.com/pytorch/pytorch/pull/185742,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149963,149963,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-31T03:54:40Z,2026-07-07T13:28:04Z,1,"Stack from ghstack (oldest at bottom): -> #185742 nonstrict_trace lowers calls through flat_apply by flattening the user args into FX node arguments. flat_apply was using torch.fx.node.base_types as the full definition of graphable leaves, but FX Node.Argument is Optional[...] and can legally carry None. As a result..."
185730,Fix dynamic shapes for variadic kwargs,https://github.com/pytorch/pytorch/pull/185730,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"150022, 150371","150022, 150371",,,,0,3,5,,,"release notes: onnx, release notes: export, agentic",2026-05-30T21:30:01Z,2026-07-07T13:27:38Z,1,"Stack from ghstack (oldest at bottom): -> #185730 torch.export validated dynamic_shapes against the structure produced by inspect.Signature.bind. For a module with **kwargs, that means all actual caller keyword inputs are hidden under the variadic parameter name, e.g. kwargs. A dict spec keyed by the real caller inp..."
188783,[xpu][fix]Fix bmm_outer_product override to support XPU tensors,https://github.com/pytorch/pytorch/pull/188783,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188773,188773,188773,,,,0,4,2,,,"triaged, open source, ciflow/trunk, topic: not user facing, merging, ciflow/xpu",2026-07-02T07:10:16Z,2026-07-07T13:27:33Z,1,"Fixes #188773 The _bmm_outer_product_cond function only checked a.is_cuda and b.is_cuda, which returns False for XPU tensors. This caused the triton override to never fire on XPU, making test_cow_inputs_accepted_by_override fail. Changes: Add _is_gpu_tensor() helper that checks both is_cuda and is_xpu Use it in _bmm..."
189043,Preload full bundled cuDNN set with RTLD_GLOBAL to prevent sublibrary version mismatch,https://github.com/pytorch/pytorch/pull/189043,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/binaries,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188892,188892,188892,,,,0,1,2,,,ciflow/binaries,2026-07-06T16:38:23Z,2026-07-07T13:27:21Z,1,"Fixes #188892 Problem cuDNN 9 is split into a dispatcher (libcudnn.so) plus a set of engine sub-libraries (libcudnn_graph.so, libcudnn_engines_*.so, libcudnn_ops.so, ...) that libcudnn/libcudnn_graph dlopen by soname at runtime. torch._C._preload_cuda_deps (in torch/__init__.py) only preloaded libcudnn.so, and with..."
186039,Fix Dynamo verbose guard checks for frame locals,https://github.com/pytorch/pytorch/pull/186039,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,CLEAN,False,0,0,,,,143308,143308,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic",2026-06-03T04:11:01Z,2026-07-07T13:27:02Z,1,"Stack from ghstack (oldest at bottom): -> #186039 RootGuardManager runtime checks already evaluate frame-local guards through FrameLocalsMapping, but verbose guard diagnostics only accepted PyObject locals. Recompilation reason analysis therefore passed frame.f_locals as a dict, so FrameLocalsGuardAccessor read valu..."
187452,Polyhedral Optimization Pass for PyTorch Inductor (Phase 1),https://github.com/pytorch/pytorch/pull/187452,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,187451,187451,187451,,,,0,13,2,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-16T16:25:56Z,2026-07-07T13:19:20Z,1,Fixes #187451 An example implementation from the RFC on dev-discuss. The goal of this PR is to demonstrate that significant speedups are achievable by applying principles of polyhedral optimization without the high overhead cost traditionally associated with it. We reused existing infrastructure when possible. Resul...
185902,Enable torch.compile for RNN modules by default,https://github.com/pytorch/pytorch/pull/185902,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149909,149909,,,,0,3,3,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-02T07:09:50Z,2026-07-07T13:19:06Z,1,"Stack from ghstack (oldest at bottom): -> #185902 Dynamo previously graph-broke on RNN, GRU, and LSTM modules unless torch._dynamo.config.allow_rnn was set manually. That made torch.compile(..., fullgraph=True) fail immediately for these modules, which is the root cause of #149909. Turn the RNN tracing path on by de..."
185752,Make OptimizedModule state_dict use public keys,https://github.com/pytorch/pytorch/pull/185752,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149894,149894,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-31T10:14:35Z,2026-07-07T13:14:15Z,1,"Stack from ghstack (oldest at bottom): -> #185752 torch.compile(module) returns an OptimizedModule wrapper that registers the original module as _orig_mod. Because the wrapper previously inherited the default nn.Module.state_dict and load_state_dict traversal, checkpoint keys exposed that implementation detail as _o..."
185759,Improve float32 flex attention backward tiling,https://github.com/pytorch/pytorch/pull/185759,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,149767,149767,,,,0,2,6,a0f1b3faa33bbae39b26509bc9d4f2b55b2f60a670d445c6637439012d345c93,2209e283aefaa200433e54e737b2d95a9bd139aded13f2e9fb62b2e5ef85c613,"topic: not user facing, module: inductor, ciflow/inductor, agentic, needs-review",2026-05-31T13:04:20Z,2026-07-07T13:11:26Z,1,"Stack from ghstack (oldest at bottom): -> #185759 CUDA float32 flex_attention backward used the conservative 16x16 fallback tile whenever Inductor did not run an expanded autotune search. For long sequence lengths with moderate head dimensions, including the issue's S=128, D=64 shape family, that produces many small..."
185762,[Dynamo] Support _AutoDispatchBelowAutograd context manager,https://github.com/pytorch/pytorch/pull/185762,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149586,149586,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-31T14:11:37Z,2026-07-07T13:07:26Z,1,"Stack from ghstack (oldest at bottom): -> #185762 Dynamo recognizes supported torch context managers through TorchCtxManagerClassVariable. torch._C._AutoDispatchBelowAutograd is a pybind context manager whose constructor enters below-autograd dispatcher state, but it was not registered there. As a result, tracing wi..."
185763,Improve disable API graph break messages,https://github.com/pytorch/pytorch/pull/185763,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149566,149566,,,,0,2,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-05-31T17:13:55Z,2026-07-07T13:03:37Z,1,"Stack from ghstack (oldest at bottom): -> #185763 Direct calls to torch._dynamo.disable() and torch.compiler.disable() inside a compiled region are public user API calls, but Dynamo represents them as SkipFunctionVariable instances because their implementation files are skipped. That made them fall through the gener..."
185766,Dynamo: support torch.jit.isinstance without graph breaks,https://github.com/pytorch/pytorch/pull/185766,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149556,149556,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-31T18:16:42Z,2026-07-07T13:00:43Z,1,"Stack from ghstack (oldest at bottom): -> #185766 Dynamo classified torch.jit.isinstance as a skipped function. Because torch.jit is also skiplisted, tracing code that used this TorchScript compatibility helper graph-broke before Dynamo could use its existing isinstance handling. Route torch.jit.isinstance through T..."
185768,Improve Dynamo error for dynamic module hook registration,https://github.com/pytorch/pytorch/pull/185768,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149047,149047,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-31T19:55:15Z,2026-07-07T12:56:11Z,1,"Stack from ghstack (oldest at bottom): -> #185768 Dynamic nn.Module hook registration during Dynamo tracing is not modeled. The registration methods mutate module hook dictionaries, create a RemovableHandle, update RemovableHandle.next_id, and use weakrefs. Existing Dynamo hook side-effect handling is for tensor hoo..."
185769,Preserve generator return values in Dynamo,https://github.com/pytorch/pytorch/pull/185769,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149037,149037,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-05-31T20:12:14Z,2026-07-07T12:46:39Z,1,"Stack from ghstack (oldest at bottom): -> #185769 Dynamo inlines generators and models generator exhaustion with StopIteration. The SEND and YIELD_FROM handlers assumed StopIteration.value was always a raw Python value and unconditionally wrapped it with ConstantVariable.create. For inlined generator returns, the va..."
185771,Support Dynamo tracing through textwrap.indent string formatting,https://github.com/pytorch/pytorch/pull/185771,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,149010,149010,,,,0,3,1,e01d30c64f8eb887b2154275a4c5cb5d6356ec148ec804b63f8bc3786af78441,7ac38334450f104986a78a3e75f50e168d40c7b4a325ef6805f1602c6c107633,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-31T21:12:55Z,2026-07-07T12:44:02Z,1,"Stack from ghstack (oldest at bottom): -> #185771 Dynamo represents str.format() calls involving symbolic shape values as a StringFormatVariable so the final string can be reconstructed after tracing. textwrap.indent() is otherwise inlineable, but it immediately calls string operations on that delayed formatted valu..."
186469,Wrap Dynamo-generated bytecode runtime errors,https://github.com/pytorch/pytorch/pull/186469,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,123374,123374,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T14:21:33Z,2026-07-07T12:40:35Z,1,"Stack from ghstack (oldest at bottom): -> #186469 Dynamo installs generated replacement code objects that inherit the user's filename and line table. If a bug in that generated no-graph-break bytecode raises at runtime, the exception escapes after compile-time error handling has finished and looks like a user except..."
185772,Fix CPU small-output bilinear upsample precision in Inductor,https://github.com/pytorch/pytorch/pull/185772,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,148859,148859,,,,0,3,1,,,"topic: not user facing, module: inductor, ciflow/inductor, agentic, needs-review",2026-05-31T22:18:11Z,2026-07-07T12:38:12Z,1,"Stack from ghstack (oldest at bottom): -> #185772 Inductor decomposes bilinear upsample through _upsample_linear. For CPU small-output 2D bilinear upsample, eager dispatches through the native channels-last kernel, which computes corner weights first and accumulates the weighted corner products. The decomposition in..."
187650,[Pytorch Test] Add skip_if_lt_x_devices and migrate FSDP autograd tests to be device-agnostic,https://github.com/pytorch/pytorch/pull/187650,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, release notes: distributed (fsdp)",2026-06-18T08:58:29Z,2026-07-07T12:35:34Z,1,Summary This PR introduces a generic @skip_if_lt_x_devices decorator in torch/testing/_internal/common_distributed.py to replace the hard-coded CUDA/HPU/XPU checks in @skip_if_lt_x_gpu. It also migrates test/distributed/_composable/fsdp/test_fully_shard_autograd.py to use the new decorator. This is a step toward mak...
189143,[Test] Make test_native_mha.py testcases device-agnostic,https://github.com/pytorch/pytorch/pull/189143,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing, module: tests/core",2026-07-07T11:17:21Z,2026-07-07T12:33:39Z,1,Stack from ghstack (oldest at bottom): -> #189143
185773,Use max dynamic bounds for Inductor autotuning,https://github.com/pytorch/pytorch/pull/185773,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,148842,148842,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-31T22:45:23Z,2026-07-07T12:33:23Z,1,"Stack from ghstack (oldest at bottom): -> #185773 Compile-time Inductor autotuning was resolving symbolic benchmark shapes with the general optimization-hint APIs. For backed dynamic symbols those APIs substitute the current concrete hint, which is the first observed input size during compilation. That made bounded..."
185774,Fix non-strict export for Tensor.numpy,https://github.com/pytorch/pytorch/pull/185774,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,148779,148779,,,,0,2,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-05-31T23:04:22Z,2026-07-07T12:30:33Z,1,"Stack from ghstack (oldest at bottom): -> #185774 Non-strict export runs user code with proxy/fake tensor subclasses under _NonStrictTorchFunctionHandler. Tensor.numpy() was intercepted there, but the handler fell through to TensorBase.numpy, which rejects tensor subclasses before make_fx can trace the rest of the c..."
185456,Fix Dynamo tracing for Tensor.random_,https://github.com/pytorch/pytorch/pull/185456,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,156614,156614,,,,0,7,1,822f3815846a107ca8563a989d8e725afe7e08b513a9c22b3b59d13c796ecba6,9209021fa02ad33ade1121bb9809ca6a3d5c626186c2baee12a60d699bcc30a1,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-28T01:55:06Z,2026-07-07T12:26:47Z,1,"Stack from ghstack (oldest at bottom): -> #185456 Tensor.random_ and uniform_(from=...) were forced to graph break in Dynamo because downstream FX code generation could produce invalid Python for ATen overload names and keyword arguments containing Python keywords. In particular, generated code could spell the rando..."
184130,Avoid torch imports in Inductor compile sidecar,https://github.com/pytorch/pytorch/pull/184130,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,148651,148651,,,,0,5,1,,,"topic: not user facing, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-17T22:48:38Z,2026-07-07T12:24:05Z,1,"Stack from ghstack (oldest at bottom): -> #184130 Move the SubprocPool worker protocol into a torch-free helper so the compile-worker sidecar can create its fork-based worker pool before any torch import starts native threads. Defer torch key, Triton path, and async-compile setup to the actual worker initializer.\n\..."
185781,Avoid unbound symbols in insignificant output strides,https://github.com/pytorch/pytorch/pull/185781,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,148199,148199,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T01:42:43Z,2026-07-07T12:20:40Z,1,"Stack from ghstack (oldest at bottom): -> #185781 With dynamic shapes, output metadata can carry stride symbols that do not belong to the realized Inductor layout. For qlinear_binary with an inplace sum and M == 1, try_match_insignificant_strides copied such a symbolic stride from FX output metadata onto the size-1..."
185408,Fix torch logging to avoid closed pytest stderr,https://github.com/pytorch/pytorch/pull/185408,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,159160,159160,,,,0,3,1,,,"topic: not user facing, agentic, needs-review",2026-05-27T17:12:41Z,2026-07-07T12:19:00Z,1,"Stack from ghstack (oldest at bottom): -> #185408 Dynamo registers a few atexit loggers for traced-frame summaries, compile-time summaries, and symbolic-shape cache stats. Pytest may replace sys.stderr while tests run and close that capture stream before these atexit callbacks fire. The torch logging setup installed..."
185526,Fix Dynamo tracing of LazyGraphModule,https://github.com/pytorch/pytorch/pull/185526,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,154282,154282,,,,0,4,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-28T14:53:55Z,2026-07-07T12:16:19Z,1,"Stack from ghstack (oldest at bottom): -> #185526 Fresh _LazyGraphModule instances leave forward pointing at _lazy_forward until the module is really recompiled. Dynamo resolves and inlines forward while tracing GraphModule calls, so a lazy graph module exposed the skipped torch.fx _lazy_forward stub before it could..."
185784,Fix flex attention captured vector gradient scatters,https://github.com/pytorch/pytorch/pull/185784,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,148112,148112,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T03:04:50Z,2026-07-07T12:07:52Z,1,"Stack from ghstack (oldest at bottom): -> #185784 FlexAttention backward lowers gradients for tensors captured by score_mod via zeros-and-scatter subgraphs. The Triton template path for these scatter subgraphs passed no loop variables to Scatter.store_output, which only worked for scalar captured gradients. A score_..."
185786,Handle non-natural SymPy powers in value range interp,https://github.com/pytorch/pytorch/pull/185786,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,"136628, 148003","136628, 148003",,,,0,2,2,,,"module: cpu, topic: not user facing, agentic, needs-review",2026-06-01T03:23:56Z,2026-07-07T12:03:15Z,1,"Stack from ghstack (oldest at bottom): -> #185786 The SymPy interpreter routed every sympy.Pow through pow_by_natural. That handler is only valid when the exponent is known to be a nonnegative integer, but SymPy also represents reciprocal expressions such as x**-1 as Pow. When Inductor's value range analysis saw a p..."
188195,Adding OpenReg allocator stats tests,https://github.com/pytorch/pytorch/pull/188195,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-06-25T21:01:50Z,2026-07-07T11:52:44Z,1,"Adds tests to verify Open Reg's DeviceMemoryAllocator stat tracking implementation (malloc/free counters, peak tracing, reset functions)"
183904,Fix CPU addcmul tensor-value decomposition numerics,https://github.com/pytorch/pytorch/pull/183904,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,176929,176929,,,,0,10,1,,,"ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), release notes: inductor, ciflow/torchtitan, ciflow/dtensor, agentic",2026-05-15T14:15:05Z,2026-07-07T11:39:48Z,1,"Stack from ghstack (oldest at bottom): -> #183904 Use a dedicated Inductor addcmul primitive so Dynamo's tensor-value addcmul_ decomposition can lower with the CPU FMA contraction order that matches eager, while preserving eager validation and autograd behavior for supported cases. Fixes #176929 Generated by my agen..."
189140,[Test] Update device type tests in TestVarlenAttention for device-generic,https://github.com/pytorch/pytorch/pull/189140,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing, module: tests/core",2026-07-07T09:41:45Z,2026-07-07T11:39:31Z,1,Stack from ghstack (oldest at bottom): -> #189140
185795,Preserve overlapping input layouts when cloning Dynamo inputs,https://github.com/pytorch/pytorch/pull/185795,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,147843,147843,,,,0,2,1,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-01T06:21:33Z,2026-07-07T11:28:13Z,1,"Stack from ghstack (oldest at bottom): -> #185795 clone_input tried to preserve strides by allocating an output view with the same sizes and strides, then copying x.clone() into it. Expanded and other internally-overlapping views make that destination invalid for copy_, so the helper fell back to plain torch.clone(x..."
185799,Avoid dataclass constructors in Dynamo and make_fx graphs,https://github.com/pytorch/pytorch/pull/185799,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,147530,147530,,,,0,4,1,98d07f810f8c4b3d66f1e3ca281119a5f56c66e82ce27e2f1733858276b54324,057f2b4e32facd46a794cadd4a06abd23ef0f7bd59f253f24402d7fe2d2d123d,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T08:12:18Z,2026-07-07T11:24:34Z,1,"Stack from ghstack (oldest at bottom): -> #185799 Dynamo and make_fx inherited the base FX dataclass lowering path from TracerBase.create_arg. That path materialized dataclass values as call_function(dataclass_type, ...) nodes, so arbitrary dataclass init or post_init user code could be represented in graphs that ar..."
185803,Fix joint export for tied parameter aliases,https://github.com/pytorch/pytorch/pull/185803,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,147380,147380,,,,0,2,1,,,"release notes: export, agentic, needs-review",2026-06-01T09:22:19Z,2026-07-07T11:21:52Z,1,"Stack from ghstack (oldest at bottom): -> #185803 Non-strict export can preserve multiple parameter input specs for the same tied Parameter. In the reported model, the graph uses the lm_head.weight placeholder for both embedding and linear while the input_embeds.weight placeholder is unused. When _export_forward_bac..."
185914,Log aot_export fw_metadata instead of printing it,https://github.com/pytorch/pytorch/pull/185914,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,147135,147135,,,,0,1,1,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-06-02T10:07:08Z,2026-07-07T11:13:26Z,1,"Stack from ghstack (oldest at bottom): -> #185914 The aot_export validation path rejects input mutations that require grad before the normal AOT structured metadata logging has a chance to run. That path embedded str(fw_metadata) directly in the RuntimeError, which can make the user-facing message extremely long whi..."
185809,Fix as_strided_copy storage bounds in Inductor,https://github.com/pytorch/pytorch/pull/185809,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,147077,147077,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T11:13:51Z,2026-07-07T11:09:45Z,1,"Stack from ghstack (oldest at bottom): -> #185809 Inductor lowered aten.as_strided_copy as an unchecked ReinterpretView followed by clone. That skips ATen's setStorage bounds check, so a compiled graph could read from storage that is too small for the requested size/stride/offset. In the issue reproducer, the clone..."
186522,Trace Tensor.requires_grad_ in pre-dispatch make_fx,https://github.com/pytorch/pytorch/pull/186522,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,113899,113899,,,,0,5,5,7565ad86b364d669df7e5329e885901a7d61bc1879557804b947d80ef507c0d6,66fcfc6f4d48eda2b32c844b9362fdfc68d7b08ece9c435229bbc32e42c1c223,"release notes: fx, ciflow/inductor, agentic, needs-review",2026-06-07T06:06:58Z,2026-07-07T11:04:20Z,1,"Stack from ghstack (oldest at bottom): -> #186522 Tensor.requires_grad_() is a Python Tensor method that does not lower through ProxyTorchDispatchMode as an aten dispatcher operator. During make_fx(..., pre_dispatch=True), PreDispatchTorchFunctionMode saw the method but did not preserve it in the FX graph, so the ca..."
185916,Fix FlexAttention nested index backward,https://github.com/pytorch/pytorch/pull/185916,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,146896,146896,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T10:26:46Z,2026-07-07T11:01:20Z,1,"Stack from ghstack (oldest at bottom): -> #185916 FlexAttention score_mod tracing lowers tensor indexing through ModIndex. For a pattern such as table[idx][0], the joint backward graph contains nested zeros_and_scatter calls: one scatter for the trailing [0] index and then one scatter for the outer tensor index. Fle..."
185823,[Dynamo] Handle class method_descriptor calls generically,https://github.com/pytorch/pytorch/pull/185823,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,146673,146673,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T15:52:30Z,2026-07-07T10:58:48Z,1,"Stack from ghstack (oldest at bottom): -> #185823 Dynamo only handled class method descriptor calls through a few ad hoc paths. For example, PR #146587 added a narrow str special case, while other valid forms such as int.bit_length(x), bytes.decode(b, encoding=...), or C extension classes represented as UserDefinedC..."
187633,Enhance documentation around the developer build,https://github.com/pytorch/pytorch/pull/187633,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,108406,108406,108406,180247,,,0,1,4,,,"triaged, open source, topic: not user facing",2026-06-18T01:41:59Z,2026-07-07T10:54:24Z,1,"The developer build documentation in README.md has several gaps that make it hard for new contributors to set up a working editable install. The legacy setup.py develop workflow is still referenced without migration guidance, users have no way to verify their local source is being loaded, and there is no troubleshoo..."
185920,[Dynamo] Detect tensor mutations when reconstructing generators,https://github.com/pytorch/pytorch/pull/185920,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,146628,146628,,,,0,5,1,b17f8bfaa8d25827f6cecb04db9553a253a023d6b9defde687971aa7c0c1b7e0,83162225e305d410f6690b5bfa47ac518cfd7ef53eafacd0b4847dbc74d84beb,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T11:21:20Z,2026-07-07T10:53:51Z,1,"Stack from ghstack (oldest at bottom): -> #185920 Dynamo reconstructs returned local generators by exhausting the remaining generator body during bytecode generation. Existing generator reconstruction side-effect checks rejected Python object mutations, but two paths still missed tensor data mutation: direct generat..."
185691,Avoid UB in float to signed integer casts,https://github.com/pytorch/pytorch/pull/185691,ready for maintainer decision,112,needs-review label present; approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,151510,151510,,,,0,5,1,,,"module: cpu, ciflow/trunk, release notes: cpu (x86), release notes: cpu (aarch64), agentic, needs-review",2026-05-30T01:41:20Z,2026-07-07T10:51:55Z,1,"Stack from ghstack (oldest at bottom): -> #185691 Fixes #151510 The repro passes torch.iinfo(torch.int64).max as the posinf replacement to nan_to_num, whose API stores replacement values as double. That value rounds to 2**63, so eager CPU later converted an out-of-range floating value to int64_t while running .long(..."
184153,[Inductor] Clone cudagraph-owned grads before replay,https://github.com/pytorch/pytorch/pull/184153,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,146569,146569,,,,0,5,1,,,"topic: bug fixes, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T03:32:10Z,2026-07-07T10:47:22Z,1,"Stack from ghstack (oldest at bottom): -> #184153 CUDAGraph tree backward outputs can be stolen into leaf gradients, so returning graph-owned tensors lets a later replay invalidate live .grad storage. Clone backward outputs before handing them to autograd so persistent grads have normal tensor lifetime. Fixes #14656..."
184161,Fix Inductor lazy negative input loads,https://github.com/pytorch/pytorch/pull/184161,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), pallas-tpu-py3.12-inductor / test (inductor-pallas-tpu, 1, 1, linux.google.tpuv7x.1)",statusCheckRollup,,145093,"145093, 186031",,,,0,8,10,072213b73c7a0128221cf59760c3c45e20b32424c6f5de9c805da09ed38fea6a,5234162e1521f05f8696ebdb69488f02881a1df3449a2a798d9567fb443ef37c,"topic: not user facing, module: inductor, ciflow/inductor, release notes: inductor, ciflow/inductor-pallas, ciflow/torchtitan, agentic, needs-review",2026-05-18T06:19:54Z,2026-07-07T10:43:09Z,1,"Stack from ghstack (oldest at bottom): -> #184161 Respect lazy negative tensor semantics when Inductor lowers graph input and constant loads, including input mutation copyback paths. Conjugated complex inputs continue to use fallback rather than generated loads. Fixes #145093 Generated by my agent cc @voznesenskym @..."
181728,[xpu][4/4]Enable MXFP8/MXFP4/NVFP4 tests for XPU,https://github.com/pytorch/pytorch/pull/181728,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,"181726, 181727, 187315, 187318",,,0,4,9,,,"open source, module: inductor, ciflow/inductor, ciflow/xpu, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-04-28T07:31:53Z,2026-07-07T10:42:46Z,1,"Enable Enable MXFP8/MXFP4 tests for XPU and inductor FP8 tests. PR Stack: Since I don't have ghstack permission, I manually created the following stacked PRs for review. I also created a combined PR #187318 to test on CI. #181726 [xpu][1/4]Implement scaled_mm_v2 for MXFP8/MXFP4/NVFP4 on XPU #181727 [xpu][2/4]Impleme..."
187315,[xpu][3/4] inductor: route MX scaled_mm_v2 fallback through v2 aten kernel,https://github.com/pytorch/pytorch/pull/187315,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,188721,"181726, 181727, 181728, 187318",,,0,3,5,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-06-15T08:13:00Z,2026-07-07T10:42:24Z,1,"The Inductor lowering for _scaled_mm_v2 ( tuned_scaled_mm_v2 ) has no Triton/extern template for the MX/NVFP4 recipes ( BlockWise1x32 / BlockWise1x16 ), multi-level (two-level NVFP4) scales, or any non-fp32 block scale. For these, it previously fell back to the v1 aten__fp8_mm choice, which expects a different scale..."
186583,Propagate contextvars to Python autograd functions,https://github.com/pytorch/pytorch/pull/186583,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,186537,186537,,,,0,5,3,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic",2026-06-08T06:29:13Z,2026-07-07T10:42:16Z,1,"Stack from ghstack (oldest at bottom): -> #186583 Reentrant activation checkpointing recomputes the checkpointed Python function from an autograd worker thread. _engine_run_backward already captures the launching thread's contextvars.Context into PyTorch's ThreadLocalState, and generated AOTAutograd runtime wrappers..."
185922,Fix fake one_hot invalid class count checks,https://github.com/pytorch/pytorch/pull/185922,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,146274,146274,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-02T11:50:36Z,2026-07-07T10:34:05Z,1,"Stack from ghstack (oldest at bottom): -> #185922 The fake/meta-friendly one_hot path builds the functional result directly with arange/eq before running eager's validation. For invalid static class counts such as num_classes=0, that allowed FakeTensorMode to return an empty class dimension instead of raising the sa..."
185829,Fix autograd.grad graph breaks under aot_eager,https://github.com/pytorch/pytorch/pull/185829,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,145899,145899,,,,0,3,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T17:28:29Z,2026-07-07T10:24:02Z,1,"Stack from ghstack (oldest at bottom): -> #185829 When trace_autograd_ops is disabled, Dynamo graph-breaks at torch.autograd.grad and normally compiles the prefix before resuming in eager. That is unsafe for AOTAutograd backends when the eager grad call consumes multiple differentiably-related outputs from the compi..."
184178,Fix masked Triton conv load pointers,https://github.com/pytorch/pytorch/pull/184178,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,143412,143412,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T08:45:34Z,2026-07-07T10:19:14Z,1,Stack from ghstack (oldest at bottom): -> #184178 Clamp load-only indices before forming masked input and weight pointers in the Triton 2D/3D convolution templates. This prevents padding and tail lanes from creating out-of-bounds pointer values while preserving the original masks and store indices. Fixes #143412 Gen...
182273,[c10d] Fix device_id not propagated through ProcessGroupWrapper,https://github.com/pytorch/pytorch/pull/182273,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,178977,178977,178977,,,,0,1,2,,,"triaged, open source, Stale, release notes: distributed (c10d)",2026-05-04T08:15:33Z,2026-07-07T09:47:10Z,1,"When TORCH_DISTRIBUTED_DEBUG=DETAIL is set, ProcessGroupWrapper wraps the actual backend but setBoundDeviceId was not forwarded to the wrapped backend. This caused NCCL to fall back to guessing the device ID based on global rank, which can hang with heterogeneous rank-to-GPU mappings. Make getBoundDeviceId and setBo..."
185586,Update README of OpenReg module to consistent with its structure and principle,https://github.com/pytorch/pytorch/pull/185586,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,1,,,"triaged, open source, topic: not user facing",2026-05-29T02:56:05Z,2026-07-07T09:43:23Z,1,Changes Update purpose and principles section to make sure they are consistent with original design Update structure section consistent with current implementation. Move Installation above Structure make sure new users have better readability.
188906,Add torch.utils.checkpoint.register_determinism_check,https://github.com/pytorch/pytorch/pull/188906,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/torchtitan,review when ready,,UNKNOWN,False,0,0,,,162980,162980,162980,,,,0,1,2,,,"triaged, open source, ciflow/torchtitan",2026-07-03T17:18:35Z,2026-07-07T09:13:44Z,1,"Fixes #162980 Summary The determinism_check argument of torch.utils.checkpoint.checkpoint looks up its metadata function in the module-private dict _allowed_determinism_checks_to_fns, so users who want a custom check (e.g. one that also compares strides, or ignores devices) have to monkey-patch the dict. As @soulitz..."
184820,[dynamo] respect custom metaclass __call__ during class construction,https://github.com/pytorch/pytorch/pull/184820,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175292,175292,,,,0,3,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-21T23:24:12Z,2026-07-07T09:07:38Z,1,"Stack from ghstack (oldest at bottom): -> #184820 Dynamo routed ordinary user-defined class calls through instantiate_user_defined_class_object. That helper intentionally models type.__call__ by running __new__ and then conditional __init__, but Python does not use those semantics when a class's metaclass overrides..."
186043,Lower symbolic bool not to torch.sym_not,https://github.com/pytorch/pytorch/pull/186043,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,143157,143157,,,,0,3,1,,,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T05:02:25Z,2026-07-07T09:04:49Z,1,"Stack from ghstack (oldest at bottom): -> #186043 Dynamo handles Python not as operator.not_ during bytecode tracing. For SymNodeVariable values that represent symbolic booleans, the old lowering also emitted operator.not_ into the FX graph. Fake evaluation of that node forced the SymBool through its example hint, s..."
186051,Fix disabled no-op module pre-hooks in compile,https://github.com/pytorch/pytorch/pull/186051,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,142358,142358,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-03T07:18:54Z,2026-07-07T09:01:09Z,1,Stack from ghstack (oldest at bottom): -> #186051 Dynamo normally graph-breaks when it calls a function wrapped with torch.compiler.disable(). For nn.Module forward pre-hooks this can happen while Dynamo is tracing nn.Module._call_impl from inside a user module's forward loop. Once that graph break occurs in the loo...
188894,Fix HAS_XPU_AND_TRITON to require triton XPU backend,https://github.com/pytorch/pytorch/pull/188894,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,4072,4072,,,,,0,2,0,,,"open source, topic: not user facing, ciflow/xpu",2026-07-03T12:40:57Z,2026-07-07T09:01:03Z,1,"Problem: HAS_XPU_AND_TRITON = torch.xpu.is_available() and HAS_TRITON evaluated True on Windows when only a generic triton package (without triton-xpu) was isntalled. This caused @requires_gpu_and_triton tests to run instead of being skipped, producing BackendCompilerFailed errors. Root cause: has_triton() returns T..."
186052,Fix Inductor C++ pybinding source literal escaping,https://github.com/pytorch/pytorch/pull/186052,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,142321,142321,,,,0,1,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T07:32:01Z,2026-07-07T08:58:18Z,1,"Stack from ghstack (oldest at bottom): -> #186052 Inductor emits CPU C++ kernels into generated Python as calls to async_compile.cpp_pybinding(). That source used to be assembled by opening a triple-quoted Python literal, splicing raw C++ text into it, and closing the literal later. On Windows, generated C++ can con..."
186554,Preserve hooks for dependent AOTAutograd outputs,https://github.com/pytorch/pytorch/pull/186554,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,91468,91468,,,,0,7,1,444652b01d61b02515432b038a028f0a10d713377f01db32894d915cce1eaafd,2c7ad3cd689e0e99b8de4c0eed9f9428be0cd7b81d51364d669d58b92c2a1ba6,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-07T17:02:25Z,2026-07-07T08:54:46Z,1,"Stack from ghstack (oldest at bottom): -> #186554 AOTAutograd wraps a Dynamo FX graph in a single custom autograd Function. When one differentiable graph output is also an ancestor of another differentiable graph output, that wrapper flattens the tensors into sibling outputs. A hook registered on the ancestor after..."
185899,Fix Dynamo data mutation on functorch tensors,https://github.com/pytorch/pytorch/pull/185899,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,149911,149911,,,,0,3,3,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-02T06:46:58Z,2026-07-07T08:50:40Z,1,"Stack from ghstack (oldest at bottom): -> #185899 Dynamo lowered Tensor .data assignment to Tensor.set_ plus version-counter repair. That bypassed the eager TensorImpl shallow_copy_from path, so GradTrackingTensor and BatchedTensor inputs under functorch transforms did not raise the RuntimeError that eager raises fo..."
187589,Make FX symbolic tracing state thread-local,https://github.com/pytorch/pytorch/pull/187589,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,126024,126024,168999,,,0,4,2,,,"release notes: fx, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-17T18:24:35Z,2026-07-07T08:44:58Z,1,"Stack from ghstack (oldest at bottom): -> #187589 FX symbolic tracing used a process-global _is_fx_tracing_flag and process-global monkeypatches of nn.Module.__call__ and __getattr__. When one thread was inside torch.fx.symbolic_trace, an unrelated thread calling torch.compile(model, dynamic=True) could observe that..."
184185,Use memory stats for persistent reductions,https://github.com/pytorch/pytorch/pull/184185,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, linux-jammy-py3.14-clang18 / test-osdc (crossref, 2, 2, mt-l-x86iavx512-8-64)",statusCheckRollup,,141916,141916,,,,0,5,10,05f712ff56f1c687d6e66f71307ed614a50bac96107b8316177b8d5b918ed267,d9c711490123794cd0b22806998e14472906d5a4dd58df973d51c9bc225ae043,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T10:45:24Z,2026-07-07T08:42:28Z,1,Stack from ghstack (oldest at bottom): -> #184185 Select persistent reduction codegen for large inner reductions when SIMDKernelFeatures memory analysis shows that keeping the reduction tile live materially reduces global memory traffic. Also fix the memory stats addition used by that heuristic.\n\nFixes #141916\nGe...
186053,Fix AOTAutograd cache save for call-free lazy backwards,https://github.com/pytorch/pytorch/pull/186053,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141790,141790,,,,0,3,1,,,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T08:21:52Z,2026-07-07T08:37:22Z,1,Stack from ghstack (oldest at bottom): -> #186053 AOTAutograd local cache entries store FXGraphCache keys for the compiled forward and backward artifacts. The issue reproducer has a lazy backward graph that contains no call_function/call_module nodes and just returns the incoming tangent. Inductor handled that backw...
184189,Make generated Triton code use runtime device properties,https://github.com/pytorch/pytorch/pull/184189,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141744,141744,,,,0,3,3,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T11:26:15Z,2026-07-07T08:34:30Z,1,"Stack from ghstack (oldest at bottom): -> #184189 Generated Triton source now resolves device properties at load time and keys async/Python caches with the runtime device properties, avoiding stale architecture metadata in standalone files. Fixes #141744 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @..."
186058,Preserve export graph metadata during normalization,https://github.com/pytorch/pytorch/pull/186058,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141640,141640,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-03T09:10:55Z,2026-07-07T08:30:24Z,1,"Stack from ghstack (oldest at bottom): -> #186058 Export's signature normalization rebuilt FX graphs while copying only a small hard-coded subset of metadata. That dropped graph-level metadata such as _param_name_to_source when it was stored in GraphModule.meta, and it also dropped arbitrary node metadata attached b..."
186063,[dynamo] Route OptimizedModule delegated methods through compiled call,https://github.com/pytorch/pytorch/pull/186063,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141473,141473,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T10:02:42Z,2026-07-07T08:27:25Z,1,"Stack from ghstack (oldest at bottom): -> #186063 torch.compile(module) wraps the original module in OptimizedModule and compiles the wrapper's forward path. Attributes that are not owned by the wrapper are delegated to the original module. For Python instance methods such as HuggingFace-style generate, that delegat..."
186069,Fix Dynamo round on NumPy scalar results,https://github.com/pytorch/pytorch/pull/186069,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141258,141258,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-03T10:54:03Z,2026-07-07T08:24:40Z,1,"Stack from ghstack (oldest at bottom): -> #186069 Dynamo represents NumPy calls with torch._numpy-backed NumpyNdarrayVariable values. For NumPy scalar-producing calls such as np.floor(3.1), eager NumPy returns a numpy.float64 scalar with scalar-only Python APIs like round, but Dynamo modeled the value as an ndarray..."
183613,[CD] Tighten ROCm manywheel build pipeline,https://github.com/pytorch/pytorch/pull/183613,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,182696,,,0,1,1,,,"module: rocm, open source, ciflow/binaries",2026-05-13T19:44:41Z,2026-07-07T08:22:23Z,1,"Summary Follow-ups to #182696 (port of ROCm manywheel build to the Python pipeline): build_env_setup.py – remove USE_STATIC_CUDNN, USE_STATIC_NCCL, ATEN_STATIC_CUDA, USE_CUDA_STATIC_LINK from ROCM_BUILD_ENV_STATIC. CUDA-only flags, no-ops in ROCm CMake. Flagged by @jithunnair-amd on the original PR. build_env_setup...."
185748,[dynamo] Model RemovableHandle hook registration side effects,https://github.com/pytorch/pytorch/pull/185748,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,149921,149921,,,,0,11,1,,,"topic: not user facing, module: dynamo, agentic",2026-05-31T07:27:12Z,2026-07-07T08:22:16Z,1,"Stack from ghstack (oldest at bottom): -> #185748 Root cause: Dynamo treated torch.utils.hooks.RemovableHandle allocation as ordinary Python state. Registering module hooks inside a compiled region captured trace-time RemovableHandle.next_id/id values, left empty nn.Module hook dicts effectively immutable, and did n..."
187318,[xpu]Implement scaled_mm for XPU - combined,https://github.com/pytorch/pytorch/pull/187318,draft/noise,13,"PR is draft; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,"181726, 181727, 181728, 187315",,,0,21,4,,,"module: cpu, triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/xpu, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-06-15T08:19:48Z,2026-07-07T08:21:54Z,1,The purpose of this PR is to test on CI only. I have split the code into the following 4 PRs for review: #181726 [xpu][1/4]Implement scaled_mm_v2 for MXFP8/MXFP4/NVFP4 on XPU #181727 [xpu][2/4]Implement scaled_mm_v1 for MXFP8/MXFP4/NVFP4 on XPU #187315 [xpu][3/4] inductor: route MX scaled_mm_v2 fallback through v2 a...
189009,[xpu][fix] Enable test_fallback_kernel_with_symexpr_output on XPU with correct head_dim,https://github.com/pytorch/pytorch/pull/189009,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188131,188131,188131,,,,0,1,2,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor",2026-07-06T06:37:25Z,2026-07-07T08:21:52Z,1,"Motivation test_fallback_kernel_with_symexpr_output was disabled on XPU via @skipIfXpu because it used tensor_shape=(4,32,4,4), giving head_dim=32//2=16 after reshape. The XPU sycltla flash attention kernel only supports specific head_dim values and crashes for unsupported ones. Fixes #188131 (DISABLED test_fallback..."
189128,Gil acquire bool conversion,https://github.com/pytorch/pytorch/pull/189128,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,,,,,,,0,5,0,,,"open source, release notes: distributed (c10d)",2026-07-07T05:19:25Z,2026-07-07T08:21:34Z,1,
178768,Pass tensor data pointer to dlDeviceToTorchDevice from maybeCopyTensor in DLConvertor.cpp,https://github.com/pytorch/pytorch/pull/178768,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,3077,3077,,3077,,,0,1,1,,,"open source, topic: not user facing",2026-03-30T12:59:36Z,2026-07-07T08:21:28Z,1,"Previously this optional data pointer was not passed to dlDeviceToTorchDevice, which is not supported on XPU. Now that is being passed along, the device can be correctly determined. Fixes intel/torch-xpu-ops#3077"
184809,Improve custom op Tensor subclass fake error,https://github.com/pytorch/pytorch/pull/184809,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,CLEAN,False,0,0,,,,175408,175408,,,,0,5,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic",2026-05-21T22:52:05Z,2026-07-07T08:21:23Z,1,"Stack from ghstack (oldest at bottom): -> #184809 During Dynamo fake-value propagation, a custom op called with a traceable Tensor subclass may receive a Subclass(FakeTensor) rather than a bare FakeTensor. If the subclass torch_dispatch passes that op through, dispatch runs the custom op implementation instead of it..."
184336,fix RPATH in libtorch extraction from wheel,https://github.com/pytorch/pytorch/pull/184336,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/binaries,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,183971,183971,183971,174753,,,0,9,4,,,"triaged, open source, ciflow/binaries, release notes: releng",2026-05-19T08:27:43Z,2026-07-07T08:21:16Z,1,"This PR adds an RPATH rewriter to the extraction logic that resets RPATH to $ORIGIN. NVIDIA deps (cudnn,nccl, nvshmem, etc.) are resolved from the user's system CUDA installation via LD_LIBRARY_PATH/ldconfig, matching the old libtorch behavior. PR #174753 switched libtorch packaging from a dedicated build to extract..."
186614,Delay RNN graph break until module call,https://github.com/pytorch/pytorch/pull/186614,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141184,141184,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-08T15:54:02Z,2026-07-07T07:24:16Z,1,"Stack from ghstack (oldest at bottom): -> #186614 Dynamo intentionally graph breaks on nn.RNN, nn.GRU, and nn.LSTM when allow_rnn is disabled. The old implementation raised the unsupported error while wrapping the recurrent module attribute. For LOAD_ATTR/LOAD_METHOD bytecode this goes through the generic step graph..."
186072,Make constant warnings reorderable in Dynamo by default,https://github.com/pytorch/pytorch/pull/186072,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141162,141162,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-03T11:45:42Z,2026-07-07T07:20:53Z,1,"Stack from ghstack (oldest at bottom): -> #186072 Dynamo already has a debug logging replay path that can defer side-effectful logging calls to the end of a compiled region. That path was only used for functions explicitly listed in torch._dynamo.config.reorderable_logging_functions, so a constant warnings.warn(...)..."
186898,Add always-on AOTI input rank checks,https://github.com/pytorch/pytorch/pull/186898,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141115,141115,,,,0,2,1,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-06-10T06:23:21Z,2026-07-07T07:18:27Z,1,"Stack from ghstack (oldest at bottom): -> #186898 AOTInductor's generated wrapper only ran its full input dtype, size, stride, and device checks when AOTI_RUNTIME_CHECK_INPUTS was enabled. That left a default runtime gap for rank mismatches: a model exported with a 1D input could be called with a 2D tensor, and if t..."
185866,Fix CPU Inductor numerics for BatchNorm erf normalize,https://github.com/pytorch/pytorch/pull/185866,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,185589,185589,,,,0,8,1,0056f49aa1c3402e8e5f838ba4e2f11e12268ba30bde4e62e0ecff5b27688aca,4f99a7d538db46d47a266928cb320b711ee7ddcfcf6ed4f7dfd5909cee34d64b,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T23:41:17Z,2026-07-07T07:15:18Z,1,"Stack from ghstack (oldest at bottom): -> #185866 The issue repro combines CPU training BatchNorm, ReLU, diag_embed, erf, and normalize(p=-1). diag_embed creates rows containing zeros, so the p=-1 normalization path clamps the denominator to eps. That magnifies otherwise tiny upstream CPU numerical differences into..."
186074,Fix multilabel_margin_loss decomposition sentinel handling,https://github.com/pytorch/pytorch/pull/186074,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,141017,141017,,,,0,2,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-03T12:10:26Z,2026-07-07T07:11:13Z,1,"Stack from ghstack (oldest at bottom): -> #186074 AOTAutograd uses the Python decomposition for aten.multilabel_margin_loss_forward. The decomposition found the first -1 sentinel in each target row, but only used that mask to make gather indices valid. Entries after the sentinel were replaced with class 0 and still..."
186078,Fix OptimizedModule _parameters access,https://github.com/pytorch/pytorch/pull/186078,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,141005,141005,,,,0,3,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T13:14:30Z,2026-07-07T07:07:26Z,1,"Stack from ghstack (oldest at bottom): -> #186078 OptimizedModule forwards ordinary attribute writes to the wrapped module, so assigning opt_mod._parameters already updated opt_mod._orig_mod._parameters. Reads were inconsistent because nn.Module.init installs a wrapper-local _parameters dict, and normal Python attri..."
184460,Remove obsolete Python wrapper symbol extraction,https://github.com/pytorch/pytorch/pull/184460,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,102870,102870,,,,0,19,1,,,"ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic",2026-05-20T01:10:28Z,2026-07-07T07:05:51Z,1,"Stack from ghstack (oldest at bottom): -> #184460 Python wrapper codegen no longer needs to recover symbolic sizes and strides by inspecting tensor inputs because free symbols are already passed as explicit graph inputs. Keep the C++ wrapper path unchanged, since it still owns its size/stride binding behavior. Fixes..."
185749,Fix Dynamo construction for classes with __getattribute__,https://github.com/pytorch/pytorch/pull/185749,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,149895,149895,,,,0,15,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic",2026-05-31T08:37:48Z,2026-07-07T07:05:24Z,1,"Stack from ghstack (oldest at bottom): -> #185749 Dynamo already knows how to trace Python-level getattribute on UserDefinedObjectVariable, but class construction was gated by SideEffects.cls_supports_mutation_side_effects. That predicate only allowed classes whose static getattribute was one of a small set of built..."
188556,[inductor] Preserve eager semantics for infinite Bessel inputs,https://github.com/pytorch/pytorch/pull/188556,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,188545,188545,188545,,,,0,2,2,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-30T15:31:33Z,2026-07-07T07:04:47Z,1,"Fixes #188545 Summary Preserve eager-mode semantics for infinite inputs in the Triton lowering of torch.special.bessel_j0 torch.special.bessel_j1 torch.special.bessel_y0 torch.special.bessel_y1 Previously, the Triton lowering called libdevice directly. For ±inf inputs, libdevice returns the mathematical limit, where..."
189020,Reorder condition to handle empty tensor the same for file_system & file_descriptor,https://github.com/pytorch/pytorch/pull/189020,draft/noise,13,PR is draft; CI lane labels: ciflow/trunk,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing",2026-07-06T11:12:45Z,2026-07-07T07:04:43Z,1,"For XPU devices, empty tensor is handled differently depends of the _sharing_strategy. For file_system test fails but for file_descriptor passed. Changing conditions order will cause that both cases will work the same for empty tensor. XPU Windows issue: intel/torch-xpu-ops#4068"
187353,Preserve signed zero in Triton neg lowering,https://github.com/pytorch/pytorch/pull/187353,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,187336,187336,,,,0,3,1,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-15T15:50:16Z,2026-07-07T07:03:49Z,1,"Stack from ghstack (oldest at bottom): -> #187353 Inductor decomposes torch.special.log_ndtr into pointwise Triton math. For large positive inputs the positive branch computes log1p(-erfc(t) / 2), so the sign of zero from the negation is observable by downstream sign-sensitive operations such as copysign. Triton una..."
185333,Graph break existing list tensor assignments in Dynamo,https://github.com/pytorch/pytorch/pull/185333,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,160247,160247,,,,0,5,9,79f13274f1f5eb4c9603884d1f402989650f1dde28ac0f92fd770b03547d2361,49637f6dc6c3846eaaa97c8602038adf7c0ebd2639097ece492dcc7de3fdc0f9,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-27T07:23:13Z,2026-07-07T07:01:15Z,1,"Stack from ghstack (oldest at bottom): -> #185333 Dynamo replays mutations to pre-existing Python lists after the compiled graph returns. When a user cache list entry is replaced with a tensor produced by the graph, that replay order keeps the old tensor entries live while all replacement outputs are materialized, w..."
186094,Fix AOTI cond predicate unbacked symbol dependency,https://github.com/pytorch/pytorch/pull/186094,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,2,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,"137520, 140842","137520, 140842, 186031",,,,0,2,9,004f2a285d71d03da3d18d82bad8c4d8369742dc072cabe93de04e1a5831b044,a5e520d30de935d6ef58cac96acf43784413216930e100c7d4da51dc41923e0a,"topic: not user facing, module: inductor, ciflow/inductor, agentic, needs-review",2026-06-03T15:54:25Z,2026-07-07T06:55:13Z,1,"Stack from ghstack (oldest at bottom): -> #186094 AOTI lowering for torch.cond can store the predicate as a SymPy boolean relation inside ExternKernel.constant_args, for example u0 > 0. The scheduler uses ExternKernel.get_free_symbol_uses() to add dependencies from users of unbacked symbols to the nodes that define..."
184207,Release max autotune benchmark artifacts,https://github.com/pytorch/pytorch/pull/184207,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,140795,140795,,,,0,3,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T14:04:14Z,2026-07-07T06:52:34Z,1,"Stack from ghstack (oldest at bottom): -> #184207 Release benchmark-only PyCodeCache modules and precompile artifacts after max-autotune selection so compiled models do not retain unnecessary host memory after compilation. Add regression coverage for module unloading, cleanup on exceptions, and precompile-cache reus..."
186111,Fix stale mutated input history in AOTAutograd,https://github.com/pytorch/pytorch/pull/186111,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,140707,140707,,,,0,1,1,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-06-03T18:27:04Z,2026-07-07T06:50:00Z,1,"Stack from ghstack (oldest at bottom): -> #186111 AOTAutograd can treat a non-grad buffer mutated with parameter-derived data as a differentiable mutated input. On a later invocation, the compiled autograd Function then receives the stale non-leaf buffer from the previous invocation. If the user-output backward does..."
187931,[XPU Graph] Refine XPU graph capture detection in XPUCachingAllocator,https://github.com/pytorch/pytorch/pull/187931,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,,,,,0,5,1,,,"triaged, open source",2026-06-23T09:35:49Z,2026-07-07T06:47:39Z,1,"1, Add deferred block handling During XPU graph capture, inserting events and querying event status are illegal operations. However, user code may still free a block that has non-empty stream_uses while capture is underway. The previous code path would unconditionally call insert_events() in this case, which is inva..."
186118,Preserve guard eval defaulting without callback churn,https://github.com/pytorch/pytorch/pull/186118,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,140683,140683,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-03T19:24:27Z,2026-07-07T06:46:54Z,1,"Stack from ghstack (oldest at bottom): -> #186118 Commit fe4fa1d moved the eval-frame callback to None before guard lookup so frames entered from guard evaluation would run through the default evaluator instead of recursively hitting Dynamo. That fixed guard-eval reentrancy, but it also made explicit run-only cache..."
186900,Improve pytree missing registration error,https://github.com/pytorch/pytorch/pull/186900,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,140615,140615,,,,0,3,1,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), agentic, needs-review",2026-06-10T06:35:31Z,2026-07-07T06:43:21Z,1,"Stack from ghstack (oldest at bottom): -> #186900 AOTI packages store input and output pytree specs in the package call spec. When a compiled artifact returns a custom output type, loading the package in a process that has not imported the package defining that type fails while deserializing the output TreeSpec. The..."
186942,[inductor] Round fmod/remainder scalars for bf16/fp16 CPU,https://github.com/pytorch/pytorch/pull/186942,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, lf-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,186875,186875,,,,0,6,2,c4951b1e1f93b2e5cb6afff7821ef5e9e59dba7005c53f8e489087ae9dc0aba2,4b7be0fa6bd0477543fa3df1b0a45c02b3eda5c6e7411f7a5b9293e74d8ca878,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-10T16:28:11Z,2026-07-07T06:40:25Z,1,"Stack from ghstack (oldest at bottom): -> #186942 CPU eager casts Python float scalar operands to the tensor dtype for bf16/fp16 fmod and remainder. Inductor only performed that rounding for comparison ops, so compiled CPU kernels kept constants like 1.7 at full precision and diverged from eager at rounding boundari..."
186119,Improve Dynamo observed exception messages,https://github.com/pytorch/pytorch/pull/186119,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,140607,140607,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-03T19:39:06Z,2026-07-07T06:36:59Z,1,"Stack from ghstack (oldest at bottom): -> #186119 Dynamo already kept the original observed exception as an ExceptionVariable and rendered it with debug_repr(), but exception_handler() only surfaced that information in the developer debug context. The primary explanation stayed as the generic ""Observed exception"", w..."
186160,Fix FakeTensor forward AD batch_norm meta rewrapping,https://github.com/pytorch/pytorch/pull/186160,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,7,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_huggingface, 1, 1, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), linux-jammy-py3.14t-clang18 / test-osdc (default, 3, 5, mt-l-x86iavx512-16-128), unit-test / inductor-cpu-build / build-osdc, inductor-cpu-build / build-osdc",statusCheckRollup,,139756,139756,,,,0,5,5,c7f82f3ff7e07a41e358d1ffe15d55b8d404228a5430b00bd9325621f3d23cd4,f71046ed6047890965e5e75ad095c03ea820f9e1fb694e22aa3f0ea5a86d0a71,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-06-04T00:09:24Z,2026-07-07T06:32:34Z,1,"Stack from ghstack (oldest at bottom): -> #186160 Forward AD can invoke fake kernels that hand FakeTensorMode plain meta tensors produced inside the fake/meta execution path. When allow_non_fake_inputs was false, FakeTensorMode rejected those tensors as ordinary non-fake inputs. The error formatting then attempted t..."
186166,Respect fake/functional modes for backend autograd kernels,https://github.com/pytorch/pytorch/pull/186166,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, linux-jammy-py3.14t-clang18 / test-osdc (default, 5, 5, mt-l-x86iavx512-16-128)",statusCheckRollup,,139707,139707,,,,0,6,1,8a9c062717d2ba8b3e4950b03a7c42946275c320143d5886bb87c1ef4e2d26aa,9667dd37d6b2088c4ca08b45023e996add777f32408e8138a9f4e69e12cd9aae,"topic: not user facing, agentic, needs-review",2026-06-04T01:17:32Z,2026-07-07T06:29:52Z,1,Stack from ghstack (oldest at bottom): -> #186166 Backend-specific Python autograd registrations such as AutogradCPU and AutogradCUDA were being invoked before active TorchDispatchMode handling. For FakeTensorMode this meant a backend override for an aten operator could execute its real Python kernel during fake pro...
186044,Stop Dynamo from importing optional libraries at import time,https://github.com/pytorch/pytorch/pull/186044,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,6,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), get-label-type / runner-determinator, get-cuda-version",statusCheckRollup,,142489,142489,,,,0,10,1,1603a5402e23bf95f048f1df5e9601bdc20d1472cb62c7595998c3d9ed37ef93,0d4a4fd1cb4c91c5cd57209502e744f8f5b19c1496b852597c5532ef029fd698,"ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-03T05:46:53Z,2026-07-07T06:24:21Z,1,Stack from ghstack (oldest at bottom): -> #186044 Dynamo still had a few import-time paths that loaded optional third-party packages. The most visible one was Triton: importing torch._dynamo could load Triton through torch._dynamo.utils.has_triton_package() and through Inductor runtime hint setup reached from guard...
186168,Fix Dynamo tracing of torch.random.fork_rng,https://github.com/pytorch/pytorch/pull/186168,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,139603,139603,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T01:37:03Z,2026-07-07T06:19:28Z,1,"Stack from ghstack (oldest at bottom): -> #186168 Dynamo was classifying torch.random.fork_rng as an in-graph torch function. That is not a valid lowering target: fork_rng is a Python context manager, and proxying the call made FX try to convert its Python arguments into graph arguments. When the devices argument wa..."
186180,Fix Dynamo cProfile under active Python profilers,https://github.com/pytorch/pytorch/pull/186180,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"137869, 139232","137869, 139232",,,,0,3,2,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T04:13:16Z,2026-07-07T06:16:58Z,1,"Stack from ghstack (oldest at bottom): -> #186180 TORCH_COMPILE_CPROFILE wraps Dynamo compilation with cProfile. On Python 3.12, cProfile owns sys.monitoring.PROFILER_ID while it is active, so a nested Dynamo cProfile cannot be enabled. The previous fallback caught the ValueError from starting cProfile, but still tr..."
186188,Fix FakeTensor wrapping for meta tensor constructor,https://github.com/pytorch/pytorch/pull/186188,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,139092,139092,,,,0,2,1,,,"topic: not user facing, agentic, needs-review",2026-06-04T05:30:09Z,2026-07-07T06:14:15Z,1,"Stack from ghstack (oldest at bottom): -> #186188 The C++ tensor constructor had a meta shortcut that returned an empty meta tensor before reaching the shared lift_fresh path. Under FakeTensorMode, CPU tensor construction reaches lift_fresh and gets wrapped as a FakeTensor, but torch.tensor(..., device='meta') bypas..."
186920,Route overrideable SDPA AOTI calls through dispatcher,https://github.com/pytorch/pytorch/pull/186920,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,184195,184195,,,,0,5,1,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic, needs-review",2026-06-10T09:14:16Z,2026-07-07T06:06:41Z,1,"Stack from ghstack (oldest at bottom): -> #186920 AOTInductor treated aten._scaled_dot_product_fused_attention_overrideable and its backward as normal AOTI fallback shim calls on devices without a real in-tree kernel. The generated CPU/CUDA/MPS shims call the CompositeExplicitAutograd implementation directly, which..."
182725,[c10d] Throttle TCPStore connect-retry warnings,https://github.com/pytorch/pytorch/pull/182725,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,1,1,,,"triaged, open source, release notes: distributed (c10d)",2026-05-06T21:58:23Z,2026-07-07T06:06:22Z,1,"Summary Add reusable throttled-log macro and applied to TCPStore retry paths to reduce startup-noise spam without losing visibility into prolonged failures. New macro C10D_WARNING_EVERY_N_ELSE_DEBUG: logs at WARNING on calls N, 2N, 3N, ... and at DEBUG otherwise. Designed for retry loops where every failure is worth..."
184367,Fix AotAutograd aliased-output replay for traceable wrapper subclasses,https://github.com/pytorch/pytorch/pull/184367,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,184002,184002,184002,,,,0,1,2,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, ciflow/inductor, ci-no-td, ciflow/dtensor",2026-05-19T12:51:27Z,2026-07-07T06:00:06Z,1,Fixes #184002 Compiled DTensor view outputs no longer fall back to unsupported as_strided. Added test to cover both generic wrapper subclass case and DTensor compile path
178844,[c10d] Support stacked output tensors in Gloo all_gather_into_tensor …,https://github.com/pytorch/pytorch/pull/178844,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,178798,178798,178798,,,,0,1,2,,,"triaged, open source, release notes: distributed (c10d)",2026-03-31T02:31:45Z,2026-07-07T05:58:57Z,1,"Summary Fixes #178798 ProcessGroupGloo::_allgather_base chunked the output tensor along dim 0 and required every chunk to match the input shape exactly. For stacked outputs (world_size, *input.shape) the chunks retained the leading dimension — e.g. (1, N) vs (N,) — so the downstream allgather() validation rejected t..."
186192,Fix CompiledFxGraph callable state across threads,https://github.com/pytorch/pytorch/pull/186192,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,138961,138961,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T06:44:36Z,2026-07-07T05:58:43Z,1,"Stack from ghstack (oldest at bottom): -> #186192 CompiledFxGraph.current_callable was shared mutable state. When two threads compiled cache-equivalent graphs, Dynamo could start calling a CompiledFxGraph object produced by another thread, and post-compile mutations such as cudagraph wrapping could overwrite the cal..."
186194,Canonicalize trunc of symbolic integer division by one,https://github.com/pytorch/pytorch/pull/186194,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,138926,138926,,,,0,1,1,,,"module: cpu, topic: not user facing, agentic, needs-review",2026-06-04T06:59:19Z,2026-07-07T05:55:45Z,1,"Stack from ghstack (oldest at bottom): -> #186194 Issue #138926 reported a torch.cond metadata mismatch where one branch returned a tensor sized with TruncToInt(IntTrueDiv(s, 1)) while the other returned the same tensor sized with s. The branches were semantically compatible, but the symbolic expressions were not ca..."
188939,[profiler][cupti] Environment + per-kernel cycle GPU counters in the JSON export,https://github.com/pytorch/pytorch/pull/188939,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,187898,,,0,13,1,,,,2026-07-04T03:31:26Z,2026-07-07T05:51:28Z,1,"Stack from ghstack (oldest at bottom): #187898 -> #188939 #188019 #188849 #188942 Add two more counter sources to the chrome-JSON exporter's GPU counters (built on the plumbing from the PM-counters PR): the environment counters (power / temperature / clocks, parsed from the CUPTI environment union) and a per-kernel..."
187898,"[profiler][cupti] Perfetto-native (.pftrace) export, full chrome-args parity",https://github.com/pytorch/pytorch/pull/187898,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,188939,,,0,27,1,,,,2026-06-23T02:43:14Z,2026-07-07T05:51:14Z,1,"Stack from ghstack (oldest at bottom): -> #187898 #188939 #188019 #188849 #188942 export_chrome_trace(path) with a .pftrace path makes the cupti_monitor backend emit a Perfetto-native trace instead of chrome JSON, encoded straight from the observer's columnar window (no chrome-dict materialization). The .pftrace out..."
184256,[inductor] Cap reduction XBLOCK to max block,https://github.com/pytorch/pytorch/pull/184256,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,138301,138301,,,,0,2,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T19:03:16Z,2026-07-07T05:50:52Z,1,Stack from ghstack (oldest at bottom): -> #184256 Limit reduction heuristic XBLOCK growth to the configured Triton maximum so generated configs preserve the indexing invariants for large grids. Fixes #138301 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe...
188867,Add USE_KLEIDIAI_SME option to skip SME/SME2 kernels on CPUs without SME,https://github.com/pytorch/pytorch/pull/188867,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,5,1,,,"triaged, open source, topic: not user facing",2026-07-03T01:27:49Z,2026-07-07T05:50:25Z,1,"Summary While rebuilding PyTorch from source on an NVIDIA DGX Spark (GB10, aarch64), I noticed a large chunk of build time going into KleidiAI's SME/SME2 kernels — kernels that GB10's Cortex-X925/A725 cores can never runtime-dispatch to, since the hardware doesn't implement SME (/proc/cpuinfo Features has sve, sve2,..."
186199,Fix functionalization return mapping for mixed mutable out ops,https://github.com/pytorch/pytorch/pull/186199,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,138264,138264,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-04T07:27:04Z,2026-07-07T05:47:07Z,1,"Stack from ghstack (oldest at bottom): -> #186199 Functionalization codegen assumed the functional replacement returned fresh outputs first and then one updated value per mutable input. That ordering is wrong for schemas that both return out= aliases and mutate additional inputs. For _native_batch_norm_legit.out, th..."
186201,Deduplicate symbolic cond inputs after export passes,https://github.com/pytorch/pytorch/pull/186201,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,138229,138229,,,,0,1,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-06-04T07:54:34Z,2026-07-07T05:44:48Z,1,"Stack from ghstack (oldest at bottom): -> #186201 Fixes #138229 Generated by my agent AOT/export runtime assertion insertion can CSE symbolic size nodes that represent the same symbolic dimension. In the failing torch.cond export, the top-level cond operand tuple was rewritten to contain the same SymInt node twice w..."
185782,fix: [PT2] [Hardening] Track recompiles alongside graph breaks in our actu...,https://github.com/pytorch/pytorch/pull/185782,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,113040,113040,113040,,,,0,7,3,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-01T02:13:07Z,2026-07-07T05:44:08Z,1,"Summary Root Cause The CI comparison script (benchmarks/dynamo/check_graph_breaks.py) only tracked graph_breaks. Recompile events were invisible: a frame recompiling with 3 graph breaks would look like 6 graph breaks (Example 1 from the issue), or a recompile with 0 graph breaks would be completely undetected (Examp..."
185486,[OpenReg] Add minimal implementation for inductor integration in openreg,https://github.com/pytorch/pytorch/pull/185486,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,181093,,,,0,3,1,,,"triaged, open source, topic: not user facing, module: openreg",2026-05-28T08:57:29Z,2026-07-07T05:41:40Z,1,"Summary Add end-to-end Inductor integration for the OpenReg reference device, exercising the register_backend_for_device and register_device_op_overrides APIs that device vendors use to plug into torch.compile(backend=""inductor""). The implementation is minimal as it delegates scheduling entirely to CppScheduling and..."
186501,Collapse foreach input mutation copies in AOTAutograd,https://github.com/pytorch/pytorch/pull/186501,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,5,1,"statusCheckRollup, inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,119191,119191,,,,0,6,9,ccd43b97cefd9f972d50e1a5e626c4f51714f22b87f29db251e44b27205f46d8,3e4b64edea50c40a8964ef030c1101e11436751d4b3cef834ebe66eef677bd3d,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-06T23:47:37Z,2026-07-07T05:40:00Z,1,"Stack from ghstack (oldest at bottom): -> #186501 AOTAutograd functionalization preserves input mutations by appending epilogue copy_ nodes to the captured graph. For foreach input mutations, every tensor in the TensorList came from the same foreach result list, but we still emitted one aten.copy_ per mutated input...."
186210,Fix cond branch-local shape assertions,https://github.com/pytorch/pytorch/pull/186210,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,138141,138141,,,,0,5,1,c4bce478a8891e3595657588b3881035802244e919f2035ab312cc44cca062d6,afc177b4f8916ac31a2825d1c72aeb8d786004ddfb4ee6fa9bef78fe3cb80235,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T08:55:04Z,2026-07-07T05:36:21Z,1,"Stack from ghstack (oldest at bottom): -> #186210 torch.cond traces both branches speculatively, but branch-only shape facts were previously evaluated against the outer ShapeEnv. A shape assert such as assert x.shape[0] == 1 in the false branch of a predicate x.shape[0] > 1 therefore tried to install or evaluate a g..."
186214,Profile lazy backward compilation with cProfile,https://github.com/pytorch/pytorch/pull/186214,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,137996,137996,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T09:18:46Z,2026-07-07T05:32:50Z,1,"Stack from ghstack (oldest at bottom): -> #186214 TORCH_COMPILE_CPROFILE only wrapped Dynamo's forward frame compilation path via _compile_inner. Lazy AOTAutograd backward lowering restores the saved CompileContext when backward() first needs to lower the backward graph, but then called the backward compiler directl..."
186217,Move FX graph cache hash details out of debug logs,https://github.com/pytorch/pytorch/pull/186217,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,137875,137875,,,,0,1,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T09:30:30Z,2026-07-07T05:28:59Z,1,"Stack from ghstack (oldest at bottom): -> #186217 compiled_fx_graph_hash computed verbose cache-key component lines and also joined the full list into the regular torch._inductor.codecache DEBUG logger. With TORCH_LOGS=+inductor this made cache-enabled compiles print every cache-key component, including large exampl..."
187604,Fix fake slice pending unbacked replacements,https://github.com/pytorch/pytorch/pull/187604,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,185185,185185,,,,0,2,1,,,"release notes: dynamo, agentic, needs-review",2026-06-17T20:28:13Z,2026-07-07T05:25:43Z,1,"Stack from ghstack (oldest at bottom): -> #187604 Multidimensional Python indexing is represented as one operator.getitem FX node, but fake evaluation may run several internal slice/select operations before bindings are computed for the final output. When a data-dependent slice or select has to mint a fresh unbacked..."
186222,Include epilogue guards in Dynamo code_parts,https://github.com/pytorch/pytorch/pull/186222,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,137388,137388,,,,0,1,1,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T10:10:32Z,2026-07-07T05:23:25Z,1,"Stack from ghstack (oldest at bottom): -> #186222 GuardManagerWrapper.populate_code_parts_for_debugging() built its debug code_parts by walking normal leaf guards under the root manager and child managers. Shape guards installed as root epilogue lambda guards are stored separately, so the guard tree could show a LAM..."
186225,Preserve NJT nested ints in AOT subclass outputs,https://github.com/pytorch/pytorch/pull/186225,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,137275,137275,,,,0,4,1,,,"module: cpu, topic: not user facing, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-04T10:50:37Z,2026-07-07T05:20:37Z,1,"Stack from ghstack (oldest at bottom): -> #186225 AOTAutograd unwraps traceable tensor subclass outputs into their inner tensors and later reconstructs the subclass at runtime. For jagged NestedTensor device transfers, the graph copies the offsets or lengths tensor. Because subclass output metadata dropped nested-in..."
184621,Fix Dynamo torch.Size from tensor shape inputs,https://github.com/pytorch/pytorch/pull/184621,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,"182649, 182651, 182652","182649, 182651, 182652",,,,0,6,19,7da79a5574b8c7b5c1ba04c653eed587ef18d61e7e92ef807dcd5b1431cb05bc,e7ca2cc4dd0e68c515af5182bee541eafe999b13f4101c2f0891152daa5dea07,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-20T22:40:27Z,2026-07-07T05:17:12Z,1,Stack from ghstack (oldest at bottom): -> #184621 Route tensor items in torch.Size construction through index semantics and avoid constructing Size example metadata from raw FakeTensor scalar values. This covers tensor-backed reshape shapes from buffers and equivalent ONNX-converted reshape patterns. Fixes #182649 F...
187549,[xpu][test]testing: enable tf32_on_and_off() for XPU devices,https://github.com/pytorch/pytorch/pull/187549,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, ciflow/xpu, module: xpu",2026-06-17T05:10:30Z,2026-07-07T05:13:31Z,1,Related: intel/torch-xpu-ops#2400 Introduce torch/testing/_internal/common.py as a device-agnostic home for TF32-related test utilities that work across both CUDA and XPU backends. Changes: New file common.py: tf32_off(): disables TF32 for cuBLAS/cuDNN (CUDA) and oneDNN/mkldnn (XPU) tf32_on(): enables TF32 for both...
186249,[dynamo] Support time.time as a runtime scalar input,https://github.com/pytorch/pytorch/pull/186249,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,137068,137068,,,,0,1,1,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-04T18:31:26Z,2026-07-07T05:05:55Z,1,"Stack from ghstack (oldest at bottom): -> #186249 Dynamo wraps module-level C functions such as time.time as SkipFunctionVariable. A zero-argument time.time() call therefore fell through to the generic skipped builtin path, which emitted the unsupported-builtin graph-break warning even though the call can be modeled..."
186261,Fix pre-dispatch tracing for saved tensor detach,https://github.com/pytorch/pytorch/pull/186261,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,137009,137009,,,,0,4,1,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-06-04T19:58:40Z,2026-07-07T05:03:06Z,1,"Stack from ghstack (oldest at bottom): -> #186261 Autograd saved-variable unpacking uses TensorImpl shallow-copy detach helpers, but those helpers only considered normal C++ TorchDispatchModeTLS and the tensor Python dispatch key before routing detach through Python. Pre-dispatch proxy mode lives in Python's pre-dis..."
187890,Fix Dynamo metadata sync after inplace tensor mutations,https://github.com/pytorch/pytorch/pull/187890,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185888,185888,186205,,,0,2,2,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-23T00:26:30Z,2026-07-07T04:57:36Z,1,"Stack from ghstack (oldest at bottom): -> #187890 TensorVariable tracks cached metadata such as size, stride, and contiguity. Dynamo refreshed those cached fields after fake execution only when an inplace tensor method had tensor arguments. That missed metadata-only mutations whose inputs are integer/list metadata,..."
186375,[CUDA] Change nvmlDeviceGetGpuFabricInfoV assertion to warning,https://github.com/pytorch/pytorch/pull/186375,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,186374,186374,186374,185946,185946,,0,1,3,,,"triaged, open source, topic: not user facing",2026-06-05T14:34:23Z,2026-07-07T04:56:49Z,1,"nvmlDeviceGetGpuFabricInfoV is only supported on Hopper and above, but get_fabric_access wraps it in an assertion, causing a crash on Orin and other pre-Hopper devices instead of falling back gracefully. Similar to #185946, this PR replaces the assertion with a warning and allows control flow to continue. Fixes #186..."
186263,Increase small fake tensor constant limit,https://github.com/pytorch/pytorch/pull/186263,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,6,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-cpu-test / test-osdc (cpu_inductor_torchbench, 1, 2, mt-l-x86iamx-8-64, unstable)",statusCheckRollup,,136642,136642,,,,3,5,4,9c26eabeaac49c28b47fca987d4327973ad3b7f95e31b407a9018415b9ea78c8,58d971fbc91cc4eac132b658bb6a4fdf5148b716d7b2f88d48e2789efdcaf5f2,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T20:28:26Z,2026-07-07T04:55:06Z,1,"Stack from ghstack (oldest at bottom): -> #186263 Fake tensor and proxy tensor tracing only kept tensor constants with one element. Shape tensors built from Python shape lists, such as torch.as_tensor([[64, 64], [32, 32]]), therefore lost their constant payload even though they are small. Later scalar uses of those..."
186280,Disable profiler Python tracing during compile,https://github.com/pytorch/pytorch/pull/186280,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,136404,136404,,,,0,1,1,,,"release notes: profiler, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-04T22:22:59Z,2026-07-07T04:50:59Z,1,"Stack from ghstack (oldest at bottom): -> #186280 Lazy torch.compile can run substantial Dynamo and Inductor compiler code inside a user torch.profiler.profile(with_stack=True) region. with_stack enables Kineto's Python tracer, so the profiler records compiler Python frames and later has to materialize a very large..."
186287,Reject non-functional export decomposition entries,https://github.com/pytorch/pytorch/pull/186287,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,135759,135759,,,,0,2,1,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-06-04T23:23:47Z,2026-07-07T04:46:25Z,1,"Stack from ghstack (oldest at bottom): -> #186287 The export core ATen decomposition table is applied after functionalization, so mutable post-autograd operators cannot appear in the graph where the table is consumed. However, _core_aten_decompositions_post_autograd() used get_decompositions() on operator packets, w..."
184294,Avoid unused storage size guards for indexing,https://github.com/pytorch/pytorch/pull/184294,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,135492,135492,,,,0,1,1,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T23:24:01Z,2026-07-07T04:40:13Z,1,"Stack from ghstack (oldest at bottom): -> #184294 Use the actual emitted indexing expressions to decide 32-bit indexing guards when possible, while keeping conservative storage-size fallback for indirect or whole-buffer dependencies. This avoids recompilation when a large tensor storage is present but the generated..."
186340,Ensure explicit compile attempts skipped code,https://github.com/pytorch/pytorch/pull/186340,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,2,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,132923,132923,,,,0,12,1,bfd0f67f14ecf57dce5b1533f793866a519747c5532c968f227e12d389edd433,8668354eeb6d926ce92281426100655c4a4d10621ac8335d9d6845263836cfde,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-05T07:29:55Z,2026-07-07T04:35:47Z,1,Stack from ghstack (oldest at bottom): -> #186340 Explicit torch.compile calls could silently run eager when the target function's code object had already been marked with FrameAction.SKIP. The C eval-frame path handled that persisted SKIP before Python convert_frame could see that this was the top-level callable th...
186291,Fix export distribution validation assertions,https://github.com/pytorch/pytorch/pull/186291,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,135061,135061,,,,0,1,1,,,"topic: not user facing, agentic, needs-review",2026-06-05T00:00:09Z,2026-07-07T04:31:47Z,1,"Stack from ghstack (oldest at bottom): -> #186291 Non-strict export traces torch.distributions validation through proxy tensors. Distribution.init and _validate_sample reduced constraint masks with torch._is_all_true(valid) and immediately consumed the result as a Python bool, which forced a data-dependent guard on..."
189070,Accept device-qualified gloo backend in monitored_barrier,https://github.com/pytorch/pytorch/pull/189070,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,"189068, 189069, 189071, 189072, 189073, 189074",,,0,3,6,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-07-06T19:23:48Z,2026-07-07T04:25:35Z,1,"Summary: Under TorchComms a gloo group is recorded with a device-qualified backend string (e.g. ""cpu:gloo"") and wrapped by BackendWrapper, so the exact ""gloo"" match rejected it even though BackendWrapper implements monitoredBarrier. Accept the group when TorchComms is enabled and its backend contains gloo. Test Plan..."
189073,Make subgroup name hash salt collective-consistent,https://github.com/pytorch/pytorch/pull/189073,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,"189068, 189069, 189070, 189071, 189072, 189074",,,0,6,6,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-07-06T19:23:56Z,2026-07-07T04:25:34Z,1,"Summary: _hash_ranks_to_str salted the group-name hash with len(_world.pg_names), which diverges across ranks after an earlier asymmetric-membership new_group (non- member ranks register fewer PGs). A later split_group then computes different names on different ranks; Gloo, which uses the group name as its connectFu..."
189074,Deduplicate comms when destroying a process group,https://github.com/pytorch/pytorch/pull/189074,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,"189068, 189069, 189070, 189071, 189072, 189073",,,0,7,6,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-07-06T19:23:59Z,2026-07-07T04:25:33Z,1,"Summary: destroy_process_group looped over pg._device_types and finalized each backend's comm. A gloo group reports both ""cuda"" and ""cpu"" device types backed by the same BackendWrapper/comm, so the loop finalized that comm twice; finalize() is not idempotent and raised ""already finalized"", crashing teardown (e.g. Me..."
189072,Create TorchComms subgroups on the correct device,https://github.com/pytorch/pytorch/pull/189072,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,"189068, 189069, 189070, 189071, 189073, 189074",,,0,5,6,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-07-06T19:23:54Z,2026-07-07T04:25:33Z,1,"Summary: Three fixes so TorchComms subgroups initialize correctly: Pass this rank's indexed device_id to new_comm instead of a device-type-only torch.device. The bootstrap otherwise defaults the device to (group-local rank % device_count), correct only for the world group; a subgroup would create its comm (and lazy..."
189071,Remove new_group to split_group delegation on the TorchComms path,https://github.com/pytorch/pytorch/pull/189071,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,"189068, 189069, 189070, 189072, 189073, 189074",,,0,4,6,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-07-06T19:23:51Z,2026-07-07T04:25:30Z,1,"Summary: Drop the migration shim that made new_group delegate to split_group under TorchComms, and delete _new_group_via_split_group. Subgroup creation goes through new_group/new_comm directly; callers that want to split a communicator call split_group explicitly. This removes a layer that only existed to express ne..."
184997,Fix TVM backend scalar tensor inputs,https://github.com/pytorch/pytorch/pull/184997,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,169188,169188,,,,0,4,5,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic",2026-05-23T17:17:30Z,2026-07-07T03:30:38Z,2,"Stack from ghstack (oldest at bottom): -> #184997 The TVM backend filtered runtime arguments by rank before binding them to the TVM graph executor. Rank-0 tensor inputs can represent scalar nn.Parameters captured by Dynamo, so active scalar inputs were never passed to TVM and the runtime used its default zero value..."
188095,[ROCm][FlexAttention] Default PRESCALE_QK and auto-detect contiguous blocks,https://github.com/pytorch/pytorch/pull/188095,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,3,2,,,"module: rocm, triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-24T20:14:22Z,2026-07-07T03:23:43Z,2,"On ROCm, opt Flex Attention's forward into two existing kernel fast paths by default: PRESCALE_QK — fold the softmax scale + change-of-base into the QK matmul feed. Enabled by default only on ROCm and only when no additive score_mod is present (score_mod is the identity); an additive bias would otherwise be exponent..."
188546,[Inductor] Fix log_sigmoid_forward flushing float32 subnormals to +0.0 on CUDA,https://github.com/pytorch/pytorch/pull/188546,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188541,188541,188541,,,,0,3,2,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-30T12:39:25Z,2026-07-07T03:07:28Z,2,"Summary Fixes #188541 Root cause: Triton kernels run with FTZ enabled by default (disable_ftz: False in triton_meta). NVIDIA libdevice does NOT call log1pf as an opaque intrinsic — LLVM inlines it as a polynomial of fma.rn.ftz.f32 PTX instructions. For subnormal inputs, the final reconstructing FMA in that polynomia..."
187876,Fix mark dynamic half bounded range,https://github.com/pytorch/pytorch/pull/187876,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, ciflow/trunk, topic: bug fixes, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-06-22T21:23:59Z,2026-07-07T03:07:05Z,2,cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
189101,[inductor] Fix data-dependent GuardOnDataDependentSymNode in get_fill_order,https://github.com/pytorch/pytorch/pull/189101,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-06T21:59:19Z,2026-07-07T02:59:08Z,2,"Stack from ghstack (oldest at bottom): -> #189101 get_stride_order / get_fill_order are called in several places without a shape_env, most notably from GraphLowering.run_node when a node needs a fixed layout: ir.get_stride_order(n.meta[""val""].stride()) With no shape_env, get_fill_order took the plain argsort path, w..."
187434,Support autograd for Tensor?[] (optional tensor list) outputs,https://github.com/pytorch/pytorch/pull/187434,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, ciflow/inductor",2026-06-16T12:20:40Z,2026-07-07T14:52:16Z,1,Support autograd for Tensor?[] (optional tensor list) outputs of custom ops Part of the effort of enabling the torch.compile support for the Transformer Engine library. Custom ops returning Tensor?[] couldn't propagate gradients — the present tensors were silently detached from the autograd graph. Three fixes: torch...
188666,Enable xpu for TestRefs and TestDecomp.,https://github.com/pytorch/pytorch/pull/188666,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/trunk, ciflow/xpu",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, ciflow/xpu",2026-07-01T13:25:43Z,2026-07-07T02:30:51Z,2,"For #114850, we will port distributed tests to Intel GPU. We could enable Intel GPU with following methods and try the best to keep the original code styles: use ""instantiate_device_type_tests"" to enabl XPU for some test path cc @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta @msaroufim @dcci @aditv..."
178493,[c10d][NCCL] avoid reusing scalable init store keys across communicat…,https://github.com/pytorch/pytorch/pull/178493,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,178473,178473,178473,,,,0,2,0,,,"triaged, open source, Stale, release notes: distributed (c10d)",2026-03-26T08:43:49Z,2026-07-07T02:20:16Z,2,"Summary Fix a scalable NCCL initialization bug where allgatherUniqueNCCLIDs() reused the same store key namespace across different communicator initialization rounds. Under scalable init, later communicator-init rounds may observe stale ncclUniqueId values left behind by earlier rounds, which can lead to mismatched..."
189107,Introduce Inductor lowering for _philox_uniform,https://github.com/pytorch/pytorch/pull/189107,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-06T22:43:55Z,2026-07-07T01:54:01Z,2,Stack from ghstack (oldest at bottom): -> #189107 #188495 Allows for fusion with surrounding ops when utilized for e.g. dropout and guaranteed bitwise-equivalent with eager. Sidesteps the perf bug for uniform fixed in #188495 since that only affects fallbacks. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-...
188000,Vectorize LayerNormSecondPass null affine cases,https://github.com/pytorch/pytorch/pull/188000,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: cpu, module: arm, module: intel, ciflow/trunk, release notes: nn",2026-06-23T23:48:59Z,2026-07-07T01:53:56Z,2,"Split LayerNormSecondPass into explicit vectorized paths for the gamma/beta null combinations instead of handling them with one scalar fallback loop. The both-present affine case already used vec::map3. This change adds vec::map/vec::map2 paths for the no-affine, beta-only, and gamma-only cases, preserving the exist..."
188312,[cpu] Address clang-tidy warnings in vec(256/512)_qint.h,https://github.com/pytorch/pytorch/pull/188312,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188311, 188451",,,0,8,0,,,"module: cpu, open source, ciflow/trunk, topic: not user facing",2026-06-26T23:03:18Z,2026-07-07T01:53:50Z,2,"Stack from ghstack (oldest at bottom): -> #188312 #188451 #188311 Silences a number of compiler warnings. The motivation for this PR stack is that cpp_wrapper mode utilizes the Vectorized interface on CPU, so my ongoing attempts to fix up tests in that mode had test failures drowning in a sea of compiler warnings. T..."
182784,"[xpu][test] Port distributed _tools,algorithms,launcher tests cases on Intel GPUs",https://github.com/pytorch/pytorch/pull/182784,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/h100-distributed,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,5,0,,,"triaged, open source, topic: not user facing, ciflow/h100-distributed",2026-05-07T06:42:51Z,2026-07-07T01:50:18Z,2,"For #114850, we will port distributed tests to Intel GPU. We will enable Intel GPU with following methods and keep the original code styles: Example: ""torch.accelerator.current_accelerator()"" to determine the accelerator backend enabled XPU for the following files: test/distributed/_tools/test_fake_collectives.py te..."
189016,[Test] Improve test reuse in test_torch.py for out-of-tree backends,https://github.com/pytorch/pytorch/pull/189016,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-07-06T08:40:46Z,2026-07-07T01:49:04Z,2,"Summary This PR is part of a series of PRs that incrementally refactor test_torch.py to improve test reuse for out-of-tree backends. It moves generic device-aware tests from TestTorch to TestTorchDeviceType, replacing manual device iteration with the injected device parameter. Changes Move the following tests from T..."
182199,[xpu][test] Port distributed _shard tests cases on Intel GPUs,https://github.com/pytorch/pytorch/pull/182199,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,15,0,,,"triaged, open source, topic: not user facing, release notes: distributed (checkpoint), keep-going, ciflow/xpu, module: xpu",2026-05-02T00:17:19Z,2026-07-07T01:48:19Z,2,"For #114850, we will port distributed tests to Intel GPU. We will enable Intel GPU with following methods and keep the original code styles: Example: ""torch.accelerator.current_accelerator()"" to determine the accelerator backend enabled XPU for the following files: test/distributed/checkpoint/_experimental/test_stag..."
182379,"[DTensor] Custom algorithm dispatch mechanism for DTensor, one-sided GEMM algorithm",https://github.com/pytorch/pytorch/pull/182379,draft/noise,13,"PR is draft; CI lane labels: ciflow/dtensor, ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,182382,,,,0,17,0,,,"open source, Stale, release notes: distributed (c10d), ciflow/inductor, release notes: distributed (dtensor), ciflow/h100-symm-mem, ciflow/torchtitan, ciflow/dtensor",2026-05-04T22:21:03Z,2026-07-07T01:08:45Z,2,This PR adds a mechanism to DTensor for selecting alternate operator implementations after normal sharding propagation has selected a strategy. This PR is a prototype to and is part of the one-sided DTensor work in #182382. Behavior This PR introduces an optional OpAlgorithm field for OutputSharding that can be set...
180608,[5.5/11][aoti] Add V2 runtime test via ctypes,https://github.com/pytorch/pytorch/pull/180608,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, topic: not user facing, module: inductor, ciflow/inductor, meta-exported",2026-04-16T21:02:48Z,2026-07-07T01:08:42Z,2,"Summary: Add test_run_v2_interface to AOTInductorArrayRefTestsTemplate which exercises the V2 minimal-arrayref interface end-to-end: compiles a model with nn.Linear to a .so, dlopens it via ctypes, creates a model instance via the C-ABI AOTInductorModelCreate (which loads constants from the embedded blob), marshals..."
180885,CUDA: unrolled kernels for upsample_linear1d forward/backward when grid is small,https://github.com/pytorch/pytorch/pull/180885,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, Stale, release notes: rocm, release notes: cuda",2026-04-20T18:03:02Z,2026-07-07T01:08:40Z,2,Summary Adds alternative CUDA implementations for upsample_linear1d forward and backward that map each output sample to its own thread (unrolled over batch × channel × output width). The existing kernels parallelize only over output_width and loop over batch and channels inside each thread; the new path improves occ...
189069,Route only TorchComms-owned backends through comms,https://github.com/pytorch/pytorch/pull/189069,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,"189068, 189070, 189071, 189072, 189073, 189074",,,0,2,0,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-07-06T19:23:46Z,2026-07-07T01:03:17Z,2,Summary: Add _torchcomms_handles_backend() and gate the TorchComms path (in init_process_group backend auto-qualification and _new_process_group_helper) on it. Backends TorchComms does not own -- custom c10d plugins such as mooncake -- then keep their native c10d creation path instead of being forced through new_com...
180283,[cuBLAS][cuBLASLt] Thread-local cuBLAS(Lt) workspace maps,https://github.com/pytorch/pytorch/pull/180283,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/h100-symm-mem, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,188721,,,,0,20,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ci-no-td, ciflow/h100, ciflow/h100-symm-mem, ciflow/b200, ciflow/torchtitan",2026-04-13T21:56:29Z,2026-07-07T00:07:43Z,2,"Following #167928 and some other more recent PRs that uncovered more than desirable CPU overhead on the cuBLASLt hot path e.g., #179022 #174594 I was thinking about whether we really needed to have cuBLAS(Lt) handles as part of the key to the map for workspaces. In fact the handle-pool with per-thread handles means..."
188466,Fix inductor scatter crash with empty index,https://github.com/pytorch/pytorch/pull/188466,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188213,188213,188213,,,,0,2,0,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-29T18:53:57Z,2026-07-07T00:07:42Z,2,"Fixes #188213 Fixes inductor scatter crash when scatter receives empty index tensors by moving dtype validation after the empty-check, which matches eager mode's early return behavior for zero-element indices cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv..."
189068,Register nccl-lazy TorchComms backend,https://github.com/pytorch/pytorch/pull/189068,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,"189069, 189070, 189071, 189072, 189073, 189074",,,0,1,0,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-07-06T19:23:43Z,2026-07-07T00:07:18Z,2,"Summary: Register NCCL_LAZY (""nccl-lazy"") as a first-class backend: add it to Backend.backend_list, mark it cuda-capable in backend_capability, and map it to ProcessGroup.BackendType.CUSTOM in backend_type_map. This lets new_group(backend=""nccl-lazy"", use_local_synchronization=True) build a members- only, per-peer l..."
182864,[donotrvw][broken][WIP] spmd_types + PP: _SpmdTypesMeta and nullcontext decorator fix,https://github.com/pytorch/pytorch/pull/182864,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,181398,,,0,3,0,,,"Stale, ciflow/torchtitan",2026-05-07T18:38:58Z,2026-07-06T23:51:34Z,2,Stack from ghstack (oldest at bottom): -> #182864 #181519 #181398 #180880 Add _SpmdTypesMeta(_TensorMeta) to pipelining/_utils.py for preserving spmd_types annotations across PP stage boundaries Update extract_tensor_meta/extract_tensor_metas with optional device_mesh Update PipelineStage.init with device_mesh param...
189099,[Profiler][NOT FOR LAND] CI test debugging,https://github.com/pytorch/pytorch/pull/189099,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,topic: not user facing,2026-07-06T21:52:11Z,2026-07-06T23:37:02Z,2,
187891,device_mesh: fall back to new_group when parent backend lacks split s…,https://github.com/pytorch/pytorch/pull/187891,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/torchtitan",2026-06-23T00:27:53Z,2026-07-06T23:07:52Z,2,"…upport init_device_mesh with eager_init=True sets bound_device_id on the default process group, which routes subgroup creation through split_group(). That call raises 'No backend for the parent process group or its backend does not support splitting' for backends whose supportsSplitting() returns false (e.g. XCCL,..."
188903,[HOP] Adding input mutation support for switch (2/2),https://github.com/pytorch/pytorch/pull/188903,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-03T16:36:19Z,2026-07-06T23:06:37Z,2,In #188199 the gen_schema function was introduced in preparation for the input_mutation. This PR is a follow-up and introduces the functionalization components in order to support the proper input mutation for switch. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @we...
182841,[CI] No-op docker change to test current build state,https://github.com/pytorch/pytorch/pull/182841,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"Stale, topic: not user facing",2026-05-07T16:29:22Z,2026-07-06T22:51:36Z,2,Adding a comment to trigger docker-builds CI and confirm whether the builds fail in their current state (before any fixes). Authored with Claude.
182838,"[CI] Fix docker builds: pin miniforge, use --extra-index-url",https://github.com/pytorch/pytorch/pull/182838,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"Stale, topic: not user facing",2026-05-07T16:03:18Z,2026-07-06T22:51:34Z,2,"Docker image builds are broken because pip install --index-url <pytorch-wheel-index> replaces PyPI as the sole package source. When pip's resolver re-evaluates the dependency graph, previously pip-installed packages (numpy, transformers, etc.) can be removed because the PyTorch wheel index carries stale or incompati..."
182183,[PT2][export] Add unsafe_allow_callable_serialization for scoped callable serialization,https://github.com/pytorch/pytorch/pull/182183,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, ciflow/trunk, ciflow/inductor, release notes: export, meta-exported",2026-05-01T20:40:18Z,2026-07-06T22:51:30Z,2,"Summary: Add unsafe_allow_callable_serialization, a context manager in torch._export.serde that temporarily registers Python callables for export serialization and unregisters them on exit. This gates serialization/deserialization of arbitrary Python callables (e.g. predispatch wrappers like JVP/vmap) behind explici..."
183759,[Testing] Expansion of non-contiguous tensor ops testing to cover more complicated and real-world applications,https://github.com/pytorch/pytorch/pull/183759,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,183758,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/mps",2026-05-14T19:16:25Z,2026-07-06T22:34:52Z,2,Issue #183758 Expanding common_utils.py::noncontiguous_like utility function with missing cases for non-contiguous tensor generation noncontiguous_channels_last_like noncontiguous_transposed_like noncontiguous_sliced_like noncontiguous_channels_last_sliced_like noncontiguous_permuted_like Adding a NoncontiguousType...
181106,Fix torch.compile crash with inference_mode and view ops in forward (#180823),https://github.com/pytorch/pytorch/pull/181106,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,180823,180823,180823,,,,0,5,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor",2026-04-22T13:18:36Z,2026-07-06T22:13:47Z,2,"Fixes #180823 Root Cause torch.compile with inference_mode() and view ops crashes during the AOT Autograd functionalization phase. When AOT Autograd replays the FX graph for functionalization, _enter_inference_mode nodes actually execute, causing tensors created during functionalization to be inference tensors. C++..."
184133,[aot_autograd] Reduce runtime_wrapper Python overhead: 4 micro-optimizations,https://github.com/pytorch/pytorch/pull/184133,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,161783,161783,161783,,,,0,7,0,,,"triaged, open source, topic: not user facing, ciflow/inductor",2026-05-17T23:55:23Z,2026-07-06T22:02:10Z,2,"Summary `torch.compile`'d functions pay a Python-layer overhead on every call through `runtime_wrapper` / `_codegen_runtime_wrapper`, even when the compiled kernel itself is fast. This PR applies four targeted micro-optimizations that cut those allocations. Fix A — Skip profiler closure when profiling is off `runtim..."
182687,Support generator arg in dirichlet and beta torch.distributions API,https://github.com/pytorch/pytorch/pull/182687,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"module: distributions, Stale, release notes: python_frontend",2026-05-06T18:53:28Z,2026-07-06T21:59:31Z,2,"Adds an optional generator: torch.Generator | None parameter to Beta and Dirichlet, enabling users to sample with a custom RNG instead of the global state. This is needed for torchvision transforms (e.g., Mixup, CutMix) to be able to support user defined generators. Authored with Claude. cc @fritzo @neerajprad @alic..."
175081,Add shim config includes for linting,https://github.com/pytorch/pytorch/pull/175081,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/binaries, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,174682,174682,,,,0,3,0,,,"open source, Stale, ciflow/binaries, ciflow/trunk, topic: not user facing",2026-02-16T11:38:51Z,2026-07-06T21:59:20Z,2,This PR adds shims for auto generated header files that may not be easily generated in some circumstances purely for linting purposes. This is done by adding a special shim directory and including it as a last back stop in the list of include directories used by the clangtidy lint adapter. Fixes #174682. Stack from...
188249,[inductor] Fix bool equality value range analysis,https://github.com/pytorch/pytorch/pull/188249,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188231,188231,188231,,,,0,1,0,,,"module: cpu, triaged, open source",2026-06-26T06:19:53Z,2026-07-06T21:30:13Z,2,"Inductor C++ codegen asks bounds analysis to interpret lowered loop bodies before emitting kernels. The torch.isclose reference path starts by comparing its inputs with eq, and bool tensor inputs can therefore reach ValueRanges.eq as boolean ranges. ValueRanges.eq only handled the numeric interval case, so it tried..."
132135,"[torch.special] Adding betainc, betaincc, betaincinv, betainccinv, betaln and beta with backward operation",https://github.com/pytorch/pytorch/pull/132135,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,132133,132133,132133,,,,0,28,0,,,"oncall: distributed, module: cpu, triaged, open source, Stale, release notes: cpp, skip-pr-sanity-checks, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2024-07-30T12:09:47Z,2026-07-06T21:23:21Z,2,"Fixes #132133 #58700 and #129836 don't support GPU, due to 'gpu_kernel_with_scalars' with only one scalar value. I changed it to 'gpu_kernel_with_scalars_tenary' which is new function I made. The contribution point in #129836 except re-indentation is #include <ATen/ops/special_betainc.h> #include <ATen/ops/special_b..."
188286,Fix out= variant recompilations for CompositeImplicitAutograd linalg ops,https://github.com/pytorch/pytorch/pull/188286,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,187953,187953,"135859, 187953",187952,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend, module: dynamo, ciflow/inductor",2026-06-26T18:42:26Z,2026-07-06T20:53:33Z,2,"Issue Fixes #187953 Related: #135859 (parent issue) and #187952, which fixed the same class of out= over-recompilation for torch.max/min/topk and explicitly deferred the CompositeImplicitAutograd linalg out= overloads handled here. Summary Under torch.compile with dynamic shapes, the out= variants of CompositeImplic..."
188451,[cpu] Address clang-tidy warnings in vec128_float_neon.h,https://github.com/pytorch/pytorch/pull/188451,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188311, 188312",,,0,5,0,,,"module: cpu, open source, ciflow/trunk, topic: not user facing",2026-06-29T18:07:01Z,2026-07-06T20:53:05Z,2,"Stack from ghstack (oldest at bottom): #188312 -> #188451 #188311 Silences a number of compiler warnings. The motivation for this PR stack is that cpp_wrapper mode utilizes the Vectorized interface on CPU, so my ongoing attempts to fix up tests in that mode had test failures drowning in a sea of compiler warnings. T..."
188311,[cpu] Address clang-tidy warnings in vec(256/512)_int.h,https://github.com/pytorch/pytorch/pull/188311,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188312, 188451",,,0,8,0,,,"module: cpu, open source, ciflow/trunk, topic: not user facing",2026-06-26T23:03:13Z,2026-07-06T20:52:58Z,2,"Stack from ghstack (oldest at bottom): #188312 #188451 -> #188311 Silences a number of compiler warnings. The motivation for this PR stack is that cpp_wrapper mode utilizes the Vectorized interface on CPU, so my ongoing attempts to fix up tests in that mode had test failures drowning in a sea of compiler warnings. T..."
188902,[ATen] Add reduction tag to missing base reduction ops,https://github.com/pytorch/pytorch/pull/188902,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,129020,129020,129020,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-07-03T15:48:40Z,2026-07-06T20:52:22Z,2,"Fixes #129020 (partial) Follow-up to #165155. Extends the reduction tag to base reduction operators that were not addressed in the original PR. Each newly tagged op is directly analogous to a currently tagged reduction, so this is a mechanical extension of the existing tag definition — no new semantics. Newly tagged..."
180408,Fix torch.eye CPU performance cliff at d=182 (#48251),https://github.com/pytorch/pytorch/pull/180408,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,48251,48251,48251,,,,0,1,0,,,"triaged, open source",2026-04-15T00:36:23Z,2026-07-06T20:31:27Z,2,"Summary Fixes #48251 torch.eye(d) on CPU has a sharp performance cliff at d=182, where it becomes roughly 10x slower. This happens because the internal zero_() call switches from a fast memset path to a slow parallel TensorIterator path when numel >= GRAIN_SIZE (32768), and 182 * 182 = 33124 just crosses that thresh..."
187755,[testintro] CI: post tests added/removed as a PR comment,https://github.com/pytorch/pytorch/pull/187755,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187754, 187756",,,0,2,0,,,topic: not user facing,2026-06-20T01:33:36Z,2026-07-06T20:30:01Z,2,"Stack from ghstack (oldest at bottom): #187756 -> #187755 #187754 Wire the introspection diff into CI so every PR gets a sticky ""Test Change Dashboard"" comment listing the tests it adds/removes, grouped by platform and linked into the PR diff. PR #187756 has an example of the rendered comment. Two-stage workflow_run..."
187675,"[inductor] Add fuse_region to set fusions,memory reuse boundaries",https://github.com/pytorch/pytorch/pull/187675,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187839, 187840, 187841, 187842, 187843, 187844, 187845, 187846",,,0,3,0,,,"topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/torchtitan",2026-06-18T16:24:11Z,2026-07-06T20:29:48Z,2,"Stack from ghstack (oldest at bottom): #187846 #187845 #187844 #187843 #187842 #187841 #187840 #187839 -> #187675 GraphTrainer full-Inductor needs outlined source regions to stay separated through lowering, scheduling, and memory planning. The earlier prototype carried this through control_deps, but that conflated o..."
188571,C++ meta kernel for torch.cat (symint aware),https://github.com/pytorch/pytorch/pull/188571,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"ciflow/inductor, suppress-bc-linter",2026-06-30T17:35:12Z,2026-07-06T20:27:56Z,2,"python meta kernels are slower than if we wrote them in C++ but existing C++ meta kernels are not symint aware. so for structured kernels like torch.cat, we keep the structured registration but have a flag to avoid actually generating a meta kernel via the structured registration. instead we define a new meta kernel..."
189081,Add tp_name property to Variable Tracker,https://github.com/pytorch/pytorch/pull/189081,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-06T20:03:53Z,2026-07-06T20:16:23Z,2,"Add a tp_name property to VariableTracker that returns CPython's exact type name string (e.g. list, NoneType, elipsis) for the Python type the VT represents. This is a step in mirroring CPython's type slot protocol in Dynamo. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blz..."
185208,[PoC] uv.lock dependency-groups for runtime + build deps,https://github.com/pytorch/pytorch/pull/185208,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel, ciflow/docker","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,183913,180250,,,0,7,0,,,"open source, release notes: releng, ciflow/binaries_wheel, ciflow/docker",2026-05-26T09:47:09Z,2026-07-06T20:14:39Z,2,"Proof of concept — not intended to land. Demonstrates the runtime + build dependency design from the requirements-unification RFC (#183913), stacked on the scikit-build-core migration (#180243–#180250). The PoC is the top two commits; everything below is the (well-known) migration stack. What it shows Runtime deps a..."
138068,Add TORCH_CHECK_INDEX in convert_indices_from_coo_to_csr_cpu,https://github.com/pytorch/pytorch/pull/138068,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2024-10-16T09:08:18Z,2026-07-06T20:13:54Z,2,"The to_sparse_csr CPU implementation convert_indices_from_coo_to_csr_cpu doesn't validate the COO indices, which may lead to illegal memory access. This PR fixes that by adding checks on the index before accessing the data. # repro num_nonzeros = 2048 #row,col dense_size = (1024, 256) row_indices = torch.randint(0,..."
188406,[dynamo] preserve Stream/Event subclass type across graph-break rec…,https://github.com/pytorch/pytorch/pull/188406,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188405,188405,188405,,,,0,1,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-06-29T12:04:08Z,2026-07-06T20:13:36Z,2,"A graph-created torch.Stream/torch.Event subclass was reconstructed as the base class after a graph break, silently dropping the subclass type. Backend extensions that rely on receiving their own subclass (the common C++ reinterpret_cast<BackendStream*> pattern) then see the wrong object type / layout and crash. Thr..."
152806,[invoke_subgraph] Force the output stride to be same as eager,https://github.com/pytorch/pytorch/pull/152806,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"Stale, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor",2025-05-05T07:03:00Z,2026-07-06T20:13:35Z,2,Stack from ghstack (oldest at bottom): -> #152806 #152675 #152770 #152772 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben
187778,Flatten all_to_all_nd copy to fix narrow-row throughput,https://github.com/pytorch/pytorch/pull/187778,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,186605,187776,,,0,2,0,,,"open source, Merged, Reverted, ciflow/trunk, release notes: distributed (c10d), ci-no-td, ciflow/h100-symm-mem",2026-06-20T21:29:10Z,2026-07-06T20:13:04Z,2,"Stack from ghstack (oldest at bottom): -> #187778 #187776 The kernel assigned one CTA's worth of threads to a single row of a peer's slot, so a row narrower than the block left most threads idle -- a copy_row_bytes-wide row used only copy_row_bytes/16 of the threads, collapsing to a single thread for 16-byte rows. I..."
183705,[SymmMem] Normalize device index in empty_strided_p2p (#183705),https://github.com/pytorch/pytorch/pull/183705,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"fb-exported, ciflow/trunk, release notes: distributed (c10d), ciflow/h100-symm-mem, meta-exported",2026-05-14T11:26:06Z,2026-07-06T20:12:51Z,2,"Summary: When a device without an explicit index (e.g. torch.device(""cuda"") instead of torch.device(""cuda:0"")) is passed to symm_mem.empty(), device.index() returns -1. This propagates through empty_strided_p2p -> allocator -> NVSHMEMAllocation.device_idx, and eventually into TeamManager::get(). The TeamManager sing..."
186370,inductor: validate square mix-order reductions by common reads,https://github.com/pytorch/pytorch/pull/186370,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-05T13:39:55Z,2026-07-06T20:12:50Z,2,"Co-authored-by: gpt-5.5 Issue Mix-order reduction currently rejects equal (numel, rnumel) groups, which blocks square-shape inner+outer reductions in non-strict mode. Simply allowing equal groups can introduce false positives: same-order reductions may look mixed-order if one node has an unrelated strided read. Fix..."
185109,Remove custom FindCUDAToolkit.cmake module,https://github.com/pytorch/pytorch/pull/185109,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, ciflow/binaries, ciflow/trunk, topic: not user facing",2026-05-25T06:00:22Z,2026-07-06T20:12:49Z,2,FindCUDAToolkit.cmake is our fork of the cmake official module. It contains many patches (some are useless for CUDA 12+) over these years. This PR extracts used patches into cmake/public/cuda.cmake and remove it.
187015,remove spmd_types dependency from nightly wheel,https://github.com/pytorch/pytorch/pull/187015,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, topic: not user facing",2026-06-11T05:05:21Z,2026-07-06T20:12:18Z,2,"Stops spmd_types being a ""hard"" dependency for torch, plan is to move it to torchannex and have guidance be pip install torch torchannex instead."
186324,[FakeTensor] Propagate backend metadata through fake tensor conversion,https://github.com/pytorch/pytorch/pull/186324,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,8,0,,,"triaged, open source, release notes: jit, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-05T04:36:24Z,2026-07-06T20:11:57Z,2,"Backends that attach c10::BackendMeta to a tensor (via the PrivateUse1 backend-metadata mechanism, the same one used for torch.save / torch.load) currently lose that metadata when the tensor is converted to a FakeTensor. FakeTensorConverter.from_real_tensor builds the fake tensor from a fresh meta tensor and never c..."
126348,[AOTAutograd] tweak min-cut partitioner to avoid saving softmax output,https://github.com/pytorch/pytorch/pull/126348,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,4,0,,,ciflow/inductor,2024-05-15T22:54:40Z,2026-07-06T20:11:52Z,2,Stack from ghstack (oldest at bottom): -> #126348 #126477 Right now the linear + cross entropy loss operation (usually to be the last part of a transformer model) does the following thing run matmul to get softmax_input load softmax_input to compute max per row. load softmax_input to compute sum per row load softmax...
188853,Register DTensorSpec as an opaque reference type for torch.export,https://github.com/pytorch/pytorch/pull/188853,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175467,175467,,,,0,2,0,,,"ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-07-02T21:54:43Z,2026-07-06T20:11:31Z,2,Stack from ghstack (oldest at bottom): -> #188853 Fixes #175467
189062,[profiler][cupti] Make CUPTI monitor buffer_size / flush_period_s config-only,https://github.com/pytorch/pytorch/pull/189062,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,topic: not user facing,2026-07-06T18:46:27Z,2026-07-06T20:11:27Z,2,"Stack from ghstack (oldest at bottom): -> #189062 Drop the TORCH_CUPTI_MONITOR_BUFFER_SIZE and TORCH_CUPTI_MONITOR_FLUSH_PERIOD_S env vars: buffer_size and flush_period_s are now plain CuptiMonitor constructor settings (defaulting to the same 4 MiB / 1.0 s), matching the move to config methods over env vars for the..."
186639,[CUDA graphs] Add resolve_and_remap convenience for kernel annotations,https://github.com/pytorch/pytorch/pull/186639,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,2,0,,,ciflow/trunk,2026-06-08T19:31:09Z,2026-07-06T20:11:11Z,2,"Stack from ghstack (oldest at bottom): -> #186639 #186638 Wrap the common end-of-capture sequence -- resolve_pending_annotations() followed by remap_to_exec_graph(graph) -- into a single resolve_and_remap(graph) call. When several graphs are captured in sequence, calling it once per graph after all captures is enoug..."
188832,[pytorch][cuda] make_graphed_callables: expose capture_error_mode (#188832),https://github.com/pytorch/pytorch/pull/188832,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, meta-exported",2026-07-02T17:30:20Z,2026-07-06T20:11:00Z,2,"Summary: Add a capture_error_mode parameter to torch.cuda.make_graphed_callables (default ""global"", preserving current behavior), forwarded to the two internal torch.cuda.graph forward/backward capture contexts. This lets callers select ""thread_local"" so that unsafe CUDA operations on other threads (e.g. an async da..."
187007,Use distinct cache variables for linker option checks,https://github.com/pytorch/pytorch/pull/187007,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: build",2026-06-11T02:54:19Z,2026-07-06T20:10:52Z,2,"Summary This fixes the linker option support helpers so different linker flags get independent CMake check results. add_link_options_if_supported and target_link_options_if_supported currently pass the same _supported variable to check_linker_flag for every flag. Since these checks are cached by variable name, a suc..."
187009,Apply libnop clang warning suppression to TensorPipe subtargets,https://github.com/pytorch/pytorch/pull/187009,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: build",2026-06-11T03:22:40Z,2026-07-06T20:10:50Z,2,Summary This PR broadens the existing TensorPipe/libnop Clang warning workaround so it also applies to TensorPipe subtargets created during add_subdirectory(third_party/tensorpipe). PyTorch already suppresses -Wmissing-template-arg-list-after-template-kw on the main tensorpipe target because vendored libnop triggers...
188976,[HOP] Inductor lowering for switch (1/2),https://github.com/pytorch/pytorch/pull/188976,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-07-05T21:09:03Z,2026-07-06T19:21:25Z,2,"This PR provides the initial Inductor lowering implementation for switch. It deliberately introduces a new IR node Switch and does not consolidate the Conditional. I believe there are several good reasons for this: Semantic differences: switch selects the branches based on an Integer, while cond selects the branch b..."
182062,[CUDA] Early-exit in `release_cached_blocks` if `captures_underway` isn't empty,https://github.com/pytorch/pytorch/pull/182062,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cuda, triaged, open source, module: cuda graphs, Stale, topic: not user facing, module: CUDACachingAllocator",2026-04-30T18:29:36Z,2026-07-06T19:08:20Z,2,"defers calling synchronize_and_free_events so that we don't call synchronize_and_free_events on non-default mempool, otherwise synchronize_and_free_events immediately asserts that it's false after authored with codex cc @ptrblck @msaroufim @jerryzh168 @tinglvv @nWEIdia @mcarilli @ezyang @eellison @penguinwu @BoyuanFeng"
182844,[CI] Pin miniforge to 26.1.1-3 to fix docker builds,https://github.com/pytorch/pytorch/pull/182844,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, topic: not user facing",2026-05-07T16:43:35Z,2026-07-06T19:08:18Z,2,"Miniforge 26.3.2-0 (released 2026-05-02) ships conda 26.3.2 which changed how pip-installed packages are tracked in the environment model (""Do not include pip-installed packages in the list of explicit packages""). This causes pip-installed packages (numpy, transformers, etc.) to go missing during subsequent conda in..."
181215,Add cooperative GPU-sharing infrastructure via CUDA process-checkpoint,https://github.com/pytorch/pytorch/pull/181215,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"Stale, release notes: distributed (c10d), ciflow/torchtitan",2026-04-23T03:34:40Z,2026-07-06T19:08:17Z,2,"Stack from ghstack (oldest at bottom): -> #181215 checkpoint/cuda_checkpoint.py wraps cuCheckpointProcess{Lock,Checkpoint, Restore,Unlock} via ctypes so a process can release its entire CUDA context (VRAM returned to the driver) and reclaim it later with all device pointers intact. checkpoint/coordinator.py + checkp..."
188947,[Bugfix][DTensor] Fix wrong shard offsets for uneven DTensor shards in TP-FSDP metadata,https://github.com/pytorch/pytorch/pull/188947,ready for maintainer decision,45,"CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: distributed (fsdp), ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-07-04T11:29:15Z,2026-07-06T19:01:04Z,2,"Fix wrong shard offsets for uneven DTensor shards in _get_box/_get_box_for Fixes a correctness bug in TP-FSDP shard metadata generation where ShardMetadata offsets are computed incorrectly for DTensors whose shard dimension is not evenly divisible by the mesh size. Root Cause In _get_box(), shard offsets are compute..."
188945,Fix assert_close to support comparisons involving Python and NumPy scalars,https://github.com/pytorch/pytorch/pull/188945,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,161863,161863,161863,,,,0,2,0,,,"triaged, module: numpy, open source, topic: not user facing",2026-07-04T07:47:12Z,2026-07-06T18:59:36Z,2,"Description This PR resolves issue #161863 where torch.testing.assert_close would raise a TypeError (no comparison pair was able to handle inputs) when comparing a torch.Tensor to a Python scalar (int, float, complex, bool) or a NumPy scalar (np.generic). Implementation Details Modified TensorLikePair._process_input..."
188923,test: add gradient health validation tests (fixes #177116),https://github.com/pytorch/pytorch/pull/188923,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,177116,177116,177116,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: mps",2026-07-03T18:27:03Z,2026-07-06T18:59:05Z,2,"Fixes #177116 Summary Adds gradient health validation tests that would catch hardware-specific gradient corruption bugs, such as the MPS gradient bug (#177116). Tests added (1 file, 3 tests) test_gradient_finite_after_backward: Ensures no NaN or Inf in gradients after a normal backward pass test_gradient_not_explodi..."
188969,[docs] document __torch_function__ behavior during autograd hooks,https://github.com/pytorch/pytorch/pull/188969,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,167223,167223,167223,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-07-05T07:53:51Z,2026-07-06T18:57:22Z,2,"Document that subclass __torch_function__ dispatch is disabled during backward autograd hooks, and show how to re-enable it when needed. Fixes #167223 Problem Tensor subclass authors expect __torch_function__ to run consistently whenever they call PyTorch ops, including inside autograd hooks. In practice, hooks regi..."
188885,[DataLoader] Clean up stale Py2 comments and pickle-error warning,https://github.com/pytorch/pytorch/pull/188885,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: dataloader",2026-07-03T08:55:49Z,2026-07-06T18:55:40Z,2,"This PR contains the following changes: Replace the obsolete ""Py2 can't serialize static methods"" rationale (min Python is 3.10) with the real reason: spawn/forkserver pickle worker callables by qualified name. Move two _IterableDatasetStopIteration/_ResumeIteration docstrings inside their classes (they were module-..."
188883,Handle empty autograd sections in backend YAML,https://github.com/pytorch/pytorch/pull/188883,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188882,188882,188882,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-07-03T08:02:08Z,2026-07-06T18:51:38Z,2,"Fixing an Issue Issue Fixes #188882 Summary Empty YAML keys parse as None. parse_backend_yaml() already normalizes empty supported: and symint: sections to empty lists, but autograd: still went through the list type check as None and raised AssertionError. What Problem This Solves External backend YAML supports seve..."
188859,Raise a clear error when linalg.eigh/svd workspace exceeds 32-bit LAPACK limit,https://github.com/pytorch/pytorch/pull/188859,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,92141,92141,92141,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend, topic: bug fixes",2026-07-02T22:30:35Z,2026-07-06T18:51:14Z,2,"Fixes #92141 Root cause The CPU LAPACK path indexes its workspace with a 32-bit int. LAPACK returns the optimal workspace size from its query call as a floating point value, which lapack_work_to_int() rounds and casts to int. For a large matrix the required workspace can exceed INT_MAX — for syevd with eigenvectors..."
188831,Fix module re-export in tensor/experimental API,https://github.com/pytorch/pytorch/pull/188831,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,171905,171905,171905,,,,0,2,0,,,"triaged, open source, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-07-02T17:24:49Z,2026-07-06T18:50:44Z,2,"Fixes #171905 Description This PR addresses the issue with type checkers and static analysis tools complaining about manual __module__ reassignments in torch.distributed.tensor.experimental. The original issue suggested using typing_extensions.TypeAliasType for this purpose. However, context_parallel, local_map, and..."
188893,[ONNX] Clarify that output_names labels outputs but does not reorder them,https://github.com/pytorch/pytorch/pull/188893,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,165758,165758,,,,,0,1,0,,,"triaged, open source, release notes: onnx",2026-07-03T12:13:38Z,2026-07-06T18:47:17Z,2,"Fixes #165758 Summary Users passing output_names to torch.onnx.export often expect it to select or reorder model outputs by name. It does not: the exporter flattens the model outputs (for a dict return, in the dict's iteration order) and applies the given names positionally, so a mismatched ordering silently mislabe..."
188874,Enable XPUGraph native recording mode,https://github.com/pytorch/pytorch/pull/188874,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, module: xpu",2026-07-03T02:42:25Z,2026-07-06T18:47:12Z,2,requires oneapi 2026.1 cc @gujinghui @EikanWang @fengyuan14 @guangyey
188857,Fix type confusion in Stream comparison against non-Stream objects,https://github.com/pytorch/pytorch/pull/188857,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188033,188033,188033,,,,0,1,0,,,"triaged, open source",2026-07-02T22:26:25Z,2026-07-06T18:46:58Z,2,"Fixes #188033 Root cause THPStream_richcompare (torch/csrc/Stream.cpp) only special-cased None and otherwise did reinterpret_cast<THPStream*>(other) unconditionally, then read stream_id / device_index / device_type off it. When other is any non-Stream object (e.g. stream == 5), those reads land at THPStream field of..."
188804,[MPS] Add solo large-tensor allocator + torch.mps.set_allocator_settings(),https://github.com/pytorch/pytorch/pull/188804,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,5,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-07-02T13:56:17Z,2026-07-06T18:45:20Z,2,"Summary Adds an opt-in ""solo"" allocation path in MPSHeapAllocatorImpl for large tensors, plus a public torch.mps.set_allocator_settings(str) API to enable it and tune the threshold. Rationale: the default MPS allocator grows a 1 GB heap block for large single-tensor allocations, which is wasteful when the caller onl..."
186294,[dynamo] Proxy OptimizedModule __class__ to wrapped module,https://github.com/pytorch/pytorch/pull/186294,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,134950,134950,,,,0,1,0,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-06-05T00:21:03Z,2026-07-06T18:38:06Z,2,"Stack from ghstack (oldest at bottom): -> #186294 torch.compile(nn.Module) returns an OptimizedModule wrapper that proxies most normal attribute access to the wrapped module. However, class was still resolved on the wrapper itself, so libraries that inspect model.class.forward saw OptimizedModule.forward instead of..."
186306,Support tensor torch._assert in export,https://github.com/pytorch/pytorch/pull/186306,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,134533,134533,,,,0,1,0,,,"module: dynamo, ciflow/inductor, release notes: export, ciflow/torchtitan, agentic, needs-review",2026-06-05T02:14:58Z,2026-07-06T18:35:10Z,2,"Stack from ghstack (oldest at bottom): -> #186306 torch._assert only had special Dynamo handling for statically true Python constants and symbolic expressions. Tensor-valued conditions fell through to the eager Python implementation, whose if not condition forced proxy tensors through Python truthiness. During expor..."
186874,Fix Dynamo resume lifetime for _EnableTorchFunction,https://github.com/pytorch/pytorch/pull/186874,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,182318,182318,,,,0,6,0,dba17529df278dcb78d18c55dc78fe05a50ce889ba140e3ed9642a4c9bf048aa,947c5852ab5fe333ea1fd60a9d66d5f52bb51430ca4c776da0862153e52eab8c,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-10T01:57:59Z,2026-07-06T18:26:10Z,2,"Stack from ghstack (oldest at bottom): -> #186874 Dynamo graph breaks at unsupported calls normally execute the unsupported instruction in the generated outer frame, then pass the resulting values into the resume function. For torch._C._EnableTorchFunction(), that constructs an RAII guard in the wrong frame. The gua..."
184303,Fix pattern matcher replacement graph cleanup safety,https://github.com/pytorch/pytorch/pull/184303,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,134363,134363,,,,0,4,0,6a2cde32ce7b86c40d406c1a2385d6a0895b4aad157366f278056d66b7eb98a2,12c1bbba278b75ff798674c467d5d0736b6cca14b07f774ec23b2f4131d5709e,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T01:42:05Z,2026-07-06T18:21:54Z,2,"Stack from ghstack (oldest at bottom): -> #184303 Skip fwd_only replacement graph cleanup when the traced graph contains mutation or side effects, and preserve nested output aliasing in remove_noop_ops. Fixes #134363 Generated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @z..."
186315,Preserve extra autocast keys on FakeTensor,https://github.com/pytorch/pytorch/pull/186315,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,133924,133924,,,,0,3,0,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-06-05T03:02:20Z,2026-07-06T18:18:35Z,2,"Stack from ghstack (oldest at bottom): -> #186315 XLA:CUDA tensors can carry an extra AutocastCUDA dispatch key even though their logical device is xla. FakeTensor reconstruction currently rebuilds wrapper tensors with device_for_backend_keys=xla, which restores the default XLA key set but drops extra keys from the..."
186319,Avoid full simplify in vector-range index analysis,https://github.com/pytorch/pytorch/pull/186319,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,133734,133734,,,,0,1,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T03:42:53Z,2026-07-06T18:15:25Z,2,"Stack from ghstack (oldest at bottom): -> #186319 The C++ vector stride helper first rewrites lane-uniform FloorDiv and ModularIndexing terms into temporary symbols. It then ran full sympy.simplify over the entire index expression before computing the stride. On dynamic-shape models such as RetinaNet, those index ex..."
186321,Fix PythonMod axiom reuse after size refinement,https://github.com/pytorch/pytorch/pull/186321,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,133655,133655,,,,0,1,0,,,"release notes: fx, module: dynamo, ciflow/inductor, agentic, needs-review",2026-06-05T04:02:57Z,2026-07-06T18:12:26Z,2,"Stack from ghstack (oldest at bottom): -> #186321 When a scalar modulo expression is checked before its operands are known size-like, ShapeEnv records the assertion with PythonMod semantics. Later _check_is_size calls refine the same symbols to non-negative ranges, so a new Python % expression can simplify to SymPy..."
186322,Add explicit schema registration for TorchBind methods,https://github.com/pytorch/pytorch/pull/186322,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,133592,133592,,,,0,1,0,,,"release notes: cpp, agentic, needs-review",2026-06-05T04:22:42Z,2026-07-06T18:10:11Z,2,"Stack from ghstack (oldest at bottom): -> #186322 TorchBind custom class methods could only be registered through torch::class_::def(name, fn), which infers the method schema from the C++ callable. Inferred schemas cannot encode Tensor aliasing and mutation annotations, so a method that mutates a tensor argument cou..."
186328,Preserve strides for functionalized in-place mutations,https://github.com/pytorch/pytorch/pull/186328,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers,review diff and decide acceptance path,,CLEAN,False,0,0,,,,133512,133512,,,,0,6,0,,,"release notes: fx, topic: not user facing, agentic, needs-review",2026-06-05T05:11:03Z,2026-07-06T18:06:16Z,2,"Stack from ghstack (oldest at bottom): -> #186328 Functionalization rewrites an in-place op such as x.add_(1) into its functional variant followed by replace_. The functional variant can return a fresh contiguous tensor even when eager in-place mutation would preserve the original tensor's size, stride, and storage..."
186330,Fix compiled autograd.Function manual vmap rules,https://github.com/pytorch/pytorch/pull/186330,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,"133493, 134784","133493, 134784",,,,0,4,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T05:50:04Z,2026-07-06T18:03:39Z,2,"Stack from ghstack (oldest at bottom): -> #186330 This consolidates two reports of the same root cause: under torch.compile, compiled autograd.Function subclasses with a manual vmap staticmethod (generate_vmap_rule=False) fall off functorch's custom_function_call path, so the user-defined vmap rule is dropped and th..."
187961,[inductor] Add compile-time per-subkernel autotuning for combo kernels,https://github.com/pytorch/pytorch/pull/187961,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"182577, 184955, 188119, 188203, 189001, 189002, 189003",,,0,8,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-23T18:04:06Z,2026-07-06T17:18:21Z,2,"Stack from ghstack (oldest at bottom): #189003 #189002 #189001 #188203 #188119 #184955 #182577 -> #187961 This PR adds per-subkernel block tuning at compile time. Each subkernel is autotuned on its own at codegen time, the winning blocks are baked into the combo as tl.constexpr, and combo then only autotunes the ker..."
184305,Fix zero-bias SDPA flash selection,https://github.com/pytorch/pytorch/pull/184305,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,3,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_models_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,133462,133462,,,,6,5,0,04bcf0db01b2e30b83f9a09988a6119c9f618d2d44e223d4a434ea9e484ebba9,b8cebc1f3dac00ba32aa47feb301b8366ef8868864ba101f1c6f2156c4be1afc,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T02:17:32Z,2026-07-06T17:13:30Z,2,"Stack from ghstack (oldest at bottom): -> #184305 Fold expanded uniform constants without requiring contiguous layout so all-zero SDPA masks can be recognized and removed. When a zero bias reaches efficient attention, switch back to flash attention only when flash is legal and dropout is disabled. Fixes #133462 Gene..."
184310,Fix flex attention shared memory config fallback,https://github.com/pytorch/pytorch/pull/184310,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,133254,133254,,,,0,3,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-19T03:14:43Z,2026-07-06T17:10:03Z,2,Stack from ghstack (oldest at bottom): -> #184310 Add fallback FlexAttention configs for low shared-memory CUDA devices and filter forward/backward configs against the final kernel shared-memory estimate before selecting a non-autotune config. This avoids selecting kernels that compile during benchmarking but fail f...
186336,Support multiword CXX commands in Inductor cpp_builder,https://github.com/pytorch/pytorch/pull/186336,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,133078,133078,,,,0,1,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-06-05T06:25:15Z,2026-07-06T17:06:55Z,2,"Stack from ghstack (oldest at bottom): -> #186336 Inductor accepted the CXX config as a string, but several compiler probe paths passed that entire string as argv[0]. A compiler command such as CXX=""zig c++"" therefore failed before compilation with InvalidCxxCompiler, even though the later build command path already..."
187596,Fall back on AOTAutograd version-check capture errors,https://github.com/pytorch/pytorch/pull/187596,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,158561,158561,184533,,,0,2,0,,,"module: dynamo, ciflow/inductor, release notes: aot autograd, agentic, needs-review",2026-06-17T19:09:24Z,2026-07-06T17:03:15Z,2,"Stack from ghstack (oldest at bottom): -> #187596 AOTAutograd chooses the training path when lifted module parameters require gradients, even if the user input itself does not. In #158561 the forward mutates a ReLU output with copy_(), so eager forward is valid and eager backward later raises the normal autograd ver..."
188944,[NCCL] try to remove the redundant copy in functional collectives,https://github.com/pytorch/pytorch/pull/188944,ready for maintainer decision,45,CI lane labels: ciflow/torchtitan; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: distributed (c10d), ciflow/torchtitan",2026-07-04T07:14:27Z,2026-07-06T16:59:48Z,2,"Hi, I am a fan of the DTensor API which uses functional collectives. However, there is redundant copies in the all_reduce and broadcast when not using compile. The NCCL interface already supports out of place APIs. Can someone take a look at this PR? The idea is to allocate output = torch.empty(...) rather than outp..."
185791,Fix cond branch shape guards leaking to export constraints,https://github.com/pytorch/pytorch/pull/185791,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,7,1,"statusCheckRollup, torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,147991,147991,,,,0,8,0,c1974a3adaa3494f7d0cb20b2c0159a1d98c4cc5a59b4d8e96e04b9d6a23f202,da157ee42d2e93d5a19b7bd8b088a680e167cf7f03cefe0d4e2bf7404d59884a,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-06-01T04:30:31Z,2026-07-06T16:59:35Z,2,"Stack from ghstack (oldest at bottom): -> #185791 torch.cond traces both branches with the same ShapeEnv as the surrounding graph. When a branch contains a shape check, such as torch._check(sequence_length > 5), tracing refines the shared ShapeEnv. Export then treats that branch-local fact as a top-level input const..."
186339,Improve torch.compile sibling output grad error,https://github.com/pytorch/pytorch/pull/186339,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,132929,132929,,,,0,1,0,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic, needs-review",2026-06-05T07:06:28Z,2026-07-06T16:55:55Z,2,"Stack from ghstack (oldest at bottom): -> #186339 AOTAutograd returns compiled graph outputs as sibling outputs of one opaque autograd Function. When user code asks torch.autograd.grad for gradients of one compiled output with respect to another compiled output that it depends on, the internal eager edge is hidden f..."
188050,Add FX graph cache custom guard replay test,https://github.com/pytorch/pytorch/pull/188050,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor",2026-06-24T10:05:15Z,2026-07-07T02:04:37Z,2,"Stack from ghstack (oldest at bottom): -> #188050 AOTAutograd passes a custom guard evaluator into the FX graph cache so it can enforce its exact-match guard contract. A previous guard-provenance change accidentally bypassed that evaluator during cache-hit post-load guard replay, which regressed vLLM through the AOT..."
186342,Preserve nested fake subclass inputs in AOTAutograd,https://github.com/pytorch/pytorch/pull/186342,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,132896,132896,,,,0,1,0,,,"topic: not user facing, ciflow/inductor, agentic, needs-review",2026-06-05T07:41:23Z,2026-07-06T16:48:47Z,2,"Stack from ghstack (oldest at bottom): -> #186342 AOTAutograd's process_inputs avoided re-fakifying wrapper subclass inputs only when their immediate tensor attrs were FakeTensors from the active FakeTensorMode. For nested traceable wrapper subclasses, those immediate attrs are wrapper subclasses themselves, so the..."
188761,MSVC compiler flag change for JIT cpp extensions on Windows,https://github.com/pytorch/pytorch/pull/188761,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-07-02T00:08:29Z,2026-07-06T16:39:28Z,2,Summary Fix Windows JIT C++ extension builds with CUDA 13+ by passing MSVC's standard-conforming preprocessor flag Zc:preprocessor through torch.utils.cpp_extension. CUDA 13.x ships CCCL headers that #error when cl.exe uses MSVC's legacy (traditional) preprocessor. PyTorch's main build already sets this flag in CMak...
176293,[CI] aarch64 manywheel and cuda image build fixes,https://github.com/pytorch/pytorch/pull/176293,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries, ciflow/docker, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"triaged, open source, module: arm, Stale, ciflow/binaries, ciflow/trunk, topic: not user facing, ciflow/docker",2026-03-03T15:33:56Z,2026-07-06T15:59:45Z,2,"This patch consolidates the non-CUDA and CUDA aarch64 manywheel Dockerfiles for pytorch, modifies the build script to use Dockerfile_2_28_aarch64 and fixes the install_openssl.sh script to not use sudo when linking the libs to /usr/lib due to there sometimes being a mismatch between the sudoers plugin system crypto..."
188656,[Test] Improve test reuse in test/test_unary_ufuncs.py for out-of-tree backends,https://github.com/pytorch/pytorch/pull/188656,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-07-01T09:22:13Z,2026-07-06T15:58:54Z,2,Summary Improve test reuse in test/test_unary_ufuncs.py for out-of-tree backends by replacing hardcoded device assumptions with the injected device parameter where applicable. Changes Deduplicate unary_mem_overlap_cases by removing per-device duplicate entries and using the injected device argument instead of the de...
188823,[Inductor][Host-side TMA] Cache host-side TMA descriptors across launches,https://github.com/pytorch/pytorch/pull/188823,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,188822,,,0,3,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-02T16:40:37Z,2026-07-06T15:48:40Z,2,"Stack from ghstack (oldest at bottom): -> #188823 #188822 #188821 This PR is a copy of #187918. Host-side TMA kernels build a TensorDescriptor and encode a CUtensorMap (cuTensorMapEncodeTiled) on every launch. For a repeated call on the same buffer this is redundant: a CUtensorMap encodes addressing (base address, s..."
188822,[Inductor][Host-side TMA] Support tensordesc<> in the static Triton launcher,https://github.com/pytorch/pytorch/pull/188822,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,188823,,,0,3,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-02T16:40:32Z,2026-07-06T15:48:29Z,2,"Stack from ghstack (oldest at bottom): #188823 -> #188822 #188821 This PR is a copy of #187917. Host-side TMA kernels take their TMA descriptor as a by-value CUtensorMap (a tensordesc<dtype[block]> signature entry). The static Triton launcher previously raised NotImplementedError for that type, so every host-side TM..."
176726,Fix Muon momentum/update numerics for XPU state_dict parity,https://github.com/pytorch/pytorch/pull/176726,draft/noise,31,PR is draft; review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, Stale, release notes: optim",2026-03-06T16:50:58Z,2026-07-06T15:46:35Z,2,Fixes: intel/torch-xpu-ops#1973 Root cause is numeric drift not a differnt Muon algorithm on XPU. CPU and XPU run the same logic but backend kernels round and accumulate differently and this case is sensitive enough to cross tolerance on XPU. CUDA can run this test upstream no skip or xfail there. This PR adjusts Mu...
180208,Update slow tests,https://github.com/pytorch/pytorch/pull/180208,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/slow, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/slow, ci-no-td",2026-04-13T08:21:39Z,2026-07-06T14:46:19Z,2,This PR is auto-generated weekly by this action. Update the list of slow tests.
189019,[MPS] Add torch.mps.MetalGraph capture/replay API,https://github.com/pytorch/pytorch/pull/189019,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180397,180397,180397,,,,0,8,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-07-06T10:07:15Z,2026-07-06T14:46:00Z,2,Adds torch.mps.metal_graph_capture() / torch.mps.metal_graph_replay() plus a torch.mps.MetalGraph class API. Records the MPS operations dispatched inside a capture context and replays them on subsequent calls without paying the per-op CPU dispatch overhead. Metal analog of torch.cuda.graph. Closes #180397. Motivatio...
188846,"Fix torch.pow docstring: scalar params are Number, not float",https://github.com/pytorch/pytorch/pull/188846,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,188378,,,0,1,0,,,"triaged, open source, release notes: python_frontend, topic: not user facing",2026-07-02T20:52:19Z,2026-07-06T20:41:24Z,2,"The torch.pow docstring types both scalar parameters as float, but they accept any Number: pow(input, exponent): the exponent Args entry reads exponent (float or tensor), but the ATen schema is pow.Tensor_Scalar(Tensor self, Scalar exponent), so a scalar exponent is a Scalar (any Number). The docstring's own example..."
188843,Fix scatter/scatter_add/scatter_reduce src gradient when index is sma…,https://github.com/pytorch/pytorch/pull/188843,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,"27614, 94336","27614, 94336","27614, 94336",,,,0,2,0,,,"triaged, open source, release notes: autograd",2026-07-02T20:09:55Z,2026-07-06T14:33:01Z,2,"Issue Fixes #94336 Fixes #27614 Summary The scatter family accepts an index smaller than src (index.size(d) <= src.size(d), an empty index of any dimensionality, and legacy rank-mismatched 0-dim/1-dim combinations where missing index dims act as size 1); only the src region covered by index participates in the outpu..."
183755,[MPS] Enable bf16/fp16 ConvTranspose3D by removing the gating,https://github.com/pytorch/pytorch/pull/183755,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,160739,183743,,,0,1,0,,,"triaged, open source, topic: improvements, release notes: mps, ciflow/mps",2026-05-14T18:54:09Z,2026-07-06T14:32:58Z,2,"Fixes issue #160739 Enables ConvTranspose3D by removing the gating preventing it. The MPSGraph supports the op but it has been blocked due to our unit tests that are unnecessarily harsh for this case. The tests are expected fails for now but will start ""unexpectedly"" succeeding once the testing change in #183743 goe..."
189005,fix(fx): filter internal frames in record_stack_traces to resolve broken stack traces (#130861),https://github.com/pytorch/pytorch/pull/189005,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,130861,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-07-06T06:06:00Z,2026-07-06T14:32:35Z,2,"This pull request addresses an issue where FX tracer stack traces were becoming broken or overly cluttered with internal framework frames. By implementing a filtering mechanism in record_stack_traces, internal frames are now excluded, resulting in cleaner and more actionable stack traces for users. This fix specific..."
188968,Fix record_stack_traces returning empty trace when no forward frame found,https://github.com/pytorch/pytorch/pull/188968,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,130861,130861,130861,,,,0,1,0,,,"triaged, open source, release notes: fx",2026-07-05T07:39:39Z,2026-07-06T14:32:32Z,2,"Fixes #130861. Summary When TracerBase._record_forward_stack_traces_only is False (the default) and the call stack contains no ""forward"" frame, _filter_traceback_frames() returns an empty list, silently discarding all stack trace information. This breaks record_stack_traces for any code path that does not pass throu..."
187926,[Test] Refactor test/test_shape_ops.py,https://github.com/pytorch/pytorch/pull/187926,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"185797, 185798, 185802, 185881, 186351, 186352",,,0,14,0,,,"open source, topic: not user facing, module: tests/core",2026-06-23T07:22:24Z,2026-07-06T13:43:44Z,2,Stack from ghstack (oldest at bottom): -> #187926 #186352 #186351 #185881 #185802 #185798 #185797
186351,[Test] Refactor test/test_scatter_gather_ops.py,https://github.com/pytorch/pytorch/pull/186351,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"185797, 185798, 185802, 185881, 186352, 187926",,,0,25,0,,,"open source, topic: not user facing, module: tests/core",2026-06-05T08:32:37Z,2026-07-06T13:43:43Z,2,Stack from ghstack (oldest at bottom): #187926 #186352 -> #186351 #185881 #185802 #185798 #185797
186352,[Test] Refactor test/test_ops.py,https://github.com/pytorch/pytorch/pull/186352,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"185797, 185798, 185802, 185881, 186351, 187926",,,0,28,0,,,"open source, topic: not user facing, module: tests/core",2026-06-05T08:43:42Z,2026-07-06T13:43:37Z,2,Stack from ghstack (oldest at bottom): #187926 -> #186352 #186351 #185881 #185802 #185798 #185797 Changes to common_methods_invocations.py are all renaming: TestCommon -> TestCommonDevice
185881,[Test] Refactor test/test_reductions.py,https://github.com/pytorch/pytorch/pull/185881,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"185797, 185798, 185802, 186351, 186352, 187926",,,0,35,0,,,"module: tests, open source, topic: not user facing, module: tests/core",2026-06-02T03:21:28Z,2026-07-06T13:43:34Z,2,Stack from ghstack (oldest at bottom): #187926 #186352 #186351 -> #185881 #185802 #185798 #185797 cc @mruberry
188156,[MPS] reduce: share the two-pass value_reduction kernel for sum and value,https://github.com/pytorch/pytorch/pull/188156,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,187455,"187787, 188333",,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-25T14:45:35Z,2026-07-06T13:39:46Z,2,Stack Position / Review Order Tracked in #187455. This is the reduce-family base and should land before the prod and var/std follow-ups. It intentionally does not close the tracking issue because #188333 and #187787 remain in the stack. Order after review: Land this PR first. Rebase/retarget #188333 so its diff is p...
188748,[MPS] Make F.linear honor PYTORCH_MPS_PREFER_METAL,https://github.com/pytorch/pytorch/pull/188748,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,188728,188728,188728,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-07-01T22:24:17Z,2026-07-06T13:00:28Z,2,Fixes #188728 Goes through reshape to 2D -> mm/addmm metal kernels instead of writing a new multi-dim linear kernel. Perf wise this should be a good approach. Lifted the env flag check into a static in utils to be able to check it across different files. On testing: The mm/addmm metal kernel should be exercised sepa...
183276,[Inductor][Refactor] 2/N Introduce shared heuristic registry with codegen_heuristic support,https://github.com/pytorch/pytorch/pull/183276,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,188721,"183277, 183278",,,0,29,0,,,"open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-05-11T12:45:31Z,2026-07-06T12:16:15Z,2,Stack from ghstack (oldest at bottom): #183278 #183277 -> #183276 #183275 Motivation In the previous PR we created the heuristics/ package with template heuristics in template/. Each heuristic category maintained its own registry (template_heuristics used template/registry.py). Adding codegen heuristics would duplic...
183277,[Inductor][Refactor] 3/N Move pointwise heuristics from triton_heuristics.py into triton_codegen/,https://github.com/pytorch/pytorch/pull/183277,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,188721,"183276, 183278",,,0,29,0,,,"open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-05-11T12:45:36Z,2026-07-06T12:16:14Z,2,Stack from ghstack (oldest at bottom): #183278 -> #183277 #183276 #183275 Motivation triton_heuristics.py contained ~130 lines of device-branching logic for pointwise autotuning configs (CUDA vs ROCm vs XPU). This makes it hard to add new device backends or modify configs for a single device without touching the sha...
188119,[inductor] Enable combo kernel compile time autotuning by default,https://github.com/pytorch/pytorch/pull/188119,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182577, 184955, 187961, 188203, 189001, 189002, 189003",,,0,6,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-25T00:04:43Z,2026-07-06T12:16:07Z,2,Stack from ghstack (oldest at bottom): #189003 #189002 #189001 #188203 -> #188119 #184955 #182577 #187961 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
188326,feat(dynamo): backport recursive dict tags,https://github.com/pytorch/pytorch/pull/188326,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,True,0,0,,,,,,,,,0,18,0,,,"open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-27T08:24:17Z,2026-07-06T11:07:14Z,2,Summary Backports the upstream recursive dict-tag guard matching path onto the 2.7.1/release-2.7 guard manager. The feature remains opt-in in this branch so reviewers can inspect the code-size and safety boundary before deciding whether to enable it for a customer build. Changes Add recursive dict-tag discovery for...
183278,[Inductor][Refactor] 4/N Move reduction heuristics from triton_heuristics.py into triton_codegen/,https://github.com/pytorch/pytorch/pull/183278,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,178153,"183276, 183277",,,0,30,0,,,"open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-05-11T12:45:41Z,2026-07-06T10:55:42Z,2,"Stack from ghstack (oldest at bottom): -> #183278 #183277 #183276 #183275 Motivation Continues the RFC #178153 refactoring of Triton heuristics into a device-scoped registry. The reduction config generation logic was previously inline in triton_heuristics.py with device-specific branches (torch.version.hip, torch.cu..."
188268,Fix positional arg overflow in TritonBenchmarkRequest when template inputs are deduplicated,https://github.com/pytorch/pytorch/pull/188268,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188069,188069,"188069, 188721",,,,0,7,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-06-26T12:51:06Z,2026-07-06T10:01:56Z,2,"Fix positional arg overflow in TritonBenchmarkRequest when template inputs are deduplicated Fixes #188069 When bmm template codegen deduplicates kernel inputs (both operands alias the same buffer, e.g. einsum(""...ab,...Ab->...aA"", grad, grad)), the compiled kernel expects fewer positional parameters than TritonBench..."
182747,Resolve PYTHON_SIX_SOURCE_DIR from installed `six` to skip pypi download,https://github.com/pytorch/pytorch/pull/182747,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"Stale, topic: not user facing",2026-05-07T00:36:41Z,2026-07-06T09:49:59Z,2,"NNPACK's cmake/DownloadSix.cmake fetches six-1.11.0.tar.gz from pypi.python.org at configure time on x86-64 unless PYTHON_SIX_SOURCE_DIR is predefined. Network-restricted runners (OSDC) can't reach pypi and fail the configure step with SSL connect error. For example, https://github.com/pytorch/pytorch/actions/runs/2..."
187317,[pr-review skill] Add device portability checks for shared accelerator code,https://github.com/pytorch/pytorch/pull/187317,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,2,0,,,"open source, topic: not user facing",2026-06-15T08:17:04Z,2026-07-06T09:39:58Z,2,"Stack from ghstack (oldest at bottom): -> #187317 Motivation: Most PyTorch code is shared across accelerator backends (CUDA,XPU, ROCm, MPS), but community contributors typically only have CUDA hardware. As a result, CUDA-specific code regularly leaks into device-generic paths (e.g. torch.cuda.* calls, hardcoded devi..."
188044,[Test] Refactor test/test_view_ops.py,https://github.com/pytorch/pytorch/pull/188044,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-06-24T09:07:30Z,2026-07-06T09:24:25Z,2,Summary Refactor test/test_view_ops.py to improve test reuse for out-of-tree backends by reducing unnecessary device-specific restrictions and making the tests more device-agnostic. Changes Extend is_view_of() to recognize PrivateUse1 devices. Remove unnecessary @onlyNativeDeviceTypes decorators. Remove @onlyCPU dec...
189006,Ensure correct sign of input for test_laplace in test_distributions.py,https://github.com/pytorch/pytorch/pull/189006,draft/noise,23,PR is draft; PR closes or references open issue numbers,wait for author to mark ready,,UNKNOWN,True,0,0,,,4021,4021,,,,,0,1,0,,,"open source, topic: not user facing",2026-07-06T06:06:27Z,2026-07-06T09:20:20Z,2,"Previously scale_1D input tensor was unconstrained, and could contain negative values, which led invalid behaviour, Partially fixes: intel/torch-xpu-ops#4021"
182577,[Inductor] Enable combo_kernel_per_subkernel_blocks by default,https://github.com/pytorch/pytorch/pull/182577,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184955, 187961, 188119, 188203, 189001, 189002, 189003",,,0,4,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-05T22:20:49Z,2026-07-06T09:15:13Z,2,Stack from ghstack (oldest at bottom): #189003 #189002 #189001 #188203 #188119 #184955 -> #182577 #187961 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo Differential...
188203,[inductor] Fix combo-kernel peak-memory KeyError when reorder_for_peak_memory is disabled,https://github.com/pytorch/pytorch/pull/188203,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182577, 184955, 187961, 188119, 189001, 189002, 189003",,,0,2,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-25T21:57:38Z,2026-07-06T09:15:11Z,2,Stack from ghstack (oldest at bottom): #189003 #189002 #189001 -> #188203 #188119 #184955 #182577 #187961
189003,[inductor] Coordinate-descent tune compile-time combo kernel blocks,https://github.com/pytorch/pytorch/pull/189003,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182577, 184955, 187961, 188119, 188203, 189001, 189002",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-06T05:33:12Z,2026-07-06T09:15:09Z,2,"Stack from ghstack (oldest at bottom): -> #189003 #189002 #189001 #188203 #188119 #184955 #182577 #187961 Compile-time per-subkernel autotuning passes each sub-kernel's block sizes as constexpr args (XBLOCK_0, XBLOCK_1, ...) instead of baking them into the body. This exposes those suffixed fields to CoordescTuner by..."
189002,[inductor] Don't combo-fuse data-independent producers feeding attention,https://github.com/pytorch/pytorch/pull/189002,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182577, 184955, 187961, 188119, 188203, 189001, 189003",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-06T05:33:09Z,2026-07-06T09:15:09Z,2,"Stack from ghstack (oldest at bottom): #189003 -> #189002 #189001 #188203 #188119 #184955 #182577 #187961 Summary: Summary Combo-kernel (horizontal) fusion can regress models with attention by hoisting per-layer causal attention masks to the front of the graph, where they get evicted from L2 before each layer's SDPA..."
189001,[inductor] Fix flaky combo-kernel peak-memory test,https://github.com/pytorch/pytorch/pull/189001,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182577, 184955, 187961, 188119, 188203, 189002, 189003",,,0,1,0,,,"topic: not user facing, module: inductor, ciflow/inductor",2026-07-06T05:33:06Z,2026-07-06T09:15:08Z,2,Stack from ghstack (oldest at bottom): #189003 #189002 -> #189001 #188203 #188119 #184955 #182577 #187961
184955,[Inductor] Exclude indirect-indexing nodes from combo kernels,https://github.com/pytorch/pytorch/pull/184955,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"182577, 187961, 188119, 188203, 189001, 189002, 189003",,,0,9,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-23T02:32:54Z,2026-07-06T09:15:06Z,2,Stack from ghstack (oldest at bottom): #189003 #189002 #189001 #188203 #188119 -> #184955 #182577 #187961 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @mlazos Diff...
186047,[Test][PrivateUse1] Enhance `get_all_devices` to include all available devices,https://github.com/pytorch/pytorch/pull/186047,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, module: testing, topic: not user facing, module: PrivateUse1",2026-06-03T06:28:26Z,2026-07-06T08:38:02Z,2,Improve DeviceTypeTestBase.get_all_devices() to return all available devices instead of just cls.primary_device cc @NmomoN @mengpenghui @fwenguang @cdzhan @1274085042 @PHLens @albanD
185797,[Test] Refactor test/test_bmm_outer_product.py,https://github.com/pytorch/pytorch/pull/185797,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"185798, 185802, 185881, 186351, 186352, 187926",,,0,10,0,,,"module: tests, open source, ciflow/trunk, topic: not user facing, module: tests/core",2026-06-01T07:22:42Z,2026-07-06T08:27:23Z,2,Stack from ghstack (oldest at bottom): #187926 #186352 #186351 #185881 #185802 #185798 -> #185797 cc @mruberry
185798,[Test] Refactor test/test_indexing.py,https://github.com/pytorch/pytorch/pull/185798,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"185797, 185802, 185881, 186351, 186352, 187926",,,0,22,0,,,"module: tests, open source, ciflow/trunk, topic: not user facing, module: tests/core",2026-06-01T08:08:49Z,2026-07-06T08:27:13Z,2,Stack from ghstack (oldest at bottom): #187926 #186352 #186351 #185881 #185802 -> #185798 #185797 cc @mruberry
185802,[Test] Refactor test/test_out_dtype_op.py,https://github.com/pytorch/pytorch/pull/185802,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"185797, 185798, 185881, 186351, 186352, 187926",,,0,22,0,,,"module: tests, open source, topic: not user facing, module: tests/core",2026-06-01T08:59:27Z,2026-07-06T08:26:56Z,2,Stack from ghstack (oldest at bottom): #187926 #186352 #186351 #185881 -> #185802 #185798 #185797 cc @mruberry
181659,[FSDP2] Add recurse= to set_symm_mem_for_comm and set_force_sum_reduction_for_comms,https://github.com/pytorch/pytorch/pull/181659,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,2,0,,,"Stale, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-04-27T21:39:03Z,2026-07-06T08:02:22Z,2,Stack from ghstack (oldest at bottom): -> #181659 Adds a recurse argument so user can call set method once at root module without iterating over all FSDPModule this follows other set methods in FSDPModule. also convinently make titan simple: pytorch/torchtitan#3105 (comment)
188886,[AOTI] Support PDL launch attributes via `cuLaunchKernelEx` in CUDA wrappers,https://github.com/pytorch/pytorch/pull/188886,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"open source, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-07-03T09:29:30Z,2026-07-06T07:51:18Z,2,"Summary Enable programmatic dependent launch (PDL) for AOTI CUDA wrappers by switching kernel launches from cuLaunchKernel to cuLaunchKernelEx / CUlaunchConfig, which can carry CU_LAUNCH_ATTRIBUTE_PROGRAMMATIC_STREAM_SERIALIZATION. PDL codegen stays CUDA-only and SM90+. Cost: forked HIP/ROCm codegen PDL is CUDA/SM90..."
188977,Fix typos in docstrings and comments across torch modules,https://github.com/pytorch/pytorch/pull/188977,ready for maintainer decision,45,"CI lane labels: ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,2,0,,,"ciflow/trunk, release notes: fx, ciflow/torchtitan",2026-07-05T21:52:42Z,2026-07-06T07:50:58Z,2,"Stack from ghstack (oldest at bottom): -> #188977 Correct spelling and grammar errors in docstrings, comments, and user-facing strings throughout various torch submodules. No functional changes."
186778,Reduce fp16 batch-1 Conv2d forward divergence on XPU,https://github.com/pytorch/pytorch/pull/186778,draft/noise,23,PR is draft; PR closes or references open issue numbers,wait for author to mark ready,,UNKNOWN,True,0,0,,,3837,3837,,,,,0,1,0,,,"module: cpu, open source, topic: not user facing",2026-06-09T11:30:00Z,2026-07-06T07:48:06Z,2,"test_conv_large_xpu showed forward mismatch between full-batch and sliced-batch execution. XPU oneDNN fp16 batch-1 path produced divergent numerics versus larger-batch path. Fix: Add a guarded fp32 compute fallback for constrained batch-1 large Conv2d path, cast result back to fp16 Fixes: intel/torch-xpu-ops#3837 cc..."
188637,[inductor] Reindex transpose-contiguous clone to enable vertical fusion,https://github.com/pytorch/pytorch/pull/188637,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,188635,188635,188635,,,,0,2,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-01T02:49:29Z,2026-07-06T07:13:02Z,2,"Fixes #188635 Summary This PR extends Inductor's loop reindexing/index-inversion fusion path to handle a layout-changing clone produced by: producer[B, S, H] -> view(B, S, heads, D) -> transpose(1, 2).contiguous() Before this change, the producer writes [B, S, H] contiguously, while the clone reads the producer buff..."
182768,fix(resize): rollback tensor metadata when storage resize fails,https://github.com/pytorch/pytorch/pull/182768,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,170298,170298,170298,,,,0,1,0,,,"open source, Stale",2026-05-07T03:55:39Z,2026-07-06T06:01:55Z,2,"Summary Provides strong exception guarantee for resize_() on non-resizable storage (e.g., numpy-backed tensors). Previously, _resize_impl_ updated tensor metadata (sizes/strides) before attempting to resize storage. When _maybe_resize_storage threw (e.g., for non-resizable numpy storage), the tensor was left in a ""z..."
181324,"[dynamo] Remove TypingVariable, trace more of typing.py",https://github.com/pytorch/pytorch/pull/181324,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,9,0,,,"open source, Stale, module: dynamo, ciflow/inductor, ciflow/dynamo, ciflow/torchtitan",2026-04-24T01:37:01Z,2026-07-06T06:01:50Z,2,"Stack from ghstack (oldest at bottom): -> #181324 TypingVariable is used to represent a number of different classes from the typing module, both builtin and in typing.py. This attempts to trace typing.py while only implmenting the builting parts of the typing module. Some caveats We still fall back to cpython for bu..."
188328,[meta] Add output_padding and bias validation to transposed convolution meta path,https://github.com/pytorch/pytorch/pull/188328,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,"178125, 178128","178125, 178128","178125, 178128",,,,0,2,0,,,"triaged, open source, ciflow/trunk, topic: bug fixes, release notes: inductor",2026-06-27T08:30:35Z,2026-07-06T05:32:45Z,2,"Summary The meta kernel for aten.convolution (meta_conv in torch/_meta_registrations.py) skipped two input validations that the eager C++ path enforces, causing the meta device to silently return incorrect shapes instead of raising errors. 1. output_padding validation (fixes #178125) For transposed convolutions, out..."
188995,[dynamo] Allow raw unbacked SymInts in strict mode (FlexAttention BlockMask) (#187547),https://github.com/pytorch/pytorch/pull/188995,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,187547,187547,187547,,,,0,1,0,,,"open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-06T04:15:40Z,2026-07-06T04:31:33Z,2,"Fixes #187547 VariableBuilder._wrap had an explicit gate requiring _is_non_strict_tracing() that blocked raw unbacked SymInts in strict Dynamo, even though the underlying ShapeEnv._transfer_foreign_expr_as_unbacked machinery already supported them. This broke FlexAttention BlockMask with raw unbacked seq_lengths und..."
188993,[dynamo] Preserve Event/Stream subclass identity across graph breaks (#188536),https://github.com/pytorch/pytorch/pull/188993,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,188536,188536,188536,,,,0,1,0,,,"open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-06T04:15:35Z,2026-07-06T04:16:23Z,2,"Fixes #188536 When a user subclasses torch.Event or torch.Stream, Dynamo's graph break path lost the subclass identity because build_event/build_stream helpers always constructed the base torch._C.Event/torch._C.Stream types. This caused record() and wait() to operate on different objects across graph breaks. Change..."
188992,[inductor][XPU] Fix AMP autocast round-trip fp16 accuracy (#188261),https://github.com/pytorch/pytorch/pull/188992,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,188261,188261,188261,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-06T04:15:32Z,2026-07-06T04:16:23Z,2,"Fixes #188261 Collapse redundant AMP autocast dtype round-trip conversions in the Inductor joint graph FX pass to prevent fp16 accuracy loss on XPU. When an operator's input and output are both in fp16 after autocast, the intermediate cast through fp32 is unnecessary and introduces precision loss. cc @voznesenskym @..."
181665,Fix UBSan in weight_norm CPU kernel for empty input (#181510),https://github.com/pytorch/pytorch/pull/181665,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,181510,181510,181510,,,,0,2,0,,,"module: cpu, triaged, open source, Stale, release notes: nn",2026-04-27T22:21:39Z,2026-07-06T02:20:21Z,3,"Fixes #181510. What torch.nn.utils.weight_norm applied to nn.Linear(0, 1) and run on an empty input tripped UBSan in an ASAN/UBSAN build: aten/src/ATen/cpu/vec/vec256/vec256_float.h:124:16: runtime error: null pointer passed as argument 2, which is declared to never be null Stack trace bottoms out in Vectorized<floa..."
181611,[don't merge] test combo kernels,https://github.com/pytorch/pytorch/pull/181611,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"Stale, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-27T16:58:06Z,2026-07-06T02:20:20Z,3,Stack from ghstack (oldest at bottom): -> #181611 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
182729,[DTensor] Preserve placement for aten.clone.default,https://github.com/pytorch/pytorch/pull/182729,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-06T22:03:03Z,2026-07-06T01:09:44Z,3,"Stack from ghstack (oldest at bottom): -> #182729 #180937 #180936 We registered clone twice, one in _tensor_ops.py and another one in _pointwise_ops.py aten.clone.default is tagged torch.Tag.pointwise and was getting auto-registered with the default pointwise single-dim strategy, which only enumerates Replicate/Shar..."
182727,[shard prop] P(*).clone()->P(*),https://github.com/pytorch/pytorch/pull/182727,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-06T22:00:46Z,2026-07-06T01:09:43Z,3,Stack from ghstack (oldest at bottom): -> #182727
182582,Add addmm quantization graph passes for PT2 readiness benchmarks (#182582),https://github.com/pytorch/pytorch/pull/182582,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, Stale, meta-exported",2026-05-05T22:53:26Z,2026-07-06T01:09:42Z,3,"Summary: Add addmm-based quantization passes to the Sigmoid graph pass registry for PT2-exported models. PT2 export decomposes nn.Linear into aten.t + aten.addmm (or bare aten.addmm when export-time constant folding removes aten.t). The existing Linear* quantization passes only match aten.linear, so they are no-ops..."
182276,[inductor] Unwrap ShapeAsConstantBuffer in tensor_constructor size,https://github.com/pytorch/pytorch/pull/182276,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,182275,182275,,,,,0,1,0,,,"triaged, open source, Stale, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-04T10:06:28Z,2026-07-06T01:09:40Z,3,"Fixes #182275. Summary When the consumer of an invoke_subgraph HOP (e.g. a function decorated with torch.compiler.nested_compile_region inside an outer torch.compile) passes a HOP-output SymInt as a size argument to aten.full (or another tensor_constructor-backed op), Inductor wraps that SymInt as a ShapeAsConstantB..."
177115,Fix opcheck to detect stride mismatches on CPU tensors,https://github.com/pytorch/pytorch/pull/177115,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,149468,149468,,,,,0,2,0,,,"module: bc-breaking, Stale, release notes: onnx, release notes: composability, release notes: python_frontend, topic: bc breaking",2026-03-11T04:51:46Z,2026-07-05T23:50:50Z,3,"Human Note torch.library.opcheck silently skipped stride validation for CPU tensors because compare_tensor_meta called check_significant_strides with only_cuda=True (the default), causing the stride check to be bypassed when both tensors are on CPU. This change passes only_cuda=False so that stride mismatches betwee..."
188642,Fix __qualname__ for OpOverloadPacket,https://github.com/pytorch/pytorch/pull/188642,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,186140,186140,186140,,,,0,1,0,,,"triaged, open source, release notes: python_frontend",2026-07-01T03:24:07Z,2026-07-05T23:21:14Z,3,"Fixes #186140 OpOverloadPacket.__getattr__ forwards dunder lookups to the underlying pybind11 operator, so __qualname__ returned the pybind function-record name instead of the op name: >>> torch.ops.aten.add.__qualname__ 'pybind11_detail_function_record_v1_system_libstdcpp_gxx_abi_1xxx_use_cxx11_abi_1.add' OpOverloa..."
188067,[ROCm] fix wait instructions,https://github.com/pytorch/pytorch/pull/188067,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: rocm, triaged, open source, ciflow/trunk, release notes: cuda",2026-06-24T15:20:48Z,2026-07-05T23:05:01Z,3,GFX 10-12 architectures have distinct counters for loading/store to/from VGPRs. Older architectures have only one counter for memory operations with VGPR. GFX 12 architecture also changes instructions that wait these counters. Therefore we must call different instructions for different architectures. cc @jeffdaily @...
179619,[PowerPC] Guard deviceTypeFromActivity with USE_KINETO check,https://github.com/pytorch/pytorch/pull/179619,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,173881,173881,173881,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-07T19:32:28Z,2026-07-05T22:54:03Z,3,"Issue Fixes #173881 Summary When building PyTorch with USE_KINETO=0 (common on ppc64le where Kineto is not supported), the deviceTypeFromActivity() function in kineto_shim.cpp and its declaration in kineto_shim.h use libkineto::ActivityType as a parameter type without #ifdef USE_KINETO guards. While the standalone A..."
182159,Fix List isSubtypeExt,https://github.com/pytorch/pytorch/pull/182159,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, Stale",2026-05-01T18:31:31Z,2026-07-05T22:49:44Z,3,"Hi, I capture the computation graph using Dynamo, and converte it to TorchScript to optimize and run it with shape annotation in the graph. Then the jit interpreter throws a runtime error. The code case is below. #!/bin/env python import torch ts_str = """""" graph(%171 : Tensor, %174 : Tensor): %50 : int = prim::Const..."
182541,add regression test for weight_norm meta initialization (#96409),https://github.com/pytorch/pytorch/pull/182541,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,96409,96409,96409,,,,0,1,0,,,"module: nn, module: tests, triaged, open source, module: norms and normalization, Stale, topic: not user facing",2026-05-05T18:38:14Z,2026-07-05T20:49:12Z,3,"Summary Fixes #96409 by adding regression coverage for nn.utils.weight_norm initialization on the Meta device. The original issue reported a NotImplementedError for aten::_weight_norm_interface when applying nn.utils.weight_norm under with torch.device(""meta""):. On current main, that repro now succeeds, so this PR d..."
182138,Update torchcomms.txt commit hash,https://github.com/pytorch/pytorch/pull/182138,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-distributed, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/trunk, topic: not user facing, ciflow/inductor, ciflow/h100-distributed",2026-05-01T16:47:18Z,2026-07-05T20:49:09Z,3,This bumps the commit hash to in theory fix the update bot. The previous hash was valid but not from main which is causing issues. meta-pytorch/torchcomms@33fbfdd https://github.com/pytorch/pytorch/actions/runs/25196853559/job/73879319979 Test plan: CI
179485,Add torchbind custom class inheritance support (#179485),https://github.com/pytorch/pytorch/pull/179485,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, Stale, topic: not user facing, meta-exported",2026-04-06T18:05:25Z,2026-07-05T20:49:08Z,3,"Summary: Add support for declaring base classes in PyTorch's torchbind custom class system via a new def_base<Base>() API on torch::class_<T>. This enables polymorphic use of derived custom classes where a base type is expected (e.g. passing a ""Animal"" where ""Dog"" is expected). Changes: ClassType: add baseType_ fiel..."
174503,[Draft]Trying to see if cpp modification would work,https://github.com/pytorch/pytorch/pull/174503,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"Stale, release notes: fx, ciflow/inductor",2026-02-06T22:30:37Z,2026-07-05T20:49:00Z,3,Stack from ghstack (oldest at bottom): -> #174503 #174445
188517,Fix/180346 dtensor clip grad mixed mesh,https://github.com/pytorch/pytorch/pull/188517,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-30T01:53:19Z,2026-07-05T20:36:01Z,3,
188516,Fix/187177 localresponsenorm validation,https://github.com/pytorch/pytorch/pull/188516,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-30T01:53:10Z,2026-07-05T20:36:00Z,3,
182530,index_add_/index_reduce_: add env-var kill switch for VecSize=4 vectorized path,https://github.com/pytorch/pytorch/pull/182530,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, fb-exported, Stale, meta-exported",2026-05-05T17:18:45Z,2026-07-05T18:51:23Z,3,"Summary: D94314062 introduced a VecSize=4 fast path in indexFuncLargeIndex that amortizes divmod and index lookup by having each thread process consecutive elements. On uniform-random workloads this gives a ~1.8-2.0x speedup. However, when the index distribution is heavy-tailed (e.g. zipf, or skewed where a few hot..."
185788,Support 64-bit indexing in channels_last bilinear upsampling,https://github.com/pytorch/pytorch/pull/185788,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: nn",2026-06-01T03:46:35Z,2026-07-05T18:30:10Z,3,"Summary torch.nn.functional.interpolate(..., mode='bilinear') with a channels_last CUDA input crashed with CUDA error: invalid configuration argument when the output had at least 2**31 elements. Repro import torch x = torch.rand((32, 64, 512, 512), dtype=torch.half, device='cuda') out = torch.nn.functional.interpola..."
184745,distributions: fix Kumaraswamy mode formula,https://github.com/pytorch/pytorch/pull/184745,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,173912,173912,173912,,,,0,3,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-05-21T16:18:52Z,2026-07-05T17:56:11Z,3,"Description torch.distributions.Kumaraswamy.mode computes the wrong formula, producing NaN for almost all valid inputs. Current (wrong): computes $(1-b)^{1/b} / (1-ab)$ Correct: $\text{Mode} = \left(\frac{a-1}{ab-1}\right)^{1/a}$ where $a = $ concentration1, $b = $ concentration0 Reproducer from the issue: import to..."
188500,Allow pre-split pipeline microbatches,https://github.com/pytorch/pytorch/pull/188500,ready for maintainer decision,45,"CI lane labels: ciflow/torchtitan, ciflow/trunk; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,8,0,,,"ciflow/trunk, release notes: distributed (pipeline), module: pipelining, ciflow/torchtitan",2026-06-30T00:32:28Z,2026-07-05T17:24:13Z,3,"Stack from ghstack (oldest at bottom): -> #188500 Pipeline schedules currently assume step() inputs are full batches: the first thing both single- and multi-stage schedules do is split args, kwargs, and target into microbatches. That makes it impossible for input pipelines that already produce microbatches to feed t..."
170571,[CUDA][cuBLASLt] addmm - Lt path improvements,https://github.com/pytorch/pytorch/pull/170571,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/nightly","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,53,0,,,"module: cuda, module: cublas, open source, Stale, ciflow/trunk, topic: not user facing, ciflow/nightly, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2025-12-16T16:42:30Z,2026-07-05T15:53:02Z,3,"Stack from ghstack (oldest at bottom): -> #170571 Allow arbitrary beta value No-copy handling of bias whenever possible (Epilogue, suitable out-of-place cases (including both 1D and 2D shapes) Retire isInputCompliesAddmmCudaLt cc @ptrblck @msaroufim @eqy @jerryzh168 @tinglvv @nWEIdia @csarofeen @xwang233"
166396,[xpu][test][1/N] Enable tests of test_nn.py on Intel GPU - instantiate TestNN with instantiate_device_type_tests,https://github.com/pytorch/pytorch/pull/166396,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/mps, ciflow/pull, ciflow/xpu",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, Merged, Reverted, Stale, release notes: nn, topic: not user facing, ciflow/mps, keep-going, ciflow/xpu, ci-no-td, ciflow/pull",2025-10-28T10:18:29Z,2026-07-05T15:52:59Z,3,"For #114850, we will port aten unit tests to Intel GPU. This PR will work on test/test_nn.py TEST_NN class for single GPU test only. We could enable Intel GPU with following methods and try the best to keep the original code styles: Use torch.accelerator to extend cude specific test to XPU. Added skipIfXPU decorator..."
187325,[Metal][memory] add configurable memory budget API for MPS allocator (#105839),https://github.com/pytorch/pytorch/pull/187325,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,105839,105839,"105839, 132086",,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-15T09:20:28Z,2026-07-05T15:46:24Z,3,"Summary Adds torch.mps.set_memory_budget(bytes) / torch.mps.get_memory_budget() for the MPS allocator. When set, the allocator caps total GPU memory and evicts cached blocks before raising a clean OOM. Default 0 keeps current behavior (unlimited). On Apple Silicon, GPU and CPU share unified memory, so an unconstrain..."
188066,test: add F.normalize zero-input correctness tests (fixes #184575),https://github.com/pytorch/pytorch/pull/188066,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184575,184575,184575,,,,0,2,0,,,"triaged, open source, module: norms and normalization, topic: not user facing",2026-06-24T14:52:28Z,2026-07-05T15:20:50Z,3,"Summary Adds tests verifying that F.normalize correctly handles zero-norm inputs. Problem F.normalize(x) computes x / ||x||. At x = 0, the norm is 0 and the result is mathematically undefined. Currently: Forward returns 0 instead of NaN Backward returns ~1e12 instead of NaN This causes silent gradient corruption in..."
181743,Remove skipXPU from test_binary_ufuncs,https://github.com/pytorch/pytorch/pull/181743,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, Stale, ciflow/trunk, topic: not user facing, ciflow/xpu",2026-04-28T14:15:11Z,2026-07-05T14:59:27Z,3,This PR is to enable test_hypot and test_ldexp on XPU devices. Fixed by: intel/torch-xpu-ops#2962 intel/torch-xpu-ops#3248.
187897,fix: Add explicit null-buffer guards and shared-memory fallback for Intel Macs with discrete GPU (MPS),https://github.com/pytorch/pytorch/pull/187897,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: cpu, triaged, open source, release notes: quantization, release notes: releng, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), ciflow/torchtitan",2026-06-23T02:27:37Z,2026-07-05T14:29:19Z,3,"Changelog: MPS Intel/Non-Apple-Silicon Compatibility Patch Full list of files changed in this branch, based on the actual diff. This replaces the earlier README, which only covered the 2 files I authored (MPSAllocator.h / MPSAllocator.mm) — this version covers all 7. 1. aten/src/ATen/mps/MPSAllocator.h +159 / -17 Us..."
187782,[dynamo] Skip guard creation for unused function inputs,https://github.com/pytorch/pytorch/pull/187782,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,187593,187593,187593,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-21T07:31:20Z,2026-07-05T14:29:15Z,3,"Fixes #187593. Summary When InstructionTranslator sets up a frame it adds every f_locals entry to symbolic_locals as a LazyVariableTracker. If that tracker is ever realized (e.g. by iteration inside compile_subgraph or side-effects replay), a guard is installed even for inputs the compiled graph never touches. This..."
188401,dtensor: include local tensor device in _stable_hash_for_caching,https://github.com/pytorch/pytorch/pull/188401,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188390,188390,188390,,,,0,1,0,,,"triaged, open source, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-29T10:19:10Z,2026-07-05T14:03:48Z,3,"Summary `DTensor._stable_hash_for_caching()` builds the AOTAutograd cache key from the `DTensorSpec` hash (mesh topology, placements, global shape/stride/dtype) and `requires_grad`, but omitted the local tensor's device index. In multi-rank training, every rank holds a `DTensor` with an identical spec -- the `Device..."
187664,[MPS] Fix `baddbmm` operation to handle zero-sized dimensions correctly,https://github.com/pytorch/pytorch/pull/187664,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,187399,187399,187399,,,,0,3,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-18T14:16:52Z,2026-07-05T12:45:29Z,3,"Fixes #187399 Problem torch.baddbmm raised an opaque internal assertion on MPS whenever an operand had a size 0 dimension RuntimeError: [srcBuf length] > 0 INTERNAL ASSERT FAILED at "".../aten/src/ATen/native/mps/OperationUtils.mm"":525, please report a bug to PyTorch. Placeholder tensor is empty! Fix added same retur..."
187453,[MPS] Add error handling for complex dtypes in max unpooling operations,https://github.com/pytorch/pytorch/pull/187453,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,187233,187233,187233,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-16T16:38:29Z,2026-07-05T12:44:29Z,3,Fixes #187233 Adds an early MPS dtype check for max_unpool2d and max_unpool3d so complex inputs fail with a clear NotImplementedError instead of falling through to Metal kernel creation and raising an internal runtime error like max_unpool_float2. Also adds MPS regression coverage for complex max_unpool2d and max_un...
186479,[dynamo] Respect mark_dynamic on Parameter inputs,https://github.com/pytorch/pytorch/pull/186479,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,135011,135011,135011,,,,0,1,0,,,"triaged, open source, release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-06T17:41:57Z,2026-07-05T09:40:32Z,3,"Fixes #135011 This PR fixes Dynamo recompilation behavior for explicitly dynamic torch.nn.Parameter inputs. Previously, Parameter inputs could still be treated as statically shaped even after torch._dynamo.mark_dynamic, which caused unnecessary recompiles and prevented Tensor and Parameter inputs from sharing guards..."
187550,[test][inductor] Fix stale `xfail` and add SM12 skip.,https://github.com/pytorch/pytorch/pull/187550,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-06-17T05:57:36Z,2026-07-05T09:39:14Z,3,"test_divisible_by_16_covers_numel_args's xfail looks stale. test_red_followed_by_transposed_pointwise relies on max_autotune_gemm kernel fusion and asserts a specific kernel that only holds on devices with enough SMs (>= 68). On SM12x (low-SM) devices the count differs, so it is now skipped via IS_BIG_GPU. Authored..."
188799,Fix callable value-opaque objects causing graph break on method access,https://github.com/pytorch/pytorch/pull/188799,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,188796,188796,188796,,,,0,1,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-02T13:20:38Z,2026-07-05T09:18:17Z,3,"Fixes #188796 When a value-opaque type registered with register_opaque_type(cls, typ=""value"") also implements __call__, accessing its methods on a tensor subclass inside torch.compile caused a graph break. Root cause: In TensorVariable.dynamic_getattr, the code handling attributes on sourceless traceable wrapper sub..."
181205,"NOT MERGE, JUST FOR TEST",https://github.com/pytorch/pytorch/pull/181205,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, Stale, topic: not user facing, ciflow/xpu",2026-04-23T01:20:54Z,2026-07-05T07:38:56Z,3,
168999,[Pt. 1] Make some aspects of Dynamo/Inductor global state thread safe,https://github.com/pytorch/pytorch/pull/168999,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,126024,126024,"126024, 168042",,,,0,7,0,,,"triaged, open source, Stale, release notes: fx, fx, module: inductor, ciflow/inductor, ciflow/torchtitan",2025-11-24T19:05:11Z,2026-07-05T05:49:00Z,3,"This is the first part of an updated patch to #168174 modifying specific parts of Dyanmo/Inductor's global state to be thread safe. Summary: Currently, Dynamo/inductor configurations are reliant on global variables. This has led to many race conditions in workflows which call torch.compile (more specifically compile..."
176619,Fix: raise RuntimeError when cudagraph partitions are empty,https://github.com/pytorch/pytorch/pull/176619,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,176611,,,,0,18,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-03-05T18:04:49Z,2026-07-05T05:09:03Z,3,"This commit ensures that 'log_cudagraph_skip_and_bump_counter' is called when 'cudagraph_partition_post_compile' returns early due to empty partitions. This prevents silent failures when 'cudagraph_or_error=True' is set, ensuring the expected RuntimeError is raised instead of silently falling back to eager execution..."
182143,[OSDC] Migrate periodic.yml to OSDC via dial-up pattern,https://github.com/pytorch/pytorch/pull/182143,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/periodic,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, topic: not user facing, ciflow/periodic",2026-05-01T17:18:58Z,2026-07-05T02:19:58Z,4,"Enable OSDC dial-up for all three periodic.yml job pairs by plumbing use-arc, python-version, compiler, and cuda-version inputs through the reusable _linux-build.yml / _linux-test.yml workflows. Authored with Claude."
182314,[OSDC] Migrate dynamo-unittest.yml to OSDC via dial-up pattern,https://github.com/pytorch/pytorch/pull/182314,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, topic: not user facing, ciflow/dynamo",2026-05-04T18:10:30Z,2026-07-05T01:09:32Z,4,"Enable OSDC dial-up for dynamo-unittest by plumbing use-arc, python-version, and compiler inputs through _linux-build.yml and _linux-test.yml. Authored with Claude."
184947,[dynamo] fix pyright typing on torch.compiler.disable,https://github.com/pytorch/pytorch/pull/184947,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-05-23T00:09:41Z,2026-07-05T00:19:46Z,4,"Add @overload to torch.compiler.disable to constrain inferred return type The outer disable() in torch._dynamo.decorators (and its public re-export torch.compiler.disable) has no return-type annotation, so pyright infers it as the union of every branch in the body. One branch returns the inner wrap function, whose f..."
186906,Update trace_rules for torch.accelerator,https://github.com/pytorch/pytorch/pull/186906,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186599, 186600, 186771, 186777, 186899, 186902",,,0,21,0,,,"open source, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/dynamo, ciflow/torchtitan",2026-06-10T07:48:06Z,2026-07-04T23:59:10Z,4,Stack from ghstack (oldest at bottom): -> #186906 #186902 #186899 #186777 #186771 #186600 #186599 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
187061,[cuda graphs] Automatic nn.Module FQN annotation in Inductor cudagraph trees,https://github.com/pytorch/pytorch/pull/187061,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,186026,,,,0,24,0,,,"triaged, open source, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-11T16:35:53Z,2026-07-04T23:34:08Z,4,This PR implements Issue #186026 . The problem it solves is that when cuda graph is enabled it is difficult to map the kernels back to the operators which created them. Since cuda graph makes it so a single kernel launch triggers multiple GPU kernels we lose information about the origin of those kernels. Then the ma...
178664,Implement igamma/igammac backward w.r.t. first argument,https://github.com/pytorch/pytorch/pull/178664,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,80025,80025,80025,,,,0,1,0,,,"module: cpu, triaged, open source, Stale, release notes: autograd",2026-03-27T22:12:19Z,2026-07-04T22:49:12Z,4,"Summary Implements the gradient of torch.special.gammainc / torch.special.gammaincc (igamma/igammac) with respect to the first argument a (shape parameter). Previously calling .backward() on these functions raised not_implemented(""igamma: input""). Uses two computation regimes based on Stan math library: asymptotic e..."
182296,fix(nn): add input dim validation to adaptive_avg_pool2d/3d,https://github.com/pytorch/pytorch/pull/182296,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,126673,126673,126673,,,,0,1,0,,,"triaged, open source, Stale",2026-05-04T15:23:56Z,2026-07-04T22:22:31Z,4,"Fixes #126673 Summary adaptive_avg_pool2d and adaptive_avg_pool3d silently accepted inputs with the wrong number of dimensions, producing incorrectly-shaped output instead of raising an error. For example, passing a 3D tensor (N, C, L) to AdaptiveAvgPool2d would silently return a tensor of shape (N, C//output_size,..."
182291,fix(nn): InstanceNorm always raises ValueError on channel mismatch,https://github.com/pytorch/pytorch/pull/182291,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,109652,109652,109652,,,,0,1,0,,,"triaged, open source, Stale, release notes: nn",2026-05-04T15:00:00Z,2026-07-04T22:21:46Z,4,Fixes #109652 Summary _InstanceNorm.forward() previously branched on self.affine when checking channel count: affine=True -- raised ValueError (correct) affine=False (the default) -- emitted a UserWarning and silently continued (bug) This means calling InstanceNorm1d(64) with an 80-channel input would silently produ...
181622,[Inductor][NVGEMM] Refactor rendering,https://github.com/pytorch/pytorch/pull/181622,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"174893, 181902, 181903, 181909",,,0,9,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-27T18:15:57Z,2026-07-04T21:41:47Z,4,Stack from ghstack (oldest at bottom): #181909 #181903 #181902 -> #181622 #174893 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
181902,[Inductor][NVGEMM] Remove duplicate nvgemm_max_profiling config,https://github.com/pytorch/pytorch/pull/181902,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"174893, 181622, 181903, 181909",,,0,3,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-29T18:56:29Z,2026-07-04T21:37:22Z,4,Stack from ghstack (oldest at bottom): #181909 #181903 -> #181902 #181622 #174893 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
181903,[Inductor][NVGEMM] Avoid double iteration over kernel cache during choice enumeration,https://github.com/pytorch/pytorch/pull/181903,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"174893, 181622, 181902, 181909",,,0,3,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-29T18:56:34Z,2026-07-04T21:37:06Z,4,Stack from ghstack (oldest at bottom): #181909 -> #181903 #181902 #181622 #174893 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
181910,[Inductor][NVGEMM] Link tracking issue in supplement configs comment,https://github.com/pytorch/pytorch/pull/181910,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"174893, 181622, 181902, 181903, 181909",,,0,1,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-29T19:15:16Z,2026-07-04T21:32:46Z,4,Stack from ghstack (oldest at bottom): -> #181910 #181909 #181903 #181902 #181622 #174893 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
188952,Fix C2872 'std' ambiguous symbol when nvcc compiles compiled_autograd.h on Windows extension builds,https://github.com/pytorch/pytorch/pull/188952,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,"148317, 173232","148317, 173232","148317, 173232",,,,0,1,0,,,"module: build, triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, module: compiled autograd",2026-07-04T16:32:08Z,2026-07-04T21:09:29Z,4,Fixes #148317 Fixes #173232 Problem Any Windows CUDA extension build that includes torch/extension.h (SageAttention-style builds via torch.utils.cpp_extension) fails to compile under nvcc with: torch/csrc/dynamo/compiled_autograd.h(1143): error C2872: 'std': ambiguous symbol C:\...\MSVC\14.42.34433\include\valarray(...
182529,Use NCCLCommProvider for TorchComm + NCCL symmetric memory,https://github.com/pytorch/pytorch/pull/182529,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, ciflow/trunk, release notes: distributed (c10d), ciflow/h100-symm-mem, meta-exported",2026-05-05T17:03:53Z,2026-07-04T20:49:04Z,4,"Summary: When the CUDA backend is a TorchComm backend (use_torchcomms=True), ProcessGroupNCCL's communicator cache is empty because TorchComm owns its own NCCL communicator. NCCLSymmetricMemory previously hard-coded ProcessGroupNCCL::getCommPtr(), which is null in this configuration and leads to ncclCommWindowRegist..."
188579,optimize split reduction heuristics on GB200,https://github.com/pytorch/pytorch/pull/188579,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,188578,188578,188578,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-30T18:40:40Z,2026-07-04T20:26:06Z,4,"Fix #188578 Summary Add an xnumel-aware inner-reduction no-split threshold for GB200. Use 32768 when xnumel < num_sm, otherwise 40960. Update affected Inductor tests for the new SM100 split behavior. Motivation PR #179729 raised the Blackwell+ no-split threshold to 524288, which helps standalone/small-model CPU wall..."
188791,Fix MKL VML first-call dispatch race producing silent low-precision results,https://github.com/pytorch/pytorch/pull/188791,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188792,188792,188792,,,,0,2,0,,,"module: cpu, triaged, open source",2026-07-02T09:34:03Z,2026-07-04T20:17:38Z,4,Issue Fixes #188792 Summary oneMKL resolves each VML function's kernel dispatch lazily on first call; concurrent first calls from intra-op worker threads can transiently execute the low-accuracy AVX2 EP kernel for one thread's chunk despite VML_HA being requested (bit-exact evidence and repro in #188792). This adds...
188473,[MPS] Add complex64 support to linalg.lu_factor (unblocks complex det/slogdet/lu),https://github.com/pytorch/pytorch/pull/188473,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,187107,187107,187107,,,,0,3,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-29T20:10:37Z,2026-07-04T19:12:03Z,4,"Fixes #187107 Summary torch.linalg.lu_factor — and the ops that decompose through it (det, slogdet, lu) — raised RuntimeError: MPS doesn't support complex types. The float path wraps Apple's MPSMatrixDecompositionLU, which is float-only, so complex had no implementation. This adds a Metal kernel implementing LU fact..."
188394,Speed up CPU quantile/nanquantile with partial selection,https://github.com/pytorch/pytorch/pull/188394,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,64947,64947,64947,,,,0,1,0,,,"triaged, open source, release notes: python_frontend, topic: performance",2026-06-29T06:56:52Z,2026-07-04T19:03:01Z,4,"Summary Closes #64947. (Part 2/2) (float32 size-limit fix landed in #187574.) torch.quantile / torch.nanquantile fully sort the reduced dimension just to read back a few order statistics. This replaces the sort with a CPU partial selection, making the common cases (median, quartiles, a handful of quantiles, batched/..."
182216,Add Half/BFloat16 copysign overloads to fix NaN sign bit loss on CUDA,https://github.com/pytorch/pytorch/pull/182216,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,181804,181804,,,,,0,1,0,,,"module: cuda, triaged, open source, Stale, release notes: cuda",2026-05-02T07:58:26Z,2026-07-04T18:50:13Z,4,"Summary Fixes #181804 torch.copysign on CUDA returns incorrect results for negative NaN inputs with float16 and bfloat16 dtypes. The sign bit is lost because CUDAMathCompat.h only has float/double copysign overloads, causing implicit conversion from Half→float which drops the NaN sign bit. This PR adds bitwise copys..."
182268,autograd: add `list_saved_tensors` utility to traverse saved tensors in autograd graph,https://github.com/pytorch/pytorch/pull/182268,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,91692,91692,91692,,,,0,1,0,,,"triaged, open source, Stale",2026-05-04T03:16:50Z,2026-07-04T18:50:12Z,4,Fixes #91692 Summary Adds a new public function torch.autograd.graph.list_saved_tensors(tensor) that traverses the autograd computation graph from a given tensor and collects all tensors saved by each node for backward propagation. This is useful for: Understanding memory usage during forward pass Comparing memory f...
182262,Use TypeAliasType for type aliases and @exposed_in for function re-exports,https://github.com/pytorch/pytorch/pull/182262,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,171905,171905,171905,,,,0,1,0,,,"triaged, open source, Stale, release notes: optim",2026-05-04T01:22:01Z,2026-07-04T18:50:11Z,4,"Summary Replace module assignment hacks in torch.optim.init with @exposed_in() decorators at the class definition sites for Adafactor and Muon. Motivation Fixes #171905 The codebase has historically used module attribute reassignment to make imported symbols appear in their public module, which breaks static analysi..."
182204,torch.cat/concat/concatenate: support dim=None (flatten before concat),https://github.com/pytorch/pytorch/pull/182204,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,70925,70925,70925,,,,0,1,0,,,"triaged, open source, Stale",2026-05-02T04:24:07Z,2026-07-04T18:50:08Z,4,"Fixes #70925 Summary The Python Array API Standard specifies that concat(tensors, dim=None) should first flatten all tensors then concatenate them, consistent with numpy.concatenate(axis=None). Previously, torch.cat([t1, t2], dim=None) raised a RuntimeError. After this PR: t1 = torch.tensor([[1, 2], [3, 4]]) t2 = to..."
182133,Allow torch.arange() to return empty tensor when bounds are inconsistent with step sign,https://github.com/pytorch/pytorch/pull/182133,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,70915,70915,70915,,,,0,1,0,,,"triaged, open source, Stale",2026-05-01T16:29:16Z,2026-07-04T18:50:05Z,4,"Fixes #70915 Per the Array API Standard, torch.arange(start, stop, step) should return an empty tensor when stop-start and step have opposite signs, rather than raising a RuntimeError. This aligns with NumPy's behavior. Before: torch.arange(1, 0) # RuntimeError: upper bound and lower bound inconsistent with step sig..."
182061,Document low-memory state_dict loading with meta device,https://github.com/pytorch/pytorch/pull/182061,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: serialization, triaged, open source, Stale, release notes: autograd, topic: docs, topic: not user facing",2026-04-30T18:15:24Z,2026-07-04T18:50:04Z,4,"Summary document the low-memory checkpoint loading pattern using orch.load(..., mmap=True, weights_only=True) show constructing modules on the meta device and loading with load_state_dict(..., assign=True) call out optimizer ordering and that this reduces load-time CPU peak, not final model size Test plan git diff -..."
182014,[inductor] Fix spmd_check deadlock in recursive post_grad_passes,https://github.com/pytorch/pytorch/pull/182014,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,3,0,,,"Stale, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-30T12:29:35Z,2026-07-04T18:50:04Z,4,"Stack from ghstack (oldest at bottom): -> #182014 ● [inductor] Fix spmd_check deadlock in recursive post_grad_passes spmd_check uses all_gather_object — a collective matched by call order across ranks, with a 600s timeout for stragglers. Before this fix, _recursive_post_grad_passes called spmd_check at every recursi..."
182053,Partition load_state_dict child keys once,https://github.com/pytorch/pytorch/pull/182053,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, Stale",2026-04-30T17:39:39Z,2026-07-04T18:50:02Z,4,Summary partition child state_dict entries in a single pass per module during load_state_dict avoid scanning all local state_dict keys once per child module add a regression test that detects cross-sibling prefix checks Test plan python -m py_compile torch\nn\modules\module.py test\nn\test_load_state_dict.py git dif...
187653,Add NanDetectMode for forward-pass NaN/Inf detection,https://github.com/pytorch/pytorch/pull/187653,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,160016,160016,160016,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-18T10:05:20Z,2026-07-04T15:13:42Z,4,Fixes #160016 Adds a TorchDispatchMode that detects NaN (and optionally Inf) in the output of every operation during the forward pass: from torch.utils.nan_detect import NanDetectMode with NanDetectMode(): out = model(x) # RuntimeError: Function aten.add.Tensor returned NaN values This complements torch.autograd.det...
187477,Add context manager support to torch.dtype,https://github.com/pytorch/pytorch/pull/187477,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,120819,120819,120819,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-16T18:46:08Z,2026-07-04T15:13:41Z,4,"Fixes #120819 Adds context manager support to torch.dtype objects, enabling: with torch.float64: model = init_model() # all params created as float64 # original dtype restored automatically This mirrors the existing with torch.device(...) pattern. Without this, users must manually save/restore via set_default_dtype:..."
182076,[Fix] Add overflow and validation checks for pixel_shuffle in decomposition and meta,https://github.com/pytorch/pytorch/pull/182076,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,171838,171838,171838,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-04-30T21:25:25Z,2026-07-04T15:13:40Z,4,"Fixes #171838 The eager C++ crash (SIGFPE) from an extremely large upscale_factor was fixed in #163154. However, the torch.compile path (Python decomposition in _refs and the meta registration) still lacked validation, which could cause incorrect shapes or crashes during compilation. This adds three checks to both _..."
181756,[Docs] Replace outdated flake8 and git-pre-commit references with lintrunner,https://github.com/pytorch/pytorch/pull/181756,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,164478,164478,164478,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-04-28T17:18:24Z,2026-07-04T15:13:39Z,4,"Fixes #164478 The ""Pre-commit tidy/linting hook"" section in CONTRIBUTING.md referenced two items that no longer exist in the repository: tools/git-pre-commit (removed in #76984) flake8 / flake8-requirements.txt (replaced by lintrunner in #121657) This PR updates the section to point to the current linting workflow u..."
183776,C++ FakeTensor,https://github.com/pytorch/pytorch/pull/183776,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,58,0,,,"module: mkldnn, release notes: fx, skip-pr-sanity-checks, module: inductor, module: dynamo, ciflow/inductor, module: compiled autograd, release notes: inductor (aoti), ciflow/torchtitan, ciflow/docker",2026-05-14T20:10:24Z,2026-07-04T13:07:15Z,4,"Reimplementing FakeTensor in C++. landing one big PR now with majority of logic. this PR doesn't actually move any dynamo/make_fx/inductor logic over to use C++, will do that in a separate PR. decomps and @register_op_impl_ handling is still using python implementations, will need to handle this in a followup PR. to..."
175082,[pytree] make `torch.utils.pytree` as public API,https://github.com/pytorch/pytorch/pull/175082,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,65761,65761,,"175083, 178051",,,0,34,0,,,"open source, ciflow/trunk, release notes: python_frontend, module: pytree, module: dynamo, ciflow/inductor, keep-going, release notes: export, ciflow/torchtitan",2026-02-16T11:53:10Z,2026-07-04T12:38:27Z,4,"Stack from ghstack (oldest at bottom): #178051 #175083 -> #175082 Create a new public module torch.utils.pytree that serves as the official user-facing API for pytree operations. Previously, users had to import from internal modules torch.utils._pytree or torch.utils._cxx_pytree. Now they can use torch.utils.pytree..."
181862,collect_env: add uv support,https://github.com/pytorch/pytorch/pull/181862,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-04-29T12:54:10Z,2026-07-04T12:16:34Z,4,"uv is becoming a very popular package manager, and may people are ditching pip and conda entirely. This PR adds support for uv-managed installations in the collect_env.py script. With uv > python3 torch/utils/collect_env.py Collecting environment information... PyTorch version: 2.11.0 Is debug build: False CUDA used..."
178051,[pytree][dynamo] add generic_pytree (torch.utils.pytree) to test_tree_map.py parametrization,https://github.com/pytorch/pytorch/pull/178051,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"175082, 175083",,,0,13,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-03-21T09:28:53Z,2026-07-04T12:10:38Z,4,"Stack from ghstack (oldest at bottom): -> #178051 #175083 #175082 Summary Add the new public torch.utils.pytree module (generic_pytree) to the pytree_modules dict in test/dynamo/test_tree_map.py, so it gets coverage in the tree_map compile tests. This is a 2-line change on top of the refactor in #178050. The generic..."
175083,[pytree] add another simplified pytree module `torch.pytree`,https://github.com/pytorch/pytorch/pull/175083,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"175082, 178051",,,0,36,0,,,"open source, ciflow/trunk, release notes: python_frontend, module: pytree",2026-02-16T11:53:23Z,2026-07-04T12:10:37Z,4,"Stack from ghstack (oldest at bottom): #178051 -> #175083 #175082 Differences between torch.pytree and torch.utils.pytree: APIs in torch.utils.pytree have a tree_ prefix: leaves, treespec = torch.utils.pytree.tree_flatten(tree) new_tree = torch.utils.pytree.tree_map(func, tree) leaevs, treespec = torch.pytree.flatte..."
187732,[DDP] Fix static_graph=True + no_sync() gradient accumulation regression,https://github.com/pytorch/pytorch/pull/187732,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: distributed (c10d)",2026-06-19T16:48:05Z,2026-07-04T11:57:08Z,4,"DDP(static_graph=True) with gradient accumulation via no_sync() crashes on the first backward with an expect_autograd_hooks_ INTERNAL ASSERT, and once that is unblocked it silently stops all-reducing gradients from the second iteration on. Both are regressions from #103487, which re-keyed the static-graph ""first ite..."
180875,[CUDA][cuBLASLt] torch.addmm - improve numerics of bias fusion with reduced precision SplitK algorithms,https://github.com/pytorch/pytorch/pull/180875,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,17,0,,,"module: cuda, module: cublas, open source, Stale, ciflow/trunk, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-04-20T16:53:17Z,2026-07-04T11:55:40Z,4,"Stack from ghstack (oldest at bottom): -> #180875 After extensive statistical testing of cuBLASLt vs cuBLAS PyTorch dispatching paths, the culprit likely to cause numerical differences is found with high statistical confidence -- bias epilogue fusions with SplitK algorithms when run with reduced precision inputs. Th..."
176265,Add backward batching rules for SDPA and fix attn_bias under vmap,https://github.com/pytorch/pytorch/pull/176265,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,"117016, 151558","117016, 151558","117016, 151558",,,,0,3,0,,,"triaged, open source, release notes: composability, topic: performance",2026-03-03T09:24:55Z,2026-07-04T10:38:32Z,4,"Adds backward batching rules for all three fused SDPA backends (efficient, flash, cuDNN), eliminating the per-sample fallback loop under vmap(grad(...)). Follows the same flatten-into-batch pattern as the forward rules added in #133964. Also fixes #151558 (attn_bias shape mismatch under vmap) by expanding non-vmappe..."
182415,[c10d] Apply record_comm names to NCCL launch annotations,https://github.com/pytorch/pytorch/pull/182415,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, release notes: distributed (c10d), ciflow/torchtitan",2026-05-05T06:35:15Z,2026-07-04T09:18:40Z,4,"Stack from ghstack (oldest at bottom): -> #182415 record_comm currently overrides the Work profiler range name. For NCCL async work, that long Work lifetime range can span unrelated compute in traces. Keep NCCL Work lifetime ranges using their backend names and apply the record_comm name at the NCCL enqueue site und..."
178500,[autograd] Support Mapping and OrderedDict inputs for autograd.grad and autograd.backward,https://github.com/pytorch/pytorch/pull/178500,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,23,0,,,"module: autograd, open source, ciflow/trunk, release notes: autograd, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-03-26T11:35:25Z,2026-07-04T08:36:05Z,4,"Stack from ghstack (oldest at bottom): -> #178500 Widen the inputs argument from dict-only (introduced in #178140) to any Mapping[str, Tensor]. OrderedDict inputs produce OrderedDict outputs; all other Mapping types produce dict. This is useful since OrderedDict is widely used as state_dict in PyTorch. See also: #17..."
182134,[dynamo_wrapped] Remove stale dynamo_expected_failures for torch_np error-path tests,https://github.com/pytorch/pytorch/pull/182134,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,182398,,,0,9,0,,,"Merged, Reverted, Stale, ciflow/trunk, topic: not user facing, keep-going, ci-no-td",2026-05-01T16:32:15Z,2026-07-04T07:27:38Z,4,"Stack from ghstack (oldest at bottom): -> #182134 #182398 #182104 fixed dynamo to graph-break at with pytest.raises(...) again under pytest 9 by re-adding the renamed RaisesExc class to is_forbidden_context_manager. That made these torch_np error-path tests genuinely pass under dynamo_wrapped, so the framework now r..."
188378,"Fix linspace/logspace docstrings: start/end are Number, not float",https://github.com/pytorch/pytorch/pull/188378,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,129338,129338,"129338, 130254",,,,0,1,0,,,"open source, release notes: python_frontend, topic: not user facing",2026-06-28T22:51:47Z,2026-07-04T07:01:07Z,4,"Fixes #129338. Also covers torch.logspace (#130254), which has the identical wording. torch.linspace and torch.logspace document start and end as float or Tensor: start (float or Tensor): the starting value for the set of points. ... end (float or Tensor): the ending value for the set of points. ... But this is narr..."
178591,fixes:division by zero check for inductor,https://github.com/pytorch/pytorch/pull/178591,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,178480,178480,,,,,0,9,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-03-27T13:18:15Z,2026-07-04T02:27:03Z,5,fixes: #178480 add check to division by zero for inductor. add code checking the divisor is zero for the function in floordiv/truncdiv/truediv in class CppOverrides cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjam...
182389,Collective buffer detection for DDP and FSDP,https://github.com/pytorch/pytorch/pull/182389,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182387, 182388",,,0,1,0,,,"Stale, release notes: distributed (c10d)",2026-05-04T23:34:03Z,2026-07-04T02:20:07Z,5,"Stack from ghstack (oldest at bottom): -> #182389 #182388 #182387 Summary This PR automatically tags communication buffer allocations (allreduce, all-gather, reduce-scatter buffers) as COLLECTIVE_BUFFER without requiring changes to each distributed strategy's implementation. This is a change on FSDP codebase, and I..."
182388,Training hooks and public API for memory component tracking,https://github.com/pytorch/pytorch/pull/182388,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182387, 182389",,,0,1,0,,,Stale,2026-05-04T23:33:58Z,2026-07-04T02:20:06Z,5,Stack from ghstack (oldest at bottom): #182389 -> #182388 #182387 Summary This PR wires the allocator primitives from PR 1 into standard training via lightweight global hooks. Users call component_tracking.enable() once — no model or optimizer reference needed — and every allocation is automatically attributed. Desi...
182387,Allocator-level memory attribution primitives,https://github.com/pytorch/pytorch/pull/182387,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182388, 182389",,,0,1,0,,,Stale,2026-05-04T23:33:53Z,2026-07-04T02:20:05Z,5,"Stack from ghstack (oldest at bottom): #182389 #182388 -> #182387 Summary This PR adds the ability to answer ""what is this GPU memory being used for?"" at the allocator level. Every CUDA allocation is now stamped with a component type (parameter, gradient, activation, optimizer state, etc.) at allocation time, with z..."
182340,Bump Windows CPU trunk test shards from 4 to 5,https://github.com/pytorch/pytorch/pull/182340,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"Stale, ciflow/trunk, topic: not user facing",2026-05-04T20:16:02Z,2026-07-04T01:04:22Z,5,"Increase the default test-matrix sharding for win-vs2022-cpu-py3 in trunk to reduce per-shard runtime as the test keeps timing out https://github.com/pytorch/pytorch/actions/runs/25328038725/job/74262204349, probably due to pytest upgrade. Authored by Claude."
182362,[PyTorch][AOTI] Add timing instrumentation to AOTI model loading pipeline,https://github.com/pytorch/pytorch/pull/182362,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, ciflow/inductor, release notes: inductor (aoti), meta-exported",2026-05-04T21:13:05Z,2026-07-03T23:50:44Z,5,"Summary: After the AOTI model .so is loaded and extern kernel nodes JSON is parsed, there is a ~30 second gap with zero logging before the model is ready. This makes it impossible to diagnose whether the time is spent in GPU memory allocation (cudaMalloc), host-to-device weight transfer (cudaMemcpy), model instance..."
182310,Add Triton scaled-mm v2 blockwise recipes,https://github.com/pytorch/pytorch/pull/182310,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-05-04T17:44:33Z,2026-07-03T22:50:18Z,5,Stack from ghstack (oldest at bottom): (to be filled) cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
187556,[MPS] Migrate smooth_l1_loss and huber_loss to native Metal kernels,https://github.com/pytorch/pytorch/pull/187556,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,187455,"187553, 187557",,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-17T07:53:57Z,2026-07-03T22:25:26Z,5,"Part of #187455. Leaf on the fused-loss base #187553 (mse_loss), sibling of #187557 (binary_cross_entropy) — this PR carries only the smooth_l1/huber delta. (Both siblings touch the shared fused_loss_bwd_fast_path helper with a few-line extension each — p1 here, the weight buffer there — so whichever lands second ta..."
178849,Enable eager tests for inductor-unittest - - Enabel more tests in test_ops.py on Intel GPU,https://github.com/pytorch/pytorch/pull/178849,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/xpu,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"176688, 176689, 179549, 179550",,,0,21,0,,,"open source, release notes: python_frontend, ciflow/xpu",2026-03-31T04:14:06Z,2026-07-03T21:38:20Z,5,"Add more cases in test_ops.py 1. Generalize TestCommon, TestCompositeCompliance, TestMathBits, TestFakeTensor, TestForwardADWithScalars with onlyAccelerator 2. Enable xpu test with instantiate_device_type_tests() allow_xpu=True augment 3. skip cases or ops with limitations in xpu. 4. remove skipIfXpu according to cu..."
177457,[DTensor] Coalesce collectives for multi-arg redistribution,https://github.com/pytorch/pytorch/pull/177457,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,170367,170367,170367,,,,0,4,0,,,"triaged, open source, Stale, release notes: distributed (dtensor)",2026-03-14T19:23:59Z,2026-07-03T21:14:35Z,5,"Fixes #170367 When multiple op arguments need the same redistribution (same mesh, same src/dst placements), batch them into a single coalesced collective call instead of issuing N separate NCCL kernel launches. This is most impactful for foreach/fused ops (e.g. foreach_add, fused_adam) during optimizer steps with ma..."
188918,[dynamo] Model list subclass __new__/__init__ arg handling to match CPython,https://github.com/pytorch/pytorch/pull/188918,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-03T18:05:59Z,2026-07-03T21:13:49Z,5,"Stack from ghstack (oldest at bottom): (to be filled) Root cause: CPython's list.new (tp_new = PyType_GenericNew) ignores all args and kwargs, so super().new(cls, seq) just returns an empty instance. Its list.init (Argument Clinic) rejects keyword arguments only when the type does not override new (Py_TYPE(self)->tp..."
182146,[caffe2] Add defensive fallback in get_logging_handler for unknown destinations (#182146),https://github.com/pytorch/pytorch/pull/182146,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, release notes: distributed (torchelastic), meta-exported, ciflow/torchtitan",2026-05-01T17:34:47Z,2026-07-03T20:59:18Z,5,"Summary: When PET_EVENT_LOG_HANDLER is set to a handler name that doesn't exist in the _log_handlers dict (e.g., stale fbpkg built before a new handler was registered), get_logging_handler() previously threw a raw KeyError that crashed the elastic agent worker. Now falls back to the ""null"" handler with a warning ins..."
176239,[dtensor][partial] just_knob error message for fixing .item() on partials,https://github.com/pytorch/pytorch/pull/176239,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,1,0,,,"Stale, ciflow/trunk, ciflow/inductor, release notes: distributed (dtensor)",2026-03-02T23:39:33Z,2026-07-03T20:59:09Z,5,Stack from ghstack (oldest at bottom): -> #176239 Differential Revision: D94975550
188646,[inductor] Add tile_n=192 and missing cluster configs to NVGEMM supplement set,https://github.com/pytorch/pytorch/pull/188646,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-07-01T05:47:55Z,2026-07-03T20:15:05Z,5,"Stack from ghstack (oldest at bottom): -> #188646 #188865 #188303 Benchmarking identified (128, 192, 1, 4) and (128, 64, 1, 4) tile configs that beat cuBLAS on tall-K decode shapes (e.g. Qwen3-32B down_proj), but nvMatmulHeuristics doesn't explore them. Add these plus surrounding cluster variants to the supplement c..."
181412,Add Claude-powered operator benchmark triage,https://github.com/pytorch/pytorch/pull/181412,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, release notes: releng, topic: not user facing",2026-04-24T18:07:06Z,2026-07-03T17:59:34Z,5,"This PR adds automated analysis for operator benchmarks using Claude Code, following the patterns from claude-issue-triage and treehugger workflows. Components 1. Benchmark Triage Skill Analyzes PR changes to determine which operator benchmarks should run Maps code changes to relevant benchmarks using benchmark_mapp..."
180083,[FSDP] Make FSDP v1 tests device-agnostic for PrivateUse1 backends,https://github.com/pytorch/pytorch/pull/180083,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem,"identify whether block is CI, merge conflict, or review gate",,DIRTY,False,0,0,,,,,,180050,,,0,2,0,,,"triaged, open source, Stale, release notes: distributed (fsdp), module: PrivateUse1, ciflow/h100-symm-mem",2026-04-10T23:16:01Z,2026-07-03T16:59:22Z,5,"Summary Replace hardcoded CUDA/NCCL references in FSDP v1 test files with device-agnostic equivalents, enabling PrivateUse1 backends to run these tests natively. This builds on top of #180050 (FSDP v2 device-agnostic changes). Changes Source changes (torch/testing/_internal/): common_fsdp.py: Cap DEVICE_COUNT with W..."
187540,"[DRAFT] Fix forward AD inconsistency for pow(abs(complex), complex)",https://github.com/pytorch/pytorch/pull/187540,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,185639,,,,0,2,0,,,open source,2026-06-17T01:54:01Z,2026-07-03T16:56:53Z,5,"Related to #185639 Summary Fix a forward AD inconsistency for pow(abs(complex), complex). The generated JVP rule for pow.Tensor_Tensor currently reuses pow_backward_self(...), which applies handle_r_to_c(...). When the base is real (e.g., torch.abs(x)), this projection removes the imaginary component and causes jacf..."
187691,[BE]: Update fmtlib to 12.2.0,https://github.com/pytorch/pytorch/pull/187691,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-18T19:45:40Z,2026-07-03T16:42:16Z,5,Big release. Lots of performance improvement changes and bugfixes: https://github.com/fmtlib/fmt/releases/tag/12.2.0
187654,[dynamo] Remove hardcoded CUDA references in test_logging.py,https://github.com/pytorch/pytorch/pull/187654,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-18T10:18:40Z,2026-07-03T16:14:17Z,5,"Summary Makes test/dynamo/test_logging.py device-agnostic by fixing two tests that had hardcoded CUDA dependencies: test_ddp_graphs: Replaced @requires_cuda_and_triton with @requires_gpu and substituted hardcoded ""cuda:0"" with the module-level device_type variable. Made device_ids conditional since it is a CUDA-spec..."
182226,[CI] Install git-core PPA without add-apt-repository,https://github.com/pytorch/pytorch/pull/182226,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, topic: not user facing",2026-05-02T17:40:25Z,2026-07-03T16:04:25Z,5,"https://status.canonical.com is showing major outage in the past few days with service still being recoverd add-apt-repository ppa:git-core/ppa resolves the PPA's signing key by calling api.launchpad.net, and the Docker base image build was failing when that endpoint returned 504. Switch to fetching the key from key..."
188887,Enable cpp_wrapper for mixed order reduction,https://github.com/pytorch/pytorch/pull/188887,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-07-03T09:29:38Z,2026-07-03T14:13:29Z,5,"Enable tirton.mix_order_reduction when cpp_wrapper=True on CUDA. Kernel signature / arg filtering. The mix-order fused kernel carries a NUM_STAGES tl.constexpr that CppWrapper.Gpu._resolve_lazy_arg_names did not filter (its internal_config_re only matched BLOCK/RSPLIT), causing an arg-count mismatch at wrapper gener..."
182426,Enable FP32 emulation via BF16 in cuSOLVER for qr and lu,https://github.com/pytorch/pytorch/pull/182426,ready for maintainer decision,45,CI lane labels: ciflow/b200; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,181849,,,0,6,0,,,"triaged, open source, module: linear algebra, release notes: cuda, topic: performance, ciflow/b200",2026-05-05T12:49:36Z,2026-07-03T13:51:18Z,5,"Continuation of #181849 for LU and QR factorization. Summary As discussed in #181849, the numerics of other operations are not reliable enough yet to enable FP32 emulation on Blackwell unconditionally. This PR therefore introduces a flag in getCurrentCUDASolverDnHandle to enable or disable emulation explicitly, maki..."
179212,feat(pipelining): add configurable auto-partitioning algorithm for pipeline parallelism,https://github.com/pytorch/pytorch/pull/179212,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: distributed (pipeline)",2026-04-03T01:09:50Z,2026-07-03T13:20:31Z,5,"Summary This PR introduces a configurable automatic pipeline partitioning algorithm (_partition.py) into torch.distributed.pipelining. The algorithm balances forward/backward computation costs across stages, enabling optimal layer assignment for various pipeline schedules (e.g., 1F1B, Interleaved1F1B, InterleavedZer..."
188789,[xpu] Add IPC memory handle sharing support to XPUCachingAllocator,https://github.com/pytorch/pytorch/pull/188789,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,188888,,,0,5,0,,,"open source, ciflow/trunk, ciflow/xpu",2026-07-02T08:44:26Z,2026-07-03T11:56:08Z,5,"Stack from ghstack (oldest at bottom): #188888 -> #188789 Motivation This PR adds IPC sharing support to XPUCachingAllocator. Currently, IPC support for expandable segments is not included because it depends on OneAPI 2026.2. Additional Context This PR requires oneAPI 2026.0"
188660,[Inductor] Enable mix-order reduction for wide reductions + tests,https://github.com/pytorch/pytorch/pull/188660,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,185534,186370,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-07-01T11:44:54Z,2026-07-03T11:52:52Z,5,"Issue Fixes part of #185534 (per suggested Fix 2). The strict-mode mix-order gate required nrow >= ncol*2, rejecting profitable wide reductions (large ncol, where the per-row persistent reduction already saturates bandwidth). Complementary to PR #186370 (which handles square shapes via a different code path), and th..."
184372,[Draft] Remove ACL,https://github.com/pytorch/pytorch/pull/184372,draft/noise,31,"PR is draft; review decision requests changes; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, open source, ciflow/trunk, release notes: quantization, release notes: releng, module: inductor, ciflow/inductor, ci-no-td, no-runner-experiments, ciflow/torchtitan, ciflow/docker",2026-05-19T13:28:39Z,2026-07-03T11:46:42Z,5,A draft PR to pull in @jondea's changes in uxlfoundation/oneDNN #5156. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coc...
188888,Add Reusable Event support to XPUEvent,https://github.com/pytorch/pytorch/pull/188888,draft/noise,13,PR is draft,wait for author to mark ready,,CLEAN,True,0,0,,,,,,188789,,,0,1,0,,,open source,2026-07-03T09:37:12Z,2026-07-03T11:07:34Z,5,Stack from ghstack (oldest at bottom): -> #188888 #188789
183323,Adding conversion from PyObject to torch::stable::tensor,https://github.com/pytorch/pytorch/pull/183323,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,9,0,,,"triaged, open source, ciflow/docker",2026-05-11T15:06:15Z,2026-07-03T11:00:36Z,5,"Hi, I would like to propose PR with adding conversions from python object to torch::stable::tensor and vice versa. Motivation I work in Transformer Engine project. We want to use torch stable ABI, but we cannot do it in recommended way with registering custom operation in libtorch like in this presentation. There ar..."
187557,[MPS] Migrate binary_cross_entropy to native Metal kernels,https://github.com/pytorch/pytorch/pull/187557,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,187455,187553,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-17T07:56:13Z,2026-07-03T10:54:23Z,5,"Part of #187455. Leaf on the fused-loss base #187553 (mse_loss) — this PR carries only the binary_cross_entropy delta; the shared infrastructure (FusedLossParams, fused_loss_pass1, fused_loss_reduce) is reviewed there. Summary Replaces the MPSGraph binary_cross_entropy forward + backward on MPS with native Metal, re..."
176420,[XPU] Ensure dense (canonical) strides for convolution inputs,https://github.com/pytorch/pytorch/pull/176420,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"module: cpu, triaged, open source",2026-03-04T10:36:28Z,2026-07-03T10:11:35Z,5,"Fix for #2239. - op_ut,third_party.torch-xpu-ops.test.xpu.functorch.test_ops_xpu.TestOperatorsXPU,test_vmapvjp_nn_functional_conv3d_xpu_float32 This commit fixes a bug where conv3d backward produces wrong results for tensors that are contiguous but have non-dense strides. Observed behavior: calling vjp on a batched..."
188760,Guard user stream compile tests with xfailIfNoAcceleratorTriton,https://github.com/pytorch/pytorch/pull/188760,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-07-01T23:52:57Z,2026-07-03T10:11:34Z,5,Summary This PR adds @xfailIfNoAcceleratorTriton at the test method level for inductor tests so torch.compile xfails instead of failing in environments without Triton test/inductor/test_user_streams.py TestStreamCodegen.test_generated_code_uses_get_stream_by_index TestStreamExternalObjectRestore.test_restore_externa...
181727,[xpu][2/4]Implement scaled_mm_v1 for MXFP8/MXFP4/NVFP4 on XPU,https://github.com/pytorch/pytorch/pull/181727,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181726, 181728, 187315, 187318",,,0,3,0,,,"module: cpu, triaged, open source",2026-04-28T07:30:15Z,2026-07-03T07:59:30Z,5,"Add blockwise scaling support (1x32), (1x16) to the v1 _scaled_mm API for XPU, enabling MXFP8/MXFP4/NVFP4. PR Stack: Since I don't have ghstack permission, I manually created the following stacked PRs for review. I also created a combined PR #187318 to test on CI. #181726 [xpu][1/4]Implement scaled_mm_v2 for MXFP8/M..."
174864,[XPU] Fix accuracy issue in addmm for bf16/f16 dtypes,https://github.com/pytorch/pytorch/pull/174864,draft/noise,41,PR is draft; review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/xpu,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,0,0,,,2837,2837,,,,,0,5,0,,,"module: cpu, open source, topic: not user facing, ciflow/xpu",2026-02-12T07:19:58Z,2026-07-03T07:46:44Z,5,Fixes intel/torch-xpu-ops#2837 Fixes precision loss in XPU addmm and baddbmm for bf16/f16 dtypes with non-trivial alpha/beta. The oneDNN 3-step post-op chain rounds intermediates to reduced precision at each stage. The fix pre-copies self into result and uses post_sum which accumulates in oneDNN's internal f32 accum...
181849,Enable FP32 emulation via BF16 in cuSOLVER,https://github.com/pytorch/pytorch/pull/181849,ready for maintainer decision,45,CI lane labels: ciflow/b200; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, module: linear algebra, Stale, release notes: cuda, topic: performance, ciflow/b200",2026-04-29T08:02:07Z,2026-07-03T07:39:16Z,5,"Summary Adds support for emulation of FP32 accuracy by emulating via 9 BF16 operations by selecting the corresponding cuSOLVER math mode for all cuSOLVER operations in linalg using the handle generated in CusolverDnHandlePool.cpp . Motivation With Blackwell, Nvidia is moving the focus of their accelerators towards l..."
188379,[HOP] None input handling for scan,https://github.com/pytorch/pytorch/pull/188379,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,11,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-28T22:53:03Z,2026-07-03T07:14:35Z,5,"Currently the scan HOP does not allows None inputs for the init as well as for the xs. This is though allowed in the JAX library and creates a gap between the torch and the JAX implementation, see for example https://gist.github.com/zou3519/80f188f5d31e2adab0e089cffc644b1c#file-scan_gaps-py-L673. This PR introduces..."
186902,Generalize dynamo RNG-related part to be device-agnostic,https://github.com/pytorch/pytorch/pull/186902,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,186867,"186599, 186600, 186771, 186777, 186899, 186906",,,0,19,0,,,"open source, ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-06-10T07:26:28Z,2026-07-03T07:06:47Z,5,Stack from ghstack (oldest at bottom): #186906 -> #186902 #186899 #186777 #186771 #186600 #186599 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98
188719,Fix Inductor KeyError on view_as_complex of non-contiguous CUDA tensors,https://github.com/pytorch/pytorch/pull/188719,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,188718,188718,188718,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-07-01T19:10:09Z,2026-07-03T07:04:20Z,5,"Issue Fixes #188718 Summary Under torch.compile, view_as_complex on a valid non-contiguous CUDA tensor crashes Inductor codegen with KeyError: 'complex64' (and 'complex128' for float64 inputs), while eager and contiguous inputs work. The fallback was registered with require_contiguous, which forces the input fully c..."
184705,[dynamo] Preserve exception chaining in InternalTorchDynamoError rewrapping,https://github.com/pytorch/pytorch/pull/184705,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,113800,113800,113800,,,,0,1,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-05-21T12:22:54Z,2026-07-07T05:47:17Z,1,"When Dynamo catches an unexpected exception during compilation and rewraps it as InternalTorchDynamoError, the original from None suppressed the entire exception chain. If the original exception had a __cause__ (e.g. a RuntimeError caused by a KeyError), that deeper chain was completely inaccessible — not in __cause..."
188858,[DO NOT MERGE] Complex Wrapper Test,https://github.com/pytorch/pytorch/pull/188858,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,open source,2026-07-02T22:29:15Z,2026-07-03T05:47:09Z,5,Stack from ghstack (oldest at bottom): -> #188858
188734,Add tracer kwarg to torch.compile for make_fx tracing,https://github.com/pytorch/pytorch/pull/188734,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: inductor, ciflow/inductor",2026-07-01T20:46:29Z,2026-07-03T04:22:06Z,5,"Stack from ghstack (oldest at bottom): -> #188734 Make the graph-capture mechanism behind torch.compile configurable via a new tracer keyword argument. The default, tracer=""dynamo"", is unchanged and routes frames through TorchDynamo. tracer=""make_fx"" bypasses Dynamo entirely: on the first call it traces the callable..."
185079,Fix invalid Dynamo setattr exception timing,https://github.com/pytorch/pytorch/pull/185079,ready for maintainer decision,67,"needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review diff and decide acceptance path,,CLEAN,False,0,0,,,,165385,165385,187707,,,0,1,0,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic, needs-review",2026-05-24T18:46:38Z,2026-07-03T03:44:16Z,6,"Stack from ghstack (oldest at bottom): -> #185079 Dynamo records user object attribute mutations as side effects and replays them later when reconstructing the frame. That is safe for ordinary instance dict and slot writes, but it changes eager semantics for opaque C getset descriptors: a write such as obj.__weakref..."
188115,[Inductor][TEST] Align `test_main_loop_scaling` with H100 support surface,https://github.com/pytorch/pytorch/pull/188115,PR blocked,98,"failed checks or blocked/unstable merge state; high-priority/blocker/regression label present; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,11,1,"statusCheckRollup, linux-jammy-aarch64-py3.10 / build-osdc, linux-jammy-cuda13.0-py3.10-gcc11-sm90 / test-osdc (smoke, 1, 1, mt-l-x86iamx-22-225-h100), linux-jammy-cuda13.0-py3.12-gcc11-sm100 / test-osdc (smoke_b200, 1, 1, mt-l-x86iamx-22-225-b200), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), lintrunner-noclang-partial / lint",statusCheckRollup,,,188721,,,,0,2,0,bd76d8d34489957ad4cc62793aa4d59b23eea7704e9e816055623cc5068ac4c1,dcac37dbf1840f716c8287ca7da5f0e97c08b2f4f638f1f342e2579035157699,"high priority, module: cuda, module: tests, triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-06-24T23:52:26Z,2026-07-03T02:51:54Z,6,"Otherwise fails with e.g., Traceback (most recent call last): File ""/usr/lib/python3.12/unittest/case.py"", line 58, in testPartExecutor yield File ""/usr/lib/python3.12/unittest/case.py"", line 634, in run self._callTestMethod(testMethod) File ""/usr/lib/python3.12/unittest/case.py"", line 589, in _callTestMethod if met..."
185115,[Test] Enable TestSDPAGpuOnly on XPU and skip cases unsupported by torch-xpu-ops,https://github.com/pytorch/pytorch/pull/185115,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"185113, 185114",,,0,4,0,,,"open source, topic: not user facing, ciflow/xpu, ci-no-td",2026-05-25T07:27:43Z,2026-07-03T02:42:08Z,6,"Stack from ghstack (oldest at bottom): -> #185115 #185114 #185113 Layer all XPU-specific changes on top of the generalization commit. Three things happen here: Instantiate TestSDPAGpuOnly on XPU. Extend instantiate_device_type_tests to only_for=('cuda', 'xpu') with allow_xpu=True. The class was renamed and de-CUDA-f..."
188633,Draft: Update CUTLASS fence proxy API usage,https://github.com/pytorch/pytorch/pull/188633,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,188477,188477,188477,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-07-01T01:49:02Z,2026-07-03T02:40:12Z,6,"Fixes #188477 The B200 smoke job installs cutlass_api from NVIDIA's floating cutlass_api branch while nvidia-cutlass-dsl is pinned to 4.5.2. The downloaded CUTLASS provider still used the older cute.arch.ProxyKind.async_shared and cute.arch.SharedSpace.shared_cta enum API, which is not available in the pinned DSL wh..."
188528,cuda: reject negative caching allocator sizes,https://github.com/pytorch/pytorch/pull/188528,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,177827,177827,177827,,,,0,3,0,,,"triaged, open source, release notes: cuda",2026-06-30T03:21:51Z,2026-07-03T01:38:20Z,6,"Fixes #177827 Summary: Adds Python-side validation in torch.cuda.memory.caching_allocator_alloc before resolving the CUDA device/stream or calling the raw caching allocator. Rejects negative integer allocation sizes with a clear ValueError. Rejects non-int sizes with TypeError, including float values and tensor scal..."
169916,[CPU] Support multiple ncores-per-instance settings in xeon/run_cpu.py,https://github.com/pytorch/pytorch/pull/169916,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,57,0,,,"module: cpu, triaged, open source, release notes: cpu (x86)",2025-12-09T03:38:55Z,2026-07-03T01:18:14Z,6,"The previous run_cpu.py assumed that the CPU cores were always evenly distributed among the sub-numa nodes. However, in some recent Xeon SKUs, the assumption does not hold. The current script would not be functioning well in --throughput-mode or --latency-mode since cross-numa cores are assigned to the respective in..."
175472,Add FA4 fp8 implementation to SDPA,https://github.com/pytorch/pytorch/pull/175472,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"182919, 182920",,,0,6,0,,,"open source, release notes: nn",2026-02-21T03:07:49Z,2026-07-03T01:01:47Z,6,"Stack from ghstack (oldest at bottom): #182920 #182919 -> #175472 Summary Added FA4 fp8 implementation using the same FA3 pathway in #172040 using the torch.ops.aten._scaled_dot_product_flash_attention.quantized overload The user interaction path is the same, using _scaled_dot_product_attention_quantized.py. This is..."
182919,Add forward-only cuDNN FP8 per-tensor attention,https://github.com/pytorch/pytorch/pull/182919,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"175472, 182920",,,0,2,0,,,"open source, ciflow/inductor, release notes: inductor (aoti)",2026-05-08T00:59:49Z,2026-07-03T01:01:36Z,6,Stack from ghstack (oldest at bottom): #182920 -> #182919 #175472 Add support for FP8 (float8_e4m3fn) scaled dot product attention through the cuDNN backend with per-tensor descaling. This is forward-only; backward pass will be added in a follow-up commit. The implementation routes through DescaleType.PER_TENSOR in...
182920,Add backward pass for cuDNN FP8 per-tensor attention,https://github.com/pytorch/pytorch/pull/182920,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"175472, 182919",,,0,2,0,,,"open source, ciflow/inductor, release notes: inductor (aoti)",2026-05-08T00:59:56Z,2026-07-03T01:01:35Z,6,"Stack from ghstack (oldest at bottom): -> #182920 #182919 #175472 Add the backward op for cuDNN FP8 per-tensor attention. The backward op is exposed as a raw ATen op without autograd wiring (no derivatives.yaml entry) — torchao's custom autograd Function will connect forward and backward, handling delayed scaling be..."
179224,spaceship operator for comparisons in Dict.h,https://github.com/pytorch/pytorch/pull/179224,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,176662,,,,0,2,0,,,"open source, Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td",2026-04-03T06:40:33Z,2026-07-03T00:20:37Z,6,"We have C++20 in PyTorch now #176662 Spaceship/Three-way comparison operator automatically generates <, <=, >, >= operator== automatically generates operator!= This reduces a little bit of code."
175947,[dynamo] Remove the dict over guarding on dict setitem,https://github.com/pytorch/pytorch/pull/175947,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175858,175858,,,,0,2,0,,,"Merged, Reverted, Stale, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ci-no-td, ciflow/dynamo, ciflow/torchtitan",2026-02-27T05:41:43Z,2026-07-03T00:10:44Z,6,Stack from ghstack (oldest at bottom): -> #175947 Fixes #175858 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo @azahed98
166388,Fixing assertion fail by checking if no_sync is true before DDPsink,https://github.com/pytorch/pytorch/pull/166388,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,143580,143580,143580,,,,0,1,0,,,"oncall: distributed, triaged, open source, topic: not user facing",2025-10-28T08:27:18Z,2026-07-02T23:22:38Z,6,"Fixes #143580 Reason is model.no_sync() sets require_backward_grad_sync as False. forward() does not call reducer.prepare_for_backward(), expect_autograd_hooks_ is true is not set. The if condition gaurding _DDPSink.apply() call in _post_forward didn't check require_backward_grad_sync unlike how the condition callin..."
188439,[CUDA] [CI] fix potential unit test failure after cuda graph xfail (resource leak),https://github.com/pytorch/pytorch/pull/188439,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-06-29T16:15:48Z,2026-07-02T23:15:38Z,6,"While running CUDA unit tests locally, I noticed that test_mempool_ctx_multithread sometimes failed at the very top when asserting that the number of active segments (before the test) should be 0. The root cause is that if an exception happens during graph capture, allocations may be ""leaked"" and cannot easily be fr..."
188762,Fix multinomial without replacement silently sampling zero-probability categories,https://github.com/pytorch/pytorch/pull/188762,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,125388,125388,125388,,,,0,1,0,,,"triaged, open source, release notes: python_frontend",2026-07-02T00:15:31Z,2026-07-02T23:01:26Z,6,"Fixes #125388 torch.multinomial(input, n_sample, replacement=False) can return categories whose probability is zero. Root cause The without-replacement path samples via Gumbel top-k (aten/src/ATen/native/Distributions.cpp): it computes q = input / Exponential(1) and takes topk(q, n_sample). A zero-weight category ge..."
186945,Enable nested Inductor configs for invoke_subgraph regions,https://github.com/pytorch/pytorch/pull/186945,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,188126,,,0,3,0,,,"ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-10T16:43:45Z,2026-07-02T22:50:21Z,6,"Stack from ghstack (oldest at bottom): -> #186945 #188126 #188700 The normal Inductor invoke_subgraph path compiled each nested region under the surrounding graph's Inductor config -- the region was lowered, ran passes, generated code, and was cache-keyed using the outer config, and any per-region (nested) config pa..."
188782,[CI] Benchmark TorchInductor perf against 2.13.0 RC,https://github.com/pytorch/pytorch/pull/188782,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,188781,,,0,1,0,,,,2026-07-02T06:35:10Z,2026-07-02T22:50:16Z,6,Temporary PR to benchmark TorchInductor against the torch==2.13.0 RC (test cu130). Not for merge. Installs the RC into the inductor-benchmarks image in test_dynamo_benchmark() (same pattern as #175777) and rebuilds torchrec/fbgemm + detectron2 for torchbench. Benchmark run: https://github.com/pytorch/pytorch/actions...
188126,[inductor] Enable cudagraph partition inside invoke_subgraph subgraphs,https://github.com/pytorch/pytorch/pull/188126,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,186945,,,0,3,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-25T02:17:06Z,2026-07-02T22:20:37Z,6,"Stack from ghstack (oldest at bottom): #186945 -> #188126 #188700 codegen_subgraph_common previously forced graph partitioning off while codegening lifted HOP subgraph bodies, so cudagraph-eligible ops inside an invoke_subgraph region were never captured into a cudagraph. This enables graph partitioning inside invok..."
174718,[inductor] fix allocation with deterministic guard,https://github.com/pytorch/pytorch/pull/174718,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,174386,174386,,,,,0,1,0,,,"triaged, module: determinism, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, release notes: inductor (aoti), ciflow/torchtitan",2026-02-10T21:26:03Z,2026-07-02T20:30:43Z,6,Fixes #174386 Co-authored with claude cc @mruberry @kurtamohler @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
188827,[AI Codemod][PerfAICT-General] Use SymDimVector to avoid heap allocs in select_symint,https://github.com/pytorch/pytorch/pull/188827,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, meta-exported",2026-07-02T16:47:47Z,2026-07-02T20:29:49Z,6,Reviewed By: DenisYaroshevskiy Differential Revision: D109912064
188200,[MPS] Vectorized contiguous cat fast path for any dimension,https://github.com/pytorch/pytorch/pull/188200,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,CLEAN,False,0,0,,,188198,188198,188198,188613,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-25T21:50:12Z,2026-07-02T20:29:24Z,6,"Fixes #188198 Note: This is staged on top of dev/joona/copy_bytes_aligned since there we'll use the refactored parts from it. So we need to land #188613. Unfortunately I worked on these in a backward order so the PR is not properly gh stacked. Routes contiguous, same-dtype torch.cat through a vectorized inner_contig..."
188826,[AI Codemod][PerfAICT-General] Use SymDimVector to avoid heap allocs in select_symint,https://github.com/pytorch/pytorch/pull/188826,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,ciflow/trunk,2026-07-02T16:47:47Z,2026-07-02T19:51:26Z,6,Reviewed By: DenisYaroshevskiy Differential Revision: D109912064
186127,[AI Codemod][PerfAICT-General] fbcode/caffe2/torch/csrc/jit/serialization/pickler.cpp (#186127),https://github.com/pytorch/pytorch/pull/186127,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,186127,,,,,0,1,0,,,"fb-exported, ciflow/trunk, release notes: jit, meta-exported",2026-06-03T20:54:05Z,2026-07-02T19:51:19Z,6,Summary: Pull Request resolved: #186127 Differential Revision: D107200378
188801,[MPS] Leak MetalShaderLibrary bundled singleton to avoid exit-time destructor crashes,https://github.com/pytorch/pytorch/pull/188801,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,188812,188812,188812,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-07-02T13:31:54Z,2026-07-02T19:51:02Z,6,"Fixes #188812. MetalShaderLibrary::getBundledLibrary() used to return a function-local static BundledShaderLibrary, so its destructor ran during __cxa_finalize. On some builds the Metal runtime is already gone at that point, and the destructor's [cpl release] / [func release] calls fault, aborting an otherwise clean..."
182810,[ROCm] Implement TORCH_HIP_VERSION,https://github.com/pytorch/pytorch/pull/182810,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/rocm-nightly, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,"181647, 181648, 181649",,,,0,1,0,,,"module: rocm, triaged, open source, ciflow/trunk, ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/rocm-nightly",2026-05-07T10:43:27Z,2026-07-02T19:47:46Z,6,This PR is to update torch.utils.cpp_extension to define TORCH_HIP_VERSION for HIP extension builds. This PR is needed as Torchaudio (https://github.com/pytorch/audio) requires TORCH_HIP_VERSION for ROCm version checks. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @narome...
181582,[AI Codemod][RlAutotuneFbsourceWearablesMemory] torch::jit::mobile::Function::append_instruction (function.cpp) (#181582),https://github.com/pytorch/pytorch/pull/181582,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,181582,,,,,0,1,0,,,"open source, fb-exported, Stale, ciflow/trunk, release notes: mobile, meta-exported",2026-04-27T13:56:46Z,2026-07-02T19:15:32Z,6,Summary: Pull Request resolved: #181582 Reviewed By: jankrom Differential Revision: D102593095
186492,[inductor] Fix stale xoffset for TMA in mix-order reduction loop,https://github.com/pytorch/pytorch/pull/186492,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,186241,186241,186241,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-06T20:59:39Z,2026-07-02T19:15:00Z,6,"Fixes #186241. Problem When TMA is enabled (triton.use_tensor_descriptor=True), the mix-order reduction codegen emits TMA descriptor loads/stores that index off the scalar xoffset, which is never advanced inside the loop. Only the xindex tensor is incremented (xindex += XBLOCK). As a result every loop iteration read..."
188229,Fix 64-bit indexing in avg_pool3d backward atomic kernel,https://github.com/pytorch/pytorch/pull/188229,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-26T01:23:20Z,2026-07-02T19:08:59Z,6,"Summary avg_pool3d backward silently corrupts the gradient on inputs with more than INT_MAX elements, even though the forward pass handles them. When the pooling windows overlap (stride < kernel_size), the backward pass scatters each output gradient into its input window with avg_pool3d_cuda_update_grad_input_atomic..."
188502,Fix/188034 autograd setup context leak,https://github.com/pytorch/pytorch/pull/188502,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-30T00:41:24Z,2026-07-02T18:49:55Z,6,
179436,Run clang-tidy with readability-container-contains,https://github.com/pytorch/pytorch/pull/179436,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,176662,,,,0,3,0,,,"triaged, open source, Merged, NNC, Reverted, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, module: compiled autograd, ci-no-td, release notes: inductor (aoti), ciflow/h100-symm-mem, ciflow/vllm",2026-04-05T23:50:18Z,2026-07-02T18:37:45Z,6,"C++20 provides a contains member function for all associative containers. Prior to C++20, the common pattern to check if an element is a member of a container was to either container.find(key) != container.end() or container.count(key) > 0. This PR replaces those instances with contains. All the changes were done by..."
173390,Add interactive glossary with hover tooltips,https://github.com/pytorch/pytorch/pull/173390,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,45,0,,,"open source, release notes: releng",2026-01-26T18:42:13Z,2026-07-02T18:18:30Z,6,"Here's an example page with glossary tooltips: https://docs-preview.pytorch.org/pytorch/pytorch/173390/user_guide/torch_compiler/export.html After this is approved, i'll add more terms across more pages, I didn't want to make the world's biggest PR."
172603,"Rewrite _get_total_norm with _powsum, add skip_root flag",https://github.com/pytorch/pytorch/pull/172603,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,"28119, 174929",,,,0,25,0,,,"Stale, release notes: nn, ciflow/inductor",2026-01-16T00:16:08Z,2026-07-02T18:08:04Z,6,"Stack from ghstack (oldest at bottom): #172620 -> #172603 For gradient clipping, the _get_total_norm function uses a 2-stage linalg.vector_norm reduction, once per parameter, then once across all parameter norms. For sharded DTensors this incurs a sqrt -> pow -> sqrt loop which degrades numerics. This rewrites it fo..."
188426,Fix incorrect CPU bf16 conv3d when innermost output dim is 1 and strided,https://github.com/pytorch/pytorch/pull/188426,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,169688,169688,169688,,,,0,1,0,,,"module: cpu, module: convolution, triaged, module: mkldnn, open source, release notes: nn, topic: bug fixes",2026-06-29T14:49:12Z,2026-07-02T17:33:35Z,6,"Summary Fixes #169688. CPU bf16 conv3d produces incorrect results when the innermost output spatial dimension is 1 and the innermost stride is > 1. Root cause On AVX512-bf16/AMX CPUs, bf16 convolution is dispatched to oneDNN in ConvParams::use_mkldnn before the reference Slow3d kernel is considered. With oneDNN v3.1..."
184795,Autogenerated operator!= in C++20,https://github.com/pytorch/pytorch/pull/184795,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/trunk, ciflow/vllm",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,176662,,,,0,5,0,,,"triaged, module: mkldnn, open source, NNC, ciflow/trunk, release notes: cpp, ciflow/vllm",2026-05-21T20:52:14Z,2026-07-02T17:23:39Z,6,The compiler can automatically generate operator!= in C++20 #176662 cc @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfeima @sanchitintel @ashokei @jingxu10 @min-jean-cho @yanbing-j @Guobing-Chen @Xia-Weiwen @snadampal @EikanWang
187553,[MPS] Migrate mse_loss to a native fused Metal kernel,https://github.com/pytorch/pytorch/pull/187553,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,187455,"187556, 187557",,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-17T07:06:12Z,2026-07-02T16:52:03Z,6,Part of #187455. Summary This is the base PR for the pointwise fused-loss stack. It migrates mse_loss forward and backward on MPS from the shape-keyed MPSGraph path to native Metal. reduction='none' uses the existing binary-op iterator with the new mse functor. mean/sum use fused_loss_pass1 to accumulate (input - ta...
188755,cudagraph_trees: honor mark_static_address outputs,https://github.com/pytorch/pytorch/pull/188755,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: cuda, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-01T23:17:06Z,2026-07-02T16:46:48Z,6,"cudagraph_trees classifies each storage that is live at the end of capture as either a private-pool allocation or a known static input. An output whose storage was flagged via torch._dynamo.mark_static_address was neither, so check_memory_pool reported it as ""not allocated in pool but should be"", and (with the slow-..."
187860,fix DDE in meta__transformer_encoder_layer_fwd,https://github.com/pytorch/pytorch/pull/187860,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-06-22T18:38:55Z,2026-07-02T16:39:26Z,6,"meta__transformer_encoder_layer_fwd checks src.numel() == 0 directly, which raises a data-dependent error (GuardOnDataDependentSymNode) when src has an unbacked symbolic numel during fake-tensor tracing / torch.compile. This is the same DDE that #187144 fixed for native_multi_head_attention_fake, applied to its un-f..."
186524,Validate conv transpose output padding,https://github.com/pytorch/pytorch/pull/186524,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review when ready,,UNKNOWN,False,0,0,,,169236,169236,169236,,,,0,1,0,,,"module: cpu, triaged, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-06-07T06:35:56Z,2026-07-02T16:36:56Z,6,Fixes #169236. Summary Move conv transpose output_padding validation into the shared transposed convolution shape check. Cover invalid output_padding for both non-empty and empty inputs through OpInfo error inputs. Update the Inductor repro expectation now that FakeTensor tracing sees the shared error message. Test...
188395,Refactor Dynamo side effect replay into registry dispatcher,https://github.com/pytorch/pytorch/pull/188395,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-29T07:06:12Z,2026-07-02T16:36:17Z,6,"This refactors Dynamo side-effect replay from a hard-coded type-dispatch block in codegen_update_mutated into an explicit registry-driven dispatcher. The existing replay branches are moved into built-in registered handlers, so the default replay behavior is intended to remain unchanged. The new registry also provide..."
188771,[inductor] Rewind block_ptr across outer loop in nested reductions (#188771),https://github.com/pytorch/pytorch/pull/188771,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-07-02T03:43:47Z,2026-07-02T16:35:35Z,6,"Summary: For a fused broadcast + reduction (e.g. mean(a * b) where a:[N] and b:[N, 1] broadcast to [N, N]), Inductor's Triton codegen emits nested for loops over the two reduction axes. Block pointers are threaded through the loops and advanced with tl.advance at the suffix of each level. The bookkeeping in TritonKe..."
188785,Upgrade submodule oneDNN to v3.12.2,https://github.com/pytorch/pytorch/pull/188785,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, module: mkldnn, open source",2026-07-02T07:42:28Z,2026-07-02T16:15:03Z,6,This PR is to upgrade oneDNN to v3.12.2. Improvements Enabled SYCL Graph record/replay mode support in Graph API on Intel GPUs Fixed some correctness and performance issues on x64 and RV64 CPUs. cc @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfeima @sanchitintel @ashokei @jingxu10 @min-jean-cho @yan...
188613,[MPS] copy for contiguous same-dtype using a compute kernel,https://github.com/pytorch/pytorch/pull/188613,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-30T22:35:16Z,2026-07-02T16:07:21Z,6,Route contiguous same-dtype MPS→MPS copies through a compute kernel instead of a blit. copy_kernel_mps's sameDataType branch now dispatches a new contiguous_byte_copy kernel. Extracted the byte-copy alignment ladder into a reusable c10::metal::copy_bytes_aligned utility (introduced in #188483) This improves performa...
187572,[Metal][ops] add EmbeddingBag offsets validation (#170370),https://github.com/pytorch/pytorch/pull/187572,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,170370,170370,170370,,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-17T14:00:48Z,2026-07-02T15:35:51Z,6,"Fixes #170370. nn.EmbeddingBag on MPS does not validate the offsets tensor invariants that CPU and CUDA already check, and silently produces wrong output when offsets[0] != 0 or offsets[-1] > num_indices. This adds the same TORCH_CHECK guards in aten/src/ATen/native/mps/operations/EmbeddingBag.mm, with identical err..."
188787,OneDNN 3.12.2 acceptance test,https://github.com/pytorch/pytorch/pull/188787,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel,"identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: mkldnn, open source, topic: not user facing, ciflow/binaries_wheel",2026-07-02T07:49:41Z,2026-07-02T15:31:04Z,6,cc @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfeima @sanchitintel @ashokei @jingxu10 @min-jean-cho @yanbing-j @Guobing-Chen @Xia-Weiwen @snadampal
188608,[ci] Re-enable cuDSS installation in CI Docker images so CUDA CI can actually build and run cuDSS tests,https://github.com/pytorch/pytorch/pull/188608,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/docker,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,171035,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/docker",2026-06-30T22:08:55Z,2026-07-02T15:06:46Z,6,"Re-enable cuDSS installation in CI Docker images so CUDA CI can actually build and run cuDSS-covered tests. This is a dependency for merging #171035. That PR adds coverage for test_linalg_solve_sparse_csr_cusolver, but the test is currently skipped on CI because PyTorch is not built with cuDSS support: install_cudss..."
186599,Introduce isInitialized&set_rng_state in torch.accelerator,https://github.com/pytorch/pytorch/pull/186599,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186600, 186771, 186777, 186899, 186902, 186906",,,0,17,0,,,"open source, ciflow/trunk, release notes: python_frontend",2026-06-08T12:33:17Z,2026-07-02T14:23:58Z,6,Stack from ghstack (oldest at bottom): #186906 #186902 #186899 #186777 #186771 #186600 -> #186599 Motivation This PR introduces the following APIs in torch.accelerator: set_rng_state set_rng_state_all
181041,[hammer] Make HSTU manual recompute policies compatible with PT2 compile,https://github.com/pytorch/pytorch/pull/181041,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-21T23:09:55Z,2026-07-02T14:16:47Z,6,"Summary: OSS Pull Request: meta-recsys/generative-recommenders#503 HSTU uses custom Triton kernels with manual activation checkpointing flags (recompute_uvqk_in_backward, recompute_normed_x_in_backward, recompute_y_in_backward). Under torch.compile, PT2's auto AC conflicts with these manual policies — it inserts red..."
179284,Fix AOT autograd leaf backward mutation for auto_functionalized V2 (#179284),https://github.com/pytorch/pytorch/pull/179284,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"fb-exported, Stale, topic: not user facing, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-03T18:47:25Z,2026-07-02T14:16:47Z,6,"Summary: When _functionalized_f_helper replays backward mutations on leaf tensors (e.g. nn.Parameters), it failed with ""leaf Variable that requires grad is being used in an in-place operation"". The fix mirrors apply_in_graph_mutations by replaying under the same autograd context (torch.no_grad). This surfaces during..."
178550,[DTensor] Support sharding propagation with meta device mesh,https://github.com/pytorch/pytorch/pull/178550,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-03-26T22:46:48Z,2026-07-02T14:16:41Z,6,"Add a way to run DTensor sharding propagation (determining output placements given input placements) without requiring distributed initialization or real devices. Previously this was impossible because DeviceMesh.__init__ always calls get_rank() even with _init_backend=False, and MeshTopoInfo.build_from_mesh calls n..."
176747,Fix root cause of extraneous termination handler error logging (eval) (#176747),https://github.com/pytorch/pytorch/pull/176747,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, release notes: jit, meta-exported",2026-03-06T21:38:50Z,2026-07-02T14:16:36Z,6,"Summary: When an exception occurs in a noexcept destructor during shutdown, it triggers std::terminate. This leads to noisy logging of errors in our dataset. It manifests as std::terminate() invoked but parsed exception is empty. Backtrace captured at c++ termination handler entry point is shown below. Backtrace: (u..."
174323,[effectful ops] get tokens working with AC,https://github.com/pytorch/pytorch/pull/174323,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"Stale, topic: not user facing, ciflow/inductor",2026-02-04T21:40:34Z,2026-07-02T14:16:33Z,6,when we need to recompute an effectful op (it's marked as PREFER_RECOMPUTE via. an AC annotation) we need to enforce that the recomputation order is maintained relative to the backward effectful ops. i ended up running into this because we don't filter intermediate tokens out of the saved_values (i assumed they woul...
188560,[MPS] Fix constant address space for scalar params in binary_dense_scalar kernels,https://github.com/pytorch/pytorch/pull/188560,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,181650,181650,181650,,,,0,1,0,,,"triaged, open source, ciflow/mps",2026-06-30T15:54:53Z,2026-07-02T12:56:21Z,6,"Summary The 8 binary_dense_scalar* kernel templates in c10/metal/indexing.h declared the CPU-scalar operand as device T*, implying read-write GPU memory. The host binds CPU 0-dim scalars via setBytes: (read-only), so the Metal Validation Layer aborts the dispatch. Change the qualifier to constant. Matches the actual..."
187327,[Metal][cache] dual-hash verification for MPSGraphCache and MPSKernelCache (#77176),https://github.com/pytorch/pytorch/pull/187327,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,77176,77176,77176,,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-15T09:48:40Z,2026-07-02T12:49:24Z,6,"Summary Adds dual-hash verification to MPSGraphCache and MPSKernelCache. The single std::hash<std::string> cache key is supplemented with an independent FNV-1a secondary hash, computed once at insertion and verified on every cache hit. Catches primary-hash collisions at runtime instead of silently returning the wron..."
187330,[Metal][memory] clear graph/kernel compilation caches in empty_cache() (#154329),https://github.com/pytorch/pytorch/pull/187330,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,154329,154329,"145374, 154329, 155060, 164299",,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-15T12:10:27Z,2026-07-02T12:23:41Z,6,Summary torch.mps.empty_cache() previously released allocator blocks but left MPSGraph and MPSKernel compilation caches intact. Long-running workloads (e.g. conv backward on M5 Max) leaked ~30 MiB of compiler memory that no public API could reclaim. This PR: Adds MPSGraphCache::clear() / MPSKernelCache::clear() and...
170396,[typing] Fix return type annotations for various functions,https://github.com/pytorch/pytorch/pull/170396,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,149639,,,,0,13,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, release notes: distributed (c10d), module: dynamo, ciflow/inductor, ci-no-td, ciflow/torchtitan",2025-12-14T05:09:44Z,2026-07-02T11:35:51Z,6,"Summary This PR fixes return type annotations across several modules, reducing functions returning Any. Overall these changes should reduce the number of Any returns by PyTorch which is useful, especially when you want to use strict type checking. I'm not an expert with all of these functions and hope I got the exac..."
188781,[CI] Benchmark TorchInductor perf against 2.12.1 release (baseline),https://github.com/pytorch/pytorch/pull/188781,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,188782,,,0,1,0,,,,2026-07-02T06:35:04Z,2026-07-02T10:36:31Z,6,"Temporary PR to benchmark TorchInductor against torch==2.12.1 (stable cu130), the baseline for the 2.13.0 RC comparison. Not for merge. Installs the release into the inductor-benchmarks image in test_dynamo_benchmark() (same pattern as #175777) and rebuilds torchrec/fbgemm + detectron2 for torchbench. Benchmark run:..."
188315,[HOP] Autograd implementation for switch,https://github.com/pytorch/pytorch/pull/188315,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,9,0,,,"triaged, open source, topic: not user facing",2026-06-26T23:28:46Z,2026-07-02T10:36:24Z,6,"This PR adds the Autograd functionality for switch. It is tightly coupled to the implementation for cond and generalizes it to multiple branches. It also consolidates a common functionality for the bw between the two into utils.py. However, there is a special handling for the int -> SymInt and the None return case f..."
185911,[dynamo] Keep relational guards out of dict-tag skips,https://github.com/pytorch/pytorch/pull/185911,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/tsan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,147212,147212,,,,0,1,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/tsan, agentic, needs-review",2026-06-02T09:50:05Z,2026-07-02T09:24:15Z,6,"Stack from ghstack (oldest at bottom): -> #185911 Dict-tag guard shortcuts can skip checking immutable child managers when a parent dictionary version has not changed. That is only correct for guards whose evaluation is independent. Relational guards keep state across multiple manager visits, so skipping one side of..."
186532,Fix index_add decomposition bounds checks,https://github.com/pytorch/pytorch/pull/186532,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,5,1,"statusCheckRollup, inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8)",statusCheckRollup,,108798,108798,,,,0,10,0,0d2cf8517f4c182d1ae0f37d77fec4a4b1e35accc6de0e341ba57391253a2023,d2b31c4f849a28df9e2be6036dbf32b1e2f4afa074e52d37602a2d583097997e,"module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), release notes: inductor, ciflow/torchtitan, ciflow/dtensor, agentic, needs-review",2026-06-07T09:25:12Z,2026-07-02T08:37:48Z,6,"Stack from ghstack (oldest at bottom): -> #186532 The index_add decomposition lowered through index_put(..., accumulate=True). That missed index value bounds during fake/meta execution, and it also gave negative indices advanced-indexing wraparound semantics that do not match eager index_add. Invalid indices could t..."
188348,[HOP] Fixes for zero-length scan dim,https://github.com/pytorch/pytorch/pull/188348,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,"186107, 186465",,,,0,9,0,,,"triaged, open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-27T21:24:23Z,2026-07-02T07:59:29Z,6,"Currently, scan as well as associative_scan require the scan dimension of the tensor to be scanned to be > 0. These gaps have been reported in #186107 as well as in #186465. Having a scan dimension being equal to 0 is supported in JAX and is a valid usecase if masking is used on the tensor to be scanned. This PR ena..."
185114,[Test] Refractor TestSDPACudaOnly into TestSDPAGpuOnly with CUDA-only assumptions removed,https://github.com/pytorch/pytorch/pull/185114,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"185113, 185115",,,0,2,0,,,"open source, topic: not user facing",2026-05-25T07:27:38Z,2026-07-02T07:39:39Z,6,Stack from ghstack (oldest at bottom): #185115 -> #185114 #185113 Rename TestSDPACudaOnly to TestSDPAGpuOnly and remove hardcoded torch.cuda.* lookups inside the class so the tests no longer assume CUDA. The class is still instantiated only on CUDA in this commit; enabling additional accelerators is left to a follow...
185113,[Test] Decouple test_transformers.py from specific hardware backends,https://github.com/pytorch/pytorch/pull/185113,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"185114, 185115",,,0,1,0,,,"open source, topic: not user facing",2026-05-25T07:27:33Z,2026-07-02T07:39:38Z,6,Stack from ghstack (oldest at bottom): #185115 #185114 -> #185113 Refactors TestTransformers and TestSDPAFailureModes to be device-agnostic Changes: Replace @onlyCUDA with @onlyAccelerator and accept a device parameter in 5 TestTransformers methods and 24 TestSDPAFailureModes methods that were hardcoded to CUDA. Rep...
182945,Add pattern matcher to deduplicate var and std calculation,https://github.com/pytorch/pytorch/pull/182945,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",review when ready,,UNKNOWN,False,0,0,,,180957,180957,180957,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-08T10:11:23Z,2026-07-08T17:58:53Z,0,"Fixes #180957 When both torch.var(x, dim) and torch.std(x, dim) appear on the same tensor, std's decomposition in torch/_refs/init.py explicitly casts the input to the compute dtype (e.g., float32) before calling var internally, while the standalone var call relies on Inductor's lowering to handle the dtype promotio..."
187938,[Inductor] Support non-Triton fused kernels in the FX backend via a high-order-operator,https://github.com/pytorch/pytorch/pull/187938,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,187649,,,,0,2,0,,,"triaged, open source, release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-23T12:29:35Z,2026-07-02T07:07:50Z,6,"Summary This is an implementation for the issues #187649 The Inductor FX backend (config.fx_wrapper / WrapperFxCodegen) wraps a graph to a torch.fx.GraphModule, embedding Triton launches as triton_kernel_wrapper_mutation nodes. But it supported only Triton kernels: # torch/_inductor/codegen/wrapper_fxir.py if not li..."
180050,[FSDPV2] Make FSDP v2 and composable tests device-agnostic for PrivateUse1 backends,https://github.com/pytorch/pytorch/pull/180050,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem,"identify whether block is CI, merge conflict, or review gate",,DIRTY,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, Stale, release notes: distributed (fsdp), module: PrivateUse1, ciflow/h100-symm-mem",2026-04-10T21:51:47Z,2026-07-02T05:49:23Z,6,"Summary Replace hard-coded CUDA/NCCL references in FSDP v2 test infrastructure and test files with device-agnostic equivalents, enabling these tests to run on any backend registered via PrivateUse1 or the accelerator API. Changes Infrastructure (torch/testing/_internal/): common_fsdp.py: Add PrivateUse1 to device de..."
172219,[XPU] Adapt test_basic_chrome_trace to XPU,https://github.com/pytorch/pytorch/pull/172219,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,2263,2263,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing, no-stale",2026-01-12T10:59:37Z,2026-07-02T05:15:03Z,6,"Fixes intel/torch-xpu-ops#2263 Allow for TestProfilerDevice suite to be run for XPU Adapt test_basic_chrome_trace for XPU. Exclude entries with tid containing __xpu_profiler__ from checks against the RECORD_END timestamp, as these entries are created after the RECORD_END timestamp is measured. This condition is neve..."
188647,torch/fx: fix _format_target to handle Python keywords and embedded quotes,https://github.com/pytorch/pytorch/pull/188647,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188538,188538,188538,,,,0,1,0,,,"triaged, open source, release notes: fx",2026-07-01T06:21:58Z,2026-07-02T05:07:39Z,6,"Fixes #188538. Root cause _format_target in torch/fx/graph.py generates Python source for module attribute accesses in FX graphs. It uses str.isidentifier() to decide between dot notation (self.foo) and getattr notation, but this has two bugs: Python keywords like ""class"" pass isidentifier() but are invalid as attri..."
183306,[AI Codemod][ClaudeCodeFbandroidWarningsAsErrors] [3] Bulk run of CodemodConfigClaudeCodeFbandroidWarningsAsErrors (10 units) (#183306),https://github.com/pytorch/pytorch/pull/183306,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, meta-exported",2026-05-11T14:29:36Z,2026-07-02T04:50:02Z,6,Summary: This diff enables warnings_as_errors = True for the Buck target and fixes any compilation warnings that arise. Reviewed By: drhill-meta Differential Revision: D104108992
186282,Fix DTensor view-unflatten when one output dim is sharded by 2+ mesh axes,https://github.com/pytorch/pytorch/pull/186282,draft/noise,13,"PR is draft; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,188766,,,0,4,0,,,"ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-04T22:37:44Z,2026-07-02T04:49:37Z,6,"Stack from ghstack (oldest at bottom): -> #186282 #188766 _ViewShardingPropagator assumed each output tensor dim hosts at most one mesh axis. When two or more axes shard one output dim (e.g. a sequence dim sharded by both cp and tp under CP+SP), unflatten was not the inverse of flatten: the second axis landed on the..."
182627,Handle bool inputs in Scalar::log,https://github.com/pytorch/pytorch/pull/182627,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,182283,182283,182283,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-05-06T08:33:46Z,2026-07-02T03:43:44Z,7,"Extend Scalar::log to handle boolean scalars consistently with the existing integral and floating-point paths, and add scalar tests that cover int, float, bool, and complex inputs. Summary add bool handling to c10::Scalar::log() add scalar_test coverage for int, float, bool, and complex inputs fix the behavior repor..."
186122,[ROCm] Enable test_inverse_errors_large & test_lu_solve_batched_many_batches linalg tests,https://github.com/pytorch/pytorch/pull/186122,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/rocm-mi200, ciflow/rocm-mi300, ciflow/rocm-nightly, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,"181647, 181648, 181649",,,,0,6,0,,,"module: rocm, triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, keep-going, ci-no-td, ciflow/rocm-mi300, ciflow/rocm-mi200, ciflow/rocm-nightly",2026-06-03T20:24:57Z,2026-07-02T03:29:51Z,7,This PR enabled inverse_errors_large linalg tests that were previously failing for ROCm due to HipBlas errors. The fix was verified locally on rocm-nightly with torch-2.13.0a0+rocm7.13.0a20260513-cp312-cp312-linux_x86_64.whl from https://rocm.nightlies.amd.com/v2-staging/gfx94X-dcgpu/torch/ and release wheels for 7....
186771,Introduce _isInBadFork and refine torch.random,https://github.com/pytorch/pytorch/pull/186771,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186599, 186600, 186777, 186899, 186902, 186906",,,0,31,0,,,"open source, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-09T06:50:44Z,2026-07-02T01:50:27Z,7,Stack from ghstack (oldest at bottom): #186906 #186902 #186899 #186777 -> #186771 #186600 #186599 Motivation This PR clarifies and refines these APIs in torch.randn cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jatay...
186816,Remove deprecated torch.chain_matmul,https://github.com/pytorch/pytorch/pull/186816,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, module: amp (automated mixed precision), ciflow/trunk, release notes: linalg_frontend, module: dynamo, ciflow/inductor, suppress-bc-linter",2026-06-09T18:11:00Z,2026-07-02T01:47:46Z,7,The time has come to remove deprecated linear algebra related functions. This PR removes torch.chain_matmul. Fresh replacement for #70978 based on current upstream/main to avoid stale/conflicting old PR state. cc @mcarilli @ptrblck @leslie-fang-intel @jgong5 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @Xiaobin...
186899,[2/N] Replace module-level RNG-related APIs with torch.accelerator,https://github.com/pytorch/pytorch/pull/186899,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186599, 186600, 186771, 186777, 186902, 186906",,,0,18,0,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/dtensor",2026-06-10T06:34:08Z,2026-07-02T01:39:12Z,7,Stack from ghstack (oldest at bottom): #186906 #186902 -> #186899 #186777 #186771 #186600 #186599 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98
188531,pooling: guard max_pool1d integer parameters,https://github.com/pytorch/pytorch/pull/188531,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,142454,142454,142454,,,,0,1,0,,,"triaged, open source, release notes: nn, topic: not user facing",2026-06-30T03:37:26Z,2026-07-02T01:32:30Z,7,"Fixes #142454 Summary: Adds safe_downcast<int, int64_t> validation for max_pool1d kernel_size, stride, padding, and dilation before output-size math. Mirrors the stale prior approach from #175223, which had no rejection and went stale, while keeping the fix to the single validation point requested on the issue. Adds..."
186968,dynamo: enforce target_values is None|Sequence.,https://github.com/pytorch/pytorch/pull/186968,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,184487,,,0,2,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-10T20:15:23Z,2026-07-02T01:26:05Z,7,"In #184487, @kshitij12345 recommended we validate that target_values is actually what it should be. This does that validation---and as Kshiteej probably suspected, it indeed flags a bug (also fixed). This bug was: FxTracebackAnnotateVariable passed the torch.fx.traceback.annotate() dict straight into target_values (..."
188744,[dynamo] fix lazy module initialization with symbolic shapes,https://github.com/pytorch/pytorch/pull/188744,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-01T22:18:03Z,2026-07-02T01:18:24Z,7,"Stack from ghstack (oldest at bottom): (to be filled) initialize_lazy_module converts proxy args to fake tensors to call _infer_parameters on the real module. In resume functions (after a graph break), fake tensors can have symbolic shapes. When _infer_parameters reads input.shape[1] to get num_features and passes i..."
188745,"[dynamo, nested graph breaks] fix empty nn.Module hook dict reconstruction",https://github.com/pytorch/pytorch/pull/188745,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-07-01T22:18:13Z,2026-07-02T01:17:42Z,7,Stack from ghstack (oldest at bottom): (to be filled) cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
188555,Fix rocm lazy context initialization in test_set_device_0,https://github.com/pytorch/pytorch/pull/188555,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"module: rocm, triaged, open source, topic: not user facing",2026-06-30T15:30:01Z,2026-07-02T00:57:00Z,7,cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
188606,Relax complex tensor consistency tolerances for CUDA binary nonlinear ops,https://github.com/pytorch/pytorch/pull/188606,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, tensor subclass, topic: not user facing",2026-06-30T21:56:27Z,2026-07-02T00:56:01Z,7,"Summary Relax CUDA consistency-test tolerances for ComplexTensor ""binary nonlinear"" ops (bmm, mm, dot, mul, convolution, etc.) whose implementations decompose complex multiply into four real BLAS calls. Native complex64 CUDA kernels use a single complex GEMM, so accumulation order differs and produces small but real..."
188581,[dynamo] Add polyfills for operator.contains and operator.call,https://github.com/pytorch/pytorch/pull/188581,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,116396,,,,0,1,0,,,"triaged, open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-30T18:59:22Z,2026-07-02T00:52:13Z,7,Fixes part of #116396 Summary operator.contains and operator.call currently cause a graph break inside torch.compile regions because they are C builtins with no traceable Python implementation: Failed to trace builtin operator contains with argument types [...] This PR adds substitute_in_graph polyfills following th...
185700,[OpenReg] Add compiler integration doc for OOT accelerators,https://github.com/pytorch/pytorch/pull/185700,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,181093,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: openreg",2026-05-30T09:29:08Z,2026-07-02T00:42:37Z,7,"Summary Adds docs/source/accelerator/compiler.md: Document for integrating a new accelerator into torch.compile. Covers both the Dynamo backend path (register_backend + DeviceInterface) and the Inductor backend path (register_backend_for_device + register_device_op_overrides), and references to the OpenReg implement..."
188732,[MPS] Enable ConvTranspose3d for fp16/bf16 via internal fp32 upcast,https://github.com/pytorch/pytorch/pull/188732,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,"160332, 160739","2435, 160332, 160739",160739,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-07-01T20:30:53Z,2026-07-02T00:38:47Z,7,"Fixes #160739 Related: #160332 (autocast crash caused by the same gate), #130256 (earlier feature request tracking the original missing-kernel state of this op; not directly closed by this PR) Summary ConvTranspose3d (5-D transposed convolution) on the MPS backend currently hard-rejects fp16 and bf16 inputs: Runtime..."
175011,[S615805][RecSysCheckpointing][PyTorch] Add strides to TensorProperties (#175011),https://github.com/pytorch/pytorch/pull/175011,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, ciflow/trunk, release notes: distributed (sharded), release notes: distributed (checkpoint), meta-exported, ciflow/torchtitan",2026-02-14T00:20:16Z,2026-07-02T00:10:07Z,7,"Summary: Root Cause Analysis: https://www.internalfb.com/intern/phabricator/paste/markdown/P2165832658/ This diff adds stride information to TensorProperties in core PyTorch DCP metadata and sharded tensor metadata. This preserves the actual tensor stride information (e.g., from unsqueeze, transpose, or broadcasting..."
188616,[AI Codemod][PerfAICT-General] Skip per-operand has_storage scan for non-PrivateUse1 devices in build,https://github.com/pytorch/pytorch/pull/188616,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, meta-exported",2026-06-30T22:59:43Z,2026-07-02T00:09:49Z,7,Reviewed By: MatzeB Differential Revision: D109916362
182691,Parameterized cuda graph launch 3,https://github.com/pytorch/pytorch/pull/182691,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,13,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-06T19:15:56Z,2026-07-01T22:03:04Z,7,"This time, it is written purely in python, thanks to cuda-bindings. The previous C++ implementation was just too intrusive, since it made so many changes to CUDAGraph.cpp. I was not happy with that. This achieves the following results on training and inference tests from the huggingface and torchbench benchmarks in..."
188563,[AI Codemod][PerfAICT-General] Enable auto-vectorization of at::vec::convert via __restrict and indexed access,https://github.com/pytorch/pytorch/pull/188563,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, ciflow/trunk, meta-exported",2026-06-30T16:04:05Z,2026-07-01T21:46:16Z,7,Differential Revision: D109918548 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
187909,[DONT MERGE] Test new xpu runners,https://github.com/pytorch/pytorch/pull/187909,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/inductor, keep-going, ciflow/xpu",2026-06-23T03:56:05Z,2026-07-01T21:29:39Z,7,
187032,[dynamo] Support member access on protobuf enums,https://github.com/pytorch/pytorch/pull/187032,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,161765,161765,161765,187707,,,0,1,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-11T12:04:30Z,2026-07-01T21:09:05Z,7,"Fixes #161765 Supersedes #174336 Adds Dynamo support for accessing protobuf enum members (e.g. EgoFeatureIndex.EGO_SPEED) inside torch.compile without graph breaks. Root cause: EnumTypeWrapper.__getattr__ traces into self._enum_type.values_by_name[name].number, which hits two C extension barriers in protobuf's UPB b..."
182092,Allow torch.meshgrid() with empty input to return empty tuple,https://github.com/pytorch/pytorch/pull/182092,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,181576,181576,181576,,,,0,1,0,,,"triaged, open source, Stale",2026-05-01T04:31:57Z,2026-07-01T21:03:34Z,7,"torch.meshgrid() with no arguments previously raised a RuntimeError. This is inconsistent with NumPy behavior and the Array API Standard, which both specify that empty input should return an empty sequence. Changes: torch/functional.py: add early return () in _meshgrid() after old-interface unwrapping torch/_refs/__..."
188242,[ROCm] Support USE_ASAN=1 build,https://github.com/pytorch/pytorch/pull/188242,draft/noise,31,PR is draft; review decision requests changes; CI lane labels: ciflow/rocm-mi300,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"module: rocm, open source, topic: not user facing, ciflow/rocm-mi300",2026-06-26T02:59:07Z,2026-07-01T20:52:52Z,7,Notable New Features: ROCm PyTorch now supports ASAN build with TheRock 7.14 SDK Use clang from TheRock 7.14 as the host compiler integration builds cleanly; tests unvalidated due to crashing during the early-initialization phase of the ROCm ASAN runtime. Build of CK SDPA with ASAN is not tested Select ASAN build of...
188349,[HOP] Adding length parameter to scan,https://github.com/pytorch/pytorch/pull/188349,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,,186107,186107,,,,0,6,0,,,"triaged, open source",2026-06-27T21:38:17Z,2026-07-02T07:59:50Z,6,"Currently, scan does not support the length argument that jax.lax.scan supports. This has been identified as a gap between torch.scan and jax.lax.scan, see #186107. This PR introduces the length argument to scan by wrapping the user-provided combine_fn and hardcoding xs=None. With this PR, the tests https://gist.git..."
188136,[BE] Build all XPU wheels on a single runner,https://github.com/pytorch/pytorch/pull/188136,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: mkldnn, open source, topic: not user facing, ciflow/binaries_wheel",2026-06-25T06:12:18Z,2026-07-01T18:10:57Z,7,"Summary Extend the unified single-runner wheel build pattern (from #184045 for CUDA and #183931 for CPU) to XPU nightly wheel builds. Instead of spawning 8 separate runners (one per Python version), a single manywheel-xpu-build job loops over all CPython versions sequentially, reusing build/ across iterations After..."
187544,Fix torch.gcd returning negative results for INT_MIN inputs,https://github.com/pytorch/pytorch/pull/187544,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187338,187338,187338,,,,0,2,0,,,"triaged, open source, topic: bug fixes, module: python frontend",2026-06-17T02:26:25Z,2026-07-01T18:10:34Z,7,"torch.gcd returns incorrect negative results when one input is the minimum value of a signed integer type. For example, gcd(-128, 124) returns -4 instead of 4 for int8. calc_gcd calls std::abs() on inputs, which is undefined behavior for INT_MIN because the positive result overflows. The fix uses unsigned arithmetic..."
180353,[MPS] Replace isin MPSGraph implementation with native Metal kernels,https://github.com/pytorch/pytorch/pull/180353,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,180833,180833,180833,,,,0,6,0,,,"triaged, open source, topic: improvements, release notes: mps, ciflow/mps",2026-04-14T15:58:49Z,2026-07-01T17:56:01Z,7,"Closes #180833 Summary Replaces the MPSGraph-based isin implementation with two native Metal kernels selected by an MPS-specific heuristic based on input shape. The previous implementation broadcast-expanded both input tensors into a 2D equality matrix of shape (ne, nt) via MPSGraph, then applied an OR reduction alo..."
188662,[ROCm] Updated mi300 runners to have 8 shards,https://github.com/pytorch/pytorch/pull/188662,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-rocm-mi300, ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, ciflow/trunk, topic: not user facing, ciflow/inductor, ciflow/rocm-mi300, ciflow/inductor-rocm-mi300",2026-07-01T12:00:29Z,2026-07-01T17:27:58Z,7,"We already use 8 shards for mi350, having 6 for mi300 can lead to flakyness and consistency issues since the test base has grown. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang"
188284,[Draft] Fix BUILD_SEPARATE_OPS cyclic dependency — register with xpu_hal.dll,https://github.com/pytorch/pytorch/pull/188284,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/xpu",2026-06-26T18:01:00Z,2026-07-01T17:27:47Z,7,Register with xpu_hal.dll via Inline Lambdas Companion PR to intel/torch-xpu-ops#4129. Registers pytorch generators with xpu_hal.dll — the function-pointer bridge DLL in torch-xpu-ops — via inline lambdas in initXPUGenVector(). Before vs After — Generator Dependency ❌ BEFORE: kernel DLLs link torch_xpu.dll directly...
170752,Support torch.accelerator.get_device_capability on CUDA,https://github.com/pytorch/pytorch/pull/170752,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,,,,,,0,17,0,,,"open source, Stale, ciflow/trunk, release notes: cuda, no-stale, ci-no-td",2025-12-18T07:07:31Z,2026-07-01T17:15:49Z,7,"Stack from ghstack (oldest at bottom): -> #170752 Motivation This PR adds support for torch.accelerator.get_device_capability on CUDA. At the current stage, it reports a limited set of basic scalar data types."
183716,Validate dim in cummax/cummin meta for empty tensors,https://github.com/pytorch/pytorch/pull/183716,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,131273,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-14T13:57:36Z,2026-07-01T17:12:31Z,7,"The eager side of this was fixed in #188361 (invalid dim on empty non-scalar tensors now raises IndexError). This PR fixes the remaining gap in the cummaxmin meta registration, which still skipped validation when numel() == 0: if self.numel() != 0 and self.ndim != 0: maybe_wrap_dim(dim, self.ndim) Because the fake/m..."
188333,[MPS] prod: migrate to the shared value_reduction<ProdOp> kernel + shape specializations,https://github.com/pytorch/pytorch/pull/188333,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,187455,"187787, 188156",,187455,0,3,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-27T13:45:24Z,2026-07-01T16:44:40Z,7,"Stack position / review order Part of #187455. Builds on #188156. Review order: #188156 lands first as the reduce-family base (value_reduction<Op, Load, TA>). This PR is then the prod follow-up: ProdOp plus prod-specific correctness/perf work. #187787 (var/std) is rebuilt after this direction is settled so it does n..."
187787,[MPS] Migrate var/std to native Metal kernels (Welford),https://github.com/pytorch/pytorch/pull/187787,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/mps",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,187455,"188156, 188333",,,0,4,0,,,"triaged, open source, release notes: mps, ciflow/mps, module: inductor, ciflow/inductor",2026-06-21T09:38:30Z,2026-07-01T16:44:40Z,7,"Part of #187455. Stack Position / Review Order This is the third reduce-family split and should be reviewed after: #188156: shared value_reduction<Op, Load, TA> base for the sum family. #188333: prod on top of that shared base. This PR: var / std / var_mean / std_mean native Metal Welford kernels. This rebuild inten..."
188629,docs: document S3-compatible object storage for distributed checkpoint,https://github.com/pytorch/pytorch/pull/188629,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-07-01T01:37:42Z,2026-07-01T14:03:19Z,7,"Documentation or Typo Fix This is a documentation-only change to docs/source/distributed.checkpoint.md. It adds no code and changes no behavior, so per the docs template it does not require a linked issue. Happy to file one if a maintainer prefers. Before submitting, I reviewed: The Ultimate Guide to PyTorch Contrib..."
188260,[nn] Fix RReLU noise tensor not being set in eval mode,https://github.com/pytorch/pytorch/pull/188260,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,176381,176381,176381,,,,0,3,0,,,"triaged, open source, release notes: nn, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-26T08:56:35Z,2026-07-01T13:56:24Z,7,"Fixes #176381 Summary In eval mode, rrelu_with_noise was leaving the noise tensor uninitialized (containing garbage values) while only computing the output via leaky_relu. This breaks the API contract where noise should reflect the slopes applied. Fixed all three dispatch paths (CPU, CUDA, inductor decomposition) to..."
187257,[DCP] Raise RuntimeError on load dtype mismatch to prevent precision …,https://github.com/pytorch/pytorch/pull/187257,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/torchtitan,review when ready,,UNKNOWN,False,0,0,,,187138,187138,187138,,,,0,1,0,,,"triaged, open source, release notes: distributed (checkpoint), ciflow/torchtitan",2026-06-13T20:13:01Z,2026-07-01T13:30:53Z,7,"Description This PR addresses issue #187138. Currently, torch.distributed.checkpoint.load silently casts tensors from the checkpoint to the template's dtype using Tensor.copy_. If the template has a lower precision (e.g., bfloat16) than the stored checkpoint (e.g., float32), precision is lost without any warning or..."
188337,[inductor] Handle non-comparable arguments in index_propagation Min/Max,https://github.com/pytorch/pytorch/pull/188337,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,"188225, 188230","188225, 188230","188225, 188230",,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-27T15:28:54Z,2026-07-01T12:50:55Z,7,"sympy.Min and sympy.Max raise ValueError when given non-comparable arguments (e.g., NaN, boolean values). This crashes torch.compile() compilation while eager mode handles these fine (NaN just propagates at runtime). Catch ValueError and return NotImplemented, which triggers the index propagation fallback path that..."
151218,Implement MKLGenerator,https://github.com/pytorch/pytorch/pull/151218,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,132395,,,,0,11,0,,,"module: cpu, triaged, open source, Merged, Reverted, Stale, ciflow/trunk, topic: not user facing, module: dynamo, ci-no-td, skip-url-lint",2025-04-14T11:06:42Z,2026-07-01T12:19:18Z,7,"This PR aims to fix the issue from #132395 by implementing a new MKLGeneratorImpl that stores a consistent, global vslStream for use in random numbers generation. This path was previously disabled due to a problem of repeating variates, caused by repeated reseeding of the MKL generator with variates from the CPUGene..."
185661,[fx] Fix pytree codegen for tuple/dict positional inputs,https://github.com/pytorch/pytorch/pull/185661,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185640,185640,185640,,,,0,5,0,,,"triaged, open source, release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-29T19:47:37Z,2026-07-01T11:59:59Z,7,"Fixes #185640. Summary stop inferring the (args, kwargs) calling convention from TreeSpec shape alone in _PyTreeCodeGen add an explicit _PyTreeInfo.is_args_kwargs flag and propagate it through export/dynamo call sites that really do flatten (args, kwargs) keep symbolic tracing / make_fx positional pytrees on the pla..."
187573,[TMA] Fix AOTI eager scratch buffer missing CACHE_TORCH_DEVICE declaration,https://github.com/pytorch/pytorch/pull/187573,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-06-17T14:02:34Z,2026-07-01T11:59:50Z,7,"When TMA is enabled and a kernel uses tensor descriptors, the AOTI C++ wrapper generates a scratch buffer allocation in a free function that references cached_torch_device_type_cuda. This file-scope static variable is produced by the CACHE_TORCH_DEVICE(cuda) macro in the wrapper prefix. The bug was that CUDADeviceOp..."
179273,[optim] support of dynamic coefficients and normalization in Muon,https://github.com/pytorch/pytorch/pull/179273,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,179250,179250,179250,,,,0,9,0,,,"triaged, open source, no-stale, release notes: optim",2026-04-03T16:56:11Z,2026-07-01T11:36:49Z,7,"This pull request improves the Muon optimizer's Newton–Schulz orthogonalization step. The main improvements are the introduction of configurable coefficient presets. (including ""jordan"" and ""polar_express""), support for multiple normalization strategies, and comprehensive tests for these new options. These changes m..."
188350,Fix Dynamo stream/event variable python_type,https://github.com/pytorch/pytorch/pull/188350,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,188325,188325,188325,,,,0,1,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-27T23:02:49Z,2026-07-01T10:55:08Z,7,Fixes #188325. Summary Make StreamVariable.python_type() and EventVariable.python_type() report the wrapped runtime value type instead of hard-coded base classes. Add a CUDA-free regression covering both wrappers with fake stream/event objects. Why SuperVariable and other type/MRO-based Dynamo paths rely on python_t...
181571,Generalize test_pooling.py to be device-agnostic and enable it on XPU,https://github.com/pytorch/pytorch/pull/181571,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,2132,2132,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing",2026-04-27T10:37:03Z,2026-07-01T09:44:24Z,7,Enable XPU testing without monkey-patching from the downstream nn/test_pooling_xpu.py file. Add allow_xpu=True to all instantiate_device_type_tests calls Add dtypesIfXPU to the test_max_pool_nan_inf test case Add dtypesIfXPU to the test_max_pool3d_ndhwc test case Generalize test_MaxUnpool_index_errors to XPU device...
188329,Optimize all-gather copy-in with a flat parameter buffer,https://github.com/pytorch/pytorch/pull/188329,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-06-27T08:49:50Z,2026-07-01T09:36:32Z,7,"FSDP2 currently packs parameter shards into a contiguous communication buffer on every foreach_all_gather. This PR places eligible sharded parameters in a single contiguous flat storage during lazy initialization. During all-gather copy-in, FSDP copies this flat buffer directly into the rank-local slice of the all-g..."
186603,Exposing MiMalloc options and stats,https://github.com/pytorch/pytorch/pull/186603,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, module: intel, topic: performance",2026-06-08T13:58:46Z,2026-07-01T08:34:53Z,7,"Some CPU backends rely on mimalloc; yet users have no programmatic access to the instrumentation provided by this library, regarding: Options: set arena size, purge policy, or many other configurations of the allocator; Stats: get a report on the count of arenas allocated, of mmaps, etc. These changes expose functio..."
185738,Make HOP alias and mutation support opt-in,https://github.com/pytorch/pytorch/pull/185738,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,149989,149989,,,,0,3,0,,,"module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan, agentic",2026-05-31T01:32:19Z,2026-07-01T08:19:03Z,7,"Stack from ghstack (oldest at bottom): -> #185738 Higher order op subgraph tracing defaulted supports_input_mutation and supports_aliasing to True. That made every caller permissive unless it explicitly opted out, so HOPs that do not actually support input mutation or aliasing could silently capture invalid body gra..."
188577,Migrate Build Triton Wheel to OSDC,https://github.com/pytorch/pytorch/pull/188577,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,topic: not user facing,2026-06-30T18:35:52Z,2026-07-01T06:46:07Z,7,"Migrate the build-wheel job (""Build Triton Wheel"") from EC2 to OSDC. OSDC/ARC runners have no host Docker daemon, so the docker run + docker exec manylinux pattern can't work there. Instead the job runs inside the manylinux builder via container:, with the build commands run directly. Runners move to OSDC labels (l-..."
188165,"[cutedsl] Pass current compute capability e.g., SM 10.3 -> `sm_103a` to CuTeDSL compile subprocess",https://github.com/pytorch/pytorch/pull/188165,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: cuda, triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, matrix multiplication, ciflow/torchtitan",2026-06-25T17:27:19Z,2026-07-01T05:14:17Z,7,"Otherwise we seem to attempt to build and run sm100a on compute-capability 10.3 which fails e.g., Traceback (most recent call last): File ""/usr/lib/python3.12/unittest/case.py"", line 58, in testPartExecutor yield File ""/usr/lib/python3.12/unittest/case.py"", line 634, in run self._callTestMethod(testMethod) File ""/us..."
187224,Duplicate concurrent alignment fixups,https://github.com/pytorch/pytorch/pull/187224,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-13T00:38:19Z,2026-07-01T05:13:59Z,7,"This hoists copy_if_misaligneds to the prologue if they are read from multiple streams. Without this we could end up generating code like this: with s1: arg0_1 = copy_if_misaligned(arg0_1) foo = op1(arg0_1) with s2: bar = op2(arg0_1) because the use at op1 is the first use by line order, this is where the alignment..."
178710,Add native torch.cuda.gds.save/load for GPUDirect Storage,https://github.com/pytorch/pytorch/pull/178710,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,12,0,,,"triaged, open source, Stale, topic: new features",2026-03-29T04:34:24Z,2026-07-01T03:50:58Z,7,"This PR enhance GDS code in the following way: It adds torch.cuda.gds.save/load/is_available. It adds storage_save_hook/storage_load_hook to serialization thread-local state. It fixes partial read/write detection, const correctness in cuFile ops, and adds CMake fallback when cuFile is not found. torch.{save,load} re..."
188543,Remove deprecated torch.utils.mobile_optimizer from API index,https://github.com/pytorch/pytorch/pull/188543,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188398,188398,188398,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-06-30T11:44:51Z,2026-07-01T03:45:16Z,8,Fixes #188398. This PR removes the deprecated torch.utils.mobile_optimizer entry from the 2.7 Python API index to resolve inconsistent navigation behavior. Removes torch.utils.mobile_optimizer from the API index list in pytorch-api.md Aligns API index with the intentional deprecation and ExecuTorch migration introdu...
188626,build-almalinux-images: migrate to OSDC via remote BuildKit,https://github.com/pytorch/pytorch/pull/188626,ready for maintainer decision,45,CI lane labels: ciflow/docker; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,188625,,,0,6,0,,,"topic: not user facing, ciflow/docker",2026-07-01T01:24:03Z,2026-07-01T03:44:03Z,8,Stack from ghstack (oldest at bottom): -> #188626 #188625 Run the almalinux-builder image build on the egress-exempt mt-rel-* OSDC release runner and opt into binary-docker-build's remote BuildKit mode (use-remote-buildkit: true) instead of the EC2 linux.9xlarge.ephemeral host daemon. .ci/docker/almalinux/build.sh g...
186652,"[ROCm] Fix ""Ignoring invalid value for boolean flag AMD_SERIALIZE_KERNEL: 3""",https://github.com/pytorch/pytorch/pull/186652,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/rocm-mi300, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,6,0,,,"module: rocm, triaged, open source, ciflow/trunk, topic: not user facing, ciflow/rocm-mi300",2026-06-08T21:24:40Z,2026-07-01T03:02:37Z,8,get_cuda_async_error_suffix used to call c10::utils::check_env to check AMD_SERIALIZE_KERNEL. However this env var is a bit mask rather than a binary flag. This PR refactors the code to generate warning messages/tips about AMD_SERIALIZE_KERNEL on ROCM platform. The behavior change can be validated with import torch...
185988,Fix FlexAttention scalar captured grad backward,https://github.com/pytorch/pytorch/pull/185988,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,145208,145208,,,,0,2,0,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan, agentic",2026-06-02T20:07:54Z,2026-07-01T02:59:52Z,8,"Stack from ghstack (oldest at bottom): -> #185988 FlexAttention backward expects captured score_mod gradients to be represented as writable buffers that the backward template can atomic-add into. Indexed captured tensors already go through zeros_and_scatter, which allocates that backing grad buffer, but an unindexed..."
188625,binary-docker-build: add opt-in remote BuildKit (OSDC) mode,https://github.com/pytorch/pytorch/pull/188625,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,188626,,,0,3,0,,,topic: not user facing,2026-07-01T01:23:55Z,2026-07-01T02:22:23Z,8,"Stack from ghstack (oldest at bottom): #188626 -> #188625 Add a use-remote-buildkit input (default false). When true, the action registers a remote buildx builder and runs /build.sh with REMOTE_BUILDKIT=1, buildx-pushing straight to docker.io with the same tag set the local-daemon path publishes (branch / commit / ...."
187786,Support int step tensors in foreach optimizer implementations,https://github.com/pytorch/pytorch/pull/187786,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,142378,142378,142378,,,,0,2,0,,,"triaged, open source, release notes: foreach_frontend, topic: not user facing",2026-06-21T09:33:39Z,2026-07-01T01:54:05Z,8,"Fixes #142378 The foreach (multi-tensor) optimizer path rejects integer step tensors at two layers: C++ grouping (ForeachUtils.h): _group_tensors_by_first_tensors_device_and_dtype only allows Float/Double/BFloat16 for step tensor dtypes, rejecting Int/Long Python CPU increment: All foreach optimizers hardcode torch...."
187207,Use abs() in lp_pool to avoid NaN for non-integer norm_type,https://github.com/pytorch/pytorch/pull/187207,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,184037,184037,184037,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-06-12T19:55:32Z,2026-07-01T01:54:04Z,8,"Fixes #184037 lp_pool{1,2,3}d compute input.pow(norm_type) directly. For non-integer norm_type (e.g. 2.3) and negative input values, (-x)^p is undefined in IEEE floating point and produces NaN: m = nn.LPPool2d(norm_type=2.3, kernel_size=1) m(torch.randn(2, 8, 8, 8)) # NaN everywhere negatives appear The standard Lp..."
185831,Use minimum-norm subgradient for clamp/clamp_min/clamp_max at boundary,https://github.com/pytorch/pytorch/pull/185831,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,184572,184572,184572,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-06-01T17:38:30Z,2026-07-01T01:54:03Z,8,"Fixes #184572 torch.clamp_min(x, 0) and torch.clamp(x, min=0) return a gradient of 1.0 at x=0, violating PyTorch's documented minimum-norm subgradient convention and inconsistent with F.relu which returns 0.0: x = torch.tensor(0.0, requires_grad=True) torch.clamp_min(x, 0.0).backward() print(x.grad) # 1.0 (should be..."
188297,"Validate target range [0, 1] in binary_cross_entropy_with_logits",https://github.com/pytorch/pytorch/pull/188297,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,110865,110865,110865,,,,0,2,0,,,"triaged, open source, release notes: nn, topic: not user facing, ciflow/inductor",2026-06-26T20:32:40Z,2026-07-01T01:53:17Z,8,"Fixes #110865 BCELoss validates that target values are in [0, 1] (per-element TORCH_CHECK on CPU, CUDA_KERNEL_ASSERT on CUDA), but BCEWithLogitsLoss does not, silently accepting any target value and producing meaningless loss values. As @mikaylagawarecki noted: ""Since we validate target for BCELoss it seems reasonab..."
186542,Support joint export with unused parameter grads,https://github.com/pytorch/pytorch/pull/186542,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,5,1,"statusCheckRollup, unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,101192,101192,,,,2,3,0,3fd53b68e8b1874969f0fc767acfec3551ac4000130bb80d9354219371890195,ce9b2ebff71121f4111206819bc27da4eb2f5f8ee01247fa355aed96a3010934,"ciflow/inductor, release notes: aot autograd, agentic, needs-review",2026-06-07T12:21:30Z,2026-07-01T01:47:15Z,8,"Stack from ghstack (oldest at bottom): -> #186542 AOTAutograd's flattened joint export path assumed that every grad-requiring forward input would produce a non-None gradient. When a parameter participated only through a detached value, autograd correctly returned None for that input, but export raised an internal as..."
188259,Add XPU caching allocator checkpoint pool state bindings,https://github.com/pytorch/pytorch/pull/188259,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,0,0,,,,,,,,,0,5,0,,,"open source, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-26T08:56:09Z,2026-07-01T01:30:26Z,8,"Adds XPU support for caching allocator private-pool checkpoint state APIs used by graph pool restoration. This change mirrors the existing CUDA private-pool checkpoint plumbing for XPU by: adding XPU allocator checkpoint state and restore helpers exposing _xpu_getCheckpointState, _xpu_setCheckpointPoolState, and _xp..."
188593,"[ROCm] Add skipIfRocmVersionAtLeast([7, 14]) skips for ROCm 7.14 known failures",https://github.com/pytorch/pytorch/pull/188593,draft/noise,13,"PR is draft; CI lane labels: ciflow/dtensor, ciflow/inductor",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: rocm, open source, release notes: distributed (fsdp), module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), ciflow/dtensor",2026-06-30T19:55:27Z,2026-07-01T01:23:01Z,8,"Skip tests that fail on ROCm 7.14 but pass on ROCm 7.2.x. These skips activate automatically when upstream CI transitions from ROCm 7.2 to ROCm 7.14, keeping CI green during the transition. Adds two new skip decorators mirroring the existing lessthan pattern: skipIfRocmVersionAtLeast(version) in common_utils.py skip..."
188368,[Inductor] Fix ValueRangeError in Inductor for zero-iteration loops,https://github.com/pytorch/pytorch/pull/188368,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188226,188226,188226,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-28T16:14:29Z,2026-07-01T01:22:00Z,8,"Fixes #188226 When inductor encounters a reduction with zero iterations (e.g., FractionalMaxPool2d with output_size=-60, it causes reduction ranges of [-60, -60]), and constructs a ValueRanges(0, upper_bound(v) - 1) object. This produces an invalid range of [0:-61] since upper_bound(v) - 1 is negative, raising a Val..."
188592,vllm wheel 0630,https://github.com/pytorch/pytorch/pull/188592,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,topic: not user facing,2026-06-30T19:52:24Z,2026-06-30T23:51:15Z,8,Stack from ghstack (oldest at bottom): -> #188592
187976,[MPS] Fix int64 amax/amin returning 0 for partial-simdgroup reductions,https://github.com/pytorch/pytorch/pull/187976,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,187455,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-23T20:27:07Z,2026-06-30T23:23:03Z,8,Part of #187455. This is an independent correctness fix for the existing MPS amax/amin value-reduction helper. It does not change the prod or var/std migration order. Summary torch.amax / torch.amin on int64 MPS tensors can return 0 instead of the true extremum when a multi-simdgroup helper reduction has a partial s...
188530,optim: remove redundant Adafactor assert,https://github.com/pytorch/pytorch/pull/188530,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,164878,,,,0,1,0,,,"triaged, open source, topic: not user facing, release notes: optim",2026-06-30T03:31:46Z,2026-06-30T22:49:40Z,8,Part of #164878 Summary: Removes a redundant plain assert from the Adafactor foreach path. Drops the now-unused TYPE_CHECKING import. Keeps runtime behavior unchanged; the removed assert only lived under TYPE_CHECKING and device_state_steps is already cast to list[Tensor]. Test Plan: git diff --check - passed python...
188526,error checking: validate _pdist_forward dimensions,https://github.com/pytorch/pytorch/pull/188526,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,145064,145064,145064,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-30T03:18:31Z,2026-06-30T22:49:04Z,8,"Fixes #145064 Summary: Adds the missing dimensionality check to the internal _pdist_forward op so non-2D inputs raise a normal RuntimeError instead of reaching the native kernel with bad indexing. Preserves the existing public torch.pdist behavior for valid 2D empty-column inputs such as shape (3, 0), which still re..."
179443,[FSDP2] Fix HSDP O(n_layers) memory accumulation of reduce-scatter output buffers,https://github.com/pytorch/pytorch/pull/179443,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,179128,179128,181664,,,0,17,0,,,"Stale, release notes: distributed (fsdp), ciflow/inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-04-06T07:15:36Z,2026-06-30T22:00:07Z,8,"Stack from ghstack (oldest at bottom): #181667 #181664 -> #179443 When reduce_dtype != orig_dtype (e.g., bf16 params with fp32 reduction), HSDP's backward pass accumulated O(num_layers) fp32 reduce-scatter output buffers instead of freeing them eagerly. Root cause: After reduce-scatter, the fp32 output is all-reduce..."
188316,[ROCm] Cap HIP stream pool per priority; add reservable streams,https://github.com/pytorch/pytorch/pull/188316,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"module: rocm, triaged, open source, release notes: distributed (c10d), ciflow/torchtitan",2026-06-27T00:03:12Z,2026-06-30T21:58:01Z,8,"PyTorch hands out 32 round-robin streams per priority from its stream pool. On ROCm each hipStream is backed by an hsa_queue drawn from a per-priority pool of at most GPU_MAX_HW_QUEUES (default 4). Once a priority's pool is full the HIP runtime stops creating new queues and shares existing ones, choosing the least-l..."
187205,Add get_device_name and memory snapshot accelerator APIs,https://github.com/pytorch/pytorch/pull/187205,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, module: accelerator",2026-06-12T19:37:17Z,2026-06-30T21:57:59Z,8,"Summary Adds torch.accelerator.get_device_name() and three memory diagnostic APIs (_record_memory_history, _dump_snapshot, _snapshot) to the generic accelerator module to replace CUDA-specific calls in Inductor's diagnostic and instrumentation code. These APIs follow the accelerator interface design principle: the a..."
187206,[WIP][no-ci] auto_custom_op,https://github.com/pytorch/pytorch/pull/187206,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,,2026-06-12T19:49:31Z,2026-06-30T21:46:47Z,8,with @ydwu4 [no-ci] Test Plan: python autogen_operators.py python -m py_compile autogen_operators.py lintrunner -a lintrunner -a fails on pre-existing ACTIONLINT SC2086 in .github/workflows/generated-macos-arm64-binary-wheel-nightly.yml. This PR was authored with the assistance of an AI coding agent.
181431,test split group,https://github.com/pytorch/pytorch/pull/181431,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"Stale, ciflow/trunk, release notes: distributed (c10d), topic: not user facing",2026-04-24T19:23:25Z,2026-06-30T21:07:50Z,8,"Summary: Add targeted coverage for dist.split_group. This tests cover calling split group natively and using torchcomms, making sure split group works the same for both approaches. Stack created with Sapling. Best reviewed with ReviewStack. #182142 -> #181431"
176894,[cuDNN][Convolution] Expose cuDNN `allow_reduced_precision_reduction` numerical note to fix `float16` tests,https://github.com/pytorch/pytorch/pull/176894,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"module: cudnn, module: convolution, triaged, open source, module: half, Stale, release notes: cudnn",2026-03-09T16:47:56Z,2026-06-30T20:13:21Z,8,"Similar to existing cuBLAS flag, useful to fix e.g., complex32 test failures observed on SM 12.0 / SM 12.1 authored mostly with Claude code cc @csarofeen @ptrblck @xwang233 @nWEIdia"
188565,[ROCm] Add initial GFX1250 architecture support,https://github.com/pytorch/pytorch/pull/188565,draft/noise,13,"PR is draft; CI lane labels: ciflow/b200, ciflow/docker, ciflow/h100, ciflow/h100-symm-mem",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,7,0,,,"module: rocm, open source, release notes: releng, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/h100-symm-mem, ciflow/b200, ciflow/torchtitan, ciflow/docker",2026-06-30T16:36:11Z,2026-06-30T20:07:06Z,8,"Summary Initial, gate-only enablement for AMD gfx1250 (Wave32). PyTorch learns to recognize the arch and routes it onto Wave32-safe paths (hipBLASLt GEMM, scaled/MX-format GEMM, shared-memory and workspace sizing). Every currently Wave64-only path (CK GEMM/SDPA, rocSHMEM collectives, int4mm MFMA kernels) is explicit..."
187878,Add c10d Watchdog: a libuv-backed timeout service,https://github.com/pytorch/pytorch/pull/187878,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-06-22T21:36:14Z,2026-06-30T19:51:50Z,8,"Summary Introduces torch.distributed._watchdog, a process-wide timer/timeout service backed by a libuv event loop running on a dedicated background thread (created lazily on first use). The C++ interface c10d::watchdog::Watchdog exposes small, reusable primitives; each returns a Handle whose cancel() removes the mon..."
188501,Fix/147170 deterministic partition ordering,https://github.com/pytorch/pytorch/pull/188501,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-30T00:41:22Z,2026-06-30T19:03:15Z,8,
188498,Fix/grammar with the same parameters,https://github.com/pytorch/pytorch/pull/188498,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-29T23:59:19Z,2026-06-30T19:02:00Z,8,
188532,Fix/threshold inplace validation,https://github.com/pytorch/pytorch/pull/188532,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-30T03:38:58Z,2026-06-30T19:01:00Z,8,
188499,Fix/typo paramater,https://github.com/pytorch/pytorch/pull/188499,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-29T23:59:30Z,2026-06-30T18:54:09Z,8,
185891,[dynamo] Guard AOTAutograd input mutation overlap changes,https://github.com/pytorch/pytorch/pull/185891,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,139789,139789,139789,,,184481,0,11,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-06-02T04:48:37Z,2026-06-30T18:31:24Z,8,"Fixes #139789. AOTAutograd currently guards storage-overlap changes only for inputs that already share storage, and StorageOverlap was only installed for symbolic-shape overlap checks. That misses a static-shape case where the first call mutates an input that does not overlap another input, then a later call passes..."
188402,Fix matrix_exp crash on non-contiguous (channels-last) inputs,https://github.com/pytorch/pytorch/pull/188402,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,188222,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend",2026-06-29T10:43:25Z,2026-06-30T18:30:39Z,8,"Summary Make matrix_exp support non-contiguous inputs. Current Situation Eager: matrix_exp performs batch squash const auto a_3d = a.view({-1, a.size(-2), a.size(-1)}); So channel_last will trigger an error. import torch x = torch.rand(8, 3, 10, 10) x = x.to(memory_format=torch.channels_last) x = torch.matrix_exp(x)..."
188511,[AI Codemod][PerfAICT-General] Replace cast with castRaw in TensorType::isSubtypeOfExt to avoid shared_ptr refcount overhead (#188511),https://github.com/pytorch/pytorch/pull/188511,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,188511,,,,,0,1,0,,,"ciflow/trunk, meta-exported",2026-06-30T01:22:49Z,2026-06-30T18:30:30Z,8,Summary: Pull Request resolved: #188511 Reviewed By: kunalspathak Differential Revision: D109921948
188524,test: centralize optimizer comparison tolerances,https://github.com/pytorch/pytorch/pull/188524,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,116203,116203,116203,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-30T03:12:26Z,2026-06-30T18:06:14Z,8,Fixes #116203 Summary: Adds a shared optimizer-test helper for widened comparison tolerances. Reuses it for the existing sparse-vs-dense comparison and foreach-vs-for-loop parameter/state comparisons. Keeps the current numeric thresholds unchanged; this is only a test refactor to make future optimizer tolerance hand...
188525,test: add Linear module error inputs,https://github.com/pytorch/pytorch/pull/188525,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,174177,174177,174177,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-30T03:15:12Z,2026-06-30T18:06:02Z,8,Fixes #174177 Summary: Adds module_error_inputs_torch_nn_Linear and wires it into the torch.nn.Linear ModuleInfo. Covers the Linear error paths for wrong trailing input features and negative in_features. Intentionally does not add the scalar-input error case because test_linear_raise_on_scalar_input already covers i...
188357,[ROCm] Add MIOpen convolution ops to FORBIDDEN_CUDAGRAPH_OPS for cudagraph safety,https://github.com/pytorch/pytorch/pull/188357,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,188338,188338,"182260, 188338",,,,0,1,0,,,"module: rocm, triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-28T09:20:27Z,2026-06-30T17:57:21Z,8,"Summary MIOpen convolution kernels are not cudagraph-safe on ROCm (see AMD documentation: https://rocm.docs.amd.com/en/latest/reference/graph-safe-support.html). When aten.convolution routes through MIOpen (e.g. transposed convs with no Triton template), it causes faults during CUDA graph replay. This adds the follo..."
188339,Fix/typo overridden applier,https://github.com/pytorch/pytorch/pull/188339,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-27T17:46:04Z,2026-06-30T17:56:08Z,8,cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
188429,[ROCm][CI] Switch rocm-preview to TheRock wheels,https://github.com/pytorch/pytorch/pull/188429,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/rocm-nightly, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,DIRTY,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,14,0,,,"module: rocm, open source, ciflow/trunk, topic: not user facing, keep-going, ci-no-td, ciflow/rocm-nightly, ciflow/docker",2026-06-29T15:03:54Z,2026-06-30T17:53:53Z,8,Replacement for #179639 to retrigger CI from a fresh branch with the same ROCm preview changes. TODO: Waiting on 2 PRs: 1. unit test skips for rocm-preview and 2. #188454 Switch the ROCm preview Docker image from ROCm tarball installs to the multi-arch TheRock wheel index. This keeps the source build targeted to gfx...
176044,[opaque object] Disallow creation of reference-type opaque objects under custom ops in torch.compile,https://github.com/pytorch/pytorch/pull/176044,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,175968,175968,175968,,,,0,4,0,,,"triaged, open source, module: inductor, module: dynamo",2026-02-28T02:46:41Z,2026-06-30T17:19:21Z,8,Fixes #175968 Creation of reference-type opaque objects is not allowed under torch.compile. But it's unclear what to do when a user creates them under a custom op which is opaque to torch.compile. This PR chooses to raise on creation even under custom ops unless the user runs the creation under the unset_fake_tempor...
188549,[AI Codemod][PerfAICT-General] Vectorize index_select bounds check via fused unsigned compare (#188549),https://github.com/pytorch/pytorch/pull/188549,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,188549,,,,,0,1,0,,,meta-exported,2026-06-30T13:22:46Z,2026-06-30T16:30:32Z,8,Summary: Pull Request resolved: #188549 Differential Revision: D110106249
176434,[export] Move schema deserialization to C++ for faster export.load(),https://github.com/pytorch/pytorch/pull/176434,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review when ready,,UNKNOWN,False,0,0,,,176428,176428,176428,,,,0,5,0,,,"triaged, open source, Stale, ciflow/inductor, release notes: export",2026-03-04T15:44:10Z,2026-06-30T16:15:57Z,8,Issue Fixes #176428 Summary See linked issue for motivation and benchmarks. The schema deserialization step in torch.export.load() is moved from Python (_dict_to_dataclass / _bytes_to_dataclass) to the existing C++ infrastructure. The code generator in schema_check.py now also produces pybind11 property bindings for...
186980,Move header-only c10/core headers to torch/headeronly,https://github.com/pytorch/pytorch/pull/186980,draft/noise,13,PR is draft; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"186981, 186982, 186983, 186984, 186985, 186986, 186987, 186988",,,0,4,0,,,ciflow/inductor,2026-06-10T22:12:44Z,2026-06-30T16:04:35Z,8,"Stack from ghstack (oldest at bottom): #186988 #186987 #186986 #186985 #186984 #186983 #186982 #186981 -> #186980 #186962 Migrate the already-header-only headers c10/core/alignment.h, c10/core/CompileTimeFunctionPointer.h, and c10/core/DeviceCapability.h into torch/headeronly/core/, leaving the old c10 paths as one-..."
188465,[Doc] Document GIL and allocation constraints for PU1 custom allocators,https://github.com/pytorch/pytorch/pull/188465,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-06-29T18:49:30Z,2026-06-30T15:36:37Z,8,"PrivateUse1 backends that register custom C allocators via REGISTER_ALLOCATOR face undocumented constraints: allocate() must be GIL-free, cannot call Python code, and cannot associate Python-side metadata with the allocation. These constraints are documented for CUDA but not in the PrivateUse1 accelerator documentat..."
182051,[docs] Link CUDA stream semantics from Tensor.to/copy_ non_blocking docstrings,https://github.com/pytorch/pytorch/pull/182051,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-30T17:32:33Z,2026-06-30T15:20:48Z,8,"Summary Adds a .. note:: admonition to the Tensor.to(non_blocking=True) and Tensor.copy_(non_blocking=True) docstrings cross-referencing the existing CUDA streams note (docs/source/notes/cuda.rst), plus a .. _cuda-stream-semantics: label so the cross-reference resolves. This is a pure docs-visibility tweak; no runti..."
187952,Fix out= variant recompilations for torch.max/min/topk,https://github.com/pytorch/pytorch/pull/187952,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,135859,172727,,,0,3,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo",2026-06-23T15:29:45Z,2026-06-30T14:01:44Z,8,"Issue Partial fix for #135859 (the torch.max / torch.min / torch.topk cases). Summary Under torch.compile, the out= variants of torch.max(x, dim, out=...), torch.min(..., out=...) and torch.topk(..., out=...) recompiled on every new input shape, while the functional variants specialize only once. With the default re..."
186617,Added profiler record for THPVariable_dealloc,https://github.com/pytorch/pytorch/pull/186617,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-08T16:40:57Z,2026-06-30T13:54:47Z,8,"This makes tensor deallocations appear in the profiler's trace. This is useful because, when some CPU backends are improperly configured, deletions of large tensors can make a non-negligible contribution to the runtime."
178305,[XPU][oneDNN] Enforce const/mutable pointer semantics for mkldnn xpu detail,https://github.com/pytorch/pytorch/pull/178305,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,9,0,,,"module: cpu, open source",2026-03-24T18:18:30Z,2026-06-30T12:13:40Z,8,"Fixes: intel/torch-xpu-ops#2919 This PR standardize the use of const_data_ptr() and mutable_data_ptr() according to oneDNN argument roles, eliminate raw data_ptr() usage in mkldnn/xpu/detail, and add lint checks based on the argument mutability classification from uxlfoundation/oneDNN#4843 cc @jgong5 @mingfeima @Xia..."
180857,Add variable-length support for Flash Attention in XPU,https://github.com/pytorch/pytorch/pull/180857,draft/noise,23,PR is draft; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for author to mark ready,,UNKNOWN,True,0,0,,,3093,3093,,,,,0,1,0,,,"module: cpu, open source, ciflow/inductor, release notes: inductor (aoti)",2026-04-20T15:05:41Z,2026-06-30T11:20:01Z,8,"Resolves intel/torch-xpu-ops#3093 Adds variable-length (varlen) support for Flash Attention on XPU, enabling efficient attention computation over batches of sequences with differing lengths without the padding overhead required by the fixed-length path. Changes flash_attn/flash_api.{h,cpp}: expose a new varlen forwa..."
182986,[Eager] [Bitwise-Equivalence] Implement an INNER_TREE sum reduction (#182986),https://github.com/pytorch/pytorch/pull/182986,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, topic: not user facing, meta-exported",2026-05-08T20:04:15Z,2026-06-30T08:48:01Z,8,"Summary: Adds the triton-style inner tree reduction kernel for sum in the caffe2/ATen CUDA backend, including partitioned launch variants, tests, and the output/input offset handling from D104676317 so TensorIterator output strides and strided outer input rows are preserved. Test Plan: arc lint -a fbcode/caffe2/aten..."
188535,[test stats] Sanitize empty-file pytest rows at the upload boundary,https://github.com/pytorch/pytorch/pull/188535,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,topic: not user facing,2026-06-30T07:00:48Z,2026-06-30T08:47:24Z,8,"Summary Test-stat rows for pytest test cases sometimes reach ClickHouse (tests.all_test_runs) with an empty file and a dotted, package-path classname (e.g. classname=""test.test_ops.TestCommonCPU"", file=""""). This re-applies the file-derivation that sanitize_pytest_xml already does in-process, but at the single upload..."
180907,[shard prop] tag more ops,https://github.com/pytorch/pytorch/pull/180907,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,181767,,,0,5,0,,,"Merged, Reverted, Stale, ciflow/trunk, ciflow/inductor, release notes: distributed (dtensor), ci-no-td, ciflow/torchtitan, ciflow/dtensor",2026-04-20T20:34:54Z,2026-06-30T05:55:38Z,8,Tags on more ops we get for free. Single-dim migrates random ops Stack from ghstack (oldest at bottom): #181767 -> #180907
176084,ci: Cleanup dead centos code,https://github.com/pytorch/pytorch/pull/176084,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/binaries_wheel,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,175994,175994,175994,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing, ciflow/binaries_wheel",2026-03-01T03:06:12Z,2026-06-30T05:55:32Z,8,Centos is no longer used by the CI in any way so remove the remaining code to simplify scripts. Closes: #175994
187714,test: Command to reproduce error is incorrect,https://github.com/pytorch/pytorch/pull/187714,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,106377,106377,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-19T07:10:55Z,2026-06-30T04:53:50Z,8,"Summary When TestEnvironment.def_flag stores a flag in repro_env_vars, the original buggy code used the Python flag name (name parameter, e.g. TEST_WITH_TORCHINDUCTOR) as the dict key instead of the actual environment variable name (env_var parameter, e.g. Root cause When TestEnvironment.def_flag stores a flag in re..."
188369,Add DTensor sharding strategy for FFT c2c,https://github.com/pytorch/pytorch/pull/188369,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,188182,188182,188182,,,,0,2,0,,,"triaged, open source, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-28T16:23:53Z,2026-06-30T04:26:57Z,8,"Summary Fixes #188182 This PR adds a DTensor sharding strategy for aten._fft_c2c.default, which is the underlying operator used by complex FFT calls such as torch.fft.fft2. The strategy preserves sharding on dimensions that are not transformed by the FFT, and avoids sharding on the FFT dimensions themselves. This ad..."
187150,Remove unused CUDA 12.4 docker build paths,https://github.com/pytorch/pytorch/pull/187150,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor-pallas","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/inductor-pallas, ciflow/docker",2026-06-12T04:15:45Z,2026-06-30T01:41:04Z,9,
188463,[PU1] Add output_size parameter to torch.bincount for eager-meta support.,https://github.com/pytorch/pytorch/pull/188463,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,188370,,,,0,1,0,,,"open source, release notes: mps, ciflow/mps",2026-06-29T18:25:45Z,2026-06-30T00:05:47Z,9,"OOT accelerators that use device=""meta"" as an eager shape oracle cannot run bincount on meta tensors because the output size depends on max(input), which is unavailable without real data. This adds an optional output_size keyword argument that lets callers specify the expected number of bins explicitly, following th..."
184545,Fix UntypedStorage.new null allocator crash,https://github.com/pytorch/pytorch/pull/184545,has linked issue,70,PR closes or references open issue numbers; high-priority/blocker/regression label present,review when ready,,UNKNOWN,False,0,0,,,184032,184032,184032,,,,0,4,0,,,"module: crash, module: internals, triaged, open source, topic: not user facing",2026-05-20T14:17:55Z,2026-06-29T23:48:11Z,9,"Summary Fixes a null allocator dereference when calling .new() on an UntypedStorage whose allocator is unset. Some storage instances, such as file-backed storages created with UntypedStorage.from_file(..., shared=False) and storages reconstructed through torch.load(..., weights_only=True), can have allocator_ == nul..."
188407,[Windows] Add triton_windows commit pin for Windows nightly builds,https://github.com/pytorch/pytorch/pull/188407,ready for maintainer decision,45,"CI lane labels: ciflow/docker, ciflow/inductor; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/inductor, ciflow/docker",2026-06-29T12:20:32Z,2026-06-29T22:39:43Z,9,"Add a Windows Triton commit pin file, mirroring the existing Linux layout for triton.txt. Add .ci/docker/ci_commit_pins/triton_windows.txt with a pinned commit on triton-lang/triton-windows Add .github/ci_commit_pins/triton_windows.txt as a pointer to the docker pin file (same pattern as triton.txt)"
187870,compile-on-one-rank: device-agnostic inductor codegen and launcher,https://github.com/pytorch/pytorch/pull/187870,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-06-22T20:11:45Z,2026-06-29T22:37:25Z,9,"Stack from ghstack (oldest at bottom): -> #187870 Inductor support for compile-on-one-rank (CooR): a graph compiled on one rank produces byte-identical code and a kernel that loads and runs correctly on any rank, so a single compiled artifact is shareable across ranks. Review in dependency order. First the post_grad..."
187927,docs: document FlopCounterMode semantics,https://github.com/pytorch/pytorch/pull/187927,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,123800,123800,123800,,,,0,3,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-06-23T07:44:26Z,2026-06-29T22:34:38Z,9,"Issue Fixes #123800 Summary Documents torch.utils.flop_counter.FlopCounterMode as a theoretical, formula-based FLOP counter and clarifies common semantic boundaries: unsupported ops, decompositions, custom/Triton kernels, module attribution, backward counting, dense formulas, sparsity/dtype/performance caveats, and..."
188321,[c10d] Integrate ncclScatter into PT,https://github.com/pytorch/pytorch/pull/188321,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,188193,188193,,188320,,,0,2,0,,,"open source, ciflow/trunk, release notes: distributed (c10d)",2026-06-27T04:47:14Z,2026-06-29T22:05:01Z,9,"Stack from ghstack (oldest at bottom): -> #188321 #188320 NCCL 2.28.3 now supports ncclScatter, this PR aims at integrating it. (release note: https://docs.nvidia.com/deeplearning/nccl/release-notes/rel_2-28-3.html#rel_2-28-3) Revive: #164267"
188320,[c10d] Integrate ncclGather into PT,https://github.com/pytorch/pytorch/pull/188320,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,188193,188193,,188321,,,0,2,0,,,"open source, ciflow/trunk, release notes: distributed (c10d)",2026-06-27T04:18:51Z,2026-06-29T22:04:19Z,9,"Stack from ghstack (oldest at bottom): #188321 -> #188320 NCCL 2.28.3 now supports ncclGather, this PR aims at integrating it. (release note: https://docs.nvidia.com/deeplearning/nccl/release-notes/rel_2-28-3.html#rel_2-28-3) Revive: #164270 which we didn't merge due to not clear use case by then."
188464,[spmd_types] infer R@FSDP axes from active mesh for empty params,https://github.com/pytorch/pytorch/pull/188464,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-06-29T18:38:11Z,2026-06-29T21:44:37Z,9,"Stack from ghstack (oldest at bottom): -> #188464 #181519 added the ability to infer missing DP/CP axes spmd type annotations. However, it only supported the case where the param was partially annotated (e.g. TP-only), and not the empty annotation case. We encountered this in TorchTitan FLUX, which only supports FSD..."
187772,perf(dynamo): add opt-in guard lookup token plan,https://github.com/pytorch/pytorch/pull/187772,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,187529,,,0,3,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-20T16:27:36Z,2026-06-29T21:27:08Z,9,"Summary Ports a conservative Last Successful Guard Memo / token-plan experiment to current PyTorch main. The path is guarded by torch._dynamo.config.enable_guard_lookup_memo and defaults to off. With the flag off, Dynamo keeps the existing guard lookup behavior: no receipt allocation, no receipt scope, no extra self..."
188410,Use typing_extensions.TypeAliasType for better reexport of __module__,https://github.com/pytorch/pytorch/pull/188410,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,171905,171905,171905,,,,0,1,0,,,"triaged, open source, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-29T12:56:17Z,2026-06-29T21:21:22Z,9,"Fixes #171905 Use typing_extensions.TypeAliasType to reexport callables from torch.distributed.tensor.experimental instead of mutating __module__ at runtime. Scope (this PR) Update function re-exports in torch/distributed/tensor/experimental/__init__.py Preserve public API: context_parallel, implicit_replication, lo..."
188399,Fix #171905: Use typing_extensions.TypeAliasType for better reexport of `__module__,https://github.com/pytorch/pytorch/pull/188399,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,171905,171905,171905,,,,0,1,0,,,"triaged, open source, release notes: quantization",2026-06-29T09:24:24Z,2026-06-29T21:15:17Z,9,"Fixes #171905 In torch/ao/quantization/__init__.py, replaced the wildcard-imported QConfigAny (which had __module__ = ""torch.ao.quantization.qconfig"") with an explicit TypeAliasType(""QConfigAny"", QConfig | None) definition, giving it the correct __module__ = ""torch.ao.quantization"" so type checkers can properly reso..."
188364,Allow from_blob / TensorMaker to attach BackendMeta at construction,https://github.com/pytorch/pytorch/pull/188364,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,5,0,,,"triaged, open source, release notes: cpp, topic: not user facing",2026-06-28T14:03:01Z,2026-06-29T21:14:50Z,9,"Today, attaching a BackendMeta to a tensor built from external memory is a two-step dance: call from_blob(...) (or for_blob(...).make_tensor()), then reach back in with set_backend_meta(): auto t = at::from_blob(ptr, sizes, strides, deleter, opts, device); t.unsafeGetTensorImpl()->set_backend_meta(c10::make_intrusiv..."
181944,[cuBLAS][cuBLASLt] Shared default workspace size across cuBLAS and Lt,https://github.com/pytorch/pytorch/pull/181944,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cuda, triaged, module: cublas, open source, Stale, ciflow/trunk, release notes: cuda, matrix multiplication, ciflow/h100, ciflow/vllm, ciflow/b200",2026-04-29T23:45:23Z,2026-06-29T21:04:40Z,9,"We've had persistent workspaces for a while now, so perhaps this helps with the peak memory allocation situation? Also motivated by @nikitaved finding a significant source of Lt - nonLt divergence was differing workspaces sizes without unified workspaces e.g., in FBCODE cc @ptrblck @msaroufim @jerryzh168 @tinglvv @n..."
188444,[Test] Refactor tensor_creation_ops and foreach.py to be device-agnostic.,https://github.com/pytorch/pytorch/pull/188444,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-29T17:44:31Z,2026-06-29T21:04:11Z,9,Add hw_classification attributes to all test classes in both files as part of the device-agnostic test refactoring initiative. test_tensor_creation_ops.py: Annotation of 6 existing classes — 4 as DEVICE_GENERIC (already wired through instantiate_device_type_tests) and 2 as GENERIC (CPU-only via only_for=cpu). test_f...
188340,Fix/indices or section typo,https://github.com/pytorch/pytorch/pull/188340,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-27T17:46:32Z,2026-06-29T21:02:55Z,9,
188296,linear_cross_entropy: support probability targets with reduction='none',https://github.com/pytorch/pytorch/pull/188296,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps, ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,188433,,,0,6,0,,,"open source, ciflow/trunk, release notes: nn, ciflow/mps, ciflow/rocm-mi300",2026-06-26T20:28:19Z,2026-06-29T21:02:34Z,9,"Stack from ghstack (oldest at bottom): #188433 -> #188296 Extends the chunked path to probability (soft-label) targets under reduction='none', which previously fell back to the reference. The no_reduction op already routes through the shared accumulator (which implements the prob math for mean/sum); this wires prob..."
188341,Fix/184845 pad docs,https://github.com/pytorch/pytorch/pull/188341,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-27T17:46:48Z,2026-06-29T21:02:09Z,9,
188448,Add DTensor sharding strategy for searchsorted,https://github.com/pytorch/pytorch/pull/188448,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,188447,,,0,1,0,,,"ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-29T18:02:25Z,2026-06-29T21:01:41Z,9,"Stack from ghstack (oldest at bottom): -> #188448 #188447 Register single-dim strategies for aten.searchsorted.Tensor and aten.searchsorted.Scalar. searchsorted(sorted_sequence, self) returns insertion indices of self into sorted_sequence (sorted along its last, ""searched"" dim); the output has self's shape. sorted_s..."
188373,Enforce 1D tensor validation for batch_norm parameters,https://github.com/pytorch/pytorch/pull/188373,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188358,188358,188358,,,,0,1,0,,,"triaged, open source",2026-06-28T18:10:28Z,2026-06-29T20:46:06Z,9,"Fixes #188358 Description This PR addresses an inconsistency where F.batch_norm would silently succeed in eager mode when provided with multidimensional (e.g., 2D/3D) tensors for bias, weight, running_mean, or running_var, but would subsequently crash with a RuntimeError deep in the native ATen kernels or during tra..."
188385,[fx] _iterate_exprs: treat device/dtype/layout/memory_format as symbol-free,https://github.com/pytorch/pytorch/pull/188385,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: fx, ciflow/inductor",2026-06-29T04:04:31Z,2026-06-29T20:24:36Z,9,"Summary _iterate_exprs (used by free_symbols / has_free_symbols in torch/fx/experimental/symbolic_shapes.py) raised AssertionError(""cannot extract sympy expressions from ..."") on torch.device, torch.dtype, torch.memory_format, and torch.layout values. These are non-symbolic scalar metadata and carry no free symbols...."
187223,Add targeted SDPA coverage to H100 CI,https://github.com/pytorch/pytorch/pull/187223,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,release notes: releng,2026-06-13T00:36:46Z,2026-06-29T20:14:11Z,9,"Add a separate h100_sdpa config to the H100 limited CI workflow so the existing smoke run remains unchanged while SDPA gets targeted coverage. The new config runs a bounded subset of inductor/test_fused_attention, and the H100 label/path routing now includes the SDPA fused-attention test and fuse pass so edits trigg..."
179020,Update third-party/rocm_composable_kernel to f1746955,https://github.com/pytorch/pytorch/pull/179020,ready for maintainer decision,45,CI lane labels: ciflow/rocm-mi300; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, fb-exported, Stale, topic: not user facing, ciflow/rocm-mi300, meta-exported",2026-04-01T17:06:25Z,2026-06-29T20:13:31Z,9,"Summary: Bumps internal CK to this commit f1746955 to unblock #178539. This diff is just mgt import, downstream CI fixes are D98344491 kept separately for easier review, but will be squashed for landing. Test Plan: mgt upgrade --src-type git https://github.com/ROCm/composable_kernel --git-commit f1746955fdaf80a3414d..."
188291,"Fix dims docstring for torch.flip and torch.rot90 (axes, element type)",https://github.com/pytorch/pytorch/pull/188291,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,130641,130641,130641,,,,0,1,0,,,"triaged, open source, release notes: python_frontend, topic: docs",2026-06-26T19:39:28Z,2026-07-04T07:01:06Z,4,"Fixes #130641. torch.flip's docstring describes dims as: dims (a list or tuple): axis to flip on dims is a list/tuple of dimensions, so ""axis"" should be plural (""axes""), and the element type can be named, consistent with how torch.permute documents its dims argument (tuple of int or list of int). torch.rot90 in the..."
188447,Add DTensor shard propagation skill,https://github.com/pytorch/pytorch/pull/188447,draft/noise,13,PR is draft,wait for author to mark ready,,CLEAN,True,0,0,,,,,,188448,,,0,1,0,,,topic: not user facing,2026-06-29T18:02:20Z,2026-06-29T19:31:11Z,9,"Stack from ghstack (oldest at bottom): #188448 -> #188447 Add a focused Codex skill for implementing and debugging DTensor sharding propagation rules. The skill captures the current preference for register_single_dim_strategy, the usual torch/distributed/tensor/_ops/ edit location, RuntimeSchemaInfo cache pitfalls,..."
182011,[WIP][PoC] Enable Arm Optimized Routines for AArch64+Linux,https://github.com/pytorch/pytorch/pull/182011,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/pull, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,182013,,,0,2,0,,,"module: cpu, open source, Stale, ciflow/trunk, topic: not user facing, ci-no-td, ciflow/pull",2026-04-30T11:01:43Z,2026-06-29T19:13:21Z,9,Stack from ghstack (oldest at bottom): -> #182011 This is just a WIP PoC to showcase the improvements we can get through Arm Optimized Routines (AOR) for a bunch of fundamental math operations in PyTorch. It still needs some cleaning + discussion - see TODO section. Performance Benchmarked performance using @malfet...
188164,Fix THPDTypeInfo_compare reading off the end of non-finfo/iinfo objects,https://github.com/pytorch/pytorch/pull/188164,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,188036,,,0,1,0,,,"triaged, open source",2026-06-25T17:17:09Z,2026-06-29T19:12:40Z,9,"Problem THPDTypeInfo_compare is registered as tp_richcompare for both THPFInfoType and THPIInfoType, but it takes a typed THPDTypeInfo* b and reads b->type without checking that the right-hand operand is actually a finfo/iinfo: static PyObject* THPDTypeInfo_compare(THPDTypeInfo* a, THPDTypeInfo* b, int op) { switch..."
188433,linear_cross_entropy: test and document bias with probability targets,https://github.com/pytorch/pytorch/pull/188433,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,188296,,,0,2,0,,,"open source, release notes: nn, topic: not user facing",2026-06-29T15:18:03Z,2026-06-29T18:51:42Z,9,"Stack from ghstack (oldest at bottom): -> #188433 #188296 Probability (soft-label) targets combined with a linear bias already work on the chunked path -- the loss adds the bias through the shared shifted_logits, and the bias gradient is the batch-sum of the dense grad_logits the prob loop already forms, correct for..."
182033,Add a utility method for interrupting long running subprocess during run_test.py,https://github.com/pytorch/pytorch/pull/182033,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale",2026-04-30T15:15:43Z,2026-06-29T18:16:28Z,9,"The Problem When attempting to interrupt a long-running subprocess, the code was explicitly sending signal.SIGINT. On Windows, the subprocess.send_signal() method does not support SIGINT properly, causing it to throw a ValueError: Unsupported signal: 2. While this doesn't necessarily crash the parent runner, it prev..."
188314,Fix tutorial link for GPU Direct Storage,https://github.com/pytorch/pytorch/pull/188314,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,False,0,0,,,,188245,188245,,,,0,1,0,,,open source,2026-06-26T23:28:27Z,2026-06-29T18:14:58Z,9,Fixes #188245 Summary Updated the tutorial link for GPU Direct Storage in the serialization notes. cc @svekars @sekyondaMeta @AlannaBurke
182030,[ROCm] Make AOTriton bundling optional via `BUILD_AOTRITON_INTO_WHEEL` flag,https://github.com/pytorch/pytorch/pull/182030,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, triaged, open source, Stale, topic: not user facing, no-stale",2026-04-30T14:33:51Z,2026-06-29T17:34:08Z,9,Previously AOTriton was always copied to Torch's output/wheel. This is unnecessary and can waste a gigabyte of space easily for distros which have AOTriton as an installable package. Adding this flag lets us avoid duplicating AOTriton in NixOS/Nixpkgs' ROCm torch output. NixOS/nixpkgs#515088 cc @jeffdaily @sunway513...
182040,Download magma & sccache install scripts over https instead of aws s3,https://github.com/pytorch/pytorch/pull/182040,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale",2026-04-30T15:41:59Z,2026-06-29T17:23:04Z,9,Removed the conditional logic that switched between curl and aws s3 cp based on the %BUILD_ENVIRONMENT% variable. The script now uses the standard curl approach exclusively to fetch the binary from S3.
188365,[MPS] Remove support for macOS 14,https://github.com/pytorch/pytorch/pull/188365,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/mps",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: binaries, release notes: releng, topic: deprecation, ciflow/mps, module: inductor, ciflow/inductor",2026-06-28T14:35:41Z,2026-06-29T17:18:35Z,9,Removes support for macos 14. No longer building binaries Removing tests for macos 14 Removing gates in kernels related to macos 14 gating cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @c...
187461,Do Not Review: Polyhedral Optimization,https://github.com/pytorch/pytorch/pull/187461,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-perf-compare, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,187451,187451,187451,187452,,,0,14,0,,,"open source, module: inductor, ciflow/inductor, ciflow/inductor-perf-compare, ciflow/torchtitan",2026-06-16T16:57:29Z,2026-06-29T17:14:26Z,9,"Fixes #187451 This PR is the same other draft PR with polyhedral_fusion = True. This is purely to be able run benchmarking workloads. As stated in the original issue the feature will be opt in. Unfortunately, I don't have access to use stack, so this is my workaround. Please review the other PR on this issue at #187..."
188363,Add kernel_size > 0 validation to FractionalMaxPool2d matching Fracti…,https://github.com/pytorch/pytorch/pull/188363,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188217,188217,"188217, 188218, 188226",,,,0,1,0,,,"triaged, open source",2026-06-28T12:58:24Z,2026-06-29T16:53:32Z,9,"Fixes #188217, #188218, #188226 Summary FractionalMaxPool3d validates that kernel_size > 0 in __init__ but FractionalMaxPool2d was missing the equivalent check. Non-positive kernel sizes were silently accepted in eager forward, producing invalid intermediate values that caused Inductor to crash in three different wa..."
188381,aten: add output_size kwarg to masked_select for meta/eager shape oracle support,https://github.com/pytorch/pytorch/pull/188381,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,188370,188370,188370,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-28T23:18:40Z,2026-06-29T16:50:49Z,9,"Summary Adds an optional output_size (SymInt?) keyword argument to masked_select, following the repeat_interleave.Tensor precedent. When output_size is provided on a device=""meta"" tensor, the meta kernel uses it to produce output metadata — enabling OOT accelerators to use masked_select in their eager shape-oracle l..."
188347,Preserve tied parameter and buffer aliasing in load_state_dict(assign=True),https://github.com/pytorch/pytorch/pull/188347,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188346,188346,,,,,0,1,0,,,"triaged, open source",2026-06-27T20:40:30Z,2026-06-29T16:28:37Z,9,"Issue Fixes #188346 Summary See the linked issue for the bug and repro. In assign mode each name is reassigned its own object via setattr, so names sharing one tied Parameter/buffer become distinct objects (copy mode preserves tying by copying in place), and _load_from_state_dict runs per module so it cannot see cro..."
188334,Handle negative integer exponents in SymPy utilities,https://github.com/pytorch/pytorch/pull/188334,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188323,188323,188323,,,,0,1,0,,,"module: cpu, triaged, open source",2026-06-27T13:58:53Z,2026-06-29T16:28:31Z,9,"Fixes #188323 Summary Fix handling of negative integer exponents in SymPy utilities. SymPy can represent expressions such as 1/x as Pow(x, -1). However, sympy.Pow is currently dispatched to pow_by_natural, which assumes a non-negative integer exponent. This could trigger assertions when symbolic expressions containi..."
188294,Fix GradientEdge None-grad error message formatting,https://github.com/pytorch/pytorch/pull/188294,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source",2026-06-26T20:20:52Z,2026-06-29T16:27:59Z,9,Summary Fix a missing space in the run_backward user-facing error message for GradientEdge when the corresponding gradient is None. Add a focused regression test to lock in the corrected error text. Co-authored-by: Cursor Test plan Attempted local targeted test in repo venv: python test/test_autograd.py -k test_grad...
188293,Allow tuple subclasses for run_backward inputs,https://github.com/pytorch/pytorch/pull/188293,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-26T20:20:31Z,2026-06-29T16:27:54Z,9,"Summary Relax run_backward input validation from PyTuple_CheckExact to PyTuple_Check so tuple subclasses are accepted, matching the existing behavior for other tuple arguments. Add an autograd regression test covering tuple-subclass inputs to torch.autograd.grad. Co-authored-by: Cursor Test plan Attempted local targ..."
182851,[dynamo] Fix free-threading race in dict recursive-tag watcher callback,https://github.com/pytorch/pytorch/pull/182851,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/tsan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, no-runner-experiments, ciflow/tsan",2026-05-07T17:30:35Z,2026-06-29T16:27:23Z,9,"The dict watcher callback on a tag-safe root previously walked the watched GuardManager's _dict_pointers map and also read/wrote _disable_dict_tag_matching, while the guard-evaluation hot path read and mutated those same members under RootGuardManager::_lock. Because the callback only held dict_to_guard_managers, bo..."
188295,Validate queue_callback input is callable,https://github.com/pytorch/pytorch/pull/188295,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source",2026-06-26T20:21:14Z,2026-06-29T16:27:19Z,9,Summary Add explicit callable validation in THPEngine_queue_callback so invalid callback arguments fail immediately on the caller thread. Add a targeted autograd regression test asserting a clear error for non-callable callback inputs. Co-authored-by: Cursor Test plan Attempted local targeted test in repo venv: pyth...
188404,Inductor: Add metadata overlap dependency implementation,https://github.com/pytorch/pytorch/pull/188404,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-29T11:47:49Z,2026-06-29T14:44:45Z,9,"Stack from ghstack (oldest at bottom): -> #188404 Overlap scheduling currently materializes ordering constraints by wrapping FX nodes in control_deps. That preserves ordering, but it hides the original nodes inside subgraphs and can force extra realization before scheduler fusion. For compute and communication overl..."
188143,Add mixed Tensor and Scalar bound support for torch.clamp and torch.clip,https://github.com/pytorch/pytorch/pull/188143,draft/noise,23,PR is draft; PR closes or references open issue numbers,wait for author to mark ready,,UNKNOWN,True,0,0,,,188088,188088,188088,,,,0,2,0,,,"open source, topic: bug fixes",2026-06-25T07:07:40Z,2026-06-29T14:05:29Z,9,"fixes #188088 Summary This PR resolves the TypeError encountered when passing mixed bound types (e.g., min=Tensor, max=Scalar) to torch.clamp and its alias torch.clip. To maintain C++ API parity, these are implemented as CompositeImplicitAutograd functions. The implementation wraps the Scalar bound into a 0-D Tensor..."
171035,Add 2D Right-hand-side support for cuDSS backend `linalg.solve`,https://github.com/pytorch/pytorch/pull/171035,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,28,0,,,"triaged, open source, Stale, release notes: sparse, module: inductor, module: dynamo",2025-12-22T03:25:59Z,2026-06-29T13:17:25Z,9,"Support for solving sparse CSR linear systems with 2D right-hand side (RHS) tensors when linear system is sparse. Fix #129856 (comment) _apply_sparse_csr_linear_solve in SparseCsrTensorMath.cu accepts both 1D and 2D RHS tensors, with additional stride and size checks for 2D cases. The function now computes leading d..."
169241,[XPU] Migrate 10 dynamo test cases for XPU,https://github.com/pytorch/pytorch/pull/169241,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,114850,114850,,,,,0,58,0,,,"triaged, open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/xpu, ciflow/torchtitan",2025-11-29T18:43:08Z,2026-06-29T11:55:14Z,9,"Description Fixes #114850, we will port dynamo, fsdp tests to Intel GPU We could enable Intel GPU with following methods and try the best to keep the original code styles: Changes Get device type with from accelerator and get_devtype helper method Replace the requires cuda statement with requires_gpu. Replace the cu..."
169798,"[XPU] Migrate 2 test cases autograd, creation_ops for XPU",https://github.com/pytorch/pytorch/pull/169798,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,114850,114850,,,,,0,25,0,,,"triaged, open source, topic: not user facing, ciflow/xpu",2025-12-08T01:27:56Z,2026-06-29T11:55:00Z,9,"Description Fixes #114850, we will port dynamo, fsdp tests to Intel GPU We could enable Intel GPU with following methods and try the best to keep the original code styles: Changes Get device type with from accelerator and get_devtype helper method Replace the requires cuda statement with requires_gpu. Replace the cu..."
188380,docs: fix broken gpu_direct_storage tutorial URL in serialization notes,https://github.com/pytorch/pytorch/pull/188380,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,188245,188245,,,,0,1,0,,,"open source, topic: not user facing",2026-06-28T22:55:27Z,2026-06-29T11:54:58Z,9,Summary Fixes #188245 The prototype path for the GPU Direct Storage tutorial was removed; the tutorial now lives under unstable. This PR updates the link in docs/source/notes/serialization.rst so the Layout Control section no longer points to a 404. Old URL (broken): https://docs.pytorch.org/tutorials/prototype/gpu_...
181996,[torch.package] Register package modules in sys.modules for Python 3.12 dataclass compat,https://github.com/pytorch/pytorch/pull/181996,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, Stale, topic: not user facing, meta-exported",2026-04-30T07:50:02Z,2026-06-29T10:04:28Z,9,"Summary: In Python 3.12, dataclasses._is_type() calls sys.modules.get(cls.__module__).__dict__ to resolve ClassVar/InitVar annotations. When a dataclass is defined inside a torch.package, cls.__module__ is set to a mangled name (e.g. <torch_package_0>.foo.bar). This mangled name is registered in the private _package..."
181678,[spmd_types] CI CPU job w/o spmd_types,https://github.com/pytorch/pytorch/pull/181678,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,181398,,,0,7,0,,,"Stale, release notes: releng",2026-04-28T00:30:40Z,2026-06-29T10:04:20Z,9,Stack from ghstack (oldest at bottom): -> #181678 #181519 #181398 #180880
177205,Fix all_gather_into_tensor with stack-shaped output under torch.compile + dynamic shapes,https://github.com/pytorch/pytorch/pull/177205,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,138795,"138795, 155632",,,,0,4,0,,,"Stale, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-03-11T21:55:58Z,2026-06-29T10:04:16Z,9,"Stack from ghstack (oldest at bottom): -> #177205 Fixes #138795 and #155632 The native dist.all_gather_into_tensor supports both concat-shaped output [world_size * dim0, ...] and stack-shaped output [world_size, dim0, ...]. Under torch.compile, Dynamo remaps this to all_gather_tensor_inplace, which calls the functio..."
185913,[BugFix] Fix torch.kron crash on non-contiguous input (Issue #185650),https://github.com/pytorch/pytorch/pull/185913,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,185650,185650,185650,,,,0,5,0,,,"triaged, open source, module: linear algebra, ciflow/trunk, release notes: linalg_frontend",2026-06-02T10:02:41Z,2026-06-29T09:47:27Z,9,"Summary: Fixes #185650 This PR resolves a RuntimeError that occurs when the second argument to torch.kron is non-contiguous (e.g., after a transpose operation). The issue also manifested in torch.compile. The Problem: The previous implementation used at::_unsafe_view, which fails for non-contiguous tensors as it can..."
188092,Fix cartesian_prod returning inconsistent shape for single tensor input,https://github.com/pytorch/pytorch/pull/188092,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,116465,116465,,,,0,1,0,,,"module: bc-breaking, triaged, open source, ciflow/trunk, topic: not user facing",2026-06-24T19:34:17Z,2026-06-29T09:22:11Z,9,"Fixes: #116465. When cartesian_prod receives multiple tensors, the multi-tensor path returns shape (N, num_tensors) via at::stack(grids, 1). But the single-tensor early return just returned the input as-is with shape (N,), which is inconsistent with both the multi-tensor case and itertools.product (which returns (N,..."
188287,Improve FX CodeGen process_inputs typing,https://github.com/pytorch/pytorch/pull/188287,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,146249,146249,,,,0,1,0,,,"open source, release notes: fx",2026-06-26T19:00:53Z,2026-06-29T07:46:43Z,9,"Summary Refs #146249. This narrows the default torch.fx.graph.CodeGen.process_inputs return annotation from Any to tuple[Any, ...]. The implementation already returns args unchanged, and Python collects *args as a tuple, so the annotation now matches the runtime behavior without changing execution: def process_input..."
181956,[do-not-review] cudagraph while node support,https://github.com/pytorch/pytorch/pull/181956,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,Stale,2026-04-30T03:09:02Z,2026-06-29T06:07:05Z,9,Stack from ghstack (oldest at bottom): -> #181956
187900,[XPU] Migrate CI Docker to Intel OMIX for oneAPI + GPU driver installation,https://github.com/pytorch/pytorch/pull/187900,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing, keep-going, ciflow/xpu, ciflow/docker",2026-06-23T02:49:30Z,2026-06-29T05:29:36Z,9,"Summary Replace the two-step GPU driver + DLE offline installer process with a single Intel OMIX apt package install for Noble CI Docker images. Add OMIX_VERSION build arg to gate install method (OMIX vs legacy DLE) When OMIX_VERSION is set, use intel-omix + intel-omix-dev apt packages from OMIX repository When unse..."
187922,[TEST] Refactor test_embedding.py with hw_classification,https://github.com/pytorch/pytorch/pull/187922,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,186918,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-06-23T05:40:40Z,2026-06-29T03:35:32Z,10,"Apply hardware classification structure following PR #186918 guidelines. Changes: Rename TestEmbeddingNN → TestEmbeddingGeneric (hw_classification = GENERIC) 18 CPU-only tests, accelerator-unrelated Rename TestEmbeddingNNDeviceType → TestEmbeddingDeviceGeneric (hw_classification = DEVICE_GENERIC) 29 device-agnostic..."
175809,[cuDNN] Migrate to new frontend API for batchnorm,https://github.com/pytorch/pytorch/pull/175809,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,18,0,,,"triaged, open source, Stale, release notes: cuda, topic: not user facing",2026-02-25T23:45:55Z,2026-06-29T01:29:15Z,10,"reopening this PR since it needs to be merged or else batchnorm will start breaking comment from old PR: Add v8 cuDNN frontend API for batchnorm Theoretically new api is a bit faster at cost of higher memory usage, but really this PR is necessary as cuDNN plans to deprecate the old API in upcoming versions. In order..."
187955,Make expandable-segment Handle a move-only RAII owner and exception-safe,https://github.com/pytorch/pytorch/pull/187955,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,188008,188008,188008,,,,0,5,0,,,"triaged, open source, topic: not user facing",2026-06-23T16:24:32Z,2026-06-29T01:21:43Z,10,"ExpandableSegment::Handle becomes a move-only RAII owner: its destructor frees the physical handle and closes the shareable fd, so the map() / unmap() / fromShared() error paths can no longer leak handles. mapAndSetAccess is made exception-safe (a partial map/setAccess failure rolls back), keeping the active() <=> m..."
181444,Scaled cublas grouped gemm sm90,https://github.com/pytorch/pytorch/pull/181444,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,177038,,,0,30,0,,,"open source, Stale, module: inductor, module: dynamo, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, ciflow/dtensor",2026-04-24T20:56:27Z,2026-06-28T23:52:40Z,10,"Adds cublasLt fp8 scaled grouped GEMM support for Hopper, compatible with CTK 13.2 update 1 and newer. To be stacked on top of #177038. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coco..."
181443,Cublas grouped gemm nvfp4,https://github.com/pytorch/pytorch/pull/181443,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,177038,,,0,32,0,,,"open source, Stale, module: inductor, module: dynamo, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan, ciflow/dtensor",2026-04-24T20:35:27Z,2026-06-28T23:52:39Z,10,"Adds support for NVFP4 grouped GEMM with cublasLt, compatible with CTK 13.2 update 1 and newer. This PR is intended to be stacked on top of #177038. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aa..."
177924,[triton hash update] update the pinned triton hash,https://github.com/pytorch/pytorch/pull/177924,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, Stale, ciflow/trunk, topic: not user facing, ciflow/inductor, ci-no-td",2026-03-20T00:37:06Z,2026-06-28T23:52:32Z,10,This PR is auto-generated nightly by this action. Update the pinned triton hash.
187871,Fix scaled_dot_product_attention: explicitly check for both attn_mask…,https://github.com/pytorch/pytorch/pull/187871,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,187801,187801,,,,,0,1,0,,,"triaged, open source, ciflow/mps",2026-06-22T20:26:50Z,2026-06-28T23:36:42Z,10,"Fixes #187801 Summary The scaled_dot_product_attention function's documentation specifies that passing both an explicit attn_mask and is_causal=True is not allowed. Previously, this constraint was only being checked inside specific backend implementations (like the math path), which could cause the function to eithe..."
188015,[SDPA] Fix 32-bit dropout RNG offset overflow in mem-efficient attention,https://github.com/pytorch/pytorch/pull/188015,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: cuda, module: sdpa",2026-06-24T02:09:45Z,2026-06-28T23:05:33Z,10,"Summary mem_eff_attention computes the per-(batch, head) dropout RNG offset as (batch_id * num_heads + head_id) * num_queries * num_keys in 32-bit. That product wraps once num_queries * num_keys reaches 2**32 (seq_len > 65536, or sooner once you multiply through several heads/batches, since the head term alone is nu..."
181561,[AI Codemod][DevmateCppModernizeFindToContains] fbcode/caffe2/torch/csrc/jit/backends/xnnpack/xnnpack_graph_builder.cpp (#181561),https://github.com/pytorch/pytorch/pull/181561,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,181561,,,,,0,1,0,,,"open source, fb-exported, Stale, ciflow/trunk, release notes: jit, meta-exported",2026-04-27T09:00:26Z,2026-06-28T20:49:58Z,10,Summary: Pull Request resolved: #181561 Reviewed By: aelatawy Differential Revision: D100773414
178283,Bump torch_tpu pin,https://github.com/pytorch/pytorch/pull/178283,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-pallas","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"Stale, topic: not user facing, ciflow/inductor, ciflow/inductor-pallas",2026-03-24T15:21:24Z,2026-06-28T20:49:52Z,10,Stack from ghstack (oldest at bottom): -> #178283
173914,[CI] [CUDA] Add CUDA ARM CI image,https://github.com/pytorch/pytorch/pull/173914,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,173950,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-01-30T09:36:39Z,2026-06-28T20:49:51Z,10,draft PR to try create an aarch64 CI image for periodic tests #173950
187999,Fix uninitialized memory warnings in CPU linear algebra kernels,https://github.com/pytorch/pytorch/pull/187999,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-06-23T23:36:44Z,2026-06-28T20:29:44Z,10,"Modifies ATen CPU linear algebra kernels in BatchLinearAlgebraKernel.cpp to value-initialize workspace query variables and .zero_() output tensors before passing their buffers to external BLAS/LAPACK functions. When building PyTorch with MemorySanitizer (msan), test suites executing CPU linear algebra operations (su..."
188031,Fix 64-bit indexing in spatial softmax backward,https://github.com/pytorch/pytorch/pull/188031,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cuda, triaged, module: 64-bit, open source, release notes: nn",2026-06-24T06:16:53Z,2026-06-28T20:23:09Z,10,"Summary The spatial softmax backward kernel still indexes with 32-bit math, so it produces wrong gradients on tensors that the forward pass already handles correctly. cunn_SpatialSoftMaxBackward runs whenever softmax/log_softmax is taken over a non-last dimension (inner_size != 1). It hardcodes uint32_t for the size..."
179820,[ATen][Native][CUDA] Use dynamic loading in CUDNN FrontEnd,https://github.com/pytorch/pytorch/pull/179820,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"module: cuda, open source, Stale, ciflow/binaries, release notes: cuda, topic: not user facing",2026-04-09T14:12:11Z,2026-06-28T19:57:15Z,10,"Per title: this PR updates CUDNN FE submodule to the latest released version of v1.22.0. The CUDNN FE introduced a dependency on LibNVRTC, and there are two ways handling it: direct linking and dynamic loading. This PR opts for dynamic loading, in order to prevent dependency LibTorch on LibNVRTC. cc @ptrblck @msarou..."
187711,Strict pyrefly checking: round 1 (2/N),https://github.com/pytorch/pytorch/pull/187711,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,187712,,,0,3,0,,,"topic: not user facing, ciflow/inductor, release notes: export",2026-06-19T04:09:55Z,2026-06-28T19:22:51Z,10,"Stack from ghstack (oldest at bottom): #187712 -> #187711 Annotates and turns on whole-file pyrefly checking (per-file [[sub-config]] entries) for torch/_export/{utils,non_strict_utils,wrappers}.py and torch/_export/serde/{schema_check,union}.py. Second batch of the effort to enable annotation checking on non-public..."
181145,[CuteDSL][Native Ops] group norm kernel,https://github.com/pytorch/pytorch/pull/181145,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,39,0,,,"open source, Stale",2026-04-22T18:07:50Z,2026-06-28T18:52:57Z,10,"Gonna park this PR here until quack rmsnorm gets merged, and then hopefully we can have a clean rebase..."
187930,Fix torch.compile validation for random ops,https://github.com/pytorch/pytorch/pull/187930,draft/noise,23,PR is draft; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for author to mark ready,,UNKNOWN,True,0,0,,,"185246, 185248","185246, 185248","185246, 185248",,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, release notes: inductor",2026-06-23T08:12:49Z,2026-06-28T02:34:01Z,11,Fixes #185246. Fixes #185248. Summary Preserve eager-style runtime validation for tensor probabilities in the compiled torch.bernoulli decomposition. Preserve eager-style runtime validation for tensor std inputs in the compiled torch.normal reference path. Add CPU Inductor regression coverage for invalid Bernoulli p...
183688,[MPS] Add torch.mps.MPSGraph capture/replay via MTLIndirectCommandBuffer,https://github.com/pytorch/pytorch/pull/183688,draft/noise,13,PR is draft; CI lane labels: ciflow/mps,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,"180397, 187455",,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-14T09:00:16Z,2026-06-27T22:52:51Z,11,"Tracking issue: #180397. Related context: #187455. Current status This branch is a monolithic reference implementation, not a final reviewable PR. Please do not review it as one patch. It needs to be split into a capture/replay stack: Capture recorder core Internal MPSStreamGraph command list and recording encoder/p..."
182749,[MPS] Zero-copy CPU->MPS via memcpy on Apple Silicon UMA,https://github.com/pytorch/pytorch/pull/182749,draft/noise,13,PR is draft; CI lane labels: ciflow/mps,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,187455,"182731, 182736, 182791",,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-07T01:00:15Z,2026-06-27T22:52:51Z,11,Tracked under #187455 until a dedicated UMA/direct-memory tracking issue exists. Current status Pending rework into the UMA stack. Please do not review this as the final standalone shape. Order: UMA base helper: shared MTLStorageModeShared pointer detection plus stream-queue synchronization. Scalar-read leaf: #18273...
182736,[MPS] Zero-copy MPS->CPU transfer via unified memory in copy_from_mps_,https://github.com/pytorch/pytorch/pull/182736,draft/noise,13,PR is draft; CI lane labels: ciflow/mps,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,"172987, 187455","182731, 182749, 182791",,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-06T23:23:53Z,2026-06-27T22:52:51Z,11,Tracked under #187455 until a dedicated UMA/direct-memory tracking issue exists. Current status Pending rework into the UMA stack. Please do not review this as the final standalone shape. Order: UMA base helper: shared MTLStorageModeShared pointer detection plus stream-queue synchronization. Scalar-read leaf: #18273...
182731,[MPS] Zero-copy scalar extraction via unified memory in _local_scalar_dense,https://github.com/pytorch/pytorch/pull/182731,draft/noise,13,PR is draft; CI lane labels: ciflow/mps,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,187455,"182736, 182749, 182791",,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-06T22:34:53Z,2026-06-27T22:52:51Z,11,Tracked under #187455 until a dedicated UMA/direct-memory tracking issue exists. Current status Pending rework into the UMA stack. Please do not review this as the final standalone shape. Order: Extract a UMA base helper first: shared MTLStorageModeShared pointer detection plus the stream-queue synchronization patte...
182791,[MPS] UMA fast path for fill/zero via CPU memset/std::fill,https://github.com/pytorch/pytorch/pull/182791,draft/noise,13,PR is draft; CI lane labels: ciflow/mps,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,187455,"182731, 182736, 182749",,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-07T07:25:56Z,2026-06-27T22:52:50Z,11,Tracked under #187455 until a dedicated UMA/direct-memory tracking issue exists. Current status Pending rework into the UMA stack. Please do not review this as the final standalone shape. Order: UMA base helper: shared MTLStorageModeShared pointer detection plus stream-queue synchronization. Scalar-read leaf: #18273...
181767,"[shard prop] autogen infra for inplace,out variants",https://github.com/pytorch/pytorch/pull/181767,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,180907,,,0,2,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-04-28T18:16:52Z,2026-06-27T22:50:21Z,11,Stack from ghstack (oldest at bottom): -> #181767 #180907
181330,[Profiler] Add a skip_build_tree metadata field to Kineto events to m…,https://github.com/pytorch/pytorch/pull/181330,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale",2026-04-24T03:52:20Z,2026-06-27T22:50:18Z,11,"In my custom kineto backend, some events are added only for state annotation within a time range and should not appear in the call tree. Including them in build_tree may lead to incorrect parent–child relationships, so we add a skip_build_tree field to skip them during build_tree."
188271,Allow PrivateUse1 backends to opt out of CEA decomposition,https://github.com/pytorch/pytorch/pull/188271,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188052,188052,188052,,,,0,1,0,,,"triaged, open source",2026-06-26T13:42:31Z,2026-06-27T22:05:19Z,11,Background Fixes #188052 PyTorch's dispatch table resolves CEA/CEANF kernels before the backend fallback. PrivateUse1 backends that register a single catch-all fallback never see composite ops intact — e.g. convolution is silently decomposed to convolution_overrideable (a stub that raises NotImplementedError) before...
181681,Fix UBSan null-pointer UB in vec256 partial-load Vec::loadu (#181510),https://github.com/pytorch/pytorch/pull/181681,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,,32502,181510,181665,,,0,5,0,,,"module: cpu, triaged, open source, Stale, release notes: nn",2026-04-28T01:05:11Z,2026-06-27T19:56:28Z,11,"What Vectorized<T>::loadu(ptr, count) in the vec256 specializations calls std::memcpy(tmp_values, ptr, count * sizeof(T)) unconditionally in its partial-load branch. When ptr == nullptr and count == 0 (which happens when a zero-numel tensor's data_ptr() is passed in), this is UB: aten/src/ATen/cpu/vec/vec256/vec256_..."
181724,S390X: update binaries labels,https://github.com/pytorch/pytorch/pull/181724,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-28T07:10:06Z,2026-06-27T18:54:05Z,11,"There's currently a big volume of generated PRs that run binaries builds. S390X CI doesn't have capacity to run too many builds at once in a timely manner. Due to that, remove non-s390x labels from s390x-specific binaries builds."
181627,[ ROCm] [ DRAFT ] 181288 + 177219,https://github.com/pytorch/pytorch/pull/181627,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/op-benchmark,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,10,0,,,"module: rocm, open source, Stale, topic: not user facing, ciflow/op-benchmark",2026-04-27T18:54:39Z,2026-06-27T18:54:03Z,11,Bundle of two changes for CI testing: #181288 + #177219 cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
181623,[torch.library] Add pallas_op for JAX/Pallas kernel registration,https://github.com/pytorch/pytorch/pull/181623,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"Stale, release notes: export",2026-04-27T18:19:28Z,2026-06-27T18:54:01Z,11,"Summary Add torch.library.pallas_op, a decorator analogous to torch.library.triton_op that registers a JAX/Pallas function as a custom PyTorch operator. This enables PrivateUse1 backends (e.g., TPU via torch_tpu) to register Pallas kernels through a first-class PyTorch API. pallas_op wraps custom_op and additionally..."
154279,Add Support for transposed convolution with Padding Mode 'same' and 'valid',https://github.com/pytorch/pytorch/pull/154279,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,"3867, 80301","3867, 80301","3867, 80301",,,,0,9,0,,,"module: cpu, triaged, open source, Stale, release notes: cpp",2025-05-23T21:12:14Z,2026-06-27T18:40:27Z,11,This pull request adds support for 'same' and 'valid' padding modes for transposed convolutions. Implementation Details for 'same' padding: Compute the minimum required padding for the left and right sides. Perform the convolution using the computed padding. Adjust the output so that its size matches the input size....
181628,[ ROCm] [ DRAFT ] 181288 + 177219 + 181615,https://github.com/pytorch/pytorch/pull/181628,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/op-benchmark,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,181615,,,0,12,0,,,"module: rocm, open source, Stale, release notes: releng, ciflow/op-benchmark",2026-04-27T18:57:12Z,2026-06-27T17:50:58Z,11,Bundle of two changes for CI testing: #181288 + #177219 + #181615 cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
181516,Fix heap-buffer-overflow in torch.linalg.eig on subnormal real inputs,https://github.com/pytorch/pytorch/pull/181516,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,162358,162358,162358,,,,0,1,0,,,"triaged, open source, Stale, release notes: linalg_frontend",2026-04-26T15:52:56Z,2026-06-27T17:50:56Z,11,Summary Fixes a heap-buffer-overflow in torch.linalg.eig triggered when LAPACK produces inconsistent eigenvalue/eigenvector outputs under subnormal arithmetic. One bounds check in the CPU unpacker plus the analogous fix in the CUDA kernel; recovers by treating the eigenvalue as real (the only interpretation consiste...
176584,[RNG] Decouple Philox subsequence from hardware topology for cross-platform consistency,https://github.com/pytorch/pytorch/pull/176584,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: distributions, module: bc-breaking, triaged, open source, Stale, topic: bc breaking",2026-03-05T09:57:17Z,2026-06-27T17:50:48Z,11,"Problem With a fixed seed, torch.randn() produces different output sequences on GPUs with different SM/CU counts. The divergence point is predictable and occurs exactly at min(grid.x * block_size) of the two devices (i.e., the smaller number of launched threads). We tested torch.randn(200000, device=""cuda"") with see..."
188183,[nn] Avoid mutating state_dict metadata in load_state_dict(assign=True),https://github.com/pytorch/pytorch/pull/188183,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188166,188166,188166,,,,0,1,0,,,"triaged, open source",2026-06-25T19:40:10Z,2026-06-27T17:47:49Z,11,"Fixes #188166. Summary Fix Module.load_state_dict(assign=True) mutating the caller's state_dict._metadata. Previously, load_state_dict() retrieved each module's metadata dictionary by reference and inserted the temporary assign_to_params_buffers flag into it. Because the metadata dictionary belonged to the caller's..."
188191,Add sm_121a (NVIDIA GB10/DGX Spark) to default CUDA arch list,https://github.com/pytorch/pytorch/pull/188191,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: build, triaged, open source",2026-06-25T20:44:04Z,2026-06-27T17:12:14Z,11,"Add sm_121a (NVIDIA GB10/DGX Spark) to default CUDA arch list Yes, I am an AI agent submitting this PR on behalf of a user who discovered this issue while training models on an NVIDIA GB10. The GB10 (sm_121 / compute capability 12.1a) is missing from the default CUDA arch list in select_compute_arch.cmake. Without i..."
181475,Add `tensor` and `from_numpy` to `torch.__all__`,https://github.com/pytorch/pytorch/pull/181475,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,131765,131765,131765,,,,0,3,0,,,"triaged, open source, Stale",2026-04-25T03:51:07Z,2026-06-27T15:53:33Z,11,Issue Fixes #131765 (one concrete instance — see this comment for the repro). Summary torch.__all__ lists Tensor (the class) but not tensor (the factory) or from_numpy — both are injected from torch._C at runtime. pyright 1.1.409 flags both as reportPrivateImportUsage where 1.1.408 did not. This adds them to __all__...
174771,Fix `sample_kwargs` in `polygamma` op info,https://github.com/pytorch/pytorch/pull/174771,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, Stale, ciflow/trunk, release notes: python_frontend, topic: not user facing",2026-02-11T14:00:12Z,2026-06-27T14:08:20Z,11,"Currently, op information UnaryUfuncInfo for polygamma function with n={1,2,3,4} is generated via a comprehension. The iteration variable n_ is used for templating the variant_test_name and the sample_kwargs. Note that the latter is a lambda: sample_kwargs=lambda device, dtype, input: ({'n': n_}, {'n': n_}), The pro..."
177612,[inductor] Add pattern match for fused gate+up GEMM + SiLU on XPU,https://github.com/pytorch/pytorch/pull/177612,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,9,0,,,"module: cpu, triaged, open source, Stale, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/xpu, ciflow/inductor-pallas, ciflow/torchtitan",2026-03-17T03:42:09Z,2026-06-27T12:55:21Z,11,"Add Inductor pattern matching that detects the Llama MLP gate+up pattern: silu(mm(input, permute(gate_weight))) * mm(input, permute(up_weight)) and replaces it with torch.ops.xpu._fused_gate_up_silu(), which runs the fused sycl-tla GEMM + SiLU kernel on Intel XPU. The pattern matches the decomposed post-grad graph —..."
186881,nn/utils: skip no-op .to(first_device) in _get_total_norm,https://github.com/pytorch/pytorch/pull/186881,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,133586,,,,0,3,0,,,"triaged, open source, release notes: nn",2026-06-10T03:24:56Z,2026-06-27T10:32:51Z,11,"Towards #133586. Replaces #184072 (accidentally closed; the branch was force-pushed afterwards so it cannot be reopened, and its history had no common ancestor with main after a shallow-clone push). Summary For N gradient tensors, the final norm aggregation in _get_total_norm always runs [norm.to(first_device) for n..."
179377,Fix gaussian_nll_loss rejecting valid multi-dim broadcastable var shapes,https://github.com/pytorch/pytorch/pull/179377,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,176244,176244,176244,,,,0,1,0,,,"triaged, open source, Stale, release notes: nn",2026-04-04T10:43:23Z,2026-06-27T10:17:30Z,11,"Fixes #176244 What gaussian_nll_loss now accepts var tensors with multiple mismatched dimensions of size 1, e.g. var=(4, 3, 1, 1) for input=(4, 3, 2, 2). Previously, the broadcastability check used sum(y for x, y in zip(...) if x != y) == 1, which summed the sizes of mismatched dimensions and required the sum to be..."
187390,Widen aot_autograd annotations to match runtime values,https://github.com/pytorch/pytorch/pull/187390,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,16,0,,,"ciflow/inductor, release notes: aotdispatcher",2026-06-15T22:26:34Z,2026-06-27T09:18:37Z,11,"Stack from ghstack (oldest at bottom): -> #187390 Runtime type-checking of torch._functorch._aot_autograd (via beartype) surfaced several parameter and return annotations that were narrower than the values actually flowing through them: list[...] annotations on functions that are handed tuples, and Callable annotati..."
187692,Skip mobilenetv2_100 eager-nondeterministic accuracy check,https://github.com/pytorch/pytorch/pull/187692,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-18T19:51:24Z,2026-06-27T09:18:31Z,11,"Stack from ghstack (oldest at bottom): -> #187692 The inductor_timm training accuracy job intermittently fails mobilenetv2_100 with status eager_two_runs_differ. This comes from the eager-vs-eager baseline self-check in benchmarks/dynamo/common.py, which on CUDA compares two unmodified eager runs at tol=0 before Ind..."
185056,Raise error when class attributes shadow registered parameters or modules,https://github.com/pytorch/pytorch/pull/185056,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,78618,78618,78618,,,,0,1,0,,,"triaged, open source, release notes: nn, topic: not user facing",2026-05-24T07:45:33Z,2026-06-27T08:37:47Z,11,"Fixes #78618 When a class defines a method or property with the same name as a parameter or submodule assigned in __init__, the registration silently succeeds but the attribute becomes inaccessible: class BadModule(nn.Module): def __init__(self): super().__init__() self.attr1 = nn.ReLU() # stored in _modules def att..."
187795,Guard release_available_cached_blocks during CUDA graph capture,https://github.com/pytorch/pytorch/pull/187795,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,159594,159594,159594,,,,0,2,0,,,"triaged, open source, release notes: cuda, topic: bug fixes",2026-06-21T14:13:56Z,2026-06-27T08:06:18Z,11,"Fixes #159594. Summary Guard release_available_cached_blocks during CUDA graph capture, matching the existing release_cached_blocks retry path. release_available_cached_blocks can release cached blocks through release_block, which may call cudaFree. Calling cudaFree during CUDA graph capture is unsafe and can invali..."
181530,[caffe2] Fix HIP build: cccl_counting_iterator needs explicit type,https://github.com/pytorch/pytorch/pull/181530,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, release notes: cuda, meta-exported",2026-04-26T22:36:27Z,2026-06-27T07:33:50Z,11,Summary: The fbpkg build for rccl-tests is broke on trunk: [2026-04-26T14:30:50.594-07:00] Stderr: buck-out/v2/art/fbcode/caffe2/__fb_aten_hipify_gen_eqsb_ATen/native/hip/EmbeddingBackwardKernel.hip__/3c9526a4a98b3c9a/out/ATen/native/hip/EmbeddingBackwardKernel.hip:309:49: error: alias template 'cccl_counting_iterat...
178906,Fix batch_isend_irecv to support ProcessGroup subclasses,https://github.com/pytorch/pytorch/pull/178906,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, Stale, ciflow/trunk, release notes: distributed (c10d), ciflow/torchtitan",2026-03-31T18:27:01Z,2026-06-27T07:33:45Z,11,"Change type(group) is ProcessGroup to isinstance(group, ProcessGroup) in batch_isend_irecv. The exact type check prevents Python-subclass backends from entering the coalescing path, falling through to individual send/recv calls. The rest of distributed_c10d.py already uses isinstance consistently for ProcessGroup ch..."
181615,[ROCm] CI: skip operator matmul microbenchmark,https://github.com/pytorch/pytorch/pull/181615,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/op-benchmark,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, Stale, ciflow/op-benchmark",2026-04-27T17:44:59Z,2026-06-27T05:48:57Z,11,"Skip pt.matmul_test in test_operator_microbenchmark() when BUILD_ENVIRONMENT matches ROCm, so CUDA/CPU coverage is unchanged. This temporary fix avoids ROCm CI failures (e.g. segfaults) while keeping the rest of the microbenchmark suite. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @h..."
187316,Support save_on_cpu offloading under first-order torch.func transforms,https://github.com/pytorch/pytorch/pull/187316,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,122729,122729,"122729, 141845, 146036",,,,0,3,0,,,"triaged, open source, release notes: torch.func, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-15T08:13:28Z,2026-06-27T05:24:16Z,11,"Issue Fixes #122729 (also #146036, #141845) Summary torch.func (grad/vjp/jacrev/vmap-over-them) disabled saved-tensor hooks for all transforms via a 2022 blanket guard, so save_on_cpu raised under any of them. This scopes the guard to higher-order only: a single first-order transform now allows the hooks; nested dif..."
181664,[FSDP2] Design notes: options for RS/AR back-edge in HSDP,https://github.com/pytorch/pytorch/pull/181664,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,179443,,,0,1,0,,,"Stale, topic: not user facing",2026-04-27T22:21:24Z,2026-06-27T02:21:55Z,12,"Stack from ghstack (oldest at bottom): #181667 -> #181664 #179443 Catalogs six narrow options (record_stream, alloc on AR stream, dedicated MemPool, post-reduce on RS stream, CPU sync, AR stream priority) plus an Alternative rotate-2 design and a parameterized rotate-k design with a user-tunable AR/RS lookahead dept..."
188161,Demote changing CUDAGraph-managed inputs to copied inputs,https://github.com/pytorch/pytorch/pull/188161,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-25T16:45:16Z,2026-06-27T02:16:25Z,12,"Stack from ghstack (oldest at bottom): -> #188161 #188078 CUDAGraph tree child nodes specialize cudagraph-managed inputs by graph-pool data pointer. When a parent produces several interchangeable outputs and a later child consumes different output slots, the child keeps re-recording even though the actual tensor val..."
188103,Enable recursive dict-tag guard optimization by default on 3.12+ non-free-threaded,https://github.com/pytorch/pytorch/pull/188103,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,185886,188101,,,0,5,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-24T21:29:24Z,2026-06-27T02:15:46Z,12,"Stack from ghstack (oldest at bottom): -> #188103 #188101 #187868 use_recursive_dict_tags_for_guards short-circuits per-module guard subtrees via a dict tag instead of a full recursive check, and is the dominant lever for guard-eval cost on models with many repeated nn.Modules (#185886 -- ~2.7x on an 82-layer stack)..."
188101,Require an installed dict watcher for the recursive dict-tag fast path on 3.12+,https://github.com/pytorch/pytorch/pull/188101,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,188103,,,0,5,0,,,"module: dynamo, ciflow/inductor",2026-06-24T21:04:10Z,2026-06-27T02:15:44Z,12,Stack from ghstack (oldest at bottom): #188103 -> #188101 #187868 The recursive dict-tag fast path (use_recursive_dict_tags_for_guards) validates an entire tag-safe-root subtree with a cheap tag check instead of a full traversal. Its safety hinges on detecting when a recorded interior dict is mutated or freed. On 3....
181414,Update FBGEMM submodule to v1.6.0,https://github.com/pytorch/pytorch/pull/181414,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"open source, Stale, topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-04-24T18:20:56Z,2026-06-27T01:21:39Z,12,"Bumps third_party/fbgemm to v1.6.0 (d08742c6) to pick up FBGEMM's bump of its own nlohmann/json submodule, which removes the unused nltk entry from docs/mkdocs/requirements.txt that was triggering Wiz CVE-2026-0846 alerts on PyTorch CI runners."
186271,[CUDA] Add TunableOp support for scaled GEMM,https://github.com/pytorch/pytorch/pull/186271,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, release notes: linalg_frontend, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-06-04T21:23:51Z,2026-06-27T01:21:00Z,12,"This PR extends CUDA TunableOp support to scaled GEMM, and is intended to be stacked on top of #186266 and #186270. cc @eqy Authored with Codex and Claude"
188252,Fix ROCm round-robin counter reset in initSingleStream,https://github.com/pytorch/pytorch/pull/188252,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, triaged, open source, topic: not user facing",2026-06-26T07:28:45Z,2026-06-27T01:02:30Z,12,"Lazily creating a stream reset priority_counters to 0, so later getStreamFromPool calls re-issued low indices and returned the same hipStream_t, defeating the pool's round-robin guarantee. The reset is a no-op on CUDA (eager init finishes before get_idx) and unrelated to the trace call above it; it only bites under..."
179625,Enable ruff rule PLW1510 codebase wide,https://github.com/pytorch/pytorch/pull/179625,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,115016,115016,115016,,,,0,2,0,,,"open source, Stale, release notes: releng, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-04-07T20:22:36Z,2026-06-27T01:01:55Z,12,Summary Fixes #115016 Enables the ruff PLW1510 lint rule by adding explicit check=False to all 89 subprocess.run() calls that were missing the check argument. This preserves existing behavior while making the intent explicit at each call site. Test plan ruff check --select PLW1510 . passes with zero violations cc @v...
182271,[ci][WIP] Cut redundant before-test fanout + trim over-eager periodic crons,https://github.com/pytorch/pytorch/pull/182271,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, release notes: releng, ciflow/mps",2026-05-04T07:05:04Z,2026-06-27T01:01:49Z,12,DRAFT / WIP — Do not merge Part of the CI-cost reduction roadmap (see internal investigation doc). Estimated savings: $90-180K/mo combined. Three changes 1. Inner get-label-type removed from llm_td_retrieval.yml and target_determination.yml. Every parent workflow already runs runner-determinator at the workflow_run...
177038,Add cublaslt backend for scaled grouped GEMM,https://github.com/pytorch/pytorch/pull/177038,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,DIRTY,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,28,0,,,"open source, Stale, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/dtensor",2026-03-10T16:24:14Z,2026-06-26T23:50:32Z,12,"This builds off of #177037, adding support for fp8 scaled grouped GEMM. Starting with tensorwise scaling only, but MXFP8 blockwise scaling is now supported too and should be added as well. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @..."
188298,[aot_autograd] Add PySourceBuilder for runtime-wrapper codegen,https://github.com/pytorch/pytorch/pull/188298,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"topic: not user facing, ciflow/inductor",2026-06-26T20:45:02Z,2026-06-26T23:49:39Z,12,"Stack from ghstack (oldest at bottom): -> #188298 The runtime-wrapper and subclass codegen paths in _aot_autograd build Python source by appending strings with hardcoded indentation prefixes to lists, then compile/exec them. This introduces a small PySourceBuilder helper (in a new codegen_utils.py) that manages inde..."
180990,[pytorch] Fix symmetric quantization scale to match C++ ChooseQuantizationParams (#180990),https://github.com/pytorch/pytorch/pull/180990,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Merged, Reverted, Stale, ciflow/trunk, release notes: quantization, ci-no-td, meta-exported",2026-04-21T15:34:21Z,2026-06-26T22:53:26Z,12,"Summary: Fix inconsistencies between Python _calculate_qparams in observer.py and C++ ChooseQuantizationParams in caffe2/aten/src/ATen/native/quantized/cpu/QuantUtils.h for symmetric quantization. Two changes: Scale formula: Replace scale = max(|min|, |max|) / 127.5 (which treats the quantization range as perfectly..."
188289,[mtia] Make torch.distributions test suite portable across acelerators,https://github.com/pytorch/pytorch/pull/188289,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,meta-exported,2026-06-26T19:17:34Z,2026-06-26T22:23:50Z,12,"Summary: Change torch.distributions test_constraints test suite to work on non-cuda devices by leveraging the torch.accelerator APIs instead of torch.cuda. Test Plan: This is a test suite only change, so if the existing passing tests keep passing, all is good. Differential Revision: D109873279"
180573,Avoid unnecessary refcount bump in inplace op Python bindings,https://github.com/pytorch/pytorch/pull/180573,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, release notes: python_frontend",2026-04-16T16:13:55Z,2026-06-26T22:01:48Z,12,"On free-threaded Python, bumping an intrusive_ptr's refcount from 1 to 2 triggers incref_pyobject, which acquires the GIL. In multi-threaded DDP workloads this is a significant source of contention. A few hot paths were performing unnecessary refcount bumps or holding the GIL across blocking I/O; this addresses them..."
181030,[MPS] Migrate SDPA from MPSGraph to Metal,https://github.com/pytorch/pytorch/pull/181030,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,1,0,,,"open source, Stale, release notes: mps, ciflow/mps",2026-04-21T22:11:19Z,2026-06-26T21:53:45Z,12,Stack from ghstack (oldest at bottom): -> #181030 Also fixes input broadcasting behavior. Claude was used to co-author the mm_simdgroup and softmax_rows functions of the Metal kernel. Performance was measured using the following script: Click to expand import torch import torch.utils.benchmark as benchmark import it...
183743,Add drift-based tolerance fallback to op tests,https://github.com/pytorch/pytorch/pull/183743,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,183621,183621,183621,,,,0,2,0,,,"triaged, open source, topic: not user facing, ciflow/mps",2026-05-14T17:41:37Z,2026-06-26T21:51:47Z,12,Fixes #183621 Used for low-precision consistency tests for MPS backend ops where accumulation order can cause discrepancies. Uses the CPU discrepancy between fp32 and the target half type to produce an estimate for the tolerable drift and checks that the MPS implementation produces an output that is within this enve...
185177,Fix DTensor nll_loss weighted mean bug (closes #185167),https://github.com/pytorch/pytorch/pull/185177,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185167,185167,185167,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-26T00:18:57Z,2026-06-26T20:50:55Z,12,"Fixes #185167 When input is batch-sharded (Shard(0)) and weight is non-None with reduction='mean', the per-rank denominator sum(w[t[n]]) varies across ranks. Averaging per-rank means does not recover the global weighted mean (global_loss_sum / global_weight_sum). Fix: add an elif branch in nll_loss_forward_strategy..."
172682,[NOT FOR MERGE] Initial test of features disabled when doing cooperative reduction,https://github.com/pytorch/pytorch/pull/172682,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,35,0,,,"open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-01-16T23:02:11Z,2026-06-26T20:50:46Z,12,Stack from ghstack (oldest at bottom): -> #172682 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @mlazos
188099,[JIT] Use native fill/zero ops for outplace tracing of fill_/zero_,https://github.com/pytorch/pytorch/pull/188099,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,"122898, 134895","122898, 134895",,,,,0,1,0,,,"triaged, open source, release notes: jit",2026-06-24T21:02:53Z,2026-06-26T20:26:35Z,12,"Previously, JIT tracing with _force_outplace=True and the RemoveMutation pass replaced aten::fill_ and aten::zero_ with aten::full_like and aten::zeros_like. This leads to problems since aten::fill_ supports both Scalar and Tensor fill values while aten::full_like only accepts a Scalar fill value. I noticed this whe..."
181614,[TEST] Fix `_debug_handlers` import error in `test_stacks_handler_all_success`,https://github.com/pytorch/pytorch/pull/181614,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/torchtitan,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"module: tests, open source, Stale, topic: not user facing, ciflow/torchtitan",2026-04-27T17:34:15Z,2026-06-26T20:10:01Z,12,Seeing AttributeError: module 'torch.distributed.debug' has no attribute '_debug_handlers' otherwise cc @mruberry
176317,[MPS] Leveraging unified memory memcpy instead of blit encoding for tensor device transfers,https://github.com/pytorch/pytorch/pull/176317,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, release notes: mps, ciflow/mps",2026-03-03T17:59:38Z,2026-06-26T20:09:48Z,12,Optimizing copy_from_mps_ and copy_to_mps_() with the following: Avoid backing buffer MPS <--> CPU blit Instead perform a Unified Mem <--> Unified Mem memcpy Adding appropriate sync points: CPU --> MPS - Synchronizing before to ensure previous GPU writes to dst_ are visible - Synchronizing after to ensure write to d...
188258,"Fix _foreach_{add,mul,div}.Tensor autograd",https://github.com/pytorch/pytorch/pull/188258,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,144580,144580,144580,,,,0,1,0,,,"triaged, open source, release notes: foreach_frontend",2026-06-26T08:52:17Z,2026-06-26T19:29:57Z,12,"Issue Fixes #144580 Summary backward through _foreach_{add,mul,div}.Tensor crashed with inconsistent range for TensorList output for lists longer than one element. The foreach codegen always emitted an N-length gradient list and copy_ranged it into the arg's index range which is right for the TensorList self (range..."
188263,(fix) NNPACK PeachPy air gapped build,https://github.com/pytorch/pytorch/pull/188263,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source",2026-06-26T09:33:53Z,2026-06-26T19:26:34Z,12,"Peachpy (a dependency of NNPACK, a dependency of Pytorch) needs a PYTHON_SIX_SOURCE_DIR to avoid fetching python-six from the web. This PR makes the build environment self contained for building pytorch"
186468,Fix incorrect LR curves in Chained/Sequential scheduler doc images,https://github.com/pytorch/pytorch/pull/186468,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, release notes: optim",2026-06-06T14:21:02Z,2026-06-26T19:21:51Z,12,"Summary The example learning-rate curve plots for the ChainedScheduler and SequentialLR schedulers in the documentation are inconsistent with the examples. As a result, the lr values listed in the ChainedScheduler docstring example were also wrong (this is not the case for SequentialLR). The cause is in docs/source/..."
188215,[dtensor] Resolve ProcessGroup from mesh in-graph under compile_on_one_rank,https://github.com/pytorch/pytorch/pull/188215,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"release notes: fx, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-06-26T00:45:51Z,2026-06-26T19:03:04Z,12,"Stack from ghstack (oldest at bottom): -> #188215 Under compile-on-one-rank (CooR) a model is traced once with make_fx into a rank-agnostic, serializable graph. Legacy in-place c10d collectives (e.g. dist.all_reduce, which lowers to c10d.allreduce_) take the ProcessGroup as a torchbind ScriptObject argument, and mak..."
188093,Fix record_stack_traces: filter fx-internal frames from recorded trac…,https://github.com/pytorch/pytorch/pull/188093,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,130861,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-24T19:43:37Z,2026-06-26T18:32:23Z,12,…es (#130861)
188149,Handle mixed Tensor/Scalar min/max bounds in torch.clamp,https://github.com/pytorch/pytorch/pull/188149,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,188088,188088,188088,,,,0,1,0,,,"triaged, open source",2026-06-25T10:37:12Z,2026-06-26T18:24:30Z,12,"Fixes #188088. The C++ clamp overloads require both min and max to be either both Scalars or both Tensors. Passing mixed types (e.g., min=Tensor, max=float) raises a TypeError with no overload matching, even though the operation has a natural elementwise meaning. Add a Python wrapper in torch/__init__.py that detect..."
188257,[RFC] S3-over-RDMA checkpoint storage backends via NVIDIA cuObject,https://github.com/pytorch/pytorch/pull/188257,draft/noise,13,PR is draft; CI lane labels: ciflow/torchtitan,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,8,0,,,"open source, release notes: distributed (checkpoint), ciflow/torchtitan",2026-06-26T08:34:48Z,2026-06-26T18:17:07Z,12,[RFC] S3-over-RDMA checkpoint storage backends via NVIDIA cuObject Draft / RFC. Opening for design feedback from the DCP (oncall: distributed) owners before polishing for merge. Authored with assistance from an AI coding assistant. Summary Add optional S3-over-RDMA storage backends for torch.distributed.checkpoint (...
188180,Reuse full-range splits for persistent CSE,https://github.com/pytorch/pytorch/pull/188180,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-25T19:34:15Z,2026-06-26T18:09:09Z,12,"Stack from ghstack (oldest at bottom): (to be filled) In the persistent reduction body, the pointwise domain is split as set_ranges([1024, 96], [128]), so the input loads use batch/group coordinates like r0_2 + 128*x0 + 24576*x1. The epilogue later asks for set_ranges([1024 * 96], [128]); without remembering the ear..."
187981,[not for land] CODA backward epilogue fusion prototype,https://github.com/pytorch/pytorch/pull/187981,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,,2026-06-23T21:01:36Z,2026-06-26T18:08:13Z,12,"Not for land — RFC / prototype. A UX prototype for CODA-style backward epilogue fusion across autograd.Function nodes (see ""CODA: Rewriting Transformer Blocks as GEMM-Epilogue Programs"", arXiv:2605.19269). Problem The backward of an autograd.Function isn't self-contained: for a chain mm -> epilogue -> mm, the epilog..."
187875,[AOTI] Refactor model container checkout ownership (#187875),https://github.com/pytorch/pytorch/pull/187875,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, ciflow/inductor, release notes: inductor (aoti), meta-exported",2026-06-22T21:06:17Z,2026-06-26T17:30:50Z,12,"Summary: Wrap checked-out AOTI models in a ModelLease RAII handle that auto-returns the model to the available pool on destruction, replacing the duplicated try/catch cleanup on exception paths; successful runs call mark_pending() instead. Also wake pool waiters when a model is returned to the available pool (fixing..."
188138,fix: Docs are little bit outdated for torch logs,https://github.com/pytorch/pytorch/pull/188138,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,137285,137285,137285,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-25T06:43:27Z,2026-06-26T17:22:45Z,12,"Summary The issue reports that the torch logs docs are outdated. The user-facing reference for torch logs is the torch._logging.set_logs API doc, which is generated from the function docstring in torch/_logging/_internal.py and is linked from the torch logs tutorial. Root cause The issue reports that the torch logs..."
188063,Enable Blackwell FP8 grouped GEMM,https://github.com/pytorch/pytorch/pull/188063,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-06-24T13:42:40Z,2026-06-26T16:56:20Z,12,"Stack from ghstack (oldest at bottom): -> #188063 FP8 grouped GEMM was disabled on SM100 because the grouped CUTLASS path was only instantiated for SM90. Removing the runtime guard alone was not enough: the SM90 schedule does not map to Blackwell's 1SM/2SM grouped-kernel schedules, and the file was only compiled for..."
184892,xpu: use stride-based conv descriptors and preserve grouped layout mapping,https://github.com/pytorch/pytorch/pull/184892,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,8,0,,,"module: cpu, open source, ciflow/xpu",2026-05-22T14:18:55Z,2026-06-26T16:49:16Z,12,"Update conv_get_md() in Conv.cpp to build oneDNN memory descriptors from actual tensor sizes/strides for src, dst, and weight, instead of relying on format tags. For grouped convolution, compute grouped weight strides for oneDNN’s grouped layout [G, OC_per_group, IC_per_group, ...] from PyTorch’s weight storage so d..."
188125,[AI Codemod][PerfAICT-General] perf: Special case non-symbolic set_sizes_contiguous,https://github.com/pytorch/pytorch/pull/188125,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,meta-exported,2026-06-25T02:13:26Z,2026-06-26T16:45:01Z,12,"Summary: Based on initial AI based proposal (as this shows up hot enough in internal profiles), but I did multiple assisted iterations on this myself now. Optimized c10::TensorImpl::set_sizes_contiguous by adding an inlined fast-path for the common non-symbolic case that eliminates redundant work: Eliminated compute..."
181285,"[TEST][ARM] Remove Pointer Authentication Code from address, fall back to `dladdr_lookup` where possible",https://github.com/pytorch/pytorch/pull/181285,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"module: tests, triaged, open source, module: arm, Stale, topic: not user facing",2026-04-23T20:46:51Z,2026-06-26T16:17:38Z,12,"authored with claude Mainly to fix failures on ARM64 systems where we were not getting stack traces e.g., python test/test_cuda.py TestCudaAllocator.test_cpp_memory_snapshot_pickle cc @mruberry @snadampal @milpuz01 @aditew01 @nikhil-arm @fadara01 @robert-hardwick @nWEIdia"
188270,Fix inplace modification error with softmax when combining forward and backward AD,https://github.com/pytorch/pytorch/pull/188270,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,162350,162350,162350,,,,0,1,0,,,"triaged, open source, release notes: autograd",2026-06-26T13:29:23Z,2026-06-26T15:54:53Z,12,"Fixes #162350 Summary logsumexp_jvp did an inplace *= on the output of .exp(), which bumped the version counter. When backward AD was also active, ExpBackward0 needed that tensor at version 0 and crashed. Added || self_p_exp.requires_grad() to the existing branch condition so the non-inplace path is taken when the t..."
182601,docs: clarify tile dims argument,https://github.com/pytorch/pytorch/pull/182601,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,129319,129319,129319,,,,0,1,0,,,"triaged, open source, release notes: python_frontend, topic: docs",2026-05-06T01:58:29Z,2026-06-26T15:38:32Z,12,"Fixes #129319 Summary Document torch.tile dims as accepting torch.Size, an int, a tuple of ints, or a list of ints. Expand the Tensor.tile docstring beyond the bare cross-reference so the tensor method makes the accepted dims forms clear. Add an integer Tensor.tile example alongside the existing tuple example. Test..."
182602,docs: cross-reference sigmoid APIs,https://github.com/pytorch/pytorch/pull/182602,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,105318,105318,105318,,,,0,1,0,,,"triaged, open source, release notes: nn, topic: docs",2026-05-06T01:58:33Z,2026-06-26T15:38:24Z,12,"Fixes #105318 Summary Make torch.nn.Sigmoid the central docstring that names the equivalent sigmoid APIs. Link torch.sigmoid, Tensor.sigmoid, torch.nn.functional.sigmoid, and torch.special.expit back to the module/formula documentation. Keep the wording scoped to documentation cross-references only. Test Plan python..."
182603,docs: clarify RNN hidden state dimensions,https://github.com/pytorch/pytorch/pull/182603,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,99421,99421,99421,,,,0,1,0,,,"triaged, open source, release notes: nn, topic: docs",2026-05-06T01:58:38Z,2026-06-26T15:38:15Z,12,"Fixes #99421 Summary Change RNN, LSTM, and GRU hidden-state shape descriptions to use num_layers * D order. Add examples showing how to split h_n by layer and direction. Add the analogous c_n split example for LSTM cell states. Test Plan python3 -m py_compile torch/nn/modules/rnn.py"
182604,docs: clarify multithreaded autograd guidance,https://github.com/pytorch/pytorch/pull/182604,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,67887,67887,67887,,,,0,1,0,,,"triaged, open source, release notes: autograd, topic: docs, topic: not user facing",2026-05-06T01:58:43Z,2026-06-26T15:38:03Z,12,Fixes #67887 Summary Add explicit multithreaded autograd guidance based on the issue discussion. Clarify that concurrent reads are safe while in-place writes to shared tensors require synchronization. Clarify why backward() can be nondeterministic through shared .grad accumulation and why torch.autograd.grad is the...
185398,docs: clean up multilabel soft margin loss args,https://github.com/pytorch/pytorch/pull/185398,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,67969,67969,67969,,,,0,1,0,,,"triaged, open source, release notes: nn, topic: docs, topic: not user facing",2026-05-27T16:45:34Z,2026-06-26T15:37:51Z,12,Fixes #67969 Summary: Stop listing the deprecated size_average and reduce arguments in torch.nn.functional.multilabel_soft_margin_loss docs. Document the existing weight argument and fix the return description typo. Test Plan: python3 -m py_compile torch/nn/functional.py git diff --check
185744,docs: clarify LSTM equation notation,https://github.com/pytorch/pytorch/pull/185744,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,131280,131280,131280,,,,0,1,0,,,"triaged, open source, release notes: nn, topic: docs",2026-05-31T04:36:39Z,2026-06-26T15:37:39Z,12,Fixes #131280 Summary: Clarifies that the LSTM gate equations use single-sample column-vector notation. Points readers to the documented Attributes section for the batched weight shapes used by the module. Test Plan: python3 -m py_compile torch/nn/modules/rnn.py git diff --check
185745,docs: clarify scalar fill values,https://github.com/pytorch/pytorch/pull/185745,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,129970,129970,129970,,,,0,1,0,,,"triaged, open source, release notes: python_frontend, topic: docs",2026-05-31T04:37:39Z,2026-06-26T15:37:31Z,12,"Fixes #129970 Summary: Documents that fill_value for torch.full, torch.full_like, and Tensor.new_full is a scalar value. Clarifies that accepted scalar values include Python numbers and zero-dimensional tensors. Test Plan: python3 -m py_compile torch/_torch_docs.py torch/_tensor_docs.py git diff --check"
182948,[CUDA] Fix reentrant deadlock in torch.cuda._lazy_call,https://github.com/pytorch/pytorch/pull/182948,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-rocm, ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, ciflow/inductor, ciflow/inductor-rocm, ciflow/rocm-mi300",2026-05-08T13:35:31Z,2026-06-26T15:20:31Z,12,"Updated _lazy_call to avoid running callbacks while _initialization_lock is held. This happens on ROCm stack while running some of the bigger models which completely deadlock the execution in torch.compile mode. New flow: fast path: if initialized, run callback immediately and return otherwise lock only for queue/in..."
182605,test: strengthen optimizer step hook ordering checks,https://github.com/pytorch/pytorch/pull/182605,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,119300,119300,119300,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-05-06T01:58:47Z,2026-06-26T14:59:02Z,12,Fixes #119300 Summary Strengthen local and global optimizer step pre/post hook tests so they assert hook order around a mocked public step implementation. Avoid relying on the private _optimizer_step_code profiler hook in the tests. Extend the combined global/local hook test to verify pre hooks run before the step b...
178751,Delay common_utils import to pytest_sessionstart to prevent faulthandler race condition,https://github.com/pytorch/pytorch/pull/178751,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"open source, Stale, ciflow/trunk, topic: not user facing, ciflow/xpu",2026-03-30T08:01:39Z,2026-06-26T14:18:26Z,12,"The Problem Currently, test/conftest.py imports torch.testing._internal.common_utils inside the pytest_configure hook. This import has the side effect of triggering device initialization (via module-level checks like torch.xpu.is_available()). During this device initialization, custom runtimes, GPU simulators, or pr..."
188253,Add XPU dispatch for linalg.polar,https://github.com/pytorch/pytorch/pull/188253,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,4162,4162,,,,,0,1,0,,,"open source, ciflow/xpu, release notes: xpu",2026-06-26T07:56:34Z,2026-06-26T12:36:49Z,12,Fixes intel/torch-xpu-ops#4162 Add XPU dispatch for newly added linalg_polar op. There are already 40 test cases on XPU backend that will be unskipped once this PR is merged
188045,[MPS] better memory allocator,https://github.com/pytorch/pytorch/pull/188045,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,"105839, 132596, 152550",,,,0,1,0,,,"triaged, open source, topic: improvements, release notes: mps, ciflow/mps",2026-06-24T09:13:01Z,2026-06-26T12:09:36Z,12,I've always believed that the memory issues we encountered either during decode or during training with variable sequences such as these issues: #132596 #152550 #105839 huggingface/transformers#33717 were caused due to MPS Graph caching of the graphs with different shapes. Thus when we used different shapes for trai...
183316,Fix/signbit cuda half nan,https://github.com/pytorch/pytorch/pull/183316,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,181806,181806,181806,,,,0,2,0,,,"triaged, open source, release notes: cuda",2026-05-11T14:48:03Z,2026-06-26T09:13:26Z,12,"Summary Fixes #181806 torch.signbit on CUDA was returning incorrect results for negative float16 NaN values. CPU correctly returns True, but CUDA was returning False. The issue was caused by signbit_kernel_cuda promoting Half and BFloat16 values to float before calling signbit(). During this conversion, CUDA canonic..."
180561,[optim] Support ndim >= 2 in Muon and add param_groups_for_muon helper,https://github.com/pytorch/pytorch/pull/180561,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,DIRTY,False,0,0,,,166209,166209,,,,,0,4,0,,,"triaged, open source, release notes: optim",2026-04-16T14:23:45Z,2026-06-26T09:09:03Z,12,"Summary Fixes pytorch/pytorch#166209 (part 2 of @janeyx99's roadmap in pytorch/pytorch#148819 (comment)). Motivation The initial Muon landing hard-rejected any parameter with ndim != 2, pushing the responsibility to users to manually filter out Conv weights, embeddings, biases, and norm scales. This is awkward for t..."
188036,Fix THPStream_richcompare reading off the end of non-Stream objects,https://github.com/pytorch/pytorch/pull/188036,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,188033,188033,188033,,,,0,1,0,,,"triaged, open source",2026-06-24T07:23:44Z,2026-06-26T08:44:36Z,12,"Fixes #188033. THPStream_richcompare casts its other argument straight to THPStream* and reads stream_id / device_index / device_type off it for any comparison except against None: if (Py_IsNone(other)) { result = Py_False; } else { switch (op) { case Py_EQ: result = THPStream_eq( reinterpret_cast<THPStream*>(self),..."
181546,Fix AssertionError: failed to set ranges (#181546),https://github.com/pytorch/pytorch/pull/181546,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,181546,,,,,0,1,0,,,"fb-exported, Stale, topic: not user facing, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-27T03:49:30Z,2026-06-26T07:48:29Z,12,Summary: Pull Request resolved: #181546 Test Plan: Fix AssertionError: failed to set ranges. A patch to D96955901 Error Example https://www.internalfb.com/intern/everpaste/?phabricator_paste_number=2266216056&handle=GJNU2Cef-HxnBvoFAOeG-aT6RflqbsIXAAAz Differential Revision: D102551863 cc @voznesenskym @penguinwu @E...
182937,Align XPU unit tests,https://github.com/pytorch/pytorch/pull/182937,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/xpu,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,2779,2779,,,,,0,9,0,,,"open source, topic: not user facing, ciflow/xpu",2026-05-08T08:30:45Z,2026-06-26T06:58:48Z,12,This PR groups XPU-related unit test skips and tolerance adjustments. Fixes: intel/torch-xpu-ops#1963 intel/torch-xpu-ops#1973 intel/torch-xpu-ops#2182 intel/torch-xpu-ops#2779 intel/torch-xpu-ops#2921
187606,"[inductor] Per-arch ROCm channels_last layout-opt multipliers (gfx942, gfx950)",https://github.com/pytorch/pytorch/pull/187606,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-rocm-mi300, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,187600,,,0,4,0,,,"module: rocm, open source, ciflow/trunk, module: inductor, ciflow/inductor, keep-going, ciflow/rocm-mi300, ciflow/inductor-perf-test-nightly-rocm-mi300, ciflow/inductor-rocm-mi300, ciflow/torchtitan",2026-06-17T21:03:29Z,2026-06-26T06:23:55Z,12,"decide_layout_opt's inference channels_last cost-multipliers (<1 => favor channels_last) were NVIDIA-measured. ROCm CDNA archs behave differently under MIOpen, so dispatch them per-arch via the device's gcnArchName: gfx942 (MI300): channels_last regresses small/grouped convs. SMALL 0.783→1.25, GROUPED 1.358→1.05 (tu..."
180758,Remove dead TORCH_HIP_VERSION guards,https://github.com/pytorch/pytorch/pull/180758,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, triaged, open source, topic: not user facing",2026-04-19T03:59:32Z,2026-06-26T06:09:24Z,12,cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
186243,"[dynamo, nested graph breaks] enable NGB on dynamo-wrapped tests",https://github.com/pytorch/pytorch/pull/186243,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"163503, 171826, 185524, 186100",,,0,21,0,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, keep-going, ci-no-td, ciflow/dynamo, ciflow/torchtitan",2026-06-04T17:23:42Z,2026-06-26T03:29:55Z,13,Stack from ghstack (oldest at bottom): #171826 #163503 #185524 #186100 #186657 #187088 #187005 -> #186243 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
181398,[spmd_types] local_map,https://github.com/pytorch/pytorch/pull/181398,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,43,0,,,"Merged, Reverted, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ci-no-td, ciflow/torchtitan, ciflow/dtensor",2026-04-24T16:57:24Z,2026-06-26T03:29:20Z,13,"Adds a local_map(spmd_types: bool = False) kwarg, that when True, runs the local_map region with spmd_types type checking. For Replicate -> R/I, infers from grad_in_placements, and if not available, defaults to R. Adds checking of output spmd_types, grad_out spmd_types with a grad hook. Stack from ghstack (oldest at..."
185077,Avoid folding mutated buffer storages,https://github.com/pytorch/pytorch/pull/185077,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,165444,165444,,,,0,4,0,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan, agentic",2026-05-24T18:15:16Z,2026-06-26T01:27:26Z,13,"Stack from ghstack (oldest at bottom): -> #185077 AOTInductor packages unlift exported module buffers into runtime-owned constants. For buffer mutation, export re-inserts a copy_ from the functional mutation output back into that buffer. The uniform-value constant folder could fold a singleton get_attr buffer into a..."
188188,[DTensor] local_map normalizes shard dim checking in_placements,https://github.com/pytorch/pytorch/pull/188188,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-25T20:40:37Z,2026-06-25T23:48:01Z,13,"titan placements will see Shard(-1) != Shard(2) for 3d tensor, just convenience"
181528,sdp: warn when enable_gqa=True silently falls back to MATH,https://github.com/pytorch/pytorch/pull/181528,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,154363,,,,0,1,0,,,"open source, Stale",2026-04-26T22:24:35Z,2026-06-25T22:59:58Z,13,"Motivation Closes part of #154363. When a user passes enable_gqa=True to F.scaled_dot_product_attention but no fused backend supports their configuration, the dispatcher silently falls back to the MATH backend. MATH is correct but unfused and often 2x+ slower than pre-expanding K/V with repeat_interleave and using a..."
186926,Remove sleef from test_aoti_abi_check linkage,https://github.com/pytorch/pytorch/pull/186926,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing, ciflow/inductor, no-stale, module: aotinductor",2026-06-10T11:15:12Z,2026-06-25T22:47:19Z,13,Actually it does not depend on sleef functions. This fix linking issues without system sleef cc @desertfire @penguinwu @yushangdi @jataylo @iupaikov-amd
188178,Guard score_mod oob indexing,https://github.com/pytorch/pytorch/pull/188178,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,150321,150321,150321,,,,0,1,0,,,"module: inductor, ciflow/inductor, module: flex attention, ciflow/torchtitan",2026-06-25T19:32:22Z,2026-06-25T22:46:40Z,13,"Human Note I will fill in Agent Report Report: #150321 Summary FlexAttention generated Triton code could read out of bounds from tensors captured by score_mod before mask_mod was applied. When the captured buffer required grad, backward could also emit out-of-bounds atomic stores into the captured-buffer grad. The i..."
186947,[functorch] Add native_group_norm_backward.multiple_grads plumbing,https://github.com/pytorch/pytorch/pull/186947,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"183980, 186414, 186946",,,0,17,0,,,"open source, ciflow/trunk, release notes: torch.func, ciflow/inductor, needs-review",2026-06-10T16:52:13Z,2026-06-25T22:08:41Z,13,"Stack from ghstack (oldest at bottom): -> #186947 #186946 #186414 #183980 Note to reviewers: this is a conditional approval, as I will not/cannot merge this until #183980 is reviewed and merged."
186145,"Spelling fix: Change ""structeur"" to ""structure"" in Lumen CLI VLLM test",https://github.com/pytorch/pytorch/pull/186145,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-06-03T22:41:06Z,2026-06-25T22:07:38Z,13,Fixed typo in comments in Lumen.
188179,Use tiling-aware hints for persistent reductions,https://github.com/pytorch/pytorch/pull/188179,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-25T19:34:09Z,2026-06-25T22:05:36Z,13,"Stack from ghstack (oldest at bottom): (to be filled) Tiling scores were already threaded into Triton kernel construction and the generated reduction_hint. The missing piece was earlier: persistent-reduction selection called get_reduction_hint() without the selected tiling scores, so a coalesced inner reduction coul..."
188184,Update definition of ROCM_VERSION in cpp_extension.py,https://github.com/pytorch/pytorch/pull/188184,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source",2026-06-25T20:21:53Z,2026-06-25T22:04:23Z,13,"Follow-up to #168097, to ensure extensions can see/use the semantically-correct version. Should have really been done in conjunction with that PR? Useful in cases where a PyTorch extension wants to compare the version of ROCm it was compiled against VS the version of ROCm installed, similar to: https://github.com/py..."
179592,Fix MPS SDPA correctness when score matrix exceeds 2^32 elements,https://github.com/pytorch/pytorch/pull/179592,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,179352,179352,,,,,0,3,0,,,"triaged, open source, Stale, release notes: mps, ciflow/mps",2026-04-07T15:56:16Z,2026-06-25T22:02:58Z,13,"Summary Fixes #179352. Apple's MPSGraph corrupts SDPA outputs when the score matrix B*H*Nq*Nkv exceeds 2^32 total elements: at the canary shape (B=1, H=8, Nq=16384, Nkv=65536, D=64), output diverges from CPU with cosine similarity ~0.44 and max relative error ~1.2. The threshold is independent of dtype, of is_causal..."
185015,Add SAC-driven fully overlapped activation CPU offload (primarily aimed at GH200/GB200/GB300 superchips),https://github.com/pytorch/pytorch/pull/185015,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,"158657, 170284, 174960",180874,,,0,5,0,,,"triaged, open source, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-23T22:07:24Z,2026-06-25T21:56:20Z,13,"SAC-Driven Activation CPU Offload Summary This PR adds a torch.compile path for activation CPU offload driven by Selective Activation Checkpointing policies: CheckpointPolicy.MUST_CPU_OFFLOAD CheckpointPolicy.PREFER_CPU_OFFLOAD The intent is that users describe what should be offloaded through SAC policy, and the co..."
186946,[mps] Add kernel for native_group_norm_backward.multiple_grads,https://github.com/pytorch/pytorch/pull/186946,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"183980, 186414, 186947",,,0,17,0,,,"open source, ciflow/trunk, release notes: mps, ciflow/mps",2026-06-10T16:52:07Z,2026-06-25T21:30:26Z,13,Stack from ghstack (oldest at bottom): #186947 -> #186946 #186414 #183980
188189,Fix split_group non-member handling regression for mixed backends,https://github.com/pytorch/pytorch/pull/188189,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, release notes: distributed (c10d), ciflow/torchtitan",2026-06-25T20:42:00Z,2026-06-25T20:57:34Z,13,"Fixing an Issue Before submitting, please review: The Ultimate Guide to PyTorch Contributions AI-Assisted Development policy Summary TestC10dTorchCommsBasicCPU::test_new_group_delegates_to_split_group_cpu was failing with RuntimeError: Current rank N is not included in the provided ranks list. TorchComms gloo's C++..."
183980,[cpu][cuda] Extend native_group_norm_backward to handle grad from all outputs,https://github.com/pytorch/pytorch/pull/183980,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,6,1,"statusCheckRollup, linux-jammy-rocm-py3.10-mi350 / test (distributed, 3, 3, amd-do-linux.rocm.gpu.gfx950.2), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,180740,180740,180740,"186414, 186946, 186947",,,1,39,0,5f45a6e650f514afa19c8fc3910c962c25f087e481c19bce22e0ffbd7611f11e,005b4f61306002c23d6230e7a0019e5f40df90e335764e75fdeeef8d066c4e14,"module: cpu, open source, ciflow/trunk, release notes: nn, ciflow/mps, module: inductor, ciflow/inductor, needs-review",2026-05-16T00:31:50Z,2026-06-25T20:55:12Z,13,"Stack from ghstack (oldest at bottom): #186947 #186946 #186414 -> #183980 Adds support for gradients from mean and rstd to native_group_norm_backward, via a new overload. Fixes a broken Inductor benchmark in the pipeline, as well as three XFAILed tests. Motivation: when running OpInfo tests with TorchInductor, gradi..."
186414,"native_group_norm: Handle non-contiguous tensors, rather than throwing",https://github.com/pytorch/pytorch/pull/186414,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"183980, 186946, 186947",,,0,22,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/inductor, needs-review",2026-06-05T20:30:53Z,2026-06-25T20:54:28Z,13,"Stack from ghstack (oldest at bottom): #186947 #186946 -> #186414 #183980 native_group_norm and native_group_norm_backward have always assumed contiguity for some tensors, and asserted it for others. The assertions in particular cause unexpected behavior for vmap, which may batch in non-contiguous ways. This PR chan..."
184568,Always reset Dynamo if it was imported,https://github.com/pytorch/pytorch/pull/184568,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,26,0,,,"ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-20T16:06:03Z,2026-06-25T20:14:25Z,13,Stack from ghstack (oldest at bottom): -> #184568 This is a more structural fix along the lines of #184556 but just resetting Dynamo ALWAYS. I think this may have some bad impact on test time because I think we sometimes intentionally rely on cache sharing across tests to reduce test time? Need to assess. One thing...
175242,Improve FX partitioner complexity,https://github.com/pytorch/pytorch/pull/175242,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, release notes: fx",2026-02-18T12:05:44Z,2026-06-25T18:44:25Z,13,Summary Optimizes torch.fx.passes.infra.partitioner.CapabilityBasedPartitioner.propose_partitions: Replaces cached transitive closures with on-demand DFS Memory: O(|V|²) → O(|V|) (sparse graphs) Adds an O(|V|) fast path when unsupported nodes cannot introduce cycles Improves worst-case time: O(|V|³) → O(|V|·|U|) Ben...
188168,Fix hipBLASLt FP32 precision degradation on MI300X (ROCM-3139),https://github.com/pytorch/pytorch/pull/188168,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-06-25T17:50:06Z,2026-06-25T18:20:01Z,13,"Summary Disable hipBLASLt fast-TF32 compute on gfx942 (MI300) where it causes ~BF16 precision loss when torch.set_float32_matmul_precision(""high"") is set. gfx950 unchanged. Test plan Tested on MI308X (gfx942) in the roc7.2-pytorch Docker image (ROCm 7.2): Built from source with USE_ROCM=1 PYTORCH_ROCM_ARCH=gfx942 py..."
187246,serialization: fix nondeterministic pickle bytes in _legacy_save (gh#39383),https://github.com/pytorch/pytorch/pull/187246,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,39383,39383,39383,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-13T12:24:06Z,2026-06-25T18:00:04Z,13,Fixes #39383. Supersedes #57536.. Root cause _legacy_save used storage._cdata (a raw C heap pointer) as the storage key written into the pickle stream. ASLR varies this address across processes which makes pickle.dumps(tensor) non deterministic. _save already carried this fix witdh its own id_map. this PR applies th...
184739,Refactor test_sdpa for out-of-tree backends,https://github.com/pytorch/pytorch/pull/184739,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, merging",2026-05-21T15:47:40Z,2026-06-25T17:53:48Z,13,"Refactors test_sdpa.py to use device-agnostic tensor creation and instantiate_device_type_tests, enabling out-of-tree backends to run these tests by default cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azah..."
178862,[FSDP] enable async H2D prefetching for world_size=1,https://github.com/pytorch/pytorch/pull/178862,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,8,0,,,"Stale, ciflow/trunk, ciflow/inductor, release notes: distributed (fsdp2), ciflow/torchtitan, ciflow/dtensor",2026-03-31T08:18:30Z,2026-06-25T17:35:10Z,13,"Stack from ghstack (oldest at bottom): -> #178862 For world_size=1 with CPU offload, FSDP's set_modules_to_forward_prefetch() was a no-op because unshard() created a dummy AllGatherResult and wait_for_unshard() did the H2D copy synchronously on the default stream. Move the H2D copy into unshard() on copy_in_stream s..."
183509,[distributed] Add fused scaled_mm_v2 + reduce_scatter for NVFP4 support,https://github.com/pytorch/pytorch/pull/183509,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/h100-symm-mem, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,182915,182915,182915,,,,0,1,0,,,"triaged, open source, ciflow/h100-symm-mem, ciflow/torchtitan",2026-05-13T04:39:32Z,2026-06-25T17:01:21Z,13,"add fused_scaled_matmul_reduce_scatter_v2 op to torch.distributed._symmetric_memory to support NVFP4 quantized models in vLLM's Sequence Parallelism and Async Tensor Parallelism pipeline. The existing v1 op uses _scaled_mm with simple scalar/row-wise scales, but NVFP4 requires _scaled_mm_v2 which takes list-based sc..."
178037,[MPS] Raise clear error when MPS is used in forked subprocess,https://github.com/pytorch/pytorch/pull/178037,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,5,0,,,"triaged, open source, Stale, module: mps, release notes: mps",2026-03-21T06:00:16Z,2026-06-25T16:24:00Z,13,"Summary MPS has _mps_is_in_bad_fork wired in C++ and a pthread_atfork handler that sets the bad-fork flag, but unlike CUDA/XPU, there is no Python-side _lazy_init() that checks it. This means using MPS in a forked subprocess silently hangs or crashes instead of raising a helpful error. This PR closes the gap: torch/..."
187501,[Stable C Shim] Use error message retrieval shim if available at runtime,https://github.com/pytorch/pytorch/pull/187501,PR blocked,108,failed checks or blocked/unstable merge state; approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,open source,2026-06-16T21:21:00Z,2026-06-25T16:14:35Z,13,"Issue Issue: #179427 Followup from PR comment: #180135 (review) Summary This introduces helpers that retrieve the error message if the necessary function shims are available to do so using a runtime check with dlsym, this ensures that good diagnostic information can be provided even if TORCH_TARGET_VERSION is lower..."
185505,[inductor] Enable ROCm CK GEMM templates under JIT cpp_wrapper,https://github.com/pytorch/pytorch/pull/185505,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/inductor-rocm-mi300, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: rocm, topic: improvements, module: inductor, ciflow/inductor, release notes: inductor, ciflow/inductor-rocm-mi300, ciflow/torchtitan",2026-05-28T11:46:03Z,2026-06-25T13:55:58Z,13,Stack from ghstack (oldest at bottom): -> #185505 Follow-up to the CUTLASS change: wire ROCm CK kernels through the same JIT cpp_wrapper path. Compiles each CK kernel's .so via ROCmCodeCache.compile and adds it to external_kernel_libs so the wrapper compile links against it. Authored with Claude. cc @jeffdaily @sunw...
185123,[c10d] Add missing runCollectiveChecks for allgather_into_tensor_coalesced in ProcessGroupWrapper,https://github.com/pytorch/pytorch/pull/185123,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"open source, release notes: distributed (c10d)",2026-05-25T08:42:50Z,2026-06-25T13:27:19Z,13,"Summary ProcessGroupWrapper::allgather_into_tensor_coalesced is the only collective operation that does not call runCollectiveChecks(), defeating the wrapper's collective fingerprint verification for this specific op. While reviewing ProcessGroupWrapper, I noticed that allgather_into_tensor_coalesced (added in #1787..."
178318,[Inductor] Add xmask unswitch for dynamic pointwise kernels,https://github.com/pytorch/pytorch/pull/178318,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-03-24T19:33:30Z,2026-06-25T13:15:40Z,13,"Summary This adds an xmask unswitch optimization for dynamic pointwise Triton kernels. When xnumel is symbolic, generated kernels normally keep xmask on every load/store, even for full blocks. That prevents Triton from seeing the common full-block path as unmasked and can block better vectorization. With config.trit..."
181307,Set device in forward input tests,https://github.com/pytorch/pytorch/pull/181307,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-04-23T23:47:50Z,2026-06-25T11:51:05Z,13,"The change sets device explicitly to avoid initializing DeviceMesh based on default behavior. Currently, since device type like XPU is not initialized for the current thread, DeviceMesh is initialized based on rank and the tests could have different expected behaviors, causing issue intel/torch-xpu-ops#3139."
184193,Preserve FX graph cache guard provenance,https://github.com/pytorch/pytorch/pull/184193,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,8,1,"statusCheckRollup, linux-jammy-rocm-py3.10-mi350 / test (distributed, 2, 3, amd-do-linux.rocm.gpu.gfx950.2), inductor-test / test-osdc (inductor_timm, 1, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_torchbench, 2, 2, mt-l-x86aavx2-29-113-a10g), inductor-test / test-osdc (inductor_timm, 2, 2, mt-l-x86aavx2-29-113-a10g), torchtitan-x-pytorch-test / test-osdc (torchtitan_features_integration, 1, 1, mt-l-x86aavx2-189-704-a10g-8), unit-test / inductor-test / test-osdc (inductor, 1, 2, mt-l-x86aavx2-29-113-a10g), unit-test / inductor-test / test-osdc (inductor, 2, 2, mt-l-x86aavx2-29-113-a10g)",statusCheckRollup,,141595,141595,,,,0,5,0,702a496ae719e6d9b0138d0f8a4d824b3bae1f6e94f364697fa49304ea40b51e,7dc7ee1aa416cf08c9a14100dcf026c5444aaf4f4d8f672b70df63c5ec0f1d56,"Merged, ciflow/trunk, release notes: fx, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T12:11:52Z,2026-06-25T11:35:14Z,13,"Stack from ghstack (oldest at bottom): -> #184193 Summary Preserve source locations for Inductor guards that are saved in the FX graph cache, so a cache-hit recompile reason still points at the original guard producer instead of the cache replay site. Fixes #141595 Generated by my agent Root Cause Without the FX gra..."
187784,[TokenSwitch] Add expert-major dispatch layout,https://github.com/pytorch/pytorch/pull/187784,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,187785,,,0,3,0,,,"open source, release notes: distributed (c10d), ciflow/h100-symm-mem, ciflow/torchtitan",2026-06-21T08:45:54Z,2026-06-25T10:06:26Z,13,"Stack from ghstack (oldest at bottom): #187785 -> #187784 Expose the receive-buffer layout through the token-switch API so callers can select it (the NCCL EP algorithm stays high_throughput). Expert-major orders the dispatched tokens so that all tokens routed to the same local expert are contiguous, which is the lay..."
186129,[dtensor] fix conv1d output placement and group scaling,https://github.com/pytorch/pytorch/pull/186129,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-03T20:59:31Z,2026-06-25T08:14:44Z,13,Fixes DTensor dispatch for depthwise convolution (groups == in_channels == out_channels) when channels are sharded across the mesh — e.g. under FSDP or TP where the weight is Shard(0) and the input is Shard(1) on the channel dim. This is the layout used by GatedDeltaNet-style depthwise Conv1d in hybrid-attention mod...
186779,Error on unsupported batch norm third derivatives,https://github.com/pytorch/pytorch/pull/186779,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,186256,186256,186256,176234,,,0,2,0,,,"triaged, open source, release notes: autograd, module: inductor, ciflow/inductor",2026-06-09T11:35:11Z,2026-06-25T07:44:26Z,13,"Fixes #186256. batchnorm_double_backward analytically accounts for the dependence of the saved training statistics on the input, so second-order gradients are correct. Those saved tensors do not carry autograd history, though, and differentiating through the manual double backward silently treats them as constants...."
186877,Use C++20 std::numbers in c10 MathConstants,https://github.com/pytorch/pytorch/pull/186877,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, better-engineering, Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td, ciflow/vllm",2026-06-10T02:15:36Z,2026-06-25T07:44:16Z,13,
187663,Avoid oversized storage for reduced mse_loss / smooth_l1_loss,https://github.com/pytorch/pytorch/pull/187663,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,185647,185647,185647,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-18T14:15:47Z,2026-06-25T07:25:40Z,13,"Fixes #185647 Summary F.mse_loss(x, y, reduction='mean') returned a scalar whose underlying storage was sized for the full elementwise result (e.g. 786432 bytes for a 3x256x256 input) rather than a single element. smooth_l1_loss had the same problem; l1_loss and huber_loss were already fine. Root cause Both ops are..."
187368,[inductor] add input validation to normal and bernoulli decompositions,https://github.com/pytorch/pytorch/pull/187368,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,"185246, 185248","185246, 185248","185246, 185248",,,,0,3,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/inductor-pallas, ciflow/torchtitan",2026-06-15T18:14:19Z,2026-06-25T06:03:51Z,13,"Summary torch.compile silently accepted invalid inputs for torch.normal and torch.bernoulli, producing garbage output instead of matching eager's RuntimeError. This is the same class of bug as #183762 / #187321 (celu_ alpha=0), where validation present in eager is missing from the Inductor decomposition. Fixes: torc..."
185812,Fix arange int64 fractional truncation,https://github.com/pytorch/pytorch/pull/185812,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/mps",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,149097,149097,149097,,,,0,2,0,,,"triaged, open source, ciflow/mps, ciflow/inductor",2026-06-01T12:44:37Z,2026-06-25T05:49:49Z,13,"[updated] Fix torch.arange int64 size computation for fractional arguments to match int32/float32/NumPy. (Fixes #149097) Added isIntegral() guard in RangeUtils.h so fractional args use double precision instead of being truncated. Deduplicated CUDA/MPS size logic via compute_arange_size, removed unused variables. Ali..."
184704,Capturable param in optimizers restored using custom vts and cleanup hooks,https://github.com/pytorch/pytorch/pull/184704,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,182706,182706,182706,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T12:13:20Z,2026-06-25T04:59:44Z,13,Fixes #182706 Optimizers capturable Params are intentionally set to true for optimization however they are never restored. There are 2 places where the restoration is to happen: ConstantVariableTracker which always have this value to be 1. I have created Custom Vt which will do restoration to original val. There is...
185258,[ROCm][DO NOT MERGE][DEBUGGING] MI355 hang reproduce and dump subprocess logs,https://github.com/pytorch/pytorch/pull/185258,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,28,0,,,"module: rocm, open source, ciflow/trunk, release notes: releng, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-26T18:08:11Z,2026-06-25T04:53:37Z,13,This PR is not intended to merge into main. It's for reproducing MI355 CI subprocess invocation hangs and check subprocess logs in the CI environment. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @...
187747,Replace CUDA APIs with generic torch.accelerator equivalents,https://github.com/pytorch/pytorch/pull/187747,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,187205,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-19T23:57:25Z,2026-06-25T03:48:22Z,13,"Summary Replace CUDA-specific API calls with generic torch.accelerator equivalents in torch._inductor and torch._functorch so that diagnostic and instrumentation code works across all backends (CUDA, XPU, MPS, third-party) without device-specific branching. Changes: numeric_utils.clean_memory(): torch.cuda.empty_cac..."
186489,Fix division by zero in range when step becomes zero after conversion,https://github.com/pytorch/pytorch/pull/186489,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/mps,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,186233,186233,186233,,,,0,3,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-06T20:18:33Z,2026-06-25T03:17:03Z,14,"Fix division-by-zero in torch.range when step underflows to zero after dtype conversion Fixes #186233 Summary Fixes a division-by-zero crash in torch.range when a nonzero floating-point step becomes 0 after conversion to the computation dtype. Example that previously crashed: torch.range( 0, 20, 1e-6, dtype=torch.in..."
186211,[Inductor] Skip nonfinite preservation for safe scaled softmax,https://github.com/pytorch/pytorch/pull/186211,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-04T09:00:51Z,2026-06-25T03:16:55Z,14,"Skip the scaled softmax nonfinite-semantics preservation logic for static safe scale factors. PR #184046 added logic in Inductor to preserve the original scaled subtraction behavior for nonfinite scaled-softmax inputs, so that Inductor does not turn eager NaNs into finite softmax results. However, for static positiv..."
176920,Add distributed runtime autotuning for Triton kernels,https://github.com/pytorch/pytorch/pull/176920,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, bot-triaged, ciflow/torchtitan, jansel-agent-skip",2026-03-09T19:25:49Z,2026-06-25T02:16:53Z,14,"Summary: This change adds optional distributed runtime autotuning, where autotuning work for Triton kernels is distributed across ranks in a process group. When enabled, each rank autotunes only 1/world_size of the kernels and results are synchronized, reducing total autotuning time. Design The design tries to optim..."
187883,Avoid per-symbol tuple allocation in the C++ symbolic shape guard,https://github.com/pytorch/pytorch/pull/187883,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,185886,185886,,,,0,5,0,,,"topic: not user facing, module: dynamo, ciflow/inductor",2026-06-22T23:16:09Z,2026-06-25T02:15:49Z,14,"Stack from ghstack (oldest at bottom): -> #187883 The C++ symbolic shape guard (SYMBOLIC_SHAPE_GUARD) collects its symbol values through IndexedGuardAccessor, which packed a fresh (index, value) tuple (PyTuple_Pack) for every symbol on every guard evaluation, then unpacked it. For models with many symbolic-shape rel..."
181481,[Codemod][MigrateInternmcUrls] fbcode/caffe2 [A] (#181481),https://github.com/pytorch/pytorch/pull/181481,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,181481,,,,,0,1,0,,,"open source, fb-exported, Stale, ciflow/trunk, topic: not user facing, meta-exported",2026-04-25T05:37:52Z,2026-06-25T01:24:33Z,14,Summary: Pull Request resolved: #181481 Reviewed By: dtolnay Differential Revision: D102268129
175526,while_loop vmap batching rule,https://github.com/pytorch/pytorch/pull/175526,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,175522,175522,175522,,,,0,6,0,,,"triaged, open source, ciflow/trunk, release notes: torch.func",2026-02-23T10:11:42Z,2026-06-25T01:18:10Z,14,Closes #175522. While_loop op integration for torch functorch hop. Allows vmap integration with it
187882,[RFC] Specialize float attributes of nn.Modules,https://github.com/pytorch/pytorch/pull/187882,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,185886,,,,0,2,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-22T21:54:02Z,2026-06-25T01:17:23Z,14,"Stack from ghstack (oldest at bottom): -> #187882 This is a proposal for the guard-eval regression in #185886, stacked on the guard-eval benchmarks PR. It changes a default for symbolic floats, so it needs sign-off from the automatic-dynamic / symbolic-shapes owner before landing. Root cause: #149564 (""fix dynamic f..."
188120,[FSDP2] Per-FQN offload policy map and cpu_offload_by_budget helper,https://github.com/pytorch/pytorch/pull/188120,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,187615,,,,0,1,0,,,"open source, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-06-25T00:22:35Z,2026-06-25T00:31:33Z,14,"Summary Adds a per-parameter CPU offload surface to FSDP2 and a helper that produces it, implementing the design agreed in #187615. Today fully_shard(..., offload_policy=...) takes a single OffloadPolicy that applies to every parameter, so offload is all-or-nothing. This splits mechanism from policy, per @weifengpy'..."
187409,[distributed] Add NCCL M2N-backed symm_mem reshard,https://github.com/pytorch/pytorch/pull/187409,draft/noise,23,"PR is draft; PR closes or references open issue numbers; CI lane labels: ciflow/h100-symm-mem, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,181061,181061,,,,0,2,0,,,"open source, release notes: distributed (c10d), ciflow/h100-symm-mem, ciflow/torchtitan",2026-06-16T02:30:20Z,2026-06-24T23:55:59Z,14,"Issue Refs #181061 Summary #181061 proposes a first-class M-to-N resharding collective for RL refit and cross-mesh layout changes, where source and destination ranks may use different DeviceMesh layouts and Shard / Replicate placements. This PR adds a preview NCCL M2N-backed implementation (https://github.com/NVIDIA..."
175246,Fix NLLLoss backward failure for non-contiguous 4D inputs,https://github.com/pytorch/pytorch/pull/175246,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,175084,175084,,,,,0,1,0,,,"triaged, open source, Merged, Reverted, Stale, release notes: nn, ci-no-td",2026-02-18T13:31:40Z,2026-06-24T21:58:09Z,14,"Fixes #175084 The nll_loss2d backward kernel requires contiguous grad_input, but the 4D case in nll_loss_nd dispatches directly to nll_loss2d without calling .contiguous() first. When the input is non-contiguous (e.g. from movedim or permute), zeros_like preserves those strides, causing the contiguity check in the b..."
184924,[Test] [CUDA] Skip test_variant_consistency_jit for grid_sampler_2d,https://github.com/pytorch/pytorch/pull/184924,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, release notes: python_frontend, topic: not user facing",2026-05-22T20:38:30Z,2026-06-24T21:45:19Z,14,"#177487 introduded gradgrad support for grid_sampler_2d, and since the backward pass is nondeterministic (uses atomicAdd) test_variant_consistency_jit_grid_sampler_2d_cuda_float32 fails flakily during the gradgrad reference comparison with PYTORCH_OPINFO_SAMPLE_INPUT_INDEX=14 on CUDA devices. This PR skips this test..."
185519,Fix ZeroTensor meta redispatch for nested hessian compile,https://github.com/pytorch/pytorch/pull/185519,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,154284,154284,,,,0,5,0,,,"module: dynamo, ciflow/inductor, release notes: dynamo, agentic",2026-05-28T14:25:03Z,2026-06-24T21:19:41Z,14,Stack from ghstack (oldest at bottom): -> #185519 ZeroTensor binary kernels use meta redispatches to borrow TensorIterator's broadcasting and dtype promotion logic before constructing the real output. Those helper inputs were produced by calling to(meta) on the original operands. In nested torch.func.hessian under t...
188086,Migrate torch_export_python to ShapesSpec API,https://github.com/pytorch/pytorch/pull/188086,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-24T18:06:38Z,2026-06-24T21:08:38Z,14,"Summary: Switch torch_export_python's dynamic-shapes surface from the legacy torch.export.Dim + dict-of-tuples form to the new torch.fx.experimental.dynamic_spec.ShapesSpec API. Public surface: Before: dynamic_shapes={""x"": (Dim(""B""), Dim.STATIC)} After: dynamic_shapes={""x"": TensorSpec([ShapeVar(""B""), STATIC])} or dy..."
184824,[Reland] Port D104346887/PR 182675 for index_add fast path (#184824),https://github.com/pytorch/pytorch/pull/184824,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td, meta-exported",2026-05-21T23:38:26Z,2026-06-24T21:08:00Z,14,"Summary: Reland of D104669063, which was reverted following SEV S664741. The underlying TMA smem stage-alignment bug in scatter_add's CuTeDSL override has since been fixed in PR #184554 (sibling commit in this stack); this diff re-applies the index_add -> scatter_add delegation unchanged from the original, with one..."
188094,[profiler][cupti] Remove unused Python reference decoder (CuptiMonitorBuffer),https://github.com/pytorch/pytorch/pull/188094,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186655, 187517, 187518, 187519, 187520, 187898",,,0,1,0,,,topic: not user facing,2026-06-24T19:57:17Z,2026-06-24T21:07:35Z,14,Stack from ghstack (oldest at bottom): #188019 #187898 #188018 #187520 #187519 #187517 #187518 #186802 #186655 -> #188094 #186812 The CUPTI monitor decodes completed activity buffers in the native GIL-free worker (the _cupti_monitor_native decode thread feeding CuptiMonitor._columns_from_native). The Python CuptiMon...
187797,Add linalg SVD operator benchmarks,https://github.com/pytorch/pytorch/pull/187797,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,187796,,,0,1,0,,,"triaged, open source, topic: not user facing, module: mps, ciflow/mps",2026-06-21T18:13:14Z,2026-06-24T20:34:10Z,14,"This PR dependent from #187796 Summary Add operator benchmarks for torch.linalg.svd and torch.linalg.svdvals. Cover CPU and MPS short configs, including small low-rank shapes, rectangular shapes, batched inputs, and a square boundary case. Register the benchmark in the aggregate operator benchmark entrypoint. Test p..."
186085,Fix transient dispatch-mode skips in Dynamo,https://github.com/pytorch/pytorch/pull/186085,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,140909,140909,,,,0,2,0,,,"ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic",2026-06-03T14:26:55Z,2026-06-24T19:58:40Z,14,"Stack from ghstack (oldest at bottom): -> #186085 Dynamo intentionally skips tracing when a non-infra TorchDispatchMode is active, including FlopCounterMode. That skip used the default ConvertFrameReturn behavior, which records a SKIP strategy on the code object. Since FlopCounterMode is transient, a first call to a..."
187625,Widen optimizer hyperparameter type hints to accept int and Tensor,https://github.com/pytorch/pytorch/pull/187625,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,11,0,,,release notes: optim,2026-06-18T00:14:50Z,2026-06-24T19:58:26Z,14,"Stack from ghstack (oldest at bottom): -> #187625 Broaden the type hints on the functional optimizer implementations so they match the values actually passed at runtime. lr now accepts float | Tensor, integer-valued hyperparameters (weight_decay, momentum, dampening, t0, lr_decay, step_size_min/max) accept float | i..."
174862,[FSDP2] Fix to differing gradient computation graphs in different ranks,https://github.com/pytorch/pytorch/pull/174862,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,171355,171355,171355,,,,0,14,0,,,"module: activation checkpointing, triaged, open source, ciflow/inductor, release notes: distributed (fsdp2), ciflow/torchtitan",2026-02-12T06:29:02Z,2026-06-24T19:51:16Z,14,Fixes #171355 Problem Models such as Qwen3 with MoE takes different forward paths in different ranks resulting in differing gradient computation graphs for each rank. This causes different input sizes to go into the NCCL reduce-scatter collective. The collective completes but sends incorrect weights/gradients to par...
187813,Generalize memory viz and cycle detection,https://github.com/pytorch/pytorch/pull/187813,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,187205,,,0,2,0,,,"triaged, open source, module: accelerator",2026-06-22T03:43:36Z,2026-06-24T19:32:34Z,14,Summary Generalize memory viz and cycle detection to work with any accelerator. torch/utils/viz/_cycles.py was previously hard-coded to CUDA. This PR replaces all CUDA-specific calls with their torch.accelerator equivalents: memory_allocated() now uses torch.accelerator.memory_allocated() _snapshot() and _record_mem...
186178,Fix matmul folding when an inner folded dim has size 1,https://github.com/pytorch/pytorch/pull/186178,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,186148,186148,186148,,,,0,4,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, matrix multiplication",2026-06-04T03:32:10Z,2026-06-24T19:21:25Z,14,"Fixes #186148 The existing matmul folding check rejected some viewable layouts when one of the folded leading dimensions has size 1. In particular, tensors shaped like [seq, batch, hidden] = [384, 1, 4096] can be reshaped to [384, 4096] without a copy, but the previous adjacent-stride check did not account for size-..."
187603,[executorch][dynamo] Import torch._utils before Dynamo variable registration,https://github.com/pytorch/pytorch/pull/187603,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-17T20:24:03Z,2026-06-24T19:15:01Z,14,"Summary: torch/_dynamo/variables/torch.py reads torch._utils at module scope (constant_fold_functions includes torch._utils._get_device_index, and the tracing-state dict maps torch._utils.is_compiling). Under lazy imports, torch/__init__'s from torch._utils import ... is deferred and torch.__getattr__ does not resol..."
185163,[aot_autograd] Handle None in saved tensors for LSTM/RNN backward; enable allow_rnn by default,https://github.com/pytorch/pytorch/pull/185163,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/docker, ciflow/dtensor, ciflow/h100",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,158007,158007,158007,,,,0,3,0,,,"module: rnn, module: cpu, triaged, open source, release notes: quantization, topic: not user facing, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), ciflow/rocm-mi300, release notes: inductor (aoti), ciflow/h100, ciflow/h100-symm-mem, ciflow/b200, ci-refresh-hf-cache, ciflow/torchtitan, ciflow/dtensor, ciflow/docker",2026-05-25T20:30:50Z,2026-06-24T19:14:57Z,14,"Fixes #158007. torch.compile now works with nn.LSTM, nn.GRU, and nn.RNN modules. Changes: AOTAutograd: save_from_forward now handles None values in the partitioned forward graph's saved-for-backward tensors. The LSTM decomposition on CPU (via MKLDNN) produces None for the optional projection weight (hr_weight when p..."
187630,fx/passes: fix non-deterministic ordering in get_source_partitions(),https://github.com/pytorch/pytorch/pull/187630,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,147170,147170,147170,,,,0,1,0,,,"triaged, open source, release notes: fx",2026-06-18T01:27:33Z,2026-06-24T19:05:24Z,14,"Summary Fixes #147170 make_partition inside get_source_partitions() collected input_nodes, output_nodes, and params using set(). Converting a set to a list produces an arbitrary iteration order that varies across Python invocations due to hash randomization (PYTHONHASHSEED). The issue reporter observed different inp..."
186777,[1/N] Replace module-level RNG-related APIs with torch.accelerator,https://github.com/pytorch/pytorch/pull/186777,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186599, 186600, 186771, 186899, 186902, 186906",,,0,30,0,,,"open source, ciflow/trunk, ciflow/mps, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-09T11:10:54Z,2026-06-24T19:04:55Z,14,Stack from ghstack (oldest at bottom): #186906 #186902 #186899 -> #186777 #186771 #186600 #186599 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
182984,[Inductor] Expose config for sm carveout of templates (#182984),https://github.com/pytorch/pytorch/pull/182984,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, fb-exported, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, meta-exported, ciflow/b200, ciflow/torchtitan",2026-05-08T18:54:05Z,2026-06-24T19:04:36Z,14,Summary: Expose SM carveout for persistent templates as a configuration option. Test Plan: test_honor_sm_carveout_with_triton_tma Differential Revision: D104442040 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjame...
188065,[AOTI] measure performance with AOTI_RUNTIME_CHECK_INPUTS=1,https://github.com/pytorch/pytorch/pull/188065,draft/noise,13,PR is draft,wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,3,0,,,,2026-06-24T13:49:53Z,2026-06-24T19:03:06Z,14,Stack from ghstack (oldest at bottom): -> #188065
188089,[forward AD] Remove obsolete __debug__ guard when loading JVP decompositions,https://github.com/pytorch/pytorch/pull/188089,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: autograd, triaged, open source, module: forward ad",2026-06-24T18:27:20Z,2026-06-24T18:51:53Z,14,"Summary Remove the __debug__ guard from torch.autograd.forward_ad._maybe_load_decompositions() and delete the associated FIXME. The FIXME was added in 2022 to avoid a TorchScript type annotation error when importing torch._decomp.decompositions_for_jvp under python -O/-OO. After investigating the current codepath, I..."
186624,Fix CLANGTIDY errors in c10/cuda,https://github.com/pytorch/pytorch/pull/186624,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,"185946, 186375",,,0,1,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td",2026-06-08T18:12:37Z,2026-06-24T18:46:28Z,14,"CI lint in #185946 and #186375 exposed CLANGTIDY failures unrelated to the changes in those PRs. This PR fixes those failures and the few other lint failures in c10/cuda, all CLANGTIDY. cc @eqy Authored with Codex"
187208,[BE][Ez]: Use CPP20 rvalue overload for ostringstream,https://github.com/pytorch/pytorch/pull/187208,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"module: cpu, open source, better-engineering, Merged, NNC, Reverted, ciflow/trunk, release notes: mobile, module: dynamo, ciflow/inductor, ci-no-td, release notes: inductor (aoti), ciflow/h100-symm-mem",2026-06-12T19:56:57Z,2026-06-24T18:46:10Z,14,Followup to #186552 . Uses the rvalue overload of ostringstream newly introduced in CPP20 to reduce the amount of string copy allowing us to steal the internal string buffer. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01 @EikanWang @voznesenskym @penguinwu @Guobing-Chen @zhuhaozhe @...
187803,fix(autograd): sinc_backward NaN for second derivative at x=0,https://github.com/pytorch/pytorch/pull/187803,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,89459,89459,89459,,,,0,3,0,,,"module: autograd, triaged, open source",2026-06-21T21:00:02Z,2026-06-24T18:38:42Z,14,"sinc_backward divided directly by self_squared_pi, which is zero at x=0. Since sinc'(x) = (pi*x*cos(pi*x) - sin(pi*x)) / (pi*x^2) has a 0/0 singularity there, NaN entered the autograd graph and sinc''(0) came out NaN instead of -pi^2/3. The fix introduces a safe denominator (ones_like where self_squared_pi == 0) to..."
187823,Fix regex for extracting gist URLs from logs,https://github.com/pytorch/pytorch/pull/187823,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-22T08:36:40Z,2026-06-24T18:35:16Z,14,cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
187713,Fix B008 flake8-bugbear violations (function calls in default arguments),https://github.com/pytorch/pytorch/pull/187713,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,106571,,,,0,1,0,,,"triaged, open source, release notes: quantization",2026-06-19T06:15:55Z,2026-06-24T18:32:01Z,14,"Fixes B008 flake8-bugbear lint violations across 32 files. Python evaluates default arguments at function definition time, so mutable defaults or function call results in signatures (e.g., opts=AllToAllOptions(), sample_shape=torch.Size()) create shared state that persists across calls. Related to #106571 Changed af..."
186600,Introduce manual_seed&seed to torch.accelerator,https://github.com/pytorch/pytorch/pull/186600,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186599, 186771, 186777, 186899, 186902, 186906",,,0,19,0,,,"open source, ciflow/trunk",2026-06-08T12:33:26Z,2026-06-24T18:20:51Z,14,Stack from ghstack (oldest at bottom): #186906 #186902 #186899 #186777 #186771 -> #186600 #186599 Motivation This PR introduces the following APIs in torch.accelerator: manual_seed manual_seed_all seed seed_all
185213,[Test] Make dynamo test_streams.py device-agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/185213,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, module: tests/graph",2026-05-26T11:35:26Z,2026-06-24T16:23:50Z,14,Refactor test/dynamo/test_streams.py to use instantiate_device_type_tests with @onlyAccelerator so stream and event tests automatically run on any registered accelerator backend. Changes Add instantiate_device_type_tests and @onlyAccelerator infrastructure Convert 58 tests from @requires_cuda to @onlyAccelerator Rep...
188058,[inductor] Avoid disjoint horizontal fusion when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/188058,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187842, 188057, 188059",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-24T12:51:17Z,2026-06-24T16:15:51Z,14,Stack from ghstack (oldest at bottom): #188059 #187842 -> #188058 #188057 Reject horizontal fusion of two pointwise nodes that each materialize a separate large external output to disjoint users (smaller output >= the fusion's shared-data score): fusing keeps both outputs live together for little reuse benefit. Acti...
188056,[inductor] Avoid disjoint horizontal fusion when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/188056,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-24T12:49:29Z,2026-06-24T16:15:45Z,14,Stack from ghstack (oldest at bottom): (to be filled) Reject horizontal fusion of two pointwise nodes that each materialize a separate large external output to disjoint users (smaller output >= the fusion's shared-data score): fusing keeps both outputs live together for little reuse benefit. Active only when allow_p...
188059,[inductor] Block only peak-crossing fusions in the peak-memory guards,https://github.com/pytorch/pytorch/pull/188059,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187842, 188057, 188058",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-24T12:51:22Z,2026-06-24T16:15:42Z,14,"Stack from ghstack (oldest at bottom): -> #188059 #187842 #188058 #188057 The disjoint-horizontal and cross-extern guards previously blocked every fusion that extends a large buffer's lifetime, giving up fusions in low-memory regions far from the peak. Make them peak-aware: build a per-step live-memory timeline (reu..."
182342,[Inductor] Deal with choices that fail to compile in async pipelined autotuning path (#182342),https://github.com/pytorch/pytorch/pull/182342,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,182342,,,,,0,1,0,,,"open source, fb-exported, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/rocm-mi300, ciflow/h100, meta-exported, ciflow/b200, ciflow/torchtitan",2026-05-04T20:17:35Z,2026-06-24T16:00:27Z,14,Summary: Pull Request resolved: #182342 Test Plan: Small refactor Differential Revision: D103725434 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
187464,[ROCm] [Operator Benchmark] Add gfx950 ROCm nightly CI lane and docker image,https://github.com/pytorch/pytorch/pull/187464,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/op-benchmark","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"module: rocm, open source, release notes: releng, topic: not user facing, ciflow/op-benchmark, ciflow/docker",2026-06-16T17:17:33Z,2026-06-24T15:35:10Z,14,"Summary Add a gfx950-specific ROCm nightly docker image and CI lane for MI350X, instead of running gfx950 hardware with the gfx942 pytorch-linux-noble-rocm-nightly-py3 stack, ran into issues with that approach here. gfx950 runners were getting gfx942 hipBLASLt libs (e.g. scaled_mm / missing gfx950 Tensile assets). A..."
187842,[inductor] Avoid cross-extern fusion when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/187842,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"188057, 188058, 188059",,,0,3,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T15:09:41Z,2026-06-24T15:24:18Z,14,Stack from ghstack (oldest at bottom): #188059 -> #187842 #188058 #188057 [inductor] Avoid cross-extern fusion when allow_peak_memory_increasing_fusion is off (PR #187842) Summary Add a generic fusion guard for cases where fusing across an extern branch would make another branch output live across that extern user....
188055,[inductor] Evaluate compile-time ProcessGroup attrs in overlap estimates,https://github.com/pytorch/pytorch/pull/188055,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-24T12:49:25Z,2026-06-24T15:24:11Z,14,"Stack from ghstack (oldest at bottom): (to be filled) Functionalized eager collectives can pass a ProcessGroup to _c10d_functional ops as a compile-time FX value. In GraphTrainer regional graphs with ATen overlap scheduling it arrives as a get_attr node with no meta[""val""], so the group-name resolver raised KeyError..."
188057,[inductor] Evaluate compile-time ProcessGroup attrs in overlap estimates,https://github.com/pytorch/pytorch/pull/188057,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187842, 188058, 188059",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-24T12:51:13Z,2026-06-24T15:24:08Z,14,"Stack from ghstack (oldest at bottom): #188059 #187842 #188058 -> #188057 Functionalized eager collectives can pass a ProcessGroup to _c10d_functional ops as a compile-time FX value. In GraphTrainer regional graphs with ATen overlap scheduling it arrives as a get_attr node with no meta[""val""], so the group-name reso..."
187832,[FX][Dynamo] Add tests for nn_module_stack and source_fn_stack metadata contracts,https://github.com/pytorch/pytorch/pull/187832,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-22T13:09:39Z,2026-06-24T15:23:57Z,14,"Summary Adds targeted tests for the nn_module_stack and source_fn_stack node metadata that FX tracing and Dynamo attach to graph nodes. These metadata are consumed by multiple downstream systems (ONNX export, quantisation, torch.export serialisation, custom compiler backends), but the existing test coverage only val..."
184867,[symbolic shapes] Avoid direct replacement cycles,https://github.com/pytorch/pytorch/pull/184867,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: fx, ciflow/inductor",2026-05-22T08:26:29Z,2026-06-24T14:47:23Z,14,"Summary: Avoid adding a direct replacement cycle in ShapeEnv._set_replacement when the target symbol already maps back to the source. This keeps replacement tracking acyclic before union-find rank reduction. Review: @laithsakka, would you mind taking a look when you have a chance? Your guidance on the symbolic shape..."
187619,Remove Dynamo guard from custom op kernels,https://github.com/pytorch/pytorch/pull/187619,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"Merged, Reverted, ciflow/trunk, release notes: composability, ci-no-td",2026-06-17T22:17:19Z,2026-06-24T14:39:30Z,14,Stack from ghstack (oldest at bottom): -> #187619 #187618 #187616 The custom-op backend wrapper decorates every registered kernel with torch._disable_dynamo. That guard was added in afb73d2 with test_dynamo_disabled_in_custom_op_kernels because Dynamo could previously turn itself back on inside custom-op implementat...
187754,[testintro] Add a test-introspection tool,https://github.com/pytorch/pytorch/pull/187754,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/docker,wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,,,"187755, 187756",,,0,2,0,,,"topic: not user facing, ciflow/docker",2026-06-20T01:32:58Z,2026-06-24T14:17:39Z,14,"Stack from ghstack (oldest at bottom): #187756 #187755 -> #187754 WARNING: AI Slop ahead - this was entirely vibe-coded and I mostly reviewed the design and methodology rather than the code itself. Add tools/testing/introspection/, which enumerates the concrete tests a CI target (platform/config) generates -- by imp..."
187720,Fix stale_issues year cutoff clamping the day to 28 in every month,https://github.com/pytorch/pytorch/pull/187720,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-06-19T09:46:11Z,2026-06-24T14:02:16Z,14,"Summary The ""year"" branch of parse_older_than in tools/stale_issues.py hardcodes the day to min(today.day, 28): elif unit == ""year"": return date(today.year - n, today.month, min(today.day, 28)) That clamp drops the day to the 28th in every month, even though only February needs it. Going back a year from a 30/31-day..."
187818,[WIP] generalize unitttests,https://github.com/pytorch/pytorch/pull/187818,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor",2026-06-22T06:34:34Z,2026-06-24T13:46:58Z,14,cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
185910,[AARCH64] Optimize AArch64 qlinear performance by triggering blocked layout in prepacked weight,https://github.com/pytorch/pytorch/pull/185910,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: cpu, triaged, open source, release notes: quantization",2026-06-02T09:14:58Z,2026-06-24T13:28:28Z,14,"By this change we update the weight prepack function pack_weight_to_onednn_tensor() which is triggered in the torch.compile + TORCHINDUCTOR_FREEZING path to trigger better blocked layout like BA24b8a for AARCH64. Currently, ideep::matmul_forward::expected_weights_desc defaults to an s32 destination for int8 weights..."
187928,test: generate native function type hint checks,https://github.com/pytorch/pytorch/pull/187928,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,159259,159259,159259,,,,0,4,0,,,"triaged, open source, topic: not user facing",2026-06-23T07:44:40Z,2026-06-24T12:29:19Z,14,"Issue Fixes #159259 Summary Adds a small generator for native-function type hint reveal tests. The generator reads native_functions.yaml through the same torchgen/pyi signature path used by tools/pyi/gen_pyi.py, derives expected return reveals from generated signatures, and emits a representative test/typing/reveal/..."
181396,Print and validate full CUDA version in smoke test,https://github.com/pytorch/pytorch/pull/181396,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"Stale, ciflow/binaries, topic: not user facing",2026-04-24T16:53:48Z,2026-06-24T12:19:52Z,14,"Summary Improve CUDA version reporting and validation in the binary smoke test: Print the full major.minor.patch CUDA version (e.g. 13.2.1) derived from torch._C._cuda_getCompiledVersion(). torch.version.cuda only reports major.minor (e.g. 13.2), which hides the patch release the wheel was actually built against. Va..."
186089,[XPU] Align XPU unit tests decorators and overrides.,https://github.com/pytorch/pytorch/pull/186089,draft/noise,23,PR is draft; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,wait for author to mark ready,,UNKNOWN,True,0,0,,,"1973, 3343, 3574, 3587, 3650, 3891","1973, 3343, 3574, 3587, 3650, 3891",,,,,0,7,0,,,"open source, ciflow/trunk, release notes: python_frontend",2026-06-03T15:07:06Z,2026-06-24T11:40:04Z,14,This PR batches several smaller changes that align the XPU unit tests with upstream PyTorch. This PR fixes or partially fixes the following torch-xpu-ops issues: intel/torch-xpu-ops#3587 intel/torch-xpu-ops#3650 intel/torch-xpu-ops#3343 intel/torch-xpu-ops#1973 intel/torch-xpu-ops#3574 intel/torch-xpu-ops#3891
187739,[ROCm][CI] Misc updates,https://github.com/pytorch/pytorch/pull/187739,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: rocm, open source, topic: not user facing",2026-06-19T19:35:40Z,2026-06-24T08:36:03Z,14,cc @jeffdaily @sunway513 @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
188037,Fix current stream selection,https://github.com/pytorch/pytorch/pull/188037,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,4091,4091,,,,,0,58,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-24T07:29:38Z,2026-06-24T07:31:38Z,14,Previous implementation searched for attributes in base torch.Stream class which do not exist. Fixes: intel/torch-xpu-ops#4091 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
181138,[Inductor][NFC] Refactor fusion legality checks for `UserDefinedTritonKernel`,https://github.com/pytorch/pytorch/pull/181138,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,15,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-22T17:26:22Z,2026-06-24T07:21:46Z,14,"Clean up user-triton related usage in scheduler, and separate concerns regarding epilogue fusion legality. Consolidate fusion legality logic Remove UserDefinedTritonKernel.can_fuse_epilogue: ir.UserDefinedTritonKernel initialisation is no longer responsible for flagging fusion legality. In turn, this also removes re..."
180455,"[AARCH64] Enable MKLDNN backend for AArch64 for INT8 Matmul, using torch op _int_mm.dtype",https://github.com/pytorch/pytorch/pull/180455,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,10,0,,,"module: cpu, triaged, open source, Merged, Reverted, ciflow/trunk, release notes: linalg_frontend, module: inductor, module: dynamo, ciflow/inductor, ci-no-td, ciflow/linux-aarch64, ciflow/torchtitan",2026-04-15T11:26:57Z,2026-06-24T06:45:05Z,14,"This PR adds a dtype variant of the existing torch op int_mm as _int_mm.dtype for AARCH64, which supports src : s8/u8 wei:s8 dst: f32/bf16 matmul triggering MKLDNN fastpath. By enabling this op variant , we will be able to leverage Accelerated INT8 matmul performance using MKLDNN. Currently Torchao's INT8DynamicActi..."
183668,[ROCm] fix flaky test_honor_sm_carveout,https://github.com/pytorch/pytorch/pull/183668,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,164271,164271,,,,,0,1,0,,,"module: rocm, triaged, open source, topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-14T07:00:53Z,2026-06-24T06:14:11Z,14,"Issue Summary On a cold hipBLASLt cache, torch.profiler.profile captures __amd_rocclr_fillBufferAligned.kd (hipBLASLt's first-call workspace zero) alongside the four matmul kernels the test expects, producing five events instead of four and breaking the 4-way unpack at line 1789. Fix Added a warmup run for scaled_mm..."
187971,[inductor] Add strict_numerics config to emit INNER_TREE reduction ordering for sum,https://github.com/pytorch/pytorch/pull/187971,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-06-23T19:39:33Z,2026-06-24T06:07:52Z,14,"Stack from ghstack (oldest at bottom): -> #187971 This PR adds a new Inductor strict_numerics mode and wires it into reduction codegen so that torch.sum reductions use Triton's deterministic, layout-independent INNER_TREE reduction order. When config.strict_numerics is enabled, Inductor emits: tl.sum(..., reduction_..."
182809,[2/N] Replace data_ptr with mutable_data_ptr,https://github.com/pytorch/pytorch/pull/182809,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: cpu, triaged, open source, release notes: quantization, release notes: linalg_frontend, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-07T10:00:52Z,2026-06-24T06:07:07Z,14,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
181199,[spmd types] fully_shard in spmd_types,https://github.com/pytorch/pytorch/pull/181199,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,181398,,,0,10,0,,,"Stale, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-04-23T01:11:04Z,2026-06-24T05:52:05Z,14,relies on D102239042 Stack from ghstack (oldest at bottom): -> #181199 #181398 #180880
187471,[inductor][mtia] Strip AMD-only autotune kwargs from MTIA Triton compile configs (#187471),https://github.com/pytorch/pytorch/pull/187471,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-16T18:18:38Z,2026-06-24T04:33:18Z,14,"Summary: The MTIA Triton backend is built on top of the AMD/HIP backend, so AMD-only autotune kwargs (waves_per_eu, matrix_instr_nonkdim, kpack) end up in cfg.kwargs for MTIA autotune configs. In the Inductor AOT compile path these are not MTIA kernel arguments, so leaving them in the compile-meta constants makes AS..."
180644,Relax align_random_eager test tolerances for Philox float precision,https://github.com/pytorch/pytorch/pull/180644,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,2,0,,,"Stale, topic: not user facing, module: inductor, ciflow/inductor",2026-04-17T04:24:31Z,2026-06-24T03:54:04Z,14,"Stack from ghstack (oldest at bottom): -> #180644 The Triton rand_eager_kernel converts Philox uint32 outputs to float32 via 1.0 - (x * inv + half), which can round differently than cuRAND's direct x * inv + half for ~1 in 10^6 elements. When such an element lands at the dropout threshold, one mask bit differs, and..."
182193,Tighten any_index_expr_overflows_int32 overflow check,https://github.com/pytorch/pytorch/pull/182193,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-01T23:07:32Z,2026-06-24T03:01:44Z,15,"[Inductor] Bound exact index expressions for int32 indexing eligibility Summary Decide int32-vs-int64 indexing by checking the actual addressing expressions a kernel will emit, instead of approximating them with a per-buffer storage-size or stride proxy. Background / Root cause When unrolled chunked-loop slices are..."
188003,[AMD][hipBLASLt] Enable streak-K work stealing for GEMM operations,https://github.com/pytorch/pytorch/pull/188003,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/rocm-mi300, ciflow/h100, meta-exported, ciflow/b200",2026-06-24T00:15:14Z,2026-06-24T03:00:52Z,15,Differential Revision: D109507257
187687,fix: aten embed bwd thread sync before proceeding to next batch,https://github.com/pytorch/pytorch/pull/187687,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: cuda",2026-06-18T18:27:13Z,2026-06-24T03:00:15Z,15,"Problem embedding_backward_feature_kernel reuses the shared-memory indices_batch buffer across outer-loop batches. The existing barriers only synchronize work inside each chunk, so one warp may start loading the next batch into indices_batch while another warp is still reading entries from the previous batch. This c..."
187696,Add DataLoader worker signal message affixes,https://github.com/pytorch/pytorch/pull/187696,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-18T20:05:13Z,2026-06-24T02:59:29Z,15,"I would appreciate feedback on the preferred test strategy for signal-handler output, since the message is emitted through low-level write() rather than Python exception handling. Problem DataLoader worker signal errors are written directly from C++ signal handlers to stderr, so Python code cannot catch and wrap the..."
187520,[profiler][cupti] Comms monitoring tests,https://github.com/pytorch/pytorch/pull/187520,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186655, 187517, 187518, 187519, 187898",,,0,38,0,,,topic: not user facing,2026-06-16T23:02:45Z,2026-06-24T02:55:15Z,15,Stack from ghstack (oldest at bottom): #188019 #187898 #188018 -> #187520 #187519 #187517 #187518 #186802 #186655 #186812 Authored with Claude.
187519,[profiler][cupti] Comms monitoring: CommsObserver + comm hook + plugins (CollTrace replacement),https://github.com/pytorch/pytorch/pull/187519,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186655, 187517, 187518, 187520, 187898",,,0,37,0,,,,2026-06-16T23:02:40Z,2026-06-24T02:55:14Z,15,Stack from ghstack (oldest at bottom): #188019 #187898 #188018 #187520 -> #187519 #187517 #187518 #186802 #186655 #186812 Authored with Claude.
187517,[profiler][cupti] CUDA-graph correlation utils: node-id remap + side-stream events,https://github.com/pytorch/pytorch/pull/187517,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186655, 187518, 187519, 187520, 187898",,,0,35,0,,,topic: not user facing,2026-06-16T23:02:30Z,2026-06-24T02:55:14Z,15,Stack from ghstack (oldest at bottom): #188019 #187898 #188018 #187520 #187519 -> #187517 #187518 #186802 #186655 #186812 Authored with Claude.
187518,[profiler][cupti] In-process NCCL metadata plugin (ncclProfiler_v6),https://github.com/pytorch/pytorch/pull/187518,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186655, 187517, 187519, 187520, 187898",,,0,38,0,,,,2026-06-16T23:02:37Z,2026-06-24T02:55:13Z,15,Stack from ghstack (oldest at bottom): #188019 #187898 #188018 #187520 #187519 #187517 -> #187518 #186802 #186655 #186812 Authored with Claude.
186655,[profiler][cupti] Benchmarks: columnar-decode roofline + monitor distortion,https://github.com/pytorch/pytorch/pull/186655,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187517, 187518, 187519, 187520, 187898",,,0,44,0,,,topic: not user facing,2026-06-08T21:33:54Z,2026-06-24T02:55:11Z,15,"Stack from ghstack (oldest at bottom): #188019 #187898 #188018 #187520 #187519 #187517 #187518 #186802 -> #186655 #186812 Three CUPTI monitor benchmarks under benchmarks/profiler_benchmark/, split out from the multiplexer change (all prefixed cupti_monitor_): cupti_monitor_bench_decode.py: a roofline for the columna..."
181051,Fix heap-buffer-overflow in quantized_max_pool2d for 1d args,https://github.com/pytorch/pytorch/pull/181051,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,162476,162476,"116254, 162476",,,,0,1,0,,,"module: cpu, triaged, open source, Stale, release notes: quantization, topic: bug fixes, no-stale",2026-04-22T01:18:13Z,2026-06-24T02:31:02Z,15,"Fixes #162476 (ASan heap-buffer-overflow). Related to #116254. Summary check_maxpool2d_params accepts kernel_size/padding/dilation of size 1 or 2, but quantized_max_pool2d then unconditionally indexes [1] on each before forwarding to q_maxpool_2d. With a 1-element list this reads past the end of the vector. Repro fr..."
181449,AOTI: Add use_fake_constants mode for zero-weight compilation,https://github.com/pytorch/pytorch/pull/181449,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-04-24T21:57:45Z,2026-06-24T01:19:07Z,15,"Stack from ghstack (oldest at bottom): (to be filled) When aot_inductor.use_fake_constants is True, the compilation pipeline accepts FakeTensor constants (metadata only, no storage) instead of real tensors. This enables zero weight memory during AOTI compilation of large models — the caller saves weights to disk bef..."
187779,Fix MPS fused RMSNorm: do the weight multiply in fp32 to match CPU/CUDA,https://github.com/pytorch/pytorch/pull/187779,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/mps,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-21T01:35:20Z,2026-06-24T01:12:50Z,15,"What does this PR do? The MPS fused RMSNorm kernel casts the normalized value to the half dtype before multiplying by the weight (out = w * static_cast<T>(x * inv)), so its output diverges from PyTorch's CPU composite and CUDA fused/reference behavior, which keep the x * inv * weight product in fp32 and cast once at..."
175486,[WIP][inductor] Layout allocator approach for symm_mem graph inputs,https://github.com/pytorch/pytorch/pull/175486,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"175450, 175476",,,0,6,0,,,"Stale, module: inductor, ciflow/inductor, release notes: inductor, ciflow/h100-symm-mem",2026-02-22T07:15:01Z,2026-06-23T23:49:13Z,15,"Stack from ghstack (oldest at bottom): -> #175486 #175476 #175450 #175449 #174856 Summary: Replace the Triton identity-copy workaround for graph inputs (InputBuffer) that need P2P memory with a Layout-based allocator constraint approach. Three paths in _maybe_realize_symm_mem, in priority order: We control allocatio..."
187328,Add fusion provenance to profiler timelines,https://github.com/pytorch/pytorch/pull/187328,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,DIRTY,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-15T10:04:39Z,2026-06-23T22:47:00Z,15,"Summary This PR extends the profiler timeline provenance path added by #186230 with an explicit fusion metadata switch. TORCH_COMPILE_DEBUG_EXTEND=1 keeps the existing stack-only behavior. TORCH_COMPILE_DEBUG_FUSION=1 enables the heavier fusion metadata path, which records the matched Inductor graph key, kernel prov..."
187367,Fix typos in comments and docstrings,https://github.com/pytorch/pytorch/pull/187367,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,19,0,,,"ciflow/trunk, release notes: quantization, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-15T18:13:05Z,2026-06-23T22:41:56Z,15,"Stack from ghstack (oldest at bottom): -> #187367 Correct assorted typos across inductor, quantization, export, nn, and docs: duplicated words (""different different"", ""cause cause"", ""to the to""), misspellings (""sqeeuze"", ""chuning"", ""bv""), and grammar slips (""typed"" -> ""were typed"", ""leave this hear"" -> ""leave this h..."
180584,Test,https://github.com/pytorch/pytorch/pull/180584,ready for maintainer decision,45,"CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,2,0,,,"open source, Stale, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-04-16T18:32:18Z,2026-06-23T22:02:02Z,15,Stack from ghstack (oldest at bottom): -> #180584 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
183468,[Inductor] Fix linter from tf32 fix,https://github.com/pytorch/pytorch/pull/183468,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, fb-exported, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, meta-exported, ciflow/torchtitan",2026-05-12T19:44:20Z,2026-06-23T22:01:56Z,15,Differential Revision: D104872670 Fixes linting issue from #183355 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @diff-train-skip-merge
187965,[inductor] Avoid disjoint horizontal fusion when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/187965,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187842, 187964",,,0,1,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-23T18:34:17Z,2026-06-23T21:34:11Z,15,Stack from ghstack (oldest at bottom): #187842 -> #187965 #187964 [inductor] Avoid disjoint horizontal fusion when allow_peak_memory_increasing_fusion is off Summary Add a fusion guard that rejects horizontal (consumer/consumer) fusion of two pointwise nodes when each side materializes a separate external output to...
187962,[inductor] Evaluate compile-time ProcessGroup attrs in overlap estimates,https://github.com/pytorch/pytorch/pull/187962,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-23T18:27:21Z,2026-06-23T21:34:05Z,15,"Stack from ghstack (oldest at bottom): -> #187962 Functionalized eager collectives can pass a ProcessGroup as a compile-time FX value to _c10d_functional ops. In GraphTrainer regional Inductor graphs with ATen distributed overlap scheduling enabled, that value can appear as a get_attr node instead of a string group..."
187964,[inductor] Evaluate compile-time ProcessGroup attrs in overlap estimates,https://github.com/pytorch/pytorch/pull/187964,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"187842, 187965",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-23T18:34:11Z,2026-06-23T21:33:50Z,15,"Stack from ghstack (oldest at bottom): #187842 #187965 -> #187964 Functionalized eager collectives can pass a ProcessGroup as a compile-time FX value to _c10d_functional ops. In GraphTrainer regional Inductor graphs with ATen distributed overlap scheduling enabled, that value can appear as a get_attr node instead of..."
186527,Support export of packed LSTM with data-dependent lengths,https://github.com/pytorch/pytorch/pull/186527,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,112044,112044,,,,0,2,0,,,"ciflow/inductor, release notes: export, agentic, needs-review",2026-06-07T07:20:14Z,2026-06-23T19:28:50Z,15,"Stack from ghstack (oldest at bottom): -> #186527 The packed sequence path produces metadata from runtime sequence lengths: _pack_padded_sequence returns packed data whose leading dimension is the sum of lengths, plus a batch_sizes tensor whose length is the max sequence length. The fake implementation only modeled..."
178584,Remove unused expected failure decorator from FakeTensorOperatorInvariants test and clean up condition in Module class,https://github.com/pytorch/pytorch/pull/178584,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,2712,2712,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, ciflow/inductor, ciflow/dynamo",2026-03-27T09:25:59Z,2026-06-23T19:24:57Z,15,"Resolves intel/torch-xpu-ops#2712 Root cause FakeTensorConverter.from_meta_and_device() stores FakeTensors in tensor_memo (a WeakValueDictionary), which creates KeyedRef weakrefs on them. When nn.Module._apply() calls swap_tensors() on FakeTensor parameters, swap_tensors rejects tensors that have weakrefs. The weakr..."
186766,fix: [BE] Deduplicate auto_functionalized and triton_kernel_wrapper_functi...,https://github.com/pytorch/pytorch/pull/186766,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,133443,133443,133443,,,,0,3,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-06-09T05:20:43Z,2026-06-23T19:23:30Z,15,Summary auto_functionalized and triton_kernel_wrapper_functional are both higher-order operators that act as functional wrappers around mutable operations. Their CompositeExplicitAutograd (dense) implementations share the same core pattern: 1. Root cause auto_functionalized and triton_kernel_wrapper_functional are b...
181392,[AC] Add example_inputs param to remat pass and mark recomputed nodes as backward,https://github.com/pytorch/pytorch/pull/181392,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-24T16:28:52Z,2026-06-23T19:14:45Z,15,"Stack from ghstack (oldest at bottom): -> #181392 Add an example_inputs parameter to remat_using_tags_for_fwd_loss_bwd_graph for consistency with other partitioner APIs, set autograd_backward=True on recomputed nodes so downstream passes can identify them as belonging to the backward region, and rewrite the pass to..."
178463,[Distributed] Make Ckpt Tests Backend Agnostic,https://github.com/pytorch/pytorch/pull/178463,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/torchtitan,wait for contributor update,CHANGES_REQUESTED,DIRTY,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, Stale, release notes: distributed (ddp), release notes: distributed (checkpoint), module: PrivateUse1, ciflow/torchtitan",2026-03-26T01:45:05Z,2026-06-23T19:14:39Z,15,This PR is part of a series to refactor distributed tests to be completely device-agnostic. Changes Source: torch/distributed/_shard/sharded_tensor/_ops/binary_cmp.py — _communicate_result: Replace torch.cuda.current_device() with torch.accelerator.current_accelerator() + torch.accelerator.current_device_index() for...
185491,Cleanup unused tensor variables,https://github.com/pytorch/pytorch/pull/185491,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, triaged, open source, ciflow/trunk, topic: not user facing, ciflow/xpu",2026-05-28T09:36:57Z,2026-06-23T18:36:01Z,15,Enable the bugprone-unused-local-non-trivial-variable clang-tidy check and remove the two unused dummy_mean_ / dummy_var_ locals in batch_norm_stats_cuda_template that it flags. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
175525,vmap compatibility with non-tensor leaves,https://github.com/pytorch/pytorch/pull/175525,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,175521,175521,175521,,,,0,6,0,,,"triaged, open source, Stale",2026-02-23T10:10:21Z,2026-06-23T18:28:04Z,15,Closes #175521
187731,Fix extract_scripts step-index zero-padding off-by-one,https://github.com/pytorch/pytorch/pull/187731,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-19T15:48:27Z,2026-06-23T16:59:39Z,15,"extract_scripts.py numbers each extracted step with a 1-based index (enumerate(steps, start=1)), so the indices run 1..len(steps). The zero-pad width, however, is computed from len(steps) - 1: index_chars = len(str(len(steps) - 1)) for i, step in enumerate(steps, start=1): ... filename = f""{i:0{index_chars}}{sanitiz..."
178069,[docs] Add examples to reduce_scatter documentation,https://github.com/pytorch/pytorch/pull/178069,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, release notes: distributed (sharded)",2026-03-21T19:46:47Z,2026-06-23T15:36:42Z,15,"Add Examples section to reduce_scatter function docstring, matching the format used in all_gather. Includes examples for both int64 and complex tensor types demonstrating the reduce-scatter operation across ranks. Also fix misleading ""We have 2 process groups, 2 ranks"" comment to ""We have 2 ranks"" across distributed..."
178068,[docs] Add comprehensive introduction to PyTorch Elastic documentation,https://github.com/pytorch/pytorch/pull/178068,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, Stale, topic: not user facing",2026-03-21T19:46:24Z,2026-06-23T15:36:40Z,15,Summary This PR adds a comprehensive introduction section to the PyTorch Elastic documentation that explains: What PyTorch Elastic is and its purpose The two key capabilities: fault tolerance and elasticity When to use PyTorch Elastic vs standard distributed training Clear explanations of the benefits for production...
185882,[PrivateUse1] Return `False` instead of `None` when PU1 backend is not available,https://github.com/pytorch/pytorch/pull/185882,PR blocked,108,failed checks or blocked/unstable merge state; approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing, module: PrivateUse1",2026-06-02T03:35:52Z,2026-06-23T15:23:08Z,15,"_is_privateuse1_backend_available() will return None if PU1 backend is not available, which is not expected. It should return False . Reference (Failed CI from my test refactor PR, has TEST_PRIVATEUSE1 in it, https://github.com/pytorch/pytorch/actions/runs/26794612675/job/78988996179?pr=185798): Traceback (most rece..."
187726,Expose public API for torch.cuda.clear_cublas_workspaces,https://github.com/pytorch/pytorch/pull/187726,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,184084,184084,184084,180283,,,0,3,0,,,"triaged, open source, release notes: distributed (fsdp), module: inductor, module: dynamo, ciflow/inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-06-19T14:31:21Z,2026-06-23T15:22:42Z,15,"Fixes: #184084 Followed these PR's: #180283, #176705, #143799 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98"
186929,[fx] fix _patch_function using wrong function when forward is wrapped,https://github.com/pytorch/pytorch/pull/186929,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,"149497, 163312, 171981","149497, 163312, 171981","149497, 163312, 171981",,,,0,2,0,,,"triaged, open source, release notes: fx",2026-06-10T12:46:50Z,2026-06-23T14:29:39Z,15,"Summary create_args_for_root uses inspect.unwrap(root_fn) to get fn_for_analysis (the innermost callable) and derives argument count and names from its code object. When forward has variadic arguments (*args or **kwargs), it then calls _patch_function(root_fn, len(args)) — but root_fn is the outer wrapper, whose co_..."
187881,Fix margin docstring rendering in HingeEmbeddingLoss,https://github.com/pytorch/pytorch/pull/187881,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-22T21:44:02Z,2026-06-23T14:26:10Z,15,"The margin argument in HingeEmbeddingLoss used a single backtick (1), which Sphinx renders as an italic title-reference rather than inline math. This makes it inconsistent with the sibling loss classes (MultiMarginLoss, MarginRankingLoss, TripletMarginLoss, CosineEmbeddingLoss), which all use :math:1``. This aligns..."
187809,Make DTensor logging and export tests accelerator-agnostic,https://github.com/pytorch/pytorch/pull/187809,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/dtensor",2026-06-22T01:57:37Z,2026-06-23T14:25:23Z,15,"Summary This PR decouples two DTensor test files from CUDA-only execution: test/distributed/tensor/test_dtensor_logging.py test/distributed/tensor/test_dtensor_export.py The tests previously used @requires_cuda and set self.device_type = ""cuda"". That made them unavailable to XPU, NPU, and other PrivateUse1 accelerat..."
187764,[MPS] Add matrix_exp support,https://github.com/pytorch/pytorch/pull/187764,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,,173247,141287,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend, ciflow/mps",2026-06-20T08:39:00Z,2026-06-23T14:24:24Z,15,"[MPS] Add linalg.matrix_exp / matrix_exp support for the MPS backend Summary torch.linalg.matrix_exp (and its alias torch.matrix_exp) currently raise NotImplementedError on the MPS backend, forcing a CPU fallback. This PR enables them natively on MPS. linalg_matrix_exp is already a device-generic composite. It compu..."
187816,Clean up mentions of deprecated nvidia-docker in README files,https://github.com/pytorch/pytorch/pull/187816,ready for maintainer decision,45,CI lane labels: ciflow/docker; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing, ciflow/docker",2026-06-22T05:46:20Z,2026-06-23T14:23:49Z,15,nvidia-docker has been deprecated as of Docker version 19.03 (2019) this PR removes or replaces confusing mentions of nvidia-docker in the README files
187847,fix(nn.utils.prune): don't crash adding a method to an empty PruningContainer,https://github.com/pytorch/pytorch/pull/187847,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-22T15:49:36Z,2026-06-23T14:22:49Z,15,"Summary PruningContainer crashes when built incrementally. __init__ only assigns self._tensor_name in the single-method branch (PruningContainer(method)); for an empty container — PruningContainer() followed by add_pruning_method(...) — _tensor_name is never set, so add_pruning_method raises AttributeError: 'Pruning..."
187814,Reject factor=0 in ConstantLR to avoid later ZeroDivisionError,https://github.com/pytorch/pytorch/pull/187814,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,185006,185006,185006,,,,0,1,0,,,"triaged, open source, release notes: optim",2026-06-22T03:49:01Z,2026-06-23T14:22:29Z,15,"Fixes #185006 Summary ConstantLR validated its multiplicative factor with factor < 0, so factor=0 was accepted at construction time. Once total_iters is reached, get_lr() scales the learning rate by 1.0 / factor, so a zero factor surfaced as a bare ZeroDivisionError several steps later — far from the offending argum..."
187796,Add native MPS SVD path,https://github.com/pytorch/pytorch/pull/187796,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-21T17:41:34Z,2026-06-23T14:21:30Z,15,"Summary Route supported MPS SVD inputs through existing MPS linalg_eigh and matmul operations so small/medium SVDs avoid generic CPU fallback. Keep controlled CPU fallback for matrices outside the current MPS staging limits. Add focused MPS SVD coverage, including pca_lowrank regression coverage for the small low-ra..."
187757,Narrow ModuleDict.__getattr__ return type to Module for type checkers,https://github.com/pytorch/pytorch/pull/187757,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,146938,146938,146938,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-06-20T02:07:36Z,2026-06-23T14:21:26Z,15,"Fixes #146938 Summary nn.ModuleDict does not override __getattr__, so attribute access such as module_dict.foo inherits Module.__getattr__'s annotated return type Union[Tensor, Module]. A type checker then reports false positives on otherwise-valid code, for example: mods = nn.ModuleDict({""foo"": nn.Linear(4, 8)}) y..."
187800,Verify Magma archive before extraction,https://github.com/pytorch/pytorch/pull/187800,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-21T18:51:07Z,2026-06-23T14:21:22Z,15,Summary: Download the Magma source archive over HTTPS Check the existing SHA256 file before extracting the archive Extract only after the checksum passes Test Plan: bash -n .ci/magma/build_magma.sh
187742,Fix calculate_out_dimension using stride[0] for the width dimension,https://github.com/pytorch/pytorch/pull/187742,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: fx",2026-06-19T20:39:06Z,2026-06-23T14:21:17Z,15,"In graph_gradual_typechecker.calculate_out_dimension, padding, dilation and kernel_size are all indexed by the index argument (0 for height, 1 for width), but the stride is hardcoded to stride[0]: n = d_in + 2 * padding[index] - dilation[index] * (kernel_size[index] - 1) - 1 return (n // stride[0]) + 1 The function..."
187799,Fix ROCm libdrm readlink buffer bounds,https://github.com/pytorch/pytorch/pull/187799,ready for maintainer decision,45,CI lane labels: ciflow/docker; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, triaged, open source, topic: not user facing, ciflow/docker",2026-06-21T18:44:05Z,2026-06-23T14:20:51Z,15,Summary: Leave room for the terminating NUL byte when the ROCm libdrm patch reads /proc/self/exe Bound readlink() with sizeof(self_path) - 1 so self_path[count] remains in bounds Test Plan: bash -n .ci/docker/common/install_rocm_drm.sh cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hon...
187769,[docs] Clarify Softplus overflow behavior for reduced-precision dtypes,https://github.com/pytorch/pytorch/pull/187769,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,187180,,,,0,1,0,,,"triaged, open source",2026-06-20T14:12:58Z,2026-06-23T14:00:44Z,15,"Summary Adds a note to the nn.Softplus documentation describing overflow behavior for reduced-precision floating-point dtypes. When the result exceeds the representable range of the output dtype (for example, torch.float16), conversion to the output dtype may overflow and produce inf according to IEEE floating-point..."
187790,[ao][pruning] Add experimental topology-aware modules,https://github.com/pytorch/pytorch/pull/187790,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187781,187781,187781,,,,0,1,0,,,"triaged, open source, release notes: quantization",2026-06-21T11:44:00Z,2026-06-23T14:00:31Z,15,Summary Fixes #187781. This PR addresses the topology-aware memory optimization prototype by adding an experimental torch.ao.pruning._experimental.topology package with four guarded training-time memory primitives: TopologyGatedLowRankLinear: starts from an exact dense nn.Linear copy and switches to low-rank factors...
187783,fix(nn): raise ValueError instead of ZeroDivisionError for norm_type=0 in lp_pool1d/2d/3d and LPPool modules,https://github.com/pytorch/pytorch/pull/187783,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187743,187743,,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-06-21T07:45:11Z,2026-06-23T13:59:01Z,15,"Summary Fixes #187743 F.lp_pool1d, F.lp_pool2d, and F.lp_pool3d all compute .pow(1.0 / norm_type) in their return expression. When norm_type=0 this produces a bare ZeroDivisionError: division by zero with no indication of which argument is invalid or what the valid range is. nn.LPPool1d / LPPool2d / LPPool3d compoun..."
187748,[MPS] Build only the reduced Q in linalg.qr (fix tall-skinny perf cliff),https://github.com/pytorch/pytorch/pull/187748,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,187567,187567,187567,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-20T00:22:33Z,2026-06-23T13:58:56Z,15,"Fixes #187567 Summary torch.linalg.qr on MPS scaled super-linearly with the row dimension for tall-skinny inputs: at (4000, 36) it was ~250x slower than CPU, and even a MPS -> CPU -> QR -> MPS round trip beat the native path by 47-170x. Root cause The Householder kernel always materialized the full m x m Q and accum..."
187717,[CUDA] solve_triangular: fix silently wrong batched results and slowness in large-RHS workaround,https://github.com/pytorch/pytorch/pull/187717,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend, topic: bug fixes",2026-06-19T07:31:39Z,2026-06-23T13:58:43Z,15,"Summary PR #117636 added a workaround for a cuBLAS crash on CUDA < 12.1 in torch.linalg.solve_triangular when B.size(-1) >= 524281 (#79191). This PR refines that implementation to address three unexpected side effects: the workaround triggers until CUDA 13.1, returns silently wrong results on batched inputs, and is..."
187153,jit: add bounds checking for textIndex in SourceRangeDeserializer,https://github.com/pytorch/pytorch/pull/187153,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,185441,185441,185441,,,,0,1,0,,,"triaged, open source, release notes: jit",2026-06-12T06:15:18Z,2026-06-23T13:57:55Z,15,Summary Fixes #185441. SourceRangeDeserializer::deserialize_source() reads a serialized IntList (textIndex) from a .pt file and uses each element as a direct index into text_table_ with no bounds validation. A crafted .pt file with out-of-bounds or negative indices in textIndex causes an out-of-bounds dereference an...
187921,Support freeing blocks with stream uses during graph capture,https://github.com/pytorch/pytorch/pull/187921,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-23T05:34:38Z,2026-06-23T13:35:14Z,15,"During XPU graph capture, inserting events and querying event status are illegal operations. However, user code may still free a block that has non-empty stream_uses while capture is underway. The previous code path would unconditionally call insert_events() in this case, which is invalid during graph recording. Thi..."
187357,Convert _scaled_grouped_mm_v2 to a structured operator,https://github.com/pytorch/pytorch/pull/187357,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-06-15T17:03:57Z,2026-06-23T13:29:24Z,15,"Stack from ghstack (oldest at bottom): -> #187357 Mirrors the recently-landed _scaled_mm_v2 work (#185271/#185272/#185273) for the grouped sibling. The shared dispatch-table helper from that stack already covers the grouped path, so this is the structured-op conversion plus validation centralization. Recipe-independ..."
186033,[CD] Enable Python 3.15 & 3.15t Windows wheel builds and Triton XPU Windows build,https://github.com/pytorch/pytorch/pull/186033,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,184352,,,,0,3,0,,,"open source, release notes: releng, ciflow/binaries_wheel",2026-06-03T03:49:02Z,2026-06-23T13:24:37Z,15,"Summary Enable Python 3.15/3.15t CD wheel builds on Windows (cpu, cuda, xpu) Add 3.15/3.15t to Windows Triton XPU wheel build matrix Add 3.15/3.15t installer support in install_python.bat (using python-3.15.0b2-amd64.exe) Works for #184352 Changes .ci/pytorch/windows/internal/install_python.bat — handle 3.15 (beta U..."
187570,Fixed heaviside nan propagation issue,https://github.com/pytorch/pytorch/pull/187570,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187296,187296,187296,,,,0,2,0,,,"module: cpu, triaged, open source",2026-06-17T13:53:12Z,2026-06-23T13:20:10Z,15,"Summary Fix torch.heaviside to propagate NaN inputs, matching NumPy behavior. Previously, torch.heaviside returned 0 for NaN inputs because NaN comparisons (a == 0 and a > 0) evaluate to false, causing the kernel to fall through to the 0 case. This change updates the CPU and CUDA kernels to preserve NaN values and a..."
187694,Unpack Python autograd saved tensors in reverse order,https://github.com/pytorch/pytorch/pull/187694,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-18T19:54:17Z,2026-06-23T13:12:02Z,15,"This PR is proposed for community discussion. The change affects observable hook call order, so I am open to feedback on whether this behavior is acceptable, whether it should be guarded by a new API, or whether another mechanism would better support offload/prefetch use cases. Problem Saved tensor hooks are often u..."
187529,perf(dynamo): add guard lookup token plan,https://github.com/pytorch/pytorch/pull/187529,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,15,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-16T23:57:55Z,2026-06-23T08:56:22Z,15,"Summary This PR proposes a conservative guard lookup fast path for stable L['self']._modules guard subtrees in TorchDynamo. The implementation records a Last Successful Guard Memo for each cache entry, trains a token plan from repeated successful guard evaluations, and reuses that plan only when the current lookup s..."
187309,Add maximize parameter to LBFGS optimizer,https://github.com/pytorch/pytorch/pull/187309,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,126642,126642,126642,,,,0,1,0,,,"triaged, open source, release notes: optim",2026-06-15T06:40:15Z,2026-07-01T01:54:07Z,8,"Fixes #126642 LBFGS is the only standard PyTorch optimizer without maximize support. A prior attempt (PR #81098) was closed because it only negated the gradient in one code path, causing the line search to fail. Unlike simpler optimizers (SGD, Adam) where negating the gradient is sufficient, LBFGS also requires nega..."
181218,FSDP2 chunked-loss ROCm race: stream-level wait_event at post_backward exit,https://github.com/pytorch/pytorch/pull/181218,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/rocm-mi300, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,10,0,,,"module: rocm, Stale, release notes: distributed (fsdp), ciflow/inductor, ciflow/rocm-mi300, ciflow/torchtitan",2026-04-23T06:17:44Z,2026-06-23T05:53:04Z,15,"Stack from ghstack (oldest at bottom): -> #181218 Summary: Closes the reduce-scatter race in the grouped chunked-loss path on ROCm (MI350X, RCCL). When a standalone per-chunk head call triggers post_backward, chunk N+1's autograd backward runs on the default stream concurrently with chunk N's accumulate on the RS st..."
181269,[draft] refactor linux workflows,https://github.com/pytorch/pytorch/pull/181269,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,28,0,,,"Stale, release notes: releng",2026-04-23T18:55:30Z,2026-06-23T05:53:03Z,15,
187597,Add return type annotation to nn.Module.named_modules,https://github.com/pytorch/pytorch/pull/187597,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,166905,166905,166905,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-17T19:15:51Z,2026-06-23T03:43:47Z,16,"Fixes #166905 nn.Module.named_modules() had no return type annotation, while its siblings named_children, named_parameters, and named_buffers are annotated, so type checkers inferred an untyped result for named_modules(). This adds -> Iterator[tuple[str, ""Module""]] to match named_children. Iterator is already import..."
187642,Add parameter-contiguous all-gather (symm_mem.all_gather_offset),https://github.com/pytorch/pytorch/pull/187642,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,187239,187750,,,0,6,0,,,"open source, Merged, Reverted, ciflow/trunk, release notes: distributed (c10d), ci-no-td, ciflow/h100-symm-mem, module: symm_mem, ciflow/torchtitan",2026-06-18T05:17:15Z,2026-06-23T03:28:37Z,16,"Stack from ghstack (oldest at bottom): #187750 -> #187642 Implements the parameter-contiguous all-gather requested in #187239, using the symmetric-memory approach preferred in the issue discussion. FSDP2 today all-gathers parameter shards in rank-major layout and then reorders them into a parameter-contiguous layout..."
187815,[Inductor][XPU] Unfuse addmm for all half-dtype inputs on XPU,https://github.com/pytorch/pytorch/pull/187815,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T05:28:40Z,2026-06-23T02:58:19Z,16,"The narrow-cast unfuse approach (PR #183680) only helps when bias is explicitly cast via convert_element_type, missing the common case where bias is already half dtype (e.g., AMP). On XPU, prefer to unfuse all half-dtype addmm so the pointwise kernel absorbs the bias add directly. CUDA/ROCm continue to keep addmm fu..."
187896,Re-land chain-based dispatch for custom_op fast path with override detection,https://github.com/pytorch/pytorch/pull/187896,draft/noise,23,PR is draft; PR closes or references open issue numbers,wait for author to mark ready,,UNKNOWN,True,0,0,,,177109,177109,177109,183508,,,0,2,0,,,open source,2026-06-23T01:31:24Z,2026-06-23T01:36:15Z,16,"Fixes #177109 Re-lands #183508 (chain-based dispatch for custom_op fast path) with a fix for the internal revert: if someone overrides a custom_op's kernel via TORCH_LIBRARY_IMPL or Library.impl, the fast path now detects it and falls back to the C++ dispatcher. Detection uses a per-operator impl_generation_ counter..."
179149,[Inductor][RFC] Symbolic Analysis of User-Defined Triton Kernels,https://github.com/pytorch/pytorch/pull/179149,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, Stale, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-02T15:39:23Z,2026-06-23T01:21:13Z,16,"cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @jansel Motivation Epilogue fusion for user-defined kernels (#173662) is currently scoped to UB tensors, with the rati..."
186880,Add FlexAttention shared memory compile hints,https://github.com/pytorch/pytorch/pull/186880,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186882,,,0,7,0,,,"topic: not user facing, module: inductor, ciflow/inductor, module: flex attention, ciflow/torchtitan",2026-06-10T02:56:57Z,2026-06-23T01:09:16Z,16,Stack from ghstack (oldest at bottom): #186882 -> #186880 Triton shared-memory failures can surface from the normal compile/launcher path before a FlexAttention kernel reaches runtime. I want to capture these errors and make them nicer / push people to unblock themselves. example error cc @voznesenskym @penguinwu @E...
187892,"[profiler][cupti] Perfetto-native (.pftrace) export, full chrome-args parity",https://github.com/pytorch/pytorch/pull/187892,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"186655, 187517, 187518, 187519, 187520",,,0,1,0,,,,2026-06-23T00:28:21Z,2026-06-23T00:28:36Z,16,"Stack from ghstack (oldest at bottom): -> #187892 #187520 #187519 #187517 #187518 #186655 #186802 #186812 #187874 #186855 #186439 #187515 #186811 export_chrome_trace(path) with a .pftrace path makes the cupti_monitor backend emit a Perfetto-native trace instead of chrome JSON, encoded straight from the observer's co..."
181252,[Codemod][MigrateInternmcUrls] fbcode/caffe2 (#181252),https://github.com/pytorch/pytorch/pull/181252,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,181252,,,,,0,1,0,,,"open source, fb-exported, Stale, ciflow/trunk, topic: not user facing, meta-exported",2026-04-23T15:29:10Z,2026-06-22T23:54:49Z,16,Summary: Pull Request resolved: #181252 Reviewed By: dtolnay Differential Revision: D102125625
187750,Tune all_gather_offset for skewed buckets,https://github.com/pytorch/pytorch/pull/187750,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,187642,,,0,1,0,,,"open source, Merged, Reverted, ciflow/trunk, release notes: distributed (c10d), ci-no-td, ciflow/h100-symm-mem, module: symm_mem",2026-06-20T00:46:40Z,2026-06-22T23:54:08Z,16,"Stack from ghstack (oldest at bottom): -> #187750 #187642 The base op drives each (parameter, peer) shard with a single CTA, so a bucket of few large shards leaves most of the GPU idle, while both the LSA-push and multimem bandwidth scale with the CTA count. This flattens the work into fixed-size byte tiles (at leas..."
187776,Speed up all_gather_offset writes with a shared unrolled copy helper,https://github.com/pytorch/pytorch/pull/187776,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,187778,,,0,2,0,,,"open source, Merged, Reverted, ciflow/trunk, release notes: distributed (c10d), ci-no-td, ciflow/h100-symm-mem, module: symm_mem",2026-06-20T18:03:10Z,2026-06-22T23:28:03Z,16,"Stack from ghstack (oldest at bottom): #187778 -> #187776 Factor the 16-byte-vectorized copy used by nccl_all_to_all_nd into a shared header (symm_mem_copy.cuh) and use it for the all_gather_offset push kernel, which previously copied each tile with a plain ld_vec/st_vec grid-stride loop. The helper batches kUnroll=..."
184398,Enable 3.15 in CI,https://github.com/pytorch/pytorch/pull/184398,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,184352,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-05-19T16:29:49Z,2026-06-22T23:27:15Z,16,Part of resolving #184352 cc @malfet @seemethere @pytorch/pytorch-dev-infra @albanD
176754,[CUDA Graph] Thread-safe RNG state for concurrent graph captures,https://github.com/pytorch/pytorch/pull/176754,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,22,0,,,"module: cuda, open source, module: cuda graphs, topic: not user facing",2026-03-06T22:48:26Z,2026-06-22T23:27:13Z,16,Stack from ghstack (oldest at bottom): -> #176754 #176753 Add proper thread safety to per-capture RNG state management: Double-checked locking in get_capture_state() with seed_ read under lock to prevent races with set_current_seed() from another thread Thread-safe replay_prologue() with mutex around seed/offset rea...
181518,[Inductor] enable MSVC precompiled headers for cpp-wrapper,https://github.com/pytorch/pytorch/pull/181518,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-04-26T16:32:35Z,2026-06-22T23:14:35Z,16,This PR fixes Windows-specific failures in Inductor cpp-wrapper builds and adds end-to-end MSVC precompiled header support in the C++ compile path. What changed Fixed CPU cpp-wrapper header emission so required device declarations are still included when graph device types are empty. Enabled Windows PCH flow in the...
183508,Chain-based dispatch for custom_op fast path,https://github.com/pytorch/pytorch/pull/183508,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,139500,"139500, 183508","139500, 177109",,,,0,59,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, ci-no-td",2026-05-13T03:55:43Z,2026-06-22T22:53:24Z,16,"Addresses #177109. Supersedes #178216. Fixes #139500 Adds a Python fast path for common custom_op eager calls. Instead of inlining the autograd/ADInplaceOrView/backend logic, the fast path chains through the existing autograd_impl and op.redispatch, with a thread-local chain that intercepts redispatch calls: fast_ca..."
181278,[spmd_types][unused] local_map for spmd_types tensors,https://github.com/pytorch/pytorch/pull/181278,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,181199,,,0,1,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-04-23T19:55:31Z,2026-06-22T22:13:56Z,16,Stack from ghstack (oldest at bottom): -> #181278 #181199 #180880
174395,Add optional out argument to F.scaled_mm,https://github.com/pytorch/pytorch/pull/174395,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,14,0,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-02-05T17:24:27Z,2026-06-22T22:12:29Z,16,"Stack from ghstack (oldest at bottom): -> #174395 Summary: Add out= argument to F.scaled_mm and basic test Properly guard MXFP4 tests where the build has CUDA but not MSLK. Gracefully refuse to run instead of hard-failing. Reland of the March 2026 merge, reverted by autorevert for breaking test_fx.py::TestFunctional..."
181265,TEMP,https://github.com/pytorch/pytorch/pull/181265,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-23T18:15:53Z,2026-06-22T21:25:13Z,16,Stack from ghstack (oldest at bottom): -> #181265 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
185446,Adding a test case using fx.wrap to test whether TLX user-defined kernel is visible in the inductor output (#185301),https://github.com/pytorch/pytorch/pull/185446,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-27T23:14:27Z,2026-06-22T20:20:57Z,16,"Summary: This diff is dependent on D96405491, where I added a test case to check whether customized kernel is fused in inductor. This diff added one more test case to it but instead of using torch.library.custom_op to wrap the kernel, it uses fx.wrap to make the kernel visible in inductor output. Test Plan: With the..."
187258,[aot_compile] Add torch.compiler.precompile,https://github.com/pytorch/pytorch/pull/187258,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-13T20:46:53Z,2026-06-22T20:16:20Z,16,"Summary Adds torch.compiler.precompile, an ahead-of-time precompile API that returns python code + serializable kernel cache that can be used to bypass kernel JIT compilation. # fn is the whole computation -- here a training step: forward, loss, backward. train_step = lambda model, x, target: loss_fn(model(x), targe..."
187134,[dynamo] Reuse ACT-traced graphs for resolved tensor inputs,https://github.com/pytorch/pytorch/pull/187134,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, release notes: distributed (c10d), module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-12T02:28:38Z,2026-06-22T19:50:06Z,16,"Summary Relax Dynamo's AsyncCollectiveTensor guard path so a graph compiled with a top-level ACT input can be reused when the runtime value is the already-resolved plain tensor. Guard the resolved tensor through a dedicated AsyncCollectiveTensorSource, while still recompiling for unrelated tensor subclasses and for..."
177985,Fix SDPA mem-efficient attention crash when num_heads >= 65536,https://github.com/pytorch/pytorch/pull/177985,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,142228,142228,142228,,,,0,2,0,,,"triaged, open source, release notes: nn",2026-03-20T17:32:25Z,2026-06-22T19:18:16Z,16,"Summary Fixes #142228 The memory-efficient attention kernel maps num_heads and num_batches directly to CUDA grid.y and grid.z dimensions, which are each limited to 65,535 blocks. When num_heads exceeds this limit the kernel launch silently fails, producing incorrect results or crashes. Existing batch chunking in _sc..."
187840,[inductor] Add allow_peak_memory_increasing_fusion config,https://github.com/pytorch/pytorch/pull/187840,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187675, 187839, 187841, 187842, 187843, 187844, 187845, 187846",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T15:09:33Z,2026-06-22T19:03:50Z,16,Stack from ghstack (oldest at bottom): #187846 #187845 #187844 #187843 #187842 #187841 -> #187840 #187839 #187675 Add a single config switch for the peak-memory-aware choices in this stack. The flag defaults to True so the old/aggressive fusion and memory-planning behavior remains unchanged unless tests or users exp...
187841,[inductor] Avoid disjoint horizontal fusion when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/187841,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187675, 187839, 187840, 187842, 187843, 187844, 187845, 187846",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T15:09:37Z,2026-06-22T19:03:46Z,16,Stack from ghstack (oldest at bottom): #187846 #187845 #187844 #187843 #187842 -> #187841 #187840 #187839 #187675 Add a horizontal-fusion guard for branches that only share a small amount of input data but would both need to materialize outputs after fusion. The guard is active only when allow_peak_memory_increasing...
187839,[inductor] Resolve ProcessGroup attrs in overlap estimates,https://github.com/pytorch/pytorch/pull/187839,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187675, 187840, 187841, 187842, 187843, 187844, 187845, 187846",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T15:09:29Z,2026-06-22T19:03:41Z,16,"Stack from ghstack (oldest at bottom): #187846 #187845 #187844 #187843 #187842 #187841 #187840 -> #187839 #187675 Functionalized eager collectives can pass a ProcessGroup get_attr node to _c10d_functional ops. The overlap scheduler's collective estimator resolved only string group names or FX nodes with meta[""val""],..."
187843,[inductor] Avoid late-output vertical fusion when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/187843,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187675, 187839, 187840, 187841, 187842, 187844, 187845, 187846",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T15:09:45Z,2026-06-22T19:03:38Z,16,Stack from ghstack (oldest at bottom): #187846 #187845 #187844 -> #187843 #187842 #187841 #187840 #187839 #187675 Add a producer-consumer fusion guard for cases where fusing a producer into one consumer would also materialize later outputs before an earlier external user of the producer output. The guard is active o...
187846,[inductor] Pick low-pressure buffer reuse when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/187846,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187675, 187839, 187840, 187841, 187842, 187843, 187844, 187845",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T15:09:58Z,2026-06-22T19:03:37Z,16,Stack from ghstack (oldest at bottom): -> #187846 #187845 #187844 #187843 #187842 #187841 #187840 #187839 #187675 Change wrapper memory planning to optionally score recent reusable buffers by the peak pressure between the free site and the allocation site. This path is active only when allow_peak_memory_increasing_f...
187845,[inductor] Prefer successor-unblocking LPMF when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/187845,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187675, 187839, 187840, 187841, 187842, 187843, 187844, 187846",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T15:09:54Z,2026-06-22T19:03:30Z,16,"Stack from ghstack (oldest at bottom): #187846 -> #187845 #187844 #187843 #187842 #187841 #187840 #187839 #187675 Make the LPMF scheduler optionally prefer a ready node that unlocks the earliest successor when all ready nodes are large. This path is active only when allow_peak_memory_increasing_fusion=False, so the..."
187844,[inductor] Realize reused pointwise add when allow_peak_memory_increasing_fusion is off,https://github.com/pytorch/pytorch/pull/187844,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187675, 187839, 187840, 187841, 187842, 187843, 187845, 187846",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-22T15:09:49Z,2026-06-22T17:48:12Z,16,"Stack from ghstack (oldest at bottom): #187846 #187845 -> #187844 #187843 #187842 #187841 #187840 #187839 #187675 Realize a reused aten.add pointwise result before reduction-like branches in peak-aware mode. The behavior is disabled when allow_peak_memory_increasing_fusion=True, so the default path preserves the pre..."
183697,Fix SVE blendv mask handling after bool fill,https://github.com/pytorch/pytorch/pull/183697,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,177244,,,,,0,3,0,,,"module: cpu, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor",2026-05-14T10:11:08Z,2026-06-22T16:19:01Z,16,"Stack from ghstack (oldest at bottom): -> #183697 This fixes the remaining SVE SDPA rewriter failures after the VecMask bit-cast fix. The failing path comes from compiled CPU vector code for boolean mask operations such as: mask.masked_fill(torch.logical_not(mask), -float(""inf"")).to(torch.float32) On SVE, blendv was..."
166876,[ci][cpu] Update AArch64 manylinux compiler to GCC14,https://github.com/pytorch/pytorch/pull/166876,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/docker, ciflow/linux-aarch64, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,168288,"152426, 168288",,,,0,32,0,,,"module: cpu, module: mkldnn, open source, module: arm, Stale, ciflow/trunk, topic: not user facing, ciflow/linux-aarch64, ciflow/docker",2025-11-03T18:33:50Z,2026-06-22T16:03:37Z,16,"Stack from ghstack (oldest at bottom): -> #166876 Realted to: #152426 Fixes: #168288 Updates manylinux compiler to gcc14 as per manylinux requirements: https://github.com/pypa/manylinux?tab=readme-ov-file#manylinux_2_28-almalinux-8-based Updates ubuntu for non-manylinux builds from jammy to noble, and updates gcc fo..."
178373,docs(nn): clarify return type for MaxPool1d/2d/3d when return_indices=True,https://github.com/pytorch/pytorch/pull/178373,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-03-25T08:17:15Z,2026-06-22T15:24:46Z,16,"Summary MaxPool1d, MaxPool2d, and MaxPool3d all support a return_indices parameter, but the documentation only mentions it briefly in the Args section. Users who aren't already familiar with the behavior may not realize: The return type changes from a Tensor to a tuple (output, indices) indices has the same shape as..."
172727,Fix torch.topk out variant recompilations by adding meta function for…,https://github.com/pytorch/pytorch/pull/172727,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,135859,135859,135859,,,,0,11,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-01-17T17:16:38Z,2026-06-22T15:22:24Z,16,"Fixes #135859 Problem When using torch.topk with the out= parameter in torch.compile, PyTorch was experiencing excessive recompilations due to a missing meta function for the aten.topk.values operator variant. Root Cause The issue occurred because PyTorch's compilation system lacked a meta function to handle dynamic..."
180986,Enable nested_compile_region for AOTInductor + tensor-subclass params,https://github.com/pytorch/pytorch/pull/180986,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-04-21T15:01:25Z,2026-06-22T15:18:58Z,16,"Enable nested_compile_region for AOTInductor + widen tensor-subclass support Three small, independent fixes that together make torch.compiler.nested_compile_region work with AOTInductor and incrementally widen the invoke_subgraph codepath's handling of traceable wrapper tensor subclasses (e.g. the torchao-quantized..."
187785,[TokenSwitch] Add symmetric-memory zero-copy dispatch/combine,https://github.com/pytorch/pytorch/pull/187785,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,187784,,,0,5,0,,,"open source, release notes: distributed (c10d), ciflow/h100-symm-mem, module: symm_mem",2026-06-21T08:45:58Z,2026-06-22T15:16:43Z,16,"Stack from ghstack (oldest at bottom): -> #187785 #187784 When a dispatch/combine tensor is backed by NCCLSymmetricMemory registered under the process group's name, address it via the registered ncclWindow + offset (RDMA zero-copy path) instead of a plain device pointer, avoiding a staging copy into a regular device..."
181557,[XPU] Enable grouped_mm and scaled_grouped_mm with OneDNN,https://github.com/pytorch/pytorch/pull/181557,draft/noise,13,"PR is draft; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,2,0,,,"module: cpu, module: mkldnn, open source, ciflow/inductor, ciflow/rocm-mi300, release notes: inductor (aoti), ciflow/h100, ciflow/b200",2026-04-27T08:16:55Z,2026-06-22T15:09:20Z,16,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @gujinghui @PenghuiCheng @jianyuh @min-jean-cho @yanbing-j @Guobing-Chen @Xia-Weiwen @snadampal
187835,[profiler] Enable OVERHEAD activity type for the XPU backend,https://github.com/pytorch/pytorch/pull/187835,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,open source,2026-06-22T14:09:58Z,2026-06-22T14:21:55Z,16,"Add ActivityType::OVERHEAD to kXpuTypes so that profiling with ProfilerActivity.XPU enables PTI's collection-overhead view, matching the CUDA backend (kCudaTypes already lists OVERHEAD). The libkineto xpupti plugin already handles the records; this was the only missing wiring on the PyTorch side. OVERHEAD events ren..."
187771,Warn when convert_sync_batchnorm drops _BatchNorm subclass behavior,https://github.com/pytorch/pytorch/pull/187771,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187298,187298,187298,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-20T15:51:37Z,2026-06-22T13:36:13Z,16,"Fixes #187298 Summary torch.nn.SyncBatchNorm.convert_sync_batchnorm rebuilds every _BatchNorm instance as a bare SyncBatchNorm, copying only the BatchNorm attributes. If the original was a _BatchNorm subclass that added behavior — an overridden forward that fuses an activation (e.g. timm's BatchNormAct2d) or extra c..."
187198,Add target range validation in binary_cross_entropy_with_logits,https://github.com/pytorch/pytorch/pull/187198,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/trunk",review when ready,,UNKNOWN,False,0,0,,,,123339,123339,,,,0,1,0,,,"triaged, open source, ciflow/trunk, release notes: nn, ciflow/dtensor",2026-06-12T18:33:54Z,2026-06-22T13:01:20Z,16,"Fixes: #123339 BCELoss raises RuntimeError: all elements of target should be between 0 and 1 for out-of-range targets, but BCEWithLogitsLoss silently accepts them and returns a meaningless (possibly negative) loss, even though the documentation of both losses requires targets in [0, 1]: >>> torch.nn.BCELoss()(torch...."
178568,Add support for activation checkpointing with nested data structures,https://github.com/pytorch/pytorch/pull/178568,ready for maintainer decision,45,CI lane labels: ciflow/torchtitan; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"module: activation checkpointing, triaged, open source, Stale, release notes: autograd, ciflow/torchtitan",2026-03-27T05:41:31Z,2026-06-22T11:27:58Z,16,"This PR fixes long-standing behavior where activation checkpointing via torch.utils.checkpoint.checkpoint non-reentrant mode only works when arguments consist purely of tensors. Despite several parts of the documentation mentioning that non-reentrant mode supports nested structures, I show that it is not the case in..."
181221,[DCP] Fix flaky test_async_process_executor by resetting global state between tests (#181221),https://github.com/pytorch/pytorch/pull/181221,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, topic: not user facing, release notes: distributed (checkpoint), meta-exported",2026-04-23T06:43:37Z,2026-06-22T11:23:21Z,16,"Summary: The test_checkpoint_save_failure_continues_serving test was disabled due to persistent flakiness (75.7% failure rate). Two root causes: The module-level _CHECKPOINT_PROCESS singleton was not reset between test classes. When TestAsyncProcessExecutorPrefixStore ran before TestAsyncProcessExecutor, it left the..."
180754,[ROCm] Fix decorator for skipping FP8 tests,https://github.com/pytorch/pytorch/pull/180754,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi200","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, Stale, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/rocm-mi200",2026-04-19T02:59:22Z,2026-06-22T11:23:18Z,16,Test is failing on MI200s when it should be skipped: https://github.com/pytorch/pytorch/actions/runs/24613964933/job/71974245080 cc @jeffdaily @sunway513 @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @Xiaob...
180184,Export auto-generated unstructured inplace meta dispatch symbols,https://github.com/pytorch/pytorch/pull/180184,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,167818,167818,167818,,,,0,2,0,,,"triaged, open source, Stale, topic: not user facing, module: PrivateUse1",2026-04-12T19:01:12Z,2026-06-22T11:23:16Z,16,"Fixes #167818 Unstructured inplace ops (like index_put_impl) get auto-generated meta kernels in gen_unstructured(), but the operator_headers() function in gen.py didn't account for these when deciding which dispatch headers to include in RegisterMeta_N.cpp. This meant the TORCH_API declaration was never seen by the..."
181189,Tensor.record_use: precise cross-stream lifetime for the CUDA caching allocator,https://github.com/pytorch/pytorch/pull/181189,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"Stale, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-22T23:25:58Z,2026-06-22T08:34:15Z,16,Stack from ghstack (oldest at bottom): -> #181189 proposal: #181191 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo @azahed98
175512,[kineto] Generate unique group_trace_id instead of reusing trace_id (#175512),https://github.com/pytorch/pytorch/pull/175512,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, Stale, meta-exported",2026-02-23T04:10:03Z,2026-06-22T08:34:02Z,16,"Summary: Previously, setTraceID() in kineto_shim.cpp set REQUEST_GROUP_TRACE_ID to the same value as REQUEST_TRACE_ID. This made the group trace id redundant and not useful for distinguishing traces that share the same trace id. This change generates a unique random uint64_t for the group trace id using std::mt19937..."
184641,Fix GroupNorm validation for non-positive num-groups,https://github.com/pytorch/pytorch/pull/184641,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184280,184280,184280,,,,0,3,0,,,"triaged, open source, release notes: nn",2026-05-21T02:32:43Z,2026-06-22T07:34:41Z,16,Summary Adds explicit value logic for num_groups <= 0 in torch.nn.GroupNorm to avoid triggering a ZeroDivisionError. New raise: ValueError: num_groups must be a positive integer Instead. Also adds tests covering num_groups=0 and negative values. Fixes #184280
162454,[xpu][feature] Integrate OneDNN SDPA training forward/backward into XPU OVERRIDEABLE Backend,https://github.com/pytorch/pytorch/pull/162454,ready for maintainer decision,45,CI lane labels: ciflow/binaries_wheel; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,9,0,,,"module: cpu, triaged, module: mkldnn, open source, Merged, Reverted, ciflow/binaries_wheel, module: inductor, keep-going, release notes: xpu, ci-no-td, release notes: inductor (aoti), jansel-agent-skip",2025-09-09T03:37:47Z,2026-06-22T06:57:10Z,16,This is the second PR split from #156272 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @gujinghui @PenghuiCheng @jianyuh @min-jean-cho @yanbing-j @Guobing-Chen @Xia-Weiwen @snadampal @voznesenskym @penguinwu @EikanWang @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @...
181187,[spmd types] fully_shard but in DTensor land,https://github.com/pytorch/pytorch/pull/181187,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,1,0,,,"Stale, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-04-22T23:14:47Z,2026-06-22T03:48:43Z,16,Stack from ghstack (oldest at bottom): -> #181187 #180880
181178,[Inductor] Allow max_autotune_gemm on Jetson Orin AGX,https://github.com/pytorch/pytorch/pull/181178,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-22T22:13:32Z,2026-06-22T03:48:42Z,16,"Summary: The is_big_gpu() check in _inductor/utils.py gates max_autotune_gemm on a minimum SM count. Master uses 16 if device.type == ""xpu"" else 68, which excludes any Nvidia Jetson Orin (16 SMs for the 64GB variant we are actively testing). This unblocks CUDA AOTInductor export of the Gemma3n audio + text decoder o..."
186332,Improved common device test module for xpu,https://github.com/pytorch/pytorch/pull/186332,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/xpu",2026-06-05T06:10:18Z,2026-06-22T02:08:33Z,17,"Description Current instantiate_device_type_tests method has overlapped parameters to control XPU test. With this PR changes we don't need to set allow_xpu to True when only_for contains ""xpu"". Changes Automatic set the allow_xpu when only_for xpu contains the key word ""xpu"" Update the onlyOn type definition Notify..."
187260,Fix static cudnn building for cudnn v9,https://github.com/pytorch/pytorch/pull/187260,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/binaries,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,186914,,,,,0,1,0,,,"triaged, open source, ciflow/binaries",2026-06-13T23:05:46Z,2026-06-22T00:32:51Z,17,"Ref #186914, #168324 Uses cuDNN.cmake in cudnn-frontend"
180867,Use SavedTensor for SAC cached tensors to enable saved tensor hooks,https://github.com/pytorch/pytorch/pull/180867,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180866, 180874",,,0,19,0,,,"Stale, ciflow/torchtitan",2026-04-20T15:51:14Z,2026-06-21T23:56:02Z,17,Stack from ghstack (oldest at bottom): #180874 -> #180867 #180866 #180919 #176455 SAC previously stored cached tensors using a custom _VersionWrapper. This switches to using _make_saved_tensor / SavedTensor so that saved tensor default hooks (pack/unpack) are applied to SAC-cached tensors. This enables composing SAC...
178099,Add TypeGuard[Tensor] return type to is_tensor_like,https://github.com/pytorch/pytorch/pull/178099,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,175324,175324,175324,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-03-22T17:30:12Z,2026-06-21T23:32:58Z,17,"Summary Annotate torch.overrides.is_tensor_like with a TypeGuard[torch.Tensor] return type, enabling type checkers to narrow variables after the guard Drop the now-redundant cast() calls that only existed to express that narrowing: tensors in torch.autograd.backward and outputs in torch.autograd.grad (grad()'s input..."
181150,update copy_misaligned to copy_if_misaligned,https://github.com/pytorch/pytorch/pull/181150,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-22T18:46:58Z,2026-06-21T22:52:20Z,17,Cherry-pick of a12ade5 onto release/2.12 for #180506. Multiple users had confusion with the prior operator. this is strictly for better readability and doesn't have any real changes. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng...
180529,[inductor] Add TORCHINDUCTOR_PERSISTENT_AUTOTUNE_DIR for persistent autotune cache (#180529),https://github.com/pytorch/pytorch/pull/180529,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, module: inductor, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-16T03:43:16Z,2026-06-21T22:52:19Z,17,"Summary: Adds a new environment variable TORCHINDUCTOR_PERSISTENT_AUTOTUNE_DIR that enables a persistent local autotune cache that survives across code changes and process restarts. This is useful when force_disable_caches=True disables all inductor caches (FxGraphCache, AOTAutogradCache, autotune cache) to get fres..."
182109,[MPS] Replace MPSGraph LayerNorm backward with native Metal kernels,https://github.com/pytorch/pytorch/pull/182109,draft/noise,31,PR is draft; review decision requests changes; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,0,0,,,,,187455,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-01T12:51:53Z,2026-06-21T22:49:50Z,17,"Part of #187455 Summary Replaces the MPSGraph-backed native_layer_norm_backward dispatch on MPS with hand-written Metal compute kernels (LayerNorm.metal), deleting the ~230-line MPSGraph block in Normalization.mm. MPSGraph caches a compiled graph per unique tensor shape; in GNN / variable-length workloads every new..."
187559,[MPS] Migrate nll_loss to native Metal kernels,https://github.com/pytorch/pytorch/pull/187559,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,187455,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-17T08:12:32Z,2026-06-21T22:38:31Z,17,"Part of #187455. Summary Replaces the MPSGraph implementation of the 1-D nll_loss path on MPS with hand-written Metal kernels (aten/src/ATen/native/mps/kernels/LossOps.metal, dispatched from LossOps.mm), removing the per-(N, C, dtype, reduction) MPSGraph shape-cache entry. Three templated kernels (nll_loss_fwd_none,..."
180955,[ROCm] Normalize AMDSMI UUID matching,https://github.com/pytorch/pytorch/pull/180955,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/rocm-mi300,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,"180123, 180227","180123, 180227",,,,,0,1,0,,,"module: rocm, triaged, open source, Stale, ciflow/rocm-mi300",2026-04-21T03:38:37Z,2026-06-21T21:58:42Z,17,Summary: Normalize ROCm UUID candidates before matching them against AMDSMI ASIC serials. Normalize rocminfo UUID output in test_raw_amdsmi_device_uuids and assert every UUID is found. Add regression coverage for uppercase ROCm GPU-... UUID candidates. Root cause: _raw_device_uuid_amdsmi() returns lower-case ASIC se...
180716,Fix torch.compile addmm with beta=0 and mismatched bias,https://github.com/pytorch/pytorch/pull/180716,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,178040,178040,"174929, 178040",,,,0,1,0,,,"triaged, open source, Stale, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-04-17T20:40:37Z,2026-06-21T21:58:41Z,17,Fixes #178040 Fixes torch.compile raising RuntimeError on torch.addmm calls where beta=0 and bias shape doesn't match output shape. When beta=0 the bias term is zeroed out so its shape is irrelevant. This now matches eager cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzhen...
180866,Add save_tensor() API for selective activation checkpointing,https://github.com/pytorch/pytorch/pull/180866,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180867, 180874",,,0,17,0,,,"Stale, ciflow/torchtitan",2026-04-20T15:51:10Z,2026-06-21T21:58:40Z,17,"Stack from ghstack (oldest at bottom): #180874 #180867 -> #180866 #180919 #176455 Adds a save_tensor() function that can be called inside a checkpointed function to unconditionally save a tensor, bypassing the policy function. This is useful when the user knows at tensor-creation time that a tensor should be saved,..."
179827,Replace BF16 matmul OpInfo checks with a feature query,https://github.com/pytorch/pytorch/pull/179827,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,,175211,175211,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-09T15:28:31Z,2026-06-21T21:58:37Z,17,"Refs #175211 Summary This adds PLATFORM_SUPPORTS_BF16_MATMUL in common_cuda.py and uses it for the BF16 dtype and test_dtypes gating on the matmul-like OpInfo entries (addbmm, baddbmm, bmm, matmul, nn.functional.bilinear, and __rmatmul__). The goal is just to replace the raw SM53OrLater or TEST_WITH_ROCM checks with..."
180675,Add test for autograd.Function fused fwd/bwd pattern (issue #180642),https://github.com/pytorch/pytorch/pull/180675,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180642,180642,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing, module: dynamo, ciflow/inductor",2026-04-17T15:11:08Z,2026-07-02T18:43:42Z,6,"The fix for this issue landed in #177368, but no test was added at the time to guard against future regressions. This adds a test covering the fused forward/backward pattern from the issue, an autograd.Function that pre-allocates gradient buffers via in-place ops and returns multiple outputs alongside saved tensors...."
180171,Fix empty changed-files handling in Filepath heuristic,https://github.com/pytorch/pytorch/pull/180171,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-12T12:17:39Z,2026-06-21T21:58:34Z,17,Summary: filter empty entries from query_changed_files when git diff --name-only has no output treat empty changed-file entries as non-doc changes instead of skipping all tests add heuristic tests for the empty-diff edge case Testing: python3 -m unittest tools/test/heuristics/test_utils.py
178800,debug server tutorial,https://github.com/pytorch/pytorch/pull/178800,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/trunk, release notes: distributed (c10d), topic: not user facing, ciflow/torchtitan",2026-03-30T19:33:52Z,2026-06-21T21:58:30Z,17,"Summary: Add a comprehensive tutorial for the PyTorch distributed debug server, providing users with end-to-end guidance on setup, available diagnostic handlers (stack traces, py-spy, flight recorder, profiling, etc.), periodic dumping, custom handler registration, and TorchElastic integration — making it easier to..."
181023,[PP] Support extra loss_fn kwargs in pipeline schedules,https://github.com/pytorch/pytorch/pull/181023,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/torchtitan",2026-04-21T21:01:13Z,2026-06-21T20:56:46Z,17,"Stack from ghstack (oldest at bottom): -> #181023 Add _loss_kwargs dict to _PipelineSchedule that is forwarded to the loss function in both _compute_loss and _backward_metadata_inference. This enables loss functions that require additional arguments beyond (output, target), such as chunked cross-entropy loss which n..."
180080,Bitwise operator tensorcompare.cpp,https://github.com/pytorch/pytorch/pull/180080,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,5,0,,,"module: cpu, triaged, open source, Stale, release notes: vulkan, module: inductor",2026-04-10T22:38:17Z,2026-06-21T20:56:43Z,17,implemented the use of bitwise operator in isclose() function. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutrube...
180781,Fix integer divide-by-zero UB in CUDA binary kernels,https://github.com/pytorch/pytorch/pull/180781,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,165650,165650,165650,,,,0,1,0,,,"module: cuda, triaged, open source, Stale",2026-04-19T20:58:32Z,2026-06-21T20:07:02Z,17,"Integer div_trunc, remainder, and fmod on CUDA previously relied on hardware behavior for division by zero, which is undefined and produced inconsistent results across dtypes and platforms (e.g. -1 for int32, mixed values for int64, different behavior on ROCm vs NVIDIA). Guard all three kernels to return 0 when the..."
187247,Type _add_docstr so type checkers see spectral ops as callable,https://github.com/pytorch/pytorch/pull/187247,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,162373,162373,162373,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-13T13:46:51Z,2026-06-21T16:33:20Z,17,"Pylint reports E1102: torch.fft.rfft is not callable (and the same for the other spectral ops) because those names are bound from _add_docstr(...) calls, where _add_docstr is the C builtin torch._C._add_docstr. Its stub in torch/_C/__init__.pyi.in was an ellipsis-bodied def, and astroid infers the call result of suc..."
186049,Fix bool argmax meta validation for Inductor,https://github.com/pytorch/pytorch/pull/186049,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185481,185481,185481,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-06-03T06:52:17Z,2026-06-21T16:27:42Z,17,"Fixes #185481 This PR adds eager-parity meta validation for bool inputs to argmax / argmin. Currently, eager mode raises RuntimeError: argmax(): does not support bool input, but Inductor can accept a bool comparison mask and use the resulting argmax index in downstream indexing. This change makes the meta path rejec..."
187250,Support NumPy 2 __array__ copy kwarg and __array_wrap__ context on Tensor,https://github.com/pytorch/pytorch/pull/187250,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,136264,136264,136264,,,,0,3,0,,,"triaged, open source, release notes: python_frontend",2026-06-13T15:24:05Z,2026-06-21T15:57:16Z,17,"PyTorch Tensors did not implement the NumPy 2 __array__ interface, which caused DeprecationWarnings when users called np.array(tensor) or performed operations mixing tensors and numpy arrays. The __array__ method now accepts the copy keyword and __array_wrap__ accepts context and return_scalar so that interactions w..."
184881,Fix SIGSEGV from functorch key contamination via Tensor.set_data,https://github.com/pytorch/pytorch/pull/184881,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,184876,184876,184876,,,,0,7,0,,,"open source, release notes: torch.func",2026-05-22T10:47:00Z,2026-06-21T15:57:14Z,17,"Fixes #184876. copy_tensor_metadata_except_version_counter already strips python_ks to avoid poisoning a plain destination. Do the same for functorch_transforms_ks, which encodes a TensorWrapper/BatchedTensorImpl subclass and triggers an unchecked cast in maybeGetTensorWrapper. Test Plan: python test/functorch/test_..."
175425,Improve error message for CPU-CUDA copy during graph capture,https://github.com/pytorch/pytorch/pull/175425,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,5,0,,,"Stale, release notes: cuda",2026-02-20T18:43:20Z,2026-06-21T15:14:30Z,17,"Stack from ghstack (oldest at bottom): -> #175425 #175421 The existing error message tells users to pin the CPU tensor, but even with pinned memory the copy will fail unless the graph was created with capture_error_mode=""thread_local"". Update the message to mention both requirements. Authored with Claude."
181047,[FSDP2] Introduce StreamHandoff; migrate AR keep-alive,https://github.com/pytorch/pytorch/pull/181047,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,179128,179443,,,0,6,0,,,"Stale, release notes: distributed (fsdp), ciflow/inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-04-22T00:05:52Z,2026-06-21T11:31:47Z,17,"Stack from ghstack (oldest at bottom): -> #181047 #179443 FSDP2's cross-stream lifetime management (the ""hold a Python ref alive across a producer->consumer stream boundary, then drop it on a stream whose FIFO already waits on the consumer-done event"" pattern) was implemented ad-hoc for each buffer: AllGatherState,..."
181010,[inductor][comms] allow compute to borrow memory from pg_alloc mempool,https://github.com/pytorch/pytorch/pull/181010,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181007, 181008, 181009",,,0,1,0,,,"Stale, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/torchtitan",2026-04-21T18:29:45Z,2026-06-21T09:48:13Z,17,"Stack from ghstack (oldest at bottom): -> #181010 #181009 #181008 #181007 Summary Allows regular compute operations to temporarily borrow idle pg_alloc buffers during memory planning. When a comm buffer is sitting idle between two collective uses, a regular allocation with matching size can reuse that memory instead..."
181009,[inductor][comms] pg_alloc config for strategies,https://github.com/pytorch/pytorch/pull/181009,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/torchtitan",2026-04-21T18:29:40Z,2026-06-21T09:48:12Z,17,"Stack from ghstack (oldest at bottom): (to be filled) Summary Adds fine-grained control over which collective types and buffer roles use pg_alloc. Instead of the all-or-nothing comms_use_pg_alloc toggle, users can now selectively enable pg_alloc for specific combinations (e.g., only all-gather outputs, only reduce-s..."
181008,[inductor][comms] memory budget for pg allocs,https://github.com/pytorch/pytorch/pull/181008,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181007, 181009, 181010",,,0,1,0,,,"Stale, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/torchtitan",2026-04-21T18:29:36Z,2026-06-21T09:48:11Z,17,"Stack from ghstack (oldest at bottom): #181010 #181009 -> #181008 #181007 Summary Adds a configurable memory budget for pg_alloc allocations. When the cumulative pg_alloc memory exceeds the budget, new allocations fall back to regular torch.empty(), preventing over-commitment to NCCL-registered memory on large model..."
186137,[dynamo] Swap @disable / @allow_in_graph order on invoke_subgraph_wrapper_unboxed,https://github.com/pytorch/pytorch/pull/186137,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, merging, ciflow/torchtitan",2026-06-03T21:29:16Z,2026-06-21T08:51:50Z,17,"The wrapper returned by invoke_subgraph_inner_compiler was decorated with @disable @torch._dynamo.allow_in_graph (disable outermost). In that order: allow_in_graph registers id(fn0) (the inner function) into trace_rules._allowed_callable_ids and returns fn0 unchanged. disable calls innermost_fn, then creates a NEW w..."
187712,Strict pyrefly checking: round 1 (3/N),https://github.com/pytorch/pytorch/pull/187712,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,187711,,,0,2,0,,,"module: cpu, topic: not user facing",2026-06-19T04:10:33Z,2026-06-21T07:29:31Z,17,"Stack from ghstack (oldest at bottom): -> #187712 #187711 #187688 Annotates and turns on whole-file pyrefly checking (per-file [[sub-config]] entries) for torch/utils/_sympy/{functions,numbers,singleton_int,interp,symbol}.py. Third batch of the effort to enable annotation checking on non-public torch files. These mo..."
180788,"Use C++20 concepts/requires in Tensor, TensorBase, and IValue",https://github.com/pytorch/pytorch/pull/180788,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-20T02:18:18Z,2026-06-21T06:27:52Z,17,
180535,"[AC remat] Per-region independent recomputation, remove user phase annotations",https://github.com/pytorch/pytorch/pull/180535,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"Stale, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-16T05:31:23Z,2026-06-21T06:14:17Z,17,"Stack from ghstack (oldest at bottom): -> #180535 #180534 Apply remat independently to each backward region that needs it, matching eager AC behavior where each backward call triggers its own recomputation via unpack hooks. Each region gets its own recomputed_nodes scope so lifetimes don't bleed across regions. Remo..."
176070,[rnn] Prevent LSTMCell segfault on input/hidden shape mismatch,https://github.com/pytorch/pytorch/pull/176070,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,175978,175978,175978,,,,0,1,0,,,"triaged, open source, Stale, release notes: nn",2026-02-28T12:29:40Z,2026-06-21T03:48:22Z,17,"aten::lstm_cell could segfault when called with mismatched input vs (hx, cx) ranks/batch sizes (e.g. unbatched input with batched hidden states), due to missing validation before downstream checks/kernels. Add explicit TORCH_CHECKs for input/hx/cx dimensionality (and batch size for batched inputs) and add a few regr..."
180947,[CUDACachingAllocator] Add test for expandable segment OOM stat,https://github.com/pytorch/pytorch/pull/180947,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"fb-exported, Stale, ciflow/trunk, meta-exported",2026-04-21T01:39:30Z,2026-06-20T23:52:44Z,18,"Summary: Adds test_expandable_segment_map_oom_stat to TestCudaAllocator that verifies the num_expandable_segment_map_oom stat counter is incremented when GPU memory is exhausted during expandable segment mapping. The test enables expandable_segments, triggers OOM by allocating 1TB, and asserts the stat is > 0. Test..."
180933,[spmd_types] fully_shard,https://github.com/pytorch/pytorch/pull/180933,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"Stale, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-04-20T22:43:53Z,2026-06-20T23:52:43Z,18,Stack from ghstack (oldest at bottom): -> #180933 #180880
186571,Fix #176298: support unsigned add and sub on CPU,https://github.com/pytorch/pytorch/pull/186571,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,176298,176298,176298,,,,0,1,0,,,"triaged, open source, module: intel, release notes: python_frontend",2026-06-07T23:04:57Z,2026-06-20T22:51:02Z,18,"Fixes #176298 Summary This PR adds CPU support for torch.add and torch.sub on torch.uint16, torch.uint32, and torch.uint64 tensors. The generated CPU add kernel did not include these unsigned dtypes, which caused torch.add to raise NotImplementedError. Since torch.sub delegates to add_stub, it failed as well. The is..."
187756,[testintro][DEMO - DO NOT LAND] exercise the test-diff CI comment,https://github.com/pytorch/pytorch/pull/187756,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"187754, 187755",,,0,2,0,,,topic: not user facing,2026-06-20T01:34:13Z,2026-06-20T21:52:03Z,18,"Stack from ghstack (oldest at bottom): -> #187756 #187755 #187754 Throwaway changes to validate the test-diff PR comment end-to-end on a real (ghstack) PR: add a new test file with plain/parametrized/onlyCUDA tests, add one test to an existing file (test_bundled_inputs), and remove one from another (test_complex). D..."
180601,"Fix pydocstyle errors in _guards.py, _ops.py, _jit_internal.py, functional.py, _tensor_str.py, library.py",https://github.com/pytorch/pytorch/pull/180601,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,112586,112586,112586,,,,0,2,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-16T19:48:57Z,2026-06-20T20:02:57Z,18,"Fixes #112586 Summary Fixes all reported pydocstyle violations across 6 files: torch/functional.py: 15 violations. Fixed imperative mood (Splits to Split, Creates to Create, etc.), added periods to summary lines, added blank lines between summary and description, removed function signatures from docstrings, fixed se..."
180924,Remove allow_cache_entry_mutation from SAC,https://github.com/pytorch/pytorch/pull/180924,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180866, 180867, 180874",,,0,2,0,,,"Stale, ciflow/torchtitan",2026-04-20T21:54:53Z,2026-06-20T20:02:56Z,18,"Stack from ghstack (oldest at bottom): #180874 #180867 #180866 -> #180924 #180919 #176455 I'm considering deleting this, because from internal and external search, no one is actually using it and it a piece of complexity that limits the design of the rest of the system. This is theoretically useful if you saved an o..."
180988,[fbandroid][buckconfig removal] removal of last files referencing tools/build_defs/android/build_mode_defs.bzl,https://github.com/pytorch/pytorch/pull/180988,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, Stale, topic: not user facing, meta-exported",2026-04-21T15:31:17Z,2026-06-20T17:57:38Z,18,Summary: removing last references to build_mode_defs.bzl so it can be deleted Test Plan: CI passes check that macro attributes did not change with `python3 arvr/scripts/daniel314/query_scripts/cquery_check.py' Reviewed By: rexzhang123 Differential Revision: D101574369
187333,[CUDA] linalg: overlap looped cuSOLVER LU (getrf) across auxiliary streams,https://github.com/pytorch/pytorch/pull/187333,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,182001,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend",2026-06-15T12:44:42Z,2026-06-20T15:58:08Z,18,"Summary lu_factor_looped_cusolver issues one cuSOLVER getrf call per batch item serially on the current CUDA stream. For small/medium matrices, these independent per-item calls cannot overlap, which can underutilize the GPU. This PR overlaps those independent GETRF calls across a small number of auxiliary CUDA strea..."
183525,[MPS] truncate fractional start/end to integer type in linspace when …,https://github.com/pytorch/pytorch/pull/183525,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,137635,137635,137635,,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-13T08:16:42Z,2026-06-20T14:23:28Z,18,"…dtype=int Closes #137635. When dtype=int, the endpoints are now truncated to int so the range is computed correctly. Adds a test to validate that it now matches the CPU linspace implementation (up to the floating point precision of MPS)."
180858,[inductor] symm memm ag/rs with cuStreamWrite/WaitValue32 sync,https://github.com/pytorch/pytorch/pull/180858,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,2,0,,,"Stale, release notes: distributed (c10d), module: inductor, ciflow/inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-04-20T15:15:11Z,2026-06-20T14:16:47Z,18,"Stack from ghstack (oldest at bottom): #180862 -> #180858 #180858 Inductor's low_contention all-gather, all-gather-v2, and reduce-scatter lowerings now allocate their input buffers from symmetric memory (empty_strided_p2p) via CommBufferLayout, matching the pattern used by one_shot_all_reduce and other symm_mem ops...."
173258,[CUDA] Math.cuh add missing return statement,https://github.com/pytorch/pytorch/pull/173258,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,10,0,,,"triaged, open source, Stale, ciflow/trunk, release notes: cuda",2026-01-24T16:02:24Z,2026-06-20T09:40:27Z,18,"I'm not an expert for this, but I'm pretty sure that NAN as a standalone phrase needs a return statement, but as always please correct me if I'm mistaken"
185319,Sort record names before hashing in writeSerializationId,https://github.com/pytorch/pytorch/pull/185319,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184219,184219,184219,,,,0,1,0,,,"caffe2, triaged, open source, topic: not user facing",2026-05-27T03:59:14Z,2026-06-20T09:15:33Z,18,closes #184219 Summary PyTorchStreamWriter::writeSerializationId builds the serialization_id from a hash of the record names plus a CRC. The goal would be for the value to be deterministic for the same set of files. The hash combine was previously done while iterating over files_written_ directly but files_written_...
174275,Adjust to use __riscv_v to detect for RISC-V Vector Extension,https://github.com/pytorch/pytorch/pull/174275,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cpu, triaged, open source, Stale, topic: not user facing, module: risc-v",2026-02-04T11:42:17Z,2026-06-20T07:48:12Z,18,"Newer compilers are now defining '__riscv_v_intrinsic' to mean that ""the compiler does support the RVV intrinsics"" instead of indicating ""currently compiling for RVV enabled"". For reference: riscv-non-isa/riscv-rvv-intrinsic-doc#376 This means that current parts that are using only '__riscv_v_intrinsic' to guard RVV..."
183811,[docs] Document 10 public APIs in torch.nn.parallel submodules,https://github.com/pytorch/pytorch/pull/183811,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,topic: not user facing,2026-05-14T23:20:25Z,2026-06-20T06:31:10Z,18,"Summary Wires 10 previously undocumented public functions in torch.nn.parallel.* into the Sphinx doctree. The functions already had docstrings in source; only doc source files change. Module Functions torch.nn.parallel.comm broadcast, broadcast_coalesced, gather, reduce_add, reduce_add_coalesced, scatter torch.nn.pa..."
180925,[dynamo] Extract dynamic graph break hints,https://github.com/pytorch/pytorch/pull/180925,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, topic: not user facing",2026-04-20T21:57:13Z,2026-06-20T02:25:37Z,19,
180930,[Inductor] Lower MAX_COMPLEX_POINTWISE_CAT from 8 to 5,https://github.com/pytorch/pytorch/pull/180930,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-20T22:26:35Z,2026-06-20T01:26:37Z,19,"Summary: Reduces the threshold for pointwise cat fusion from 8 to 5 inputs. When cat has >= 6 inputs and is fused into a reduction kernel (e.g., layer_norm), the generated code has 5+ conditional branches per element in the reduction inner loop to select the source tensor. This destroys warp coherence and memory coa..."
174826,add mem_get_info usage notes to CUDA memory management docs,https://github.com/pytorch/pytorch/pull/174826,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"open source, Stale, ciflow/trunk, topic: docs, topic: not user facing",2026-02-11T23:56:18Z,2026-06-20T01:26:30Z,19,"Supersedes #174822. The previous PR was opened from an outdated branch, which caused unrelated commits to be included. This PR contains only the intended documentation change."
175490,Fix CPU `NotImplementedError` for bf16/fp16 antialiased upsampling (`interpolate`),https://github.com/pytorch/pytorch/pull/175490,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,175489,175489,175489,,,,0,5,0,,,"module: cpu, triaged, open source, Stale, release notes: nn",2026-02-22T13:49:32Z,2026-06-20T01:26:27Z,19,"Fixes : #175489 In this PR I am trying to fix a CPU gap in torch.nn.functional.interpolate where antialias=True (supported for mode=""bilinear""/""bicubic"" per the PyTorch Docs) raised NotImplementedError for torch.bfloat16 and torch.float16 inputs. This fix covers both forward and backward. The root cause is that the..."
180665,"[FSDP2] Fix HSDP + CPUOffloadPolicy grad corruption race, add stream-behavior tests",https://github.com/pytorch/pytorch/pull/180665,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,179443,,,0,4,0,,,"Stale, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-04-17T11:29:11Z,2026-06-19T23:52:43Z,19,"Stack from ghstack (oldest at bottom): -> #180665 #179443 #180666 Bug Under HSDP + CPUOffloadPolicy, the fp32 reduce-scatter output buffer (reduce_output) is allocated on reduce_scatter_stream but the D2H memcpy that offloads the grad runs on all_reduce_stream (post_reduce_stream in the HSDP path). The CUDA caching..."
187595,Fix torch.gcd/lcm integer overflow on INT_MIN inputs,https://github.com/pytorch/pytorch/pull/187595,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187338,187338,187338,,,,0,1,0,,,"triaged, open source, release notes: python_frontend",2026-06-17T19:04:05Z,2026-06-19T23:33:16Z,19,"Fixes #187338 Problem torch.gcd returns a wrong, negative result when an input is the minimum value of a signed integer dtype. For example, gcd(-128, 124) returns -4 instead of 4 for int8: >>> import torch >>> torch.gcd(torch.tensor([124, -128, -126], dtype=torch.int8), ... torch.tensor([124, 124, 126], dtype=torch...."
180686,test: relax error tolerances for complex64 tests with tan and tanh,https://github.com/pytorch/pytorch/pull/180686,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-17T16:26:01Z,2026-06-19T20:55:26Z,19,"There is an issue caused by precision limitations in the standard std::tanh (and std::tan) implementation used by PyTorch on CPU when evaluated near complex poles. For example, TestUnaryUfuncsCPU.test_reference_numerics_extremal__refs_nn_functional_tanhshrink_cpu_complex64: File ""/third_party/py/torch/test/test_unar..."
180188,Validate opt_dtype in nanmean to reject non-floating-point types,https://github.com/pytorch/pytorch/pull/180188,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,131043,131043,,,,,0,1,0,,,"triaged, open source, Stale",2026-04-12T22:31:25Z,2026-06-19T17:18:04Z,19,"Fixing an Issue Issue Fixes #131043 Summary nanmean() validated only self.scalar_type() but ignored the explicit dtype= parameter. Two broken cases: torch.nanmean(torch.tensor([0., 1., 2.]), dtype=torch.int64) hit kernel dispatch failure with confusing ""nansum_cpu"" not implemented for 'Long' instead of a dtype error..."
180187,add OneCycleLR degenerate phase cases,https://github.com/pytorch/pytorch/pull/180187,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,168044,,,,0,2,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-12T21:01:47Z,2026-06-19T17:18:03Z,19,"This PR adds failing tests that highlight degenerate phase handling in OneCycleLR for very small schedules. When total_steps is tiny, certain values of pct_start produce phases with zero width (for example, total_steps=2, pct_start=0.5 or total_steps=1, pct_start=1.0). In these cases, the scheduler proceeds with pha..."
177455,Add nn.SwiGLU activation function,https://github.com/pytorch/pytorch/pull/177455,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,128712,128712,128712,,,,0,1,0,,,"triaged, open source, Stale, no-stale",2026-03-14T17:46:06Z,2026-06-19T16:40:56Z,19,"Issue Fixes #128712 Summary Adds torch.nn.SwiGLU and torch.nn.functional.swiglu, the SiLU-gated linear unit used in LLaMA, Mistral, Gemma, and most current transformer architectures. m = torch.nn.SwiGLU(dim=-1) output = m(input) # splits input in half, applies SiLU gating The math: SwiGLU(a, b) = SiLU(a) * b, where..."
180764,cpp_extension: add CUDA dlink flags to JIT load APIs,https://github.com/pytorch/pytorch/pull/180764,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,,180762,180762,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-19T08:20:44Z,2026-06-19T13:44:47Z,19,"Summary This PR adds an opt-in CUDA device-link path for JIT cpp extension builds. Today, torch.utils.cpp_extension.CUDAExtension supports RDC/device-link workflows, but torch.utils.cpp_extension.load() and load_inline() do not expose an equivalent capability. This change adds a JIT-side API for forwarding CUDA devi..."
180783,"docs: fix index_add_ `index` description (points into self, not source)",https://github.com/pytorch/pytorch/pull/180783,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180119,180119,180119,,,,0,1,0,,,"triaged, open source, Stale",2026-04-19T22:51:09Z,2026-06-19T13:44:46Z,19,"Summary Fixes #180119. torch.Tensor.index_add_'s docstring currently describes the index argument as: index (Tensor): indices of ``source`` to select from, should have dtype either `torch.int64` or `torch.int32` That's backwards. As the docstring's own 3-D indexing block makes clear: self[index[i], :, :] += alpha *..."
180782,docs: remove unsupported out= from torch.linalg._powsum docstring,https://github.com/pytorch/pytorch/pull/180782,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180649,180649,180649,,,,0,1,0,,,"triaged, open source, Stale, release notes: linalg_frontend",2026-04-19T22:45:32Z,2026-06-19T13:44:45Z,19,"Summary Fixes #180649. torch.linalg._powsum's Python docstring advertises an out=None keyword argument in both the signature line and the Keyword args section, but the underlying aten schema has no out parameter: - func: linalg__powsum(Tensor self, Scalar ord=2, int[1]? dim=None, bool keepdim=False, *, ScalarType? d..."
180319,[ROCm] Handle profiler trace format differences in combo kernel tests,https://github.com/pytorch/pytorch/pull/180319,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,"180017, 180018, 180019","180017, 180018, 180019",,,,,0,1,0,,,"module: rocm, triaged, open source, Merged, Reverted, Stale, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ci-no-td, ciflow/rocm-mi300, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355, ciflow/torchtitan",2026-04-14T06:18:52Z,2026-06-19T13:44:45Z,19,"Fixes #180017. Fixes #180018. Fixes #180019. Three combo kernel tests access event[""args""][""grid""] from profiler chrome traces. The grid key structure differs between CUPTI and ROCTracer trace formats, so this throws KeyError on ROCm. Use .get() and only assert on grid values when present. All other validation (nume..."
174491,fix: prevent issue with destruction of std::shared_timed_mutex,https://github.com/pytorch/pytorch/pull/174491,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, Stale, ciflow/trunk, topic: not user facing",2026-02-06T20:39:01Z,2026-06-19T13:44:36Z,19,"The issue was found by a memory analyzer: c10::utils::get_env uses a global std::shared_timed_mutex, which may have already been destroyed when get_env is called during process exit via static/thread-local object destruction. Accessing a destroyed mutex results in undefined behavior. This change allocates the mutex..."
180315,Add _Float16 support for Tensor::item() C++ API,https://github.com/pytorch/pytorch/pull/180315,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,157776,157776,157776,,,,0,3,0,,,"triaged, open source, Stale, ciflow/mps",2026-04-14T04:09:14Z,2026-06-19T08:36:53Z,19,"Fixes #157776 Adds a conversion operator from c10::Half to the compiler-builtin _Float16 type, and an item<_Float16>() specialization. This follows the existing pattern used for ARM's native float16_t on aarch64. All changes are guarded by a new C10_HAS_FLOAT16_TYPE macro (requires FLT16_MAX, excluded on aarch64/CUD..."
178338,[inductor] add option to autotune at fusion time for mm,https://github.com/pytorch/pytorch/pull/178338,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"Stale, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-03-24T23:05:20Z,2026-06-19T08:21:12Z,19,"For epilogue fusion with max-autotuning enabled, we first autotune gemms at lowering time to get an initial selection of configs before benchmarking them with fused epilogue. Because we autotune prior to fusion benchmarking, we potentially miss out on configs that are optimal for fusion. In this PR, we add an option..."
185355,[pipelining] Fix backward hook accumulation on recv buffer activations,https://github.com/pytorch/pytorch/pull/185355,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185331,185331,185331,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: pipelining, ciflow/torchtitan",2026-05-27T11:09:09Z,2026-06-19T07:58:02Z,19,"Fixes #185331 Root cause: _retrieve_recv_activations passed info.buffer directly as the stage activation via requires_grad_(), which is in-place and returns self. This aliased the persistent recv buffer as the activation tensor, so backward hooks attached by user code (e.g. via module forward pre-hooks for gradient..."
180317,[xpu][inductor] Add Dual GEMM SiLU-Mul fusion pass for XPU,https://github.com/pytorch/pytorch/pull/180317,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: cpu, open source, Stale, release notes: releng, module: inductor",2026-04-14T04:25:08Z,2026-06-19T06:18:01Z,19,"Summary Adds an Inductor post-grad FX pass (xpu_dual_gemm_pass) that recognises the pattern: out = F.silu(x @ w1.t()) * (x @ w3.t()) and replaces it with a single fused kernel call: out = torch.ops.xpu_ops.dual_gemm_silu_mul(x, w1, w3) The kernel is implemented in intel/torch-xpu-ops#3323 using the sycl-tla (CUTLASS..."
187656,Reject complex scale/zero_point tensors in quantize_per_tensor with an explicit dtype error instead of value cannot be converted to type double without overflow.,https://github.com/pytorch/pytorch/pull/187656,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,137468,137468,137468,,,,0,1,0,,,"triaged, open source, release notes: quantization",2026-06-18T11:01:19Z,2026-06-19T06:01:06Z,19,"Reject complex scale/zero_point tensors in quantize_per_tensor with an explicit dtype error instead of value cannot be converted to type double without overflow. (Fixes #137468) Added shared checkPerTensorQParamTensors() in Quantizer.h and wired it into tensor-qparams, list, fake-quant, and Vulkan entry points. Adde..."
186346,Handle invalid CUDA JITerator cache entries,https://github.com/pytorch/pytorch/pull/186346,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,132756,132756,,,,0,4,0,,,"Merged, Reverted, ciflow/trunk, release notes: cuda, ci-no-td, agentic",2026-06-05T08:01:29Z,2026-06-19T05:59:07Z,19,"Stack from ghstack (oldest at bottom): -> #186346 JITerator trusted any readable persistent kernel cache file. If a cache entry was truncated or otherwise invalid, CUDA module loading failed with a driver-level invalid argument error and PyTorch surfaced that failure directly instead of rebuilding the kernel. This m..."
187062,[inductor] Expose compile_worker_malloc_conf to tune the compile-worker allocator (#187062),https://github.com/pytorch/pytorch/pull/187062,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-11T17:51:58Z,2026-06-19T05:59:00Z,19,Summary: Adds an Inductor config (env TORCHINDUCTOR_COMPILE_WORKER_MALLOC_CONF / config.compile_worker_malloc_conf) Test Plan: buck2 build fbcode//mode/opt -c fbcode.platform010_cuda_version=12.8 -c fbcode.nvcc_arch=h100a -c fbcode.enable_gpu_sections=true fbsource//users/sa/sashko/inductor:compile_worker_memory_rep...
187396,Fix integer linspace decomposition to match eager,https://github.com/pytorch/pytorch/pull/187396,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,137546,137546,137546,,,,0,3,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-06-15T23:00:05Z,2026-06-19T05:12:07Z,19,"Picked this up as a good first issue. Stale 2024 report that was never addressed. Fixes #137546. Summary torch._refs.linspace keeps its float endpoints and truncates only the final result, while eager linspace (linspace_kernel) casts the start/end scalars to the integer dtype before interpolating. For integer dtypes..."
174667,Update tolerances for tan* functions w/ complex dtypes on XPU,https://github.com/pytorch/pytorch/pull/174667,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"open source, Stale, release notes: python_frontend",2026-02-10T10:47:56Z,2026-06-19T03:48:39Z,19,"On XPU, several ops (including tan, tanh, tanhshrink) need the same tolerance overrides in tests as on CUDA. This commit adds the DecorateInfo's to the common_methods_invocations.py file."
185173,Run cpp/test_jit and cpp/test_lazy sequentially in test_libtorch_jit,https://github.com/pytorch/pytorch/pull/185173,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Merged, Reverted, topic: not user facing, ci-no-td",2026-05-25T23:02:29Z,2026-06-19T02:38:39Z,20,"Running both binaries in a single run_test.py call puts them in the parallel-test multiprocessing.Pool simultaneously, which has been observed to hang pytest-xdist worker teardown on some runners. See https://github.com/pytorch/pytorch/actions/runs/26416445385/job/77763032718 for an example. Splitting the invocation..."
174940,[BE] Delete old Metal ops,https://github.com/pytorch/pytorch/pull/174940,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, Stale, ciflow/trunk, release notes: mobile, skip-pr-sanity-checks",2026-02-13T02:01:42Z,2026-06-19T01:02:01Z,20,I suspect their use shoudl have been mirgated to ExecuTorch cc @albanD
178781,[MPS] Remove stale @skipIfMPS from test_nn.py,https://github.com/pytorch/pytorch/pull/178781,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/mps, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,179259,179259,,,,0,7,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/mps",2026-03-30T15:23:30Z,2026-06-19T01:01:59Z,20,Stack from ghstack (oldest at bottom): #186990 -> #178781 #186989 Several tests were skipped on MPS but now pass: test_upsamplingNearestExact1d_correctness test_upsamplingNearestExact2d_correctness test_upsamplingBiMode2d_antialias_True test_clip_grad_value (foreach=True self-skips internally) test_clip_grad_norm (f...
178804,Merge clang18 Python-version Docker images into one multi-env container,https://github.com/pytorch/pytorch/pull/178804,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"open source, Stale, topic: not user facing",2026-03-30T20:02:10Z,2026-06-19T01:01:58Z,20,"Stack from ghstack (oldest at bottom): -> #178804 #178803 Instead of building 6 nearly-identical Docker images differing only in Python version (py3.10-clang18 through py3.14t-clang18), build a single pytorch-linux-jammy-py3.10-clang18 image containing all Python versions as separate conda environments (py_3.11, py_..."
178812,[BE][MPS] Delete stale skips from common_modules.py,https://github.com/pytorch/pytorch/pull/178812,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,178781,,,0,2,0,,,"open source, ciflow/mps",2026-03-30T21:31:50Z,2026-06-19T01:01:57Z,20,Stack from ghstack (oldest at bottom): -> #178812 #178781 Or replace then with expectedFailure when possible
180181,[CI] Split benchmark docker builds into three-phase build,https://github.com/pytorch/pytorch/pull/180181,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,11,0,,,"open source, topic: not user facing, ciflow/inductor",2026-04-12T18:21:19Z,2026-06-19T01:01:55Z,20,Stack from ghstack (oldest at bottom): -> #180181 Split benchmark docker builds into three jobs that help avoid rebuilding expensive torchbench/huggingface/timm dependencies on every .ci folder change: The workflow now has three jobs: docker-build-torchbench runs in parallel with docker-build. It builds a standalone...
182025,[MPS] Use MetalPerformancePrimitives for F.linear on macOS 26+,https://github.com/pytorch/pytorch/pull/182025,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,182038,,,0,3,0,,,"open source, release notes: mps, ciflow/mps",2026-04-30T13:48:50Z,2026-06-19T01:01:52Z,20,"Stack from ghstack (oldest at bottom): #182038 -> #182025 MPSNDArrayMatrixMultiplication and MPSGraph matrixMultiplication produce non-deterministic results for >2D fp16/bf16 inputs on M5 (macOS 26). Replace them with a custom MPP matmul2d kernel on macOS 26+ that handles arbitrary strides (via contiguify+flatten),..."
182038,[MPS] Migrate backward to MPP on MacOS-26,https://github.com/pytorch/pytorch/pull/182038,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,182025,,,0,1,0,,,"open source, release notes: mps, ciflow/mps",2026-04-30T15:34:56Z,2026-06-19T01:01:51Z,20,Stack from ghstack (oldest at bottom): -> #182038 #182025
182323,[Mergerules][Temp] Restrict docker changes to just maintainers,https://github.com/pytorch/pytorch/pull/182323,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,182227,,,,,0,1,0,,,"open source, Merged, Reverted, topic: not user facing, ci-no-td",2026-05-04T18:25:24Z,2026-06-19T01:01:50Z,20,Stack from ghstack (oldest at bottom): -> #182323 #182322 While outage is happening See #182227
184404,[BE] Exclude `c10/metal` from HEADER_ONLY wraps,https://github.com/pytorch/pytorch/pull/184404,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-19T17:01:17Z,2026-06-19T01:01:48Z,20,"StableABI protects only CPU code part, and headers when used are fully migrated into the compiled metallib"
186012,Add ownership headers to GitHub workflows and a linter to enforce them,https://github.com/pytorch/pytorch/pull/186012,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/docker, ciflow/inductor-pallas",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,186844,,,0,3,0,,,"open source, topic: not user facing, ciflow/inductor-pallas, ciflow/docker",2026-06-02T23:47:47Z,2026-06-19T01:01:43Z,20,"Stack from ghstack (oldest at bottom): -> #186012 #186844 #186843 #186842 #186841 #186840 #186839 #186838 Mirror the test-file # Owner(s): [...] convention for every workflow under .github/workflows/ so that workflow failures can be routed to a responsible team, and add a WORKFLOWOWNERS lint that rejects any workflo..."
186844,Add owner headers to docs/lint/benchmark/vllm workflows,https://github.com/pytorch/pytorch/pull/186844,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186012,,,0,1,0,,,"open source, topic: not user facing",2026-06-09T20:50:41Z,2026-06-19T01:01:41Z,20,"Stack from ghstack (oldest at bottom): #186012 -> #186844 #186843 #186842 #186841 #186840 #186839 #186838 Attribute the remaining developer-tooling workflows under .github/workflows/ to their owning module (docs, lint, benchmark, vllm), following the test-file # Owner(s): [...] convention. Part of a stack that adds..."
187130,[BE] Do not try to install CUDA-12.8 wheels in torchbench images,https://github.com/pytorch/pytorch/pull/187130,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/docker,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"open source, topic: not user facing, ciflow/docker",2026-06-12T02:11:03Z,2026-06-19T01:01:39Z,20,Because latest torch is 13.0 by default
187545,[BE] Use `std::numbers` in `c10/util/MathConstants.h`,https://github.com/pytorch/pytorch/pull/187545,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, ciflow/docker",2026-06-17T02:30:57Z,2026-06-19T01:01:33Z,20,Stack from ghstack (oldest at bottom): -> #187545 #187525 Replace the hardcoded math literals with the bit-identical std::numbers constants (1/sqrt(2) has no equivalent and is kept) and drop the file from the CLANGTIDY burn-down list. Authored with assistance from Claude Code.
187479,Add Python c10d Backend trampoline,https://github.com/pytorch/pytorch/pull/187479,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-06-16T18:54:46Z,2026-06-19T00:29:35Z,20,"Summary Add PyBackend, a pybind trampoline for c10d::Backend, and bind the Backend APIs needed for Python subclasses to implement Backend-level collectives, p2p operations, fault-tolerance hooks, coalescing, metadata, tensor allocation, and lifecycle methods. Custom backend registration now owns the Python-object li..."
187419,Add numerics-debugging skill,https://github.com/pytorch/pytorch/pull/187419,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,topic: not user facing,2026-06-16T06:22:47Z,2026-06-18T22:27:01Z,20,"Stack from ghstack (oldest at bottom): -> #187419 Adds a numerics-debugging skill for localizing numeric divergence between two PyTorch runs by capturing plain torch.utils._debug_mode.DebugMode.debug_string() dumps with tensor hashes enabled. The workflow focuses on comparing the first mismatching hash / input_hash,..."
187624,[DRAFT]codeowner updates finer grained groups,https://github.com/pytorch/pytorch/pull/187624,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,topic: not user facing,2026-06-17T23:53:22Z,2026-06-18T22:26:02Z,20,Stack from ghstack (oldest at bottom): -> #187624
182074,[TEST][CUDA] Bump reflection_pad_2d decomp test to use `bfloat16` threshold,https://github.com/pytorch/pytorch/pull/182074,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cuda, module: tests, open source, module: padding, ciflow/trunk, topic: not user facing",2026-04-30T21:10:51Z,2026-06-18T22:06:23Z,20,"Otherwise we see data-dependent unlucky failures the 3d version of the test already uses the newer threshold targeting e.g., PYTORCH_OPINFO_SAMPLE_INPUT_INDEX=8 python test/test_decomp.py TestDecompCUDA.test_comprehensive_nn_functional_pad_reflect_cuda_bfloat16 authored with codex cc @ptrblck @msaroufim @jerryzh168..."
187674,Preserve boxed lifetimes for regional compiled calls,https://github.com/pytorch/pytorch/pull/187674,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,187675,,,0,1,0,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-18T16:24:04Z,2026-06-18T19:37:05Z,20,"Stack from ghstack (oldest at bottom): #187675 -> #187674 Graph Trainer exposed a lifetime regression in the regional compiled-call path. The compiled subgraphs themselves use boxed calling conventions, but regional_inductor and invoke_subgraph were rebuilding Python argument tuples around those calls. That kept the..."
187661,Return empty result for FFT with empty batch on MKL backend,https://github.com/pytorch/pytorch/pull/187661,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,174984,174984,174984,,,,0,1,0,,,"triaged, open source",2026-06-18T14:15:17Z,2026-06-18T19:36:04Z,20,"Fixes #174984 Summary On x86 (MKL), torch.fft.rfft (and the other FFTs) raised an opaque MKL FFT error: Intel oneMKL DFTI ERROR: Inconsistent configuration parameters when given a tensor with an empty batch dimension but a valid signal length, e.g. torch.zeros([0, 2, 16]). On ARM (pocketfft) the same input already r..."
181906,"[lint][RAWTHROW][3/5] aten, c10/util, distributed, inductor",https://github.com/pytorch/pytorch/pull/181906,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181904, 181905, 181907",,,0,4,0,,,"module: cpu, release notes: mobile, release notes: quantization, ciflow/inductor, release notes: inductor (aoti), ciflow/docker",2026-04-29T18:59:31Z,2026-06-18T19:35:00Z,20,"Stack from ghstack (oldest at bottom): #181907 -> #181906 #181905 #181904 Remove 4 RAWTHROW exclude patterns. Annotate vendored code (flat_hash_map, qnnpack), standalone runtime (aoti_runtime), and STL-convention throws (string_view). Convert std::invalid_argument/logic_error/runtime_error to TORCH_CHECK in distribu..."
181904,[lint][RAWTHROW][1/5] Small patterns sweep: annotate + fix 22 exclude groups,https://github.com/pytorch/pytorch/pull/181904,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181905, 181906, 181907",,,0,4,0,,,"release notes: distributed (c10d), module: dynamo, ciflow/inductor, module: compiled autograd, ciflow/docker",2026-04-29T18:59:21Z,2026-06-18T19:34:59Z,20,"Stack from ghstack (oldest at bottom): #181907 #181906 #181905 -> #181904 Remove 22 RAWTHROW exclude patterns from .lintrunner.toml covering the smallest violation groups. Most files get a per-file // @allow-raw-throw annotation for intentional throws (python_error, py::type_error, rethrows, test code). A handful of..."
181907,[lint][RAWTHROW][4/5] Live JIT: python + runtime,https://github.com/pytorch/pytorch/pull/181907,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181904, 181905, 181906",,,0,4,0,,,"release notes: jit, ciflow/docker",2026-04-29T18:59:37Z,2026-06-18T19:28:04Z,20,"Stack from ghstack (oldest at bottom): -> #181907 #181906 #181905 #181904 Replace the blanket torch/csrc/jit/** exclude with per-subdirectory permanent excludes for the 9 dead TorchScript subdirs. Fix the 2 live subdirs (jit/python, jit/runtime) with per-file annotations. Authored with Claude."
181905,"[lint][RAWTHROW][2/5] torch/csrc autograd, utils, profiler, fx",https://github.com/pytorch/pytorch/pull/181905,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181904, 181906, 181907",,,0,4,0,,,ciflow/docker,2026-04-29T18:59:26Z,2026-06-18T19:25:32Z,20,"Stack from ghstack (oldest at bottom): #181907 #181906 -> #181905 #181904 Remove 4 RAWTHROW exclude patterns. Most files get @allow-raw-throw for intentional patterns (python_error, PythonError, UnwindError). Convert a few std::logic_error/invalid_argument throws to TORCH_CHECK. Authored with Claude."
187548,Update ck aiter gfx1250,https://github.com/pytorch/pytorch/pull/187548,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, ciflow/trunk, ciflow/rocm-mi300",2026-06-17T04:48:56Z,2026-06-23T20:56:50Z,15,Changes to update CK and AITER while incorporating new fields and keeping backwards compatibility
187685,Name AOTAutograd backward gradient outputs grad_<primal> (#110700),https://github.com/pytorch/pytorch/pull/187685,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,110700,,,,0,2,0,,,"open source, topic: not user facing, ciflow/inductor, ciflow/torchtitan",2026-06-18T18:00:52Z,2026-06-18T18:36:46Z,20,"Summary Addresses part 2 of #110700. The joint/backward graph returns its gradients as anonymous compute nodes (e.g. mul_2), so TORCH_LOGS=aot_graphs output gives no indication of which returned gradient corresponds to which forward input. This names each returned gradient grad_<primal_name> using the GradAOTOutput..."
187414,[Inductor][cond] Allow output-input aliasing in torch.cond branches (re-land D106106380 + dev-nosan test fix),https://github.com/pytorch/pytorch/pull/187414,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-16T05:44:12Z,2026-06-18T18:34:15Z,20,"Summary: Re-land of D106106380 (reverted by D108642140 - ""not approved in OSS, needs more discussion""), folded together with its follow-up dev-nosan test fix D108479020. Allow output-input aliasing in torch.cond branches. torch.cond branches are mutually exclusive — only one branch executes at runtime. When the true..."
180641,Prune binary size from flash,https://github.com/pytorch/pytorch/pull/180641,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cuda, Stale, topic: not user facing",2026-04-17T04:02:49Z,2026-06-18T17:36:11Z,20,Stack from ghstack (oldest at bottom): -> #180641 Human Note: I was doing some investigations with codex using cubloaty: https://gist.github.com/drisspg/d47f0de45a7ccdd9a27f2f7056cbef13 this one feels like a no brainer Summary This PR stops building PyTorch’s vendored flash-attention CUDA kernels for sm_75. PyTorch...
186445,[AOTI] Avoid serializing full storage for shared-storage constants,https://github.com/pytorch/pytorch/pull/186445,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,186176,186176,186176,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-06-06T04:48:20Z,2026-06-18T17:12:23Z,20,"Fixes #186176 AOTI's binary-blob constant serialization wrote each constant's entire untyped_storage(). When constants share one storage (e.g. RNN params after flatten_parameters()), every view serialized the whole buffer, bloating the blob to ~N× for N shared views. For contiguous constants, this computes the seria..."
184886,[inductor] Diagonal-symmetric mm TritonTemplate for Gram matrices,https://github.com/pytorch/pytorch/pull/184886,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-05-22T12:53:17Z,2026-06-18T17:03:20Z,20,"Stack from ghstack (oldest at bottom): -> #184886 #184370 Tag Gram mm nodes in decomp_comms with math properties (meta[""math""][""symmetric""] = True) and add a TritonTemplate that exploits the symmetry: compute only upper-triangle tiles, mirror to lower. The template takes two inputs A, B (C = A @ B.T) and integrates..."
187662,Reject aliased Q/R out tensors in torch.linalg.qr,https://github.com/pytorch/pytorch/pull/187662,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,180377,180377,180377,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend",2026-06-18T14:15:31Z,2026-06-18T16:58:06Z,20,"Fixes #180377 Summary torch.linalg.qr(X, out=(B, B)) silently accepted the same tensor for both Q and R when X is square, so R overwrote Q and the Q result was lost. For non-square inputs it errored only incidentally, via a shape mismatch. Root cause Nothing checked that the two distinct outputs Q and R refer to non..."
133297,Update cmake,https://github.com/pytorch/pytorch/pull/133297,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,132368,132368,"133289, 133296, 133315, 133392, 133419, 133423, 133667, 133753",,,0,20,0,,,"module: cpu, module: mkldnn, open source, NNC, ciflow/trunk, release notes: quantization, release notes: releng, ciflow/inductor, no-stale, ciflow/linux-aarch64, release notes: inductor (aoti)",2024-08-13T03:42:24Z,2026-06-18T16:40:32Z,20,"MKLDNN was renamed to oneDNN in 2020. And it's available on Arm of course. We can gradually switch to the new name oneDNN. We plan to rename all ‘mkldnn’ to ‘onednn’, meanwhile, keep user facing APIs with ‘mkldnn’ and create user facing APIs with ‘onednn’. At one point in the near future, we will deprecate user faci..."
187435,[inductor] Support control_deps as fusion barriers,https://github.com/pytorch/pytorch/pull/187435,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, keep-going, ciflow/torchtitan",2026-06-16T12:21:54Z,2026-06-18T16:30:22Z,20,"Stack from ghstack (oldest at bottom): -> #187435 Teach control_deps to carry a compiler-only no_fuse_region flag. The flag reuses control_deps subgraph outlining, eager, fake, autograd, and lowering plumbing, but it does not create ordering dependencies: lowering materializes region boundary inputs, tags only ops c..."
175706,add C++20 bit operations and contains to C10,https://github.com/pytorch/pytorch/pull/175706,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, Stale, topic: not user facing",2026-02-25T02:34:11Z,2026-06-18T15:46:18Z,20,This PR moves some C10 code to use C++20 STL. It also updates torch targets to C++20.
133753,"Keep mkldnn_rnn_layer, add onednn_rnn_layer",https://github.com/pytorch/pytorch/pull/133753,draft/noise,31,"PR is draft; review decision requests changes; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"133289, 133296, 133297, 133315, 133392, 133419, 133423, 133667",,,0,33,0,,,"module: cpu, module: mkldnn, open source, module: amp (automated mixed precision), ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, no-stale, ciflow/linux-aarch64, release notes: inductor (aoti), ciflow/torchtitan",2024-08-17T05:10:11Z,2026-06-18T15:22:00Z,20,Stack from ghstack (oldest at bottom): -> #133753 #133667 #133423 #133419 #133392 #133315 #133297 #133296 #133289 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01 @gujinghui @PenghuiCheng @jianyuh @min-jean-cho @Guobing-Chen @Xia-Weiwen @snadampal @mcarilli @ptrblck @leslie-fang-intel...
133667,Update non user APIs - 2,https://github.com/pytorch/pytorch/pull/133667,draft/noise,31,"PR is draft; review decision requests changes; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/mps, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"133289, 133296, 133297, 133315, 133392, 133419, 133423, 133753",,,0,27,0,,,"module: cpu, module: mkldnn, open source, module: amp (automated mixed precision), NNC, ciflow/trunk, release notes: quantization, release notes: linalg_frontend, ciflow/mps, module: inductor, ciflow/inductor, no-stale, ciflow/linux-aarch64, release notes: inductor (aoti), ciflow/torchtitan",2024-08-16T06:26:37Z,2026-06-18T15:21:56Z,20,Stack from ghstack (oldest at bottom): #133753 -> #133667 #133423 #133419 #133392 #133315 #133297 #133296 #133289 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01 @gujinghui @PenghuiCheng @jianyuh @min-jean-cho @Guobing-Chen @Xia-Weiwen @snadampal @mcarilli @ptrblck @leslie-fang-intel...
133423,"Keep torch.backends.mkldnn, add torch.backends.onednn",https://github.com/pytorch/pytorch/pull/133423,draft/noise,31,"PR is draft; review decision requests changes; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/mps, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"133289, 133296, 133297, 133315, 133392, 133419, 133667, 133753",,,0,30,0,,,"module: cpu, module: mkldnn, open source, module: amp (automated mixed precision), ciflow/trunk, release notes: quantization, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, no-stale, ciflow/linux-aarch64, ciflow/torchtitan",2024-08-14T08:17:06Z,2026-06-18T15:21:52Z,20,Stack from ghstack (oldest at bottom): #133753 #133667 -> #133423 #133419 #133392 #133315 #133297 #133296 #133289 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01 @gujinghui @PenghuiCheng @jianyuh @min-jean-cho @Guobing-Chen @Xia-Weiwen @snadampal @mcarilli @ptrblck @leslie-fang-intel...
185930,[Elastic][Distributed] Flush dirty rendezvous state immediately,https://github.com/pytorch/pytorch/pull/185930,ready for maintainer decision,45,CI lane labels: ciflow/torchtitan; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: distributed (torchelastic), ciflow/torchtitan",2026-06-02T13:47:00Z,2026-06-18T14:33:39Z,20,"In current PyTorch, _DistributedRendezvousOpExecutor.run() syncs the shared rendezvous state at the beginning of each loop iteration. When an action mutates local rendezvous state, the executor marks the state as dirty, but the actual write-back to the backend is deferred until the next loop iteration. At very large..."
186566,Fix ldexp gradient for negative integer exponents,https://github.com/pytorch/pytorch/pull/186566,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,186556,186556,186556,,,,0,3,0,,,"triaged, open source, ciflow/trunk, release notes: autograd, topic: bug fixes, merging",2026-06-07T20:10:07Z,2026-06-18T14:09:56Z,20,"Fixes #186556 Summary torch.ldexp(x, exponent) returned an incorrect gradient w.r.t. x when exponent was an integer tensor with negative values. The gradient came back as 0 instead of 2**exponent. Root cause The backward formula for ldexp.Tensor in tools/autograd/derivatives.yaml computes the multiplier as at::pow(2..."
184696,Normalize export slice default bounds from getitem,https://github.com/pytorch/pytorch/pull/184696,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,178618,178618,,,,0,8,0,,,"Merged, Reverted, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: export, ci-no-td, ciflow/torchtitan, agentic",2026-05-21T11:05:03Z,2026-06-18T13:50:10Z,20,"Stack from ghstack (oldest at bottom): -> #184696 When tensor indexing lowers omitted slice bounds to aten.slice.Tensor, C++ fills in 0/sys.maxsize sentinel values. Convert getitem-originated unbounded defaults back to None in exported graphs while leaving explicit aten.slice calls untouched. Fixes #178618 Generated..."
170679,Fix FileBaton leaving dangling lock files that cause compilation hang…,https://github.com/pytorch/pytorch/pull/170679,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, Merged, Reverted, Stale, ciflow/trunk, topic: not user facing, suppress-bc-linter, ci-no-td",2025-12-17T15:17:09Z,2026-06-18T12:58:30Z,20,"The FileBaton implementation may leave a dangling lock file when the program exits abnormally. The next time it starts compiling, it will detect the lock file and enter a permanent wait state. We should use a real file lock."
133419,"Keep is_mkldnn, add is_onednn",https://github.com/pytorch/pytorch/pull/133419,draft/noise,41,"PR is draft; review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,132368,132368,"133289, 133296, 133297, 133315, 133392, 133423, 133667, 133753",,,0,29,0,,,"oncall: distributed, module: cpu, module: mkldnn, open source, module: amp (automated mixed precision), NNC, ciflow/trunk, release notes: quantization, module: inductor, ciflow/inductor, no-stale, ciflow/linux-aarch64, release notes: inductor (aoti), ciflow/torchtitan",2024-08-14T07:34:53Z,2026-06-18T12:58:26Z,20,"MKLDNN was renamed to oneDNN in 2020. And it's available on Arm of course. We can gradually switch to the new name oneDNN. We plan to rename all ‘mkldnn’ to ‘onednn’, meanwhile, keep user facing APIs with ‘mkldnn’ and create user facing APIs with ‘onednn’. At one point in the near future, we will deprecate user faci..."
133392,Update DispatchKey and namespace,https://github.com/pytorch/pytorch/pull/133392,draft/noise,41,"PR is draft; review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,132368,132368,"133289, 133296, 133297, 133315, 133419, 133423, 133667, 133753",,,0,31,0,,,"module: cpu, module: mkldnn, open source, module: amp (automated mixed precision), NNC, ciflow/trunk, release notes: quantization, module: inductor, ciflow/inductor, no-stale, ciflow/linux-aarch64, release notes: inductor (aoti), ciflow/torchtitan",2024-08-14T01:21:03Z,2026-06-18T12:58:24Z,20,"MKLDNN was renamed to oneDNN in 2020. And it's available on Arm of course. We can gradually switch to the new name oneDNN. We plan to rename all ‘mkldnn’ to ‘onednn’, meanwhile, keep user facing APIs with ‘mkldnn’ and create user facing APIs with ‘onednn’. At one point in the near future, we will deprecate user faci..."
133296,Update file name with onednn,https://github.com/pytorch/pytorch/pull/133296,draft/noise,41,"PR is draft; review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,132368,132368,"133289, 133297, 133315, 133392, 133419, 133423, 133667, 133753",,,0,17,0,,,"module: cpu, open source, NNC, ciflow/trunk, release notes: quantization, release notes: releng, module: inductor, ciflow/inductor, no-stale, release notes: inductor (aoti), ciflow/torchtitan",2024-08-13T03:42:16Z,2026-06-18T12:58:20Z,20,"MKLDNN was renamed to oneDNN in 2020. And it's available on Arm of course. We can gradually switch to the new name oneDNN. We plan to rename all ‘mkldnn’ to ‘onednn’, meanwhile, keep user facing APIs with ‘mkldnn’ and create user facing APIs with ‘onednn’. At one point in the near future, we will deprecate user faci..."
133315,Update non user APIs,https://github.com/pytorch/pytorch/pull/133315,draft/noise,41,"PR is draft; review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,132368,132368,"133289, 133296, 133297, 133392, 133419, 133423, 133667, 133753",,,0,26,0,,,"module: cpu, module: mkldnn, open source, module: amp (automated mixed precision), NNC, ciflow/trunk, release notes: quantization, release notes: linalg_frontend, module: inductor, ciflow/inductor, no-stale, ciflow/linux-aarch64, release notes: inductor (aoti), ciflow/torchtitan",2024-08-13T12:10:20Z,2026-06-18T12:58:17Z,20,"MKLDNN was renamed to oneDNN in 2020. And it's available on Arm of course. We can gradually switch to the new name oneDNN. We plan to rename all ‘mkldnn’ to ‘onednn’, meanwhile, keep user facing APIs with ‘mkldnn’ and create user facing APIs with ‘onednn’. At one point in the near future, we will deprecate user faci..."
187647,Remove XPU skips on test_partitioner_unbacked_symints and test_sdpa_s…,https://github.com/pytorch/pytorch/pull/187647,draft/noise,13,PR is draft; CI lane labels: ciflow/inductor,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-18T08:42:05Z,2026-06-18T11:25:08Z,20,Fixes: intel/torch-xpu-ops#3860 intel/torch-xpu-ops#3835 Remove @skipIfXpu decorators from two tests that now pass on XPU backend. Test commands: python -m pytest test_repros.py::ReproTestsDevice::test_partitioner_activation_memory_budget_with_unbacked_symints -xvs python -m pytest test_repros.py::ReproTestsDevice::...
170051,Add pivoted QR decomposition to ATen and torch.linalg,https://github.com/pytorch/pytorch/pull/170051,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,10454,,,,0,7,0,,,"triaged, open source, module: amp (automated mixed precision), Stale, release notes: linalg_frontend, module: inductor, ciflow/inductor, ciflow/torchtitan",2025-12-10T06:59:22Z,2026-06-18T11:06:22Z,20,"Pull Request: Add pivoted QR decomposition (qr_piv) to ATen and torch.linalg This PR adds a rank-revealing QR decomposition with column pivoting (pivoted QR) to ATen and the PyTorch torch.linalg namespace. Pivoted QR is widely used in numerical linear algebra, particularly in rank-revealing algorithms, subset select..."
184939,Support SymInt steps for linspace/logspace export,https://github.com/pytorch/pytorch/pull/184939,ready for maintainer decision,112,needs-review label present; approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,172930,172930,,,,0,4,0,,,"Merged, Reverted, ciflow/trunk, release notes: export, ci-no-td, agentic, needs-review",2026-05-22T22:47:15Z,2026-06-18T10:35:58Z,20,"Stack from ghstack (oldest at bottom): -> #184939 torch.export can represent data-dependent tensor scalars as unbacked SymInts when the user provides the needed runtime checks. That already works for shape-producing operators such as arange, but linspace and logspace still declared steps as a plain int. Passing a te..."
178150,[12/12] Enforce C++20 minimum in header guards (#178150),https://github.com/pytorch/pytorch/pull/178150,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Merged, Reverted, Stale, ciflow/trunk, release notes: cpp, topic: bc breaking, topic: improvements, module: inductor, ciflow/inductor, ci-no-td, meta-exported, ciflow/torchtitan",2026-03-23T15:15:33Z,2026-06-18T09:57:27Z,20,"Summary: Raise the compile-time enforcement from C++17 to C++20 in the header guards of c10/util/C++17.h, ATen/ATen.h, and torch/all.h. This is the enforcement gate — all consumers must already be building with -std=c++20 before this lands. #buildmore Test Plan: Sandcastle Reviewed By: huydhn Differential Revision:..."
178088,Use more structure binding,https://github.com/pytorch/pytorch/pull/178088,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,9,0,,,"module: cpu, triaged, open source, Stale, release notes: quantization, ciflow/inductor, release notes: inductor (aoti)",2026-03-22T10:47:27Z,2026-06-18T08:15:24Z,20,This PR applies structure binding to C++ loops and returning values to improve the readability and performance. It also simplifies capture lists of lambdas. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
187323,Fix Tensor.to() ignoring memory_format for non-contiguous tensors,https://github.com/pytorch/pytorch/pull/187323,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,"62027, 86558, 132020","62027, 86558, 132020","62027, 86558, 132020",,,,0,1,0,,,"triaged, open source",2026-06-15T09:11:30Z,2026-06-18T07:16:17Z,20,"Root cause: to_will_alias() used suggest_memory_format() which returns Contiguous as a default for anything not channels-last shaped — incorrectly skipping the copy for stride-0 (expand) and permuted-stride (transpose) tensors when contiguous_format was requested. Fix: For contiguous_format, replace suggest_memory_f..."
187636,[AOTI] Flip the default value of autotune_at_compile_time to False,https://github.com/pytorch/pytorch/pull/187636,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-18T02:22:50Z,2026-06-18T07:07:07Z,20,Stack from ghstack (oldest at bottom): -> #187636 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
187558,Fix torch.sign to propagate NaN instead of returning 0,https://github.com/pytorch/pytorch/pull/187558,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,"41245, 187295","41245, 187295","41245, 187295",,,,0,1,0,,,"module: bc-breaking, module: cpu, triaged, open source, topic: bc breaking, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-17T08:11:32Z,2026-06-18T05:05:22Z,20,"Fixes #187295 Fixes #41245 Summary torch.sign incorrectly returns 0 for NaN inputs due to the comparison-based implementation (0 < x) - (x < 0), where IEEE 754 comparisons with NaN are always false. This diverges from NumPy's np.sign, which propagates NaN. Changes: Fix the eager CPU kernel (scalar + SIMD vectorized..."
187640,Support in-tree XNNPACK build on Windows ARM64,https://github.com/pytorch/pytorch/pull/187640,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,open source,2026-06-18T04:34:24Z,2026-06-18T04:46:24Z,20,"The PR aims at enabling XNNPACK for Windows ARM64 builds. Three changes are required. Windows ARM64 reports CMAKE_SYSTEM_PROCESSOR=ARM64 (uppercase), which the lowercase-only {Q/X}NNPACK architecture guard in cmake/Dependencies.cmake did not match, so USE_XNNPACK was silently force-disabled there. Add ARM64 to the g..."
187612,docker-builds: move to default runner group (regular OSDC pool),https://github.com/pytorch/pytorch/pull/187612,draft/noise,13,PR is draft; CI lane labels: ciflow/docker,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,2,0,,,"topic: not user facing, ciflow/docker",2026-06-17T21:39:37Z,2026-06-17T23:59:57Z,21,"TODO: This requires the roll out of cache-enforcer removal on ci-infra Moves the docker-build job off the release-runners group onto the default group (the arc-cbr-production cluster, which deploys BuildKit), using the regular OSDC runner pools: Arch Before After x86 mt-rel-l-x86iavx512-8-64 mt-l-x86iavx512-2-4 arm6..."
187405,[dtensor] Propagate `_scaled_mm` tensor meta without eager scale checks,https://github.com/pytorch/pytorch/pull/187405,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,186997,,,,,0,1,0,,,"triaged, open source, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-16T01:48:47Z,2026-06-17T23:32:38Z,21,"DTensor sharding propagation computes output metadata by running the op in fake mode before applying the registered sharding strategy. For _scaled_mm this calls the eager meta implementation, which validates row-wise scale tensor shapes against the global tensor shapes. That rejects the DTensor row-parallel row-wise..."
187620,[FSDP2] Add PartialOffloadPolicy for fractional CPU parameter offload,https://github.com/pytorch/pytorch/pull/187620,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,"114299, 174960, 187615",,,,0,1,0,,,"open source, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-06-17T22:36:33Z,2026-06-17T22:49:51Z,21,"Validation status Opening as a draft to get maintainer direction on the API (per #187615) before investing in the full multi-rank test matrix. The pure offload selector has a standalone unit test passing locally: 26/26 (boundary equivalence at ratio 0.0 and 1.0, never-overshoot of the target fraction, set-monotonici..."
180527,Add thread_time_ns and idle_pct to dynamo_timed chromium events (#180527),https://github.com/pytorch/pytorch/pull/180527,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, Stale, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-16T03:42:08Z,2026-06-17T21:25:15Z,21,"Summary: Track CPU thread time alongside wall time in every dynamo_timed span. This enables identifying compilation phases where the CPU is idle (e.g., waiting on GPU synchronization, blocked on syscalls like munmap, or waiting on subprocess compilation workers). Two new fields are added to every chromium event's ar..."
182794,[inductor][rocm] Fix AOTInductor autotune int64 Triton kernel test and enable FP8 platform-aware handling,https://github.com/pytorch/pytorch/pull/182794,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,168566,168566,168566,,,,0,1,0,,,"module: rocm, triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor",2026-05-07T07:48:33Z,2026-06-17T21:24:04Z,21,"Fixes #168566 This PR fixes and enables AOTInductor tests to run correctly on ROCm devices by addressing platform-specific dtype mismatches, removing ROCM skip decorators, and fixing issues related to FP8 and int64 Triton kernel autotuning, where incorrect dtype handling led to int8 overflow behavior and memory-rela..."
180570,Fix torch.cond crash with FunctionalTensor constants (#180354),https://github.com/pytorch/pytorch/pull/180570,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180354,180354,"174929, 174930, 180354",,,,0,2,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: export",2026-04-16T15:39:23Z,2026-06-17T21:23:49Z,21,Fixes #180354 Summary Fixed a crash in torch.cond when branch functions contain tensor constants that become FunctionalTensor attributes in the GraphModule during export and decomposition. Problem When ExportedProgram.run_decompositions() was called on a program containing torch.cond with tensor constants in branche...
181465,Fix torch.fx.Tracer.record_stack_traces regression (user frames filtered by internal PyTorch frames),https://github.com/pytorch/pytorch/pull/181465,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,130861,130861,130861,,,,0,2,0,,,"triaged, open source, release notes: fx",2026-04-24T23:34:25Z,2026-06-17T21:23:24Z,21,Description Fixes #130861 The record_stack_traces feature in torch.fx.Tracer was broken after PR #121449 replaced the old _find_user_frame approach (which walked the callstack to find the first non-PyTorch frame) with CapturedTraceback.extract().summary(). The new approach returned ALL frames including internal PyTo...
180810,[inductor][static launcher] validate pointer arg dtypes before launch,https://github.com/pytorch/pytorch/pull/180810,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180807,180807,180807,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-20T05:21:04Z,2026-06-17T21:22:49Z,21,"Summary This PR adds runtime dtype validation for pointer-typed Triton arguments in the static launcher path. Previously, the static launcher compressed pointer-typed Triton args to ""O"" and did not preserve enough information to validate the runtime tensor dtype before launch. As a result, if compile-time metadata/f..."
179546,Add structured call_hierarchy metadata to FX nodes,https://github.com/pytorch/pytorch/pull/179546,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, release notes: fx, module: dynamo",2026-04-07T04:25:38Z,2026-06-17T21:21:31Z,21,"Add a new 'call_hierarchy' metadata field to FX nodes that provides a unified, ordered list of module and function call entries for each operation. Built during Dynamo's tx chain walk at proxy creation time, it uses nn_module_stack depth changes to distinguish module entries from function entries without relying on..."
179422,[inductor] Force contiguous inputs for custom ops in FallbackKernel,https://github.com/pytorch/pytorch/pull/179422,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: composability, module: inductor",2026-04-05T19:13:34Z,2026-06-17T21:19:45Z,21,"Summary Custom C++ ops registered via torch.library that call .data_ptr() on their inputs crash when torch.compile is used, because Inductor passes reinterpret_tensor views that lack accessible storage. Works on PyTorch 2.9.1, crashes on 2.11.0. Error RuntimeError: Cannot access data pointer of Tensor that doesn't h..."
175914,[Export] Fix unflatten crash when mutation intermediate has list args,https://github.com/pytorch/pytorch/pull/175914,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, fb-exported, Stale, module: inductor, release notes: export, meta-exported",2026-02-26T23:23:15Z,2026-06-17T21:17:39Z,21,"Summary: torch.export.unflatten crashes with AttributeError: 'immutable_list' object has no attribute 'graph' when processing mutation intermediates that have list-type arguments. In _IVals.read(), the code calls remap_input(node.args[0]) assuming args[0] is always a single torch.fx.Node. For ops like torch.cat or t..."
186579,Fix mse_loss/smooth_l1_loss allocating oversized storage for scalar output,https://github.com/pytorch/pytorch/pull/186579,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185647,185647,185647,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-08T04:50:01Z,2026-06-17T21:10:39Z,21,"Summary Fix F.mse_loss() and F.smooth_l1_loss() with reduction='mean' or 'sum' allocating storage equal to the full input tensor size instead of scalar size (4 bytes for float32). Root cause: the meta function called build_borrowing_binary_op() (full-size alloc) then resize_({}) (only changes shape, not storage). Fi..."
186104,[Inductor] Allow output-input aliasing in torch.cond branches (#186104),https://github.com/pytorch/pytorch/pull/186104,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-03T17:20:19Z,2026-06-17T21:07:05Z,21,"Summary: Allow output-input aliasing in torch.cond branches. torch.cond branches are mutually exclusive — only one branch executes at runtime. When the true branch returns an operand directly (e.g., identity/skip path), the output aliases the input. This is safe because the non-taken branch never runs, so there is n..."
187154,Use TORCH_INSTALL_LIB_DIR instead of hardcoded path 'lib',https://github.com/pytorch/pytorch/pull/187154,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,89489,89489,89489,,,,0,1,0,,,"triaged, open source, release notes: distributed (c10d), module: inductor, ciflow/inductor",2026-06-12T07:13:31Z,2026-06-17T21:06:49Z,21,Fixes #89489 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
180442,fix(mps): skip nograph linear on AMD RDNA2 GPUs when input_features>=256,https://github.com/pytorch/pytorch/pull/180442,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, Stale, release notes: mps, ciflow/mps",2026-04-15T07:20:07Z,2026-06-17T19:19:22Z,21,Summary The MPSNDArrayMatrixMultiplication kernel (nograph path) crashes on AMD RDNA2/RDNA3 discrete GPUs when input_features >= 256. Previous Fix (incorrect) The previous PR #178754 added a Python-level fallback in that replaced the entire function for all MPS devices when . This was wrong because: It affected all...
186570,[distributed] Fix Windows gloo allreduce crash by moving ProcessGroupGlooCuda.cpp to base_sources,https://github.com/pytorch/pytorch/pull/186570,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,186535,186535,186535,,,,0,1,0,,,"triaged, open source, release notes: distributed (c10d)",2026-06-07T21:10:27Z,2026-06-17T17:28:56Z,21,"Fixes #186535 What's wrong? The dist.all_reduce() function in a CUDA tensor with the gloo backend on Windows causes the process to fail due to an access violation 0xC0000005. No Python exception or stack trace is thrown; the process simply stops. Since NCCL is not available on Windows, gloo is the only multi-GPU opt..."
187566,[docs] Improve autograd anomaly mode documentation,https://github.com/pytorch/pytorch/pull/187566,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source",2026-06-17T10:09:27Z,2026-06-17T17:17:48Z,21,"Added automodule/autosummary/autoclass directives for torch.autograd.anomaly_mode Improved the module-level docstring in anomaly_mode.py Fixed the set_detect_anomaly docstring to mark check_nan as optional with a documented default value, matching PyTorch's docstring type formatting conventions. Test Plan: cd docs &..."
186797,"Fix mm_plus_mm XPU: guard broken Triton template, enable all mm_plus_mm tests",https://github.com/pytorch/pytorch/pull/186797,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,"173473, 184490","173473, 184490",,,,,0,3,0,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/xpu, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-06-09T15:16:28Z,2026-06-17T16:02:37Z,21,"Root Cause The Triton mm_plus_mm_template produces incorrect results on XPU. The benchmark() verification (which checks each autotuning choice against a reference) catches the mismatch and crashes the test. All three mm_plus_mm tests (test_mm_plus_mm, test_mm_plus_mm2, test_mm_plus_mm3) were affected. This is a know..."
180737,[inductor] Support pin_memory for torch.rand and torch.tensor,https://github.com/pytorch/pytorch/pull/180737,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,173939,174929,,,,0,1,0,,,"Stale, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-18T12:25:02Z,2026-06-17T15:47:07Z,21,Stack from ghstack (oldest at bottom): -> #180737 Fixes #173939 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
186836,Fix Cauchy/Gumbel icdf boundary values at q=0/q=1 (#186824),https://github.com/pytorch/pytorch/pull/186836,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,186824,186824,186824,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-06-09T20:40:13Z,2026-06-17T14:56:14Z,21,"Fixes #186824 Summary Cauchy.icdf and Gumbel.icdf return finite values at the support boundaries q=0 / q=1, where the unbounded support means they should return -inf / +inf. For Cauchy in float32 the returned value also has the wrong sign: import torch torch.distributions.Cauchy(0., 1.).icdf(torch.tensor(0.0)) # ten..."
187286,nn: add torch.nn.functional.rotary_embedding and torch.nn.RotaryEmbedding,https://github.com/pytorch/pytorch/pull/187286,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,149534,149534,149534,,,,0,3,0,,,"triaged, open source",2026-06-14T23:19:40Z,2026-06-17T14:54:27Z,21,"Fixes #149534 Summary Rotary Position Embedding (RoPE) is a foundational positional encoding technique widely used across modern large language models. Despite its importance, PyTorch lacked a canonical implementation in core, leading to divergent and incompatible ad-hoc versions across the ecosystem (torchtune, tor..."
185578,Fix addmm max-autotune SliceView bias guard,https://github.com/pytorch/pytorch/pull/185578,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185533,185533,185533,,,,0,3,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-05-29T01:42:46Z,2026-06-17T14:54:20Z,21,Fixes #185533 Summary: Guard the aten_bias_addmm max-autotune probe with maybe_get_stride() so lazy view inputs that do not expose concrete strides do not crash while checking the optional cuBLASLt bias candidate. Keep the bias fast path available for concrete 2D row-broadcast inputs; unrealized views skip that extr...
183673,[ROCm][Windows] torchgen: fix Windows DLL linkage for native and namespaced forward declarations,https://github.com/pytorch/pytorch/pull/183673,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,10,0,,,"module: rocm, triaged, open source",2026-05-14T07:59:08Z,2026-06-17T14:34:13Z,21,"This PR fixes the following inconsistent DLL issues on Windows: Namespaced declaration codegen always prefixed symbols with TORCH_API, regardless of dispatch backend. On Windows, TORCH_API, TORCH_CUDA_CPP_API, and TORCH_XPU_API encode which DLL owns the symbol’s import/export contract. Using the wrong macro on a for..."
187554,Fix flaky gradcheck in _gradcheck_log_prob for simplex distributions.,https://github.com/pytorch/pytorch/pull/187554,draft/noise,23,PR is draft; PR closes or references open issue numbers,wait for author to mark ready,,UNKNOWN,True,0,0,,,4022,4022,,,,,0,1,0,,,"open source, topic: not user facing",2026-06-17T07:36:16Z,2026-06-17T12:08:55Z,21,"_gradcheck_log_prob samples from the distribution under test and runs torch.autograd.gradcheck on its log_prob. For simplex-constrained distributions like RelaxedOneHotCategorical, low temperatures concentrate samples toward one-hot vectors, frequently producing components near zero. When gradcheck finite-difference..."
183530,feat: add XPUCodeCache tests for compilation and execution,https://github.com/pytorch/pytorch/pull/183530,draft/noise,23,"PR is draft; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,3095,3095,,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-13T08:57:28Z,2026-06-17T12:02:41Z,21,"Summary Harden and test the XPU (Intel GPU / SYCL) compile path in TorchInductor: add a dedicated XPUCompileError, wire it through XPUCodeCache, fix cache_clear() to also clear cached DLL wrappers, and add test coverage. Fixes intel/torch-xpu-ops#3095. Motivation XPUCodeCache inherited CUDA behavior: compile failure..."
184680,Inductor Lowering of Trunc Div should use div_rn,https://github.com/pytorch/pytorch/pull/184680,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,184408,184408,184408,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-21T07:59:21Z,2026-06-17T12:01:27Z,21,"Fixes: #184408 Use div_rn (IEEE round-to-nearest) instead of truediv here because Triton's default division uses an approximate reciprocal, which can nudge the quotient past an integer boundary and cause trunc() to return the wrong integer. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper..."
187564,"Register `linalg_qr`, `geqrf`, and `linalg_householder_product` for XPU",https://github.com/pytorch/pytorch/pull/187564,draft/noise,13,PR is draft; CI lane labels: ciflow/inductor,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, ciflow/inductor, release notes: inductor (aoti)",2026-06-17T09:03:21Z,2026-06-17T09:16:14Z,21,"Add XPU dispatch entries for: geqrf / geqrf.a linalg_householder_product / linalg_householder_product.out linalg_qr.out These ops reuse the existing shared implementations which internally dispatch via geqrf_stub and orgqr_stub, now registered for XPU in torch-xpu-ops. The XPU kernel implementations are in: intel/to..."
179235,[Inductor] Add _debug_assert_fused option to foreach_map,https://github.com/pytorch/pytorch/pull/179235,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,158970,158970,"158968, 158970",,,,0,1,0,,,"triaged, open source, ciflow/trunk, release notes: foreach_frontend, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-03T09:16:13Z,2026-06-17T08:52:35Z,21,"Summary Fixes #158970. Part of #158968. Add an _debug_assert_fused option to torch._higher_order_ops.foreach_map. When _debug_assert_fused=True, foreach_map now errors in lowering if the subgraph cannot remain on the foreach fused path. When _debug_assert_fused=False, behavior is unchanged. This change also adds tes..."
172539,[torchgen] Enable Structured Kernels and Out-as-Primary Wrappers for PrivateUse1,https://github.com/pytorch/pytorch/pull/172539,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,172538,172538,172538,,,,0,27,0,,,"triaged, open source, topic: not user facing",2026-01-15T08:42:58Z,2026-06-17T07:00:03Z,21,"Fix #172538 Summary This PR makes gen_backend_stubs.py support full in-tree aten structured features for out-of-tree PrivateUse1 backends. It enhances the PrivateUse1 (custom backend) infrastructure by reusing native Structured Kernels and an automated ""Out-as-Primary"" dispatch pattern. This allows external hardware..."
186050,Output Shape Check added to transposed Conv,https://github.com/pytorch/pytorch/pull/186050,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,185581,185581,185581,,,,0,1,0,,,"module: cpu, triaged, open source, topic: not user facing",2026-06-03T06:55:30Z,2026-06-17T05:12:18Z,21,"Fixes #185581 Add output size validation for transposed convolution in the common check_shape_forward path in Convolution.cpp. Previously, the ""output size too small"" check only existed in NaiveConvolutionTranspose2d.cpp (the SlowTranspose2d backend and was also missing in cpu_template), which is never reached on CP..."
185807,Stable sorting for consistent gradient for torch.quantile on tied values,https://github.com/pytorch/pytorch/pull/185807,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,185543,185543,185543,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-01T10:14:40Z,2026-06-17T05:11:36Z,21,"Fixes #185543 quantile_compute (device-agnostic code in Sorting.cpp) calls self.sort() without the stable flag. For tied values, CUDA eager's has different sorting mechanism (doesn't gaurantees the original relative order) and produces a different element ordering than Inductor's compiled path . The quantile value i..."
186543,Fix Windows debug build issues,https://github.com/pytorch/pytorch/pull/186543,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/xpu",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,186544,186544,186544,,,,0,7,0,,,"triaged, open source, module: inductor, module: dynamo, ciflow/inductor, ciflow/xpu, ciflow/torchtitan",2026-06-07T12:32:48Z,2026-06-17T05:11:10Z,21,"Fix #186544 Summary Fix Windows Debug build (CMAKE_BUILD_TYPE=Debug) failing with linker error python310_d.lib not found when using a standard (non-debug) CPython distribution from conda/pip. Problem When building PyTorch with MSVC in Debug mode, the compiler defines _DEBUG. This causes CPython's Python.h to: Emit #..."
187391,[export][serde] Serialize SymBoolArgument in graph input/output specs (#187391),https://github.com/pytorch/pytorch/pull/187391,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, module: dynamo, ciflow/inductor, release notes: export, meta-exported",2026-06-15T22:28:36Z,2026-06-17T05:10:33Z,21,"Summary: GraphModuleSerializer.serialize_argument_spec and its deserialize_argument_spec counterpart handled TensorArgument, SymIntArgument, SymFloatArgument, ConstantArgument, and CustomObjArgument, but not SymBoolArgument. When an exported program has a SymBool graph input/output (e.g. a module that returns x.shap..."
187430,Add Kaiser-Bessel Derived window,https://github.com/pytorch/pytorch/pull/187430,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,85366,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-06-16T10:45:08Z,2026-06-17T05:09:59Z,21,Relates to #85366 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98
181167,remove size oblivious reasoning core,https://github.com/pytorch/pytorch/pull/181167,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, keep-going, suppress-bc-linter, ciflow/torchtitan",2026-04-22T20:25:39Z,2026-06-17T05:09:46Z,21,Stack from ghstack (oldest at bottom): -> #181167 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98
180722,Friendler error messages,https://github.com/pytorch/pytorch/pull/180722,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-04-17T22:39:42Z,2026-06-17T03:48:46Z,21,"Stack from ghstack (oldest at bottom): -> #180722 More helpful error messages: #!/usr/bin/env python3 import argparse import traceback from collections.abc import Callable import torch from torch.nn.functional import ScalingType, scaled_grouped_mm def make_low_precision(shape: tuple[int, ...], *, transpose_last_two:..."
179747,Add rotary positional embeddings support to nn.MultiheadAttention,https://github.com/pytorch/pytorch/pull/179747,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,6,0,,,"triaged, open source, Stale, release notes: quantization",2026-04-08T19:09:42Z,2026-06-17T03:48:39Z,21,"Adds rotary positional embeddings support to nn.MultiheadAttention via a new use_rotary option, integrates RoPE application into the attention path for query and key projections, and includes typing, CPU, and CUDA test coverage to validate the new feature end to end."
187490,[Dynamo] Trace raw unbacked SymInt inputs,https://github.com/pytorch/pytorch/pull/187490,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,187272,,181167,,,0,9,0,,,"module: cpu, release notes: fx, module: dynamo, ciflow/inductor, keep-going, ciflow/torchtitan",2026-06-16T19:36:36Z,2026-06-17T03:27:22Z,22,Stack from ghstack (oldest at bottom): #187491 #181167 -> #187490 #187489 Non-strict tracing can pass raw unbacked SymInt values into a nested Dynamo trace. FlexAttention hits this when BlockMask.seq_lengths are derived from an unbacked query/key sequence dimension and then passed through block_mask.as_tuple() to th...
133289,Update native/mkldnn to native/onednn,https://github.com/pytorch/pytorch/pull/133289,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,132368,132368,"133296, 133297, 133315, 133392, 133419, 133423, 133667, 133753",,,0,14,0,,,"module: cpu, open source, NNC, ciflow/trunk, release notes: quantization, release notes: releng, ciflow/inductor, no-stale, release notes: inductor (aoti), ciflow/docker",2024-08-13T02:18:25Z,2026-06-17T01:25:48Z,22,"MKLDNN was renamed to oneDNN in 2020. And it's available on Arm of course. We can gradually switch to the new name oneDNN. We plan to rename all ‘mkldnn’ to ‘onednn’, meanwhile, keep user facing APIs with ‘mkldnn’ and create user facing APIs with ‘onednn’. At one point in the near future, we will deprecate user faci..."
187492,Avoid std::views::reverse in CUDACachingAllocator,https://github.com/pytorch/pytorch/pull/187492,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,topic: not user facing,2026-06-16T19:39:33Z,2026-06-17T01:16:14Z,22,"PR #186308 replaced explicit reverse-iterator loops with std::views::reverse(...) in trimHandles() and the two LIFO lookups over allocation_scopes_. That fails to compile with clang-15 + libstdc++ 11.x (Meta's FRL toolchain): std::views::reverse over those containers does not satisfy the range concepts, yielding ""no..."
187472,[dynamo] don't reconstruct live torch.Stream values as StreamContext across graph breaks,https://github.com/pytorch/pytorch/pull/187472,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,184487,,,0,1,0,,,"ciflow/trunk, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-16T18:21:37Z,2026-06-17T00:00:58Z,22,Summary: A plain torch.Stream value that is live across a graph break occurring inside a with torch.cuda.stream(side): block was being silently corrupted into a torch.cuda.StreamContext. A subsequent eager Tensor.record_stream(s) then crashed with RuntimeError: unknown parameter type (the C++ arg parser has no STREA...
180707,Improve cpp_extension and drop setuptools runtime requirement,https://github.com/pytorch/pytorch/pull/180707,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,180624,180624,180624,180250,,,0,8,0,,,open source,2026-04-17T18:54:30Z,2026-06-16T23:41:24Z,22,"This implements the changes proposed in #180624, clarifying the different facets of torch.utils.cpp_extension and removing setuptools as a runtime requirement. Depends on the stack #177641--#180250. Closes #180624."
185316,[dynamo] Resolve autograd.Function static attrs in getattr,https://github.com/pytorch/pytorch/pull/185316,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,160757,160757,,,,0,2,0,,,"ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, agentic",2026-05-27T03:15:56Z,2026-06-18T01:24:49Z,21,"Stack from ghstack (oldest at bottom): -> #185316 Dynamo already represents torch.autograd.Function subclasses with AutogradFunctionVariable and routes apply through special handling that traces forward and backward into the autograd_function_apply higher-order op. However, plain attribute reads on the subclass stil..."
187410,[distributed] Use node.target instead of node.name in apply_manual_re…,https://github.com/pytorch/pytorch/pull/187410,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-16T02:36:00Z,2026-06-16T21:31:07Z,22,"…ordering_and_get_graph FX node names are assigned by tracing order and are not stable identifiers. Any upstream change that adds, removes, or reorders collective ops silently shifts the names (_1, _2 suffixes), causing nn_module_stack metadata to be injected onto the wrong node. The dependent TestManualOverlapBucke..."
187245,update backend parametrization for varlen attn,https://github.com/pytorch/pytorch/pull/187245,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-13T12:13:05Z,2026-06-16T21:26:56Z,22,"This pull request updates the test parameterization logic in test/test_varlen_attention.py to use a helper function for backend selection, improving maintainability and consistency. The main changes are: Test Parameterization Improvements: Replaced manual construction of the backend parameter list in test_batch_inva..."
184968,Fix autocast state propagation for normalization layers in functorch batched autograd,https://github.com/pytorch/pytorch/pull/184968,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,184890,184890,184890,,,,0,2,0,,,"triaged, open source",2026-05-23T06:43:32Z,2026-06-16T21:25:19Z,22,"Fixes #184890 Description When calling torch.vmap over torch.autograd.grad on a model containing normalization layers (BatchNorm, InstanceNorm, LayerNorm, GroupNorm) run under torch.autocast, the backward crashes with RuntimeError: expected scalar type Half but found Float. Root Cause: The functorch batched autograd..."
187253,Align dynamic qparams test helper with C++ overflow check,https://github.com/pytorch/pytorch/pull/187253,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187048,187048,187048,,,,0,1,0,,,"triaged, open source, release notes: quantization, topic: not user facing",2026-06-13T17:57:54Z,2026-06-16T21:10:10Z,22,"Problem _calculate_dynamic_qparams in test_quantized_tensor.py used double-precision reciprocal overflow detection, while ChooseQuantizationParams in C++ checks std::isinf(1.0f / float(scale)). For degenerate tensors where min == max == float32.tiny, the Python helper and backend can disagree on whether to fall back..."
180528,Add dynamo_timed wrappers for AOT autograd compilation stages (#180528),https://github.com/pytorch/pytorch/pull/180528,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-16T03:42:33Z,2026-06-16T20:31:38Z,22,"Summary: Wrap additional AOT autograd stages with dynamo_timed for compile-time visibility in tlparse chromium traces. These stages were previously invisible in the trace, making it difficult to understand where time is spent during AOT dispatch. New instrumented stages: aot_dispatch_base_graph — the full base graph..."
180684,AOTI: Write tensor data directly to file without Python bytes copy,https://github.com/pytorch/pytorch/pull/180684,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180682, 180683",,,0,1,0,,,"Stale, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-17T16:12:21Z,2026-06-16T20:31:37Z,22,"Stack from ghstack (oldest at bottom): -> #180684 #180683 #180682 When streaming constants to the external weights file, use a ctypes buffer wrapping the tensor's raw memory and write it directly to the file. This avoids creating a Python bytes object per tensor (~GB for large weight tensors), reducing transient mem..."
179700,[torch][inductor] Add RAIIAtenRecordFunctionHandle profiling to remaining uncovered kernel paths (#179700),https://github.com/pytorch/pytorch/pull/179700,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"fb-exported, Stale, module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-04-08T08:59:43Z,2026-06-16T20:31:36Z,22,"Summary: This adds RAIIAtenRecordFunctionHandle profiling to 4 kernel call paths that previously had no profiler events, making them visible in GPU traces with ""External id"" linkage: generate_fallback_kernel_with_runtime_lookup_aot (cpp_wrapper_cpu.py): AOT ProxyExecutor dispatch via aoti_torch_proxy_executor_call_f..."
184487,dynamo: round-trip torch.cuda.stream ctx mgr across graph breaks,https://github.com/pytorch/pytorch/pull/184487,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,11,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, ci-no-td, ciflow/torchtitan",2026-05-20T05:46:08Z,2026-06-16T19:42:01Z,22,"Upstream #166472 ('[user-streams] Switch to fx annotations at trace time') changed StreamContextVariable.target_values from a positional list [stream_var] to a dict {""stream"": idx} so that StreamContextVariable.enter() (inherited from FxTracebackAnnotateVariable.enter) could feed it to torch.fx.traceback.annotate. T..."
182431,[Inductor] Elide redundant .to(tl.float32) in Triton codegen when value already has target dtype,https://github.com/pytorch/pytorch/pull/182431,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,182430,182430,182430,,,,0,1,0,,,"open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-05T13:32:00Z,2026-06-16T19:05:49Z,22,TritonOverrides.to_dtype() unconditionally emits .to(out_dtype) even when the CSE variable already has the target dtype. This produces redundant no-op casts (e.g. tmp3 = tmp2.to(tl.float32) where tmp2 is already float32) when codegen_upcast_to_fp32=True and the IR has an explicit to_dtype node whose logical source d...
180683,AOTI: Stream external weights to disk instead of in-memory blob,https://github.com/pytorch/pytorch/pull/180683,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180682, 180684",,,0,1,0,,,"Stale, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-17T16:12:16Z,2026-06-16T18:53:41Z,22,"Stack from ghstack (oldest at bottom): #180684 -> #180683 #180682 When using binary_blob format with external weights (the Metal/MPS path), serialized_weights was built as a single Python bytes object holding all model constants (~22GB for large models), then written to disk. This doubles memory: tensors + bytes blo..."
180682,AOTI: Only clone mutated buffers in _unlift_graph,https://github.com/pytorch/pytorch/pull/180682,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180683, 180684",,,0,1,0,,,"Stale, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-17T16:12:12Z,2026-06-16T18:53:41Z,22,"Stack from ghstack (oldest at bottom): #180684 #180683 -> #180682 Previously, _unlift_graph cloned ALL buffers via clone_preserve_strides to preserve their pre-tracing values. For large models with many non-mutated buffer weights (e.g. quantized MoE with 256 experts), this wastes ~20GB of memory by cloning read-only..."
178209,Fix CUDA batch_norm eval returning populated save_mean/save_invstd,https://github.com/pytorch/pytorch/pull/178209,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,85960,85960,85960,,,,0,1,0,,,"triaged, open source, Stale, release notes: cuda",2026-03-23T23:00:48Z,2026-06-16T18:53:29Z,22,"Fixes #85960 Summary In eval mode, the CUDA native_batch_norm kernel was copying running_mean into save_mean and computing save_invstd from running_var, returning populated tensors of shape [num_features]. The CPU kernel returns empty (size-0) tensors for both, since they are not needed for the backward pass in eval..."
187369,[Test] Add test hardware classification for test_complex.py and test_kernel_launch_checks.py,https://github.com/pytorch/pytorch/pull/187369,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,"185142, 185590",186918,,,0,1,0,,,"open source, topic: not user facing",2026-06-15T18:20:23Z,2026-06-16T18:51:59Z,22,"…kernel_launch_checks.py Classify test classes in test/autograd/test_complex.py and test/test_kernel_launch_checks.py by hardware scope per the test classification RFC (#185142, #185590). Both files contain tests that exercise pure CPU-side logic with no accelerator dependency, autograd complex view mechanics and CU..."
183638,[inductor] Fuse NVFP4 nested-reduction packing,https://github.com/pytorch/pytorch/pull/183638,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,13,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-14T00:35:08Z,2026-06-16T18:50:10Z,22,"Stack from ghstack (oldest at bottom): -> #183638 Add the half-resolution nested-reduction epilogue needed by the NVFP4 pair-packing pattern. The grouped amax/scale stage still runs at reduced resolution, while the inline-asm pack body runs over the pair-resolution view and reuses the parent-tile values already load..."
178004,[ROCm] enable compiler-driven loop unrolling,https://github.com/pytorch/pytorch/pull/178004,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, Stale, ciflow/trunk, topic: not user facing, ciflow/rocm-mi300, ciflow/rocm-mi355",2026-03-20T20:48:28Z,2026-06-16T17:27:05Z,22,"LLVM is now capable of unrolling the modified normalization loop, thanks to the following change: llvm/llvm-project#181241 cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang"
176657,[CUDAGraph] Detect and error on freeing parent-capture tensors inside CUDA graph conditional nodes,https://github.com/pytorch/pytorch/pull/176657,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,176670,,,0,3,0,,,"module: cuda, open source, module: cuda graphs, Stale, topic: not user facing, module: CUDACachingAllocator",2026-03-05T23:09:03Z,2026-06-16T16:35:47Z,22,"Stack from ghstack (oldest at bottom): #176670 -> #176657 If a tensor allocated during parent CUDA graph capture is freed inside a conditional node’s child graph capture, the allocator can incorrectly return the block to the free pool. This happens because the child capture runs on a different stream and therefore h..."
185645,Add batched small-matrix CUDA kernel for linalg.eig,https://github.com/pytorch/pytorch/pull/185645,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,183806,183806,183806,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend",2026-05-29T17:07:06Z,2026-06-16T15:55:48Z,22,Summary Fixes #183806 cuSolver's Xgeev has no batched API. Right now PyTorch loops over each matrix with one kernel launch per element which kills performance for small batched matrices since the launch overhead dwarfs the actual math. I wrote a custom CUDA kernel that handles the whole batch in one launch for n <=...
179702,[torch][inductor] Add dump_python_module config for AOT single-pass mode (#179702),https://github.com/pytorch/pytorch/pull/179702,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"fb-exported, Stale, module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-04-08T09:00:16Z,2026-06-16T15:36:46Z,22,"Summary: When autotune_at_compile_time=True (the default for AOT compilation), the single-pass codegen flow goes directly from C++ wrapper generation to .so compilation, skipping compile_to_module() entirely. This means no Python compiled module (.py) is written to PyCodeCache — unlike the two-pass flow (autotune_at..."
185612,[WIP][CI][CRCR] Update mergebot rule to adapt CRCR L3,https://github.com/pytorch/pytorch/pull/185612,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-05-29T08:35:24Z,2026-06-16T15:35:35Z,22,This WIP PR is a placeholder for now. CRCR L3 is under developing and the implementation is not settled down.
187238,Add docs pages for torch.e and torch.pi,https://github.com/pytorch/pytorch/pull/187238,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,134964,134964,134964,,,,0,1,0,,,"triaged, open source, release notes: python_frontend, topic: docs",2026-06-13T05:28:43Z,2026-06-16T14:49:12Z,22,"Fixes #134964 Problem: torch.e and torch.pi are public constants, but the docs only listed them in the constants table and did not generate dedicated API pages for them. Fix: Add e and pi to the torch.md constants autosummary so Sphinx generates generated/torch.e and generated/torch.pi pages. Override the autodoc do..."
169249,[ARM] Disable SVE for GCC 11.5.0 to workaround compiler bug,https://github.com/pytorch/pytorch/pull/169249,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,162422,162422,162422,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2025-11-30T04:26:52Z,2026-06-19T14:20:18Z,19,"GCC 11.5.0 has a known regression causing internal compiler errors when compiling SVE intrinsics. This manifests as segmentation faults during the GIMPLE optimization pass. This change detects GCC 11.5.0 and automatically disables SVE support, falling back to NEON vectorization. This prevents build failures while ma..."
186589,Fix serialization errors in SubTest calls,https://github.com/pytorch/pytorch/pull/186589,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,3797,3797,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-08T08:27:29Z,2026-06-16T14:25:09Z,22,"Change subTest method of passing functions in test_prunning.py and test_autograd.py to not cause serialization errors when running with xdist. This aligns with methods used in the rest of the repo. Surfaced in: intel/torch-xpu-ops#3793, intel/torch-xpu-ops#3797"
187329,Raise ValueError for non-positive kernel_size/dilation/stride in unfold/fold,https://github.com/pytorch/pytorch/pull/187329,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,187299,187299,187299,,,,0,1,0,,,"triaged, open source, release notes: nn, ciflow/mps",2026-06-15T11:57:59Z,2026-06-16T14:24:24Z,22,"Fixes #187299 Problem torch.nn.functional.unfold / fold forward non-positive kernel_size, dilation, or stride straight into the im2col/col2im kernels, which surface a RuntimeError from deep in the implementation: F.unfold(torch.randn(1, 3, 8, 8), kernel_size=(3, 3), stride=(0, 1)) # RuntimeError: stride should be gr..."
187243,Reject p < 1 in triplet_margin_loss to prevent NaN from p-norm overflow,https://github.com/pytorch/pytorch/pull/187243,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187178,187178,187178,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-13T10:45:07Z,2026-06-16T14:24:23Z,22,"Summary F.triplet_margin_loss silently returns NaN when p < 1 because the p-norm computation overflows. For very small p, each (|a - b| + eps)^p is close to 1, so the sum across the feature dimension is roughly N, and N^(1/p) blows past fp32 range. The loss then becomes inf - inf + margin = NaN with no error. The do..."
187242,Validate LocalResponseNorm parameters to prevent silent inf output,https://github.com/pytorch/pytorch/pull/187242,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187177,187177,187177,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-13T10:38:02Z,2026-06-16T14:24:23Z,22,"Summary LocalResponseNorm silently produces +/-inf when both k and alpha are set to non-positive values. The denominator (k + alpha/size * sum(x^2)) becomes zero in this case, and 0^(-beta) evaluates to infinity for any positive beta. This adds a ValueError in both LocalResponseNorm.__init__ and F.local_response_nor..."
173519,Add on-demand ppc64le wheel build support,https://github.com/pytorch/pytorch/pull/173519,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,5,0,,,"module: cpu, triaged, open source, Stale, release notes: quantization, release notes: releng",2026-01-27T15:12:47Z,2026-06-16T13:06:41Z,22,"Fixes #ISSUE_NUMBEPR Description This PR introduces initial support for building ppc64le wheels in the CI/CD pipeline using a self-hosted runner provisioned under the existing s390x account, following guidance from the CI team. The goal is to enable Power (ppc64le) architecture compatibility for wheel builds while k..."
177649,BLAS sparse ops: dense arguments -- extend support beyond just dense and non-overlapping inputs.,https://github.com/pytorch/pytorch/pull/177649,draft/noise,13,PR is draft; CI lane labels: ciflow/trunk,wait for author to mark ready,,CLEAN,True,0,0,,,,,176274,"177648, 178167, 178298",,,0,33,0,,,"open source, ciflow/trunk, release notes: sparse",2026-03-17T14:32:49Z,2026-06-16T12:24:10Z,22,Stack from ghstack (oldest at bottom): -> #177649 #178298 #178167 #177648 Test for the utils from the previous PR -- not a batched case yet. Part of #176274.
178167,triangular_solve_stub: expand leading dim support beyond dense and non-overlapping inputs,https://github.com/pytorch/pytorch/pull/178167,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,176274,"177648, 177649, 178298",,,0,26,0,,,"open source, ciflow/trunk, release notes: linalg_frontend",2026-03-23T17:29:01Z,2026-06-16T12:23:46Z,22,Stack from ghstack (oldest at bottom): #177649 #178298 -> #178167 #177648 Tackle #176274 for torch.triangular_solve.
177648,LinAlgUtils -- update the toolkit to address the MathLibs dispatch with redundant memory allocations/movements.,https://github.com/pytorch/pytorch/pull/177648,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,176274,"177649, 178167, 178298",,,0,31,0,,,"open source, topic: not user facing, ciflow/inductor",2026-03-17T14:32:44Z,2026-06-16T12:23:27Z,22,Stack from ghstack (oldest at bottom): #177649 #178298 #178167 -> #177648 Toolkit for tackling #176274
178298,torch.linalg.solve_triangular: support for eligible non-dense args,https://github.com/pytorch/pytorch/pull/178298,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,178379,"176274, 178379","177648, 177649, 178167",,,0,27,0,,,"module: cuda, module: cpu, module: cublas, open source, module: linear algebra, ciflow/trunk, release notes: linalg_frontend, ciflow/inductor",2026-03-24T17:23:23Z,2026-06-16T12:23:01Z,22,Stack from ghstack (oldest at bottom): #177649 -> #178298 #178167 #177648 This PR rewrites the logic behind linalg.solve_triangular to address: memory issues as per #176274 Fixes #178379 conj never triggers a clone (conditioned on inputs with reasonable strides). conj materialization is deterred and only materialize...
186930,Fix NaN propagation behavior in torch.sign for CPU and CUDA backends,https://github.com/pytorch/pytorch/pull/186930,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,186901,186901,186901,,,,0,8,0,,,"module: cpu, triaged, open source",2026-06-10T13:35:38Z,2026-06-16T10:13:22Z,22,Description This PR fixes an issue where torch.sign incorrectly handles NaN floating-point inputs by returning 0 instead of propagating the NaN value through the operator kernel. The current behavior violates modern numerical parity expectations (aligning with NumPy and TensorFlow 2.x) and standard IEEE 754 floating...
185790,Fix torch.nn.functional.gelu returning NaN for infinite inputs,https://github.com/pytorch/pytorch/pull/185790,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,185770,,,,0,1,0,,,"module: cpu, triaged, open source, release notes: nn",2026-06-01T04:21:18Z,2026-06-16T09:23:20Z,22,"#185770 gelu(+inf) should evaluate to +inf and gelu(-inf) should evaluate to 0, but both cases currently return NaN. The -inf issue comes from IEEE 754 behavior (-inf * 0 = NaN) and affects all GELU implementations. The +inf issue is specific to the MKLDNN path. This change adds explicit handling for infinite inputs..."
186752,[Inductor][Triton] Add host-side TMA support for B200 GEMM templates,https://github.com/pytorch/pytorch/pull/186752,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"module: mkldnn, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-06-09T01:44:25Z,2026-06-16T08:19:37Z,22,"Adds host-side TMA support for the B200 GEMM template in Inductor, building off #185825 (issue tracked in #185819), turned on with ENABLE_HOST_SIDE_TMA=1. This PR uses a new {{tma_descriptor()}} Jinja hook to create TMA descriptors, either on the host (via registers in host_tma_descriptor_args) or on device (by emit..."
186756,[Inductor][Host-side TMA GEMM] Enable host-side TMA for TMA epilogue store,https://github.com/pytorch/pytorch/pull/186756,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-09T03:34:41Z,2026-06-16T08:19:28Z,22,"Extend host-side TMA to cover the output store descriptor. When enable_host_side_tma=True and enable_template_tma_store=True, the output buffer's TMA descriptor is now created on the host (via the launcher) instead of device-side inside the kernel. The tma_descriptor() hook registers the output buffer in host_tma_de..."
181570,Fix batched LU factorization grain_size heuristic for large matrices (#181570),https://github.com/pytorch/pytorch/pull/181570,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, release notes: linalg_frontend, meta-exported",2026-04-27T10:17:15Z,2026-06-16T08:18:26Z,22,"Summary: The parallelization heuristic in apply_lu_factor computes chunk_size_per_thread as min(1.0, 3200 / rank^3) and then grain_size = chunk_size_per_thread * num_threads. For large matrices (rank > ~14), this truncates to chunk_size_per_thread = 0, producing grain_size = 0. However, at::parallel_for interprets g..."
186988,Move header-only c10/util Layer 1 utils to torch/headeronly,https://github.com/pytorch/pytorch/pull/186988,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"186980, 186981, 186982, 186983, 186984, 186985, 186986, 186987",,,0,4,0,,,"ciflow/inductor, ciflow/docker",2026-06-10T22:13:27Z,2026-06-16T06:07:48Z,22,"Stack from ghstack (oldest at bottom): -> #186988 #186987 #186986 #186985 #186984 #186983 #186982 #186981 #186980 #186962 Migrate three already-header-only c10/util headers that depend on headers migrated earlier in this wave into torch/headeronly/util/, leaving the old c10 paths as one-line forwarders (fully backwa..."
186987,Move header-only c10/util large/vendored headers to torch/headeronly,https://github.com/pytorch/pytorch/pull/186987,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"186980, 186981, 186982, 186983, 186984, 186985, 186986, 186988",,,0,4,0,,,"ciflow/inductor, ciflow/docker",2026-06-10T22:13:21Z,2026-06-16T06:07:36Z,22,"Stack from ghstack (oldest at bottom): #186988 -> #186987 #186986 #186985 #186984 #186983 #186982 #186981 #186980 #186962 Migrate four large, already-header-only c10/util headers into torch/headeronly/util/, leaving the old c10 paths as one-line forwarders (fully backward compatible, no consumer include changes): fl..."
186986,Move header-only c10/util string/filesystem utils to torch/headeronly,https://github.com/pytorch/pytorch/pull/186986,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"186980, 186981, 186982, 186983, 186984, 186985, 186987, 186988",,,0,4,0,,,ciflow/inductor,2026-06-10T22:13:16Z,2026-06-16T06:06:59Z,22,"Stack from ghstack (oldest at bottom): #186988 #186987 -> #186986 #186985 #186984 #186983 #186982 #186981 #186980 #186962 Migrate three already-header-only c10/util headers into torch/headeronly/util/, leaving the old c10 paths as one-line forwarders (fully backward compatible, no consumer include changes): string_u..."
186984,Move header-only c10/util container/storage utils to torch/headeronly,https://github.com/pytorch/pytorch/pull/186984,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"186980, 186981, 186982, 186983, 186985, 186986, 186987, 186988",,,0,4,0,,,ciflow/inductor,2026-06-10T22:13:05Z,2026-06-16T06:06:56Z,22,"Stack from ghstack (oldest at bottom): #186988 #186987 #186986 #186985 -> #186984 #186983 #186982 #186981 #186980 #186962 Migrate four already-header-only c10/util headers into torch/headeronly/util/, leaving the old c10 paths as one-line forwarders (fully backward compatible, no consumer include changes): Array.h,..."
187326,Fix torch.special.zeta returning NaN for q=+inf,https://github.com/pytorch/pytorch/pull/187326,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,187294,187294,187294,,,,0,2,0,,,"triaged, open source, ciflow/trunk, release notes: python_frontend, topic: bug fixes",2026-06-15T09:42:00Z,2026-06-16T06:06:42Z,22,"zeta(s, +inf) returned NaN instead of 0 for s > 1. The tail expansion computes b * w where b = 0 and w = inf, producing NaN. Added isinf(q) early return in both CPU (Math.h) and CUDA jiterator (Math.cuh) paths. Added test_zeta_with_inf_q covering all-inf and mixed finite/inf inputs. Fixes #187294"
186982,Move header-only c10/util concurrency/lifetime utils to torch/headeronly,https://github.com/pytorch/pytorch/pull/186982,draft/noise,13,PR is draft; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"186980, 186981, 186983, 186984, 186985, 186986, 186987, 186988",,,0,4,0,,,ciflow/inductor,2026-06-10T22:12:54Z,2026-06-16T06:06:41Z,22,"Stack from ghstack (oldest at bottom): #186988 #186987 #186986 #186985 #186984 #186983 -> #186982 #186981 #186980 #186962 Migrate four already-header-only c10/util headers into torch/headeronly/util/, leaving the old c10 paths as one-line forwarders (fully backward compatible, no consumer include changes): ScopeExit..."
186983,Move header-only c10/util math utils to torch/headeronly,https://github.com/pytorch/pytorch/pull/186983,draft/noise,13,PR is draft; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"186980, 186981, 186982, 186984, 186985, 186986, 186987, 186988",,,0,4,0,,,ciflow/inductor,2026-06-10T22:13:00Z,2026-06-16T06:06:39Z,22,"Stack from ghstack (oldest at bottom): #186988 #186987 #186986 #186985 #186984 -> #186983 #186982 #186981 #186980 #186962 Migrate two already-header-only c10/util headers into torch/headeronly/util/, leaving the old c10 paths as one-line forwarders (fully backward compatible, no consumer include changes): copysign.h..."
186981,Move header-only c10/util metaprogramming/type utils to torch/headeronly,https://github.com/pytorch/pytorch/pull/186981,draft/noise,13,PR is draft; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"186980, 186982, 186983, 186984, 186985, 186986, 186987, 186988",,,0,4,0,,,ciflow/inductor,2026-06-10T22:12:49Z,2026-06-16T06:06:39Z,22,"Stack from ghstack (oldest at bottom): #186988 #186987 #186986 #186985 #186984 #186983 #186982 -> #186981 #186980 #186962 Migrate six already-header-only c10/util headers into torch/headeronly/util/, leaving the old c10 paths as one-line forwarders (fully backward compatible, no consumer include changes): overloaded..."
186985,Move header-only c10/util low-level utils to torch/headeronly,https://github.com/pytorch/pytorch/pull/186985,draft/noise,13,PR is draft; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"186980, 186981, 186982, 186983, 186984, 186986, 186987, 186988",,,0,4,0,,,ciflow/inductor,2026-06-10T22:13:11Z,2026-06-16T06:06:31Z,22,"Stack from ghstack (oldest at bottom): #186988 #186987 #186986 -> #186985 #186984 #186983 #186982 #186981 #186980 #186962 Migrate five already-header-only c10/util headers into torch/headeronly/util/, leaving the old c10 paths as one-line forwarders (fully backward compatible, no consumer include changes): irange.h,..."
183366,[FSDP] Support replicated shard placement params,https://github.com/pytorch/pytorch/pull/183366,draft/noise,23,"PR is draft; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,183207,183207,,,,0,1,0,,,"ciflow/inductor, release notes: distributed (fsdp2), ciflow/torchtitan",2026-05-11T21:03:40Z,2026-06-16T05:40:51Z,22,"Stack from ghstack (oldest at bottom): -> #183366 Allow fully_shard shard_placement_fn to return Replicate() for 1D meshes, mapping those parameters to a DDP-style all-reduce param group while keeping Shard placements on the FSDP reduce-scatter path. Add coverage for both direct Replicate() and explicit ShardPlaceme..."
180423,Fix checkpoint DeviceContext leak onto TorchFunction mode stack,https://github.com/pytorch/pytorch/pull/180423,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"Stale, ciflow/torchtitan",2026-04-15T03:54:17Z,2026-06-16T03:48:36Z,22,"Stack from ghstack (oldest at bottom): -> #180423 #180422 #180424 Checkpoint's recompute_fn was capturing the live DeviceContext object from the forward pass's mode stack and re-entering it during recomputation. This caused prev_mode linkage between DeviceContext objects, so exit would push a zombie DeviceContext ba..."
179698,[torch][inductor] Fix numel variable scoping with enable_kernel_profile (#179698),https://github.com/pytorch/pytorch/pull/179698,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"fb-exported, Stale, module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-04-08T08:58:56Z,2026-06-16T03:48:36Z,22,"Summary: When TORCHINDUCTOR_CPP_ENABLE_KERNEL_PROFILE=1 is set, the AOTInductor C++ wrapper codegen wraps regular kernel calls in {} scope blocks containing a KernelContextGuard for profiling. The previous fix (D94577289) addressed an ""undeclared identifier"" error caused by numel variables declared inside these {} b..."
177087,[torch][inductor] Get shape information for extern/Aten kernels with enable_kernel_profile (#177087),https://github.com/pytorch/pytorch/pull/177087,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-03-10T22:20:25Z,2026-06-16T03:48:30Z,22,"Summary: Problem When profiling AOT Inductor compiled models with TORCHINDUCTOR_CPP_ENABLE_KERNEL_PROFILE=1, the generated GPU trace showed input tensor shapes for Triton kernels but not for extern/ATen kernels (e.g., aten::mm, aten::addmm). Root Cause The issue was in cpp_wrapper_cpu.py, in the generate_c_shim_exte..."
183579,[OpenReg][Quantization] Register dequantize_tensor_per_tensor_affine stub for PrivateUse1,https://github.com/pytorch/pytorch/pull/183579,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, topic: not user facing",2026-05-13T17:14:02Z,2026-06-16T02:38:34Z,23,
179699,[torch][inductor] Add KernelContextGuard profiling to all kernel codegen paths (#179699),https://github.com/pytorch/pytorch/pull/179699,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"fb-exported, Stale, module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-04-08T08:59:15Z,2026-06-16T00:30:54Z,23,"Summary: When TORCHINDUCTOR_CPP_ENABLE_KERNEL_PROFILE=1 is set, AOTInductor wraps kernel calls in { KernelContextGuard ...; } scope blocks so the PyTorch profiler can link CUDA runtime events (e.g. cudaLaunchKernel) to their originating CPU operations via the ""External id"" field in the trace. Previously, only two co..."
179587,[pytorch] Fix resource loading for PAR compatibility using importlib.resources (#179587),https://github.com/pytorch/pytorch/pull/179587,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-07T15:31:15Z,2026-06-16T00:05:33Z,23,"Summary: Several functions use Path(__file__).parent, os.listdir(os.path.dirname(__file__)), or get_file_path_2() which fail when running from PAR files because __file__ resolves to /proc/self/fd/N/... - a file descriptor path that doesn't support filesystem traversal. This change uses importlib.resources.files() wh..."
185741,[PyTorch][Profiler] Fix use-after-free in global RecordFunction callback teardown (#185741),https://github.com/pytorch/pytorch/pull/185741,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, meta-exported",2026-05-31T03:42:52Z,2026-06-15T23:19:55Z,23,"Summary: A MAST training job crashed with a Python segfault on rank 114: https://www.internalfb.com/mlhub/pipelines/runs/mast/v4_midtrain_rec_v1_cpt_TIER2_2p3B-n5gwz7fk?job_attempt=0&version=3&tab=execution_details. The faulthandler stack showed the DataLoader pin_memory thread dispatching aten::pin_memory, and the..."
186102,[inductor] Fast paths in sizevars for concrete integer comparisons,https://github.com/pytorch/pytorch/pull/186102,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-03T16:41:25Z,2026-06-15T22:52:58Z,23,"Stack from ghstack (oldest at bottom): (to be filled) statically_known_equals and check_leq create sympy.Eq/sympy.Le objects on every call. For concrete integers, use direct Python comparisons. Profiled on 20K-node graph (10K params, H100), on top of previous fixes: Total before: 91.9s Total after: 87.8s (1.05x) Tes..."
186101,[inductor] Avoid O(n) set union in decide_inplace_update,https://github.com/pytorch/pytorch/pull/186101,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186095, 186096, 186097, 186102, 186106",,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-03T16:41:21Z,2026-06-15T22:52:28Z,23,"Stack from ghstack (oldest at bottom): #186102 #186106 -> #186101 #186097 #186096 #186095 decide_inplace_update creates a union of three OrderedSets on every call during codegen. Since OrderedSet union uses dict.fromkeys which is O(n), the total cost is O(n^2) for large graphs. Replace the set union with three separ..."
186097,[inductor] Add early exits to can_fusion_increase_peak_memory,https://github.com/pytorch/pytorch/pull/186097,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186095, 186096, 186101, 186102, 186106",,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-03T16:03:10Z,2026-06-15T22:51:36Z,23,"Stack from ghstack (oldest at bottom): #186102 #186106 #186101 -> #186097 #186096 #186095 Return early when either node has no single-user inputs or when the reuse-key intersection is empty -- the memory overhead is guaranteed zero in both cases, so fusion is always allowed. This is the common case for vertical (pro..."
186096,[inductor] Skip combined tiling computation for <=2D pointwise fusion,https://github.com/pytorch/pytorch/pull/186096,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186095, 186097, 186101, 186102, 186106",,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-03T16:03:06Z,2026-06-15T22:51:32Z,23,Stack from ghstack (oldest at bottom): #186102 #186106 #186101 #186097 -> #186096 #186095 The tiling compatibility check (tiling_prevents_pointwise_fusion) only rejects fusion when at least one node has >2D tiling. For <=2D nodes the check unconditionally passes and tiling3 is never consulted. Skip the expensive com...
183270,[inductor][overlap] Post-scheduling foreach_mm for optimizer mm batching,https://github.com/pytorch/pytorch/pull/183270,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182631, 185339",,,0,2,0,,,"ciflow/trunk, release notes: foreach_frontend, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/torchtitan",2026-05-11T12:30:36Z,2026-06-15T22:49:47Z,23,"Stack from ghstack (oldest at bottom): #185530 #185339 -> #183270 #182631 [inductor][overlap] Post-scheduling foreach_mm for optimizer mm batching Summary Inductor pass that batches independent same-shape mm ops into aten._foreach_mm after overlap scheduling, preserving compute/comm overlap decisions while enabling..."
182631,[aten] Add aten._foreach_mm operator,https://github.com/pytorch/pytorch/pull/182631,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"183270, 185339",,,0,11,0,,,"ciflow/trunk, release notes: foreach_frontend, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-05-06T09:16:09Z,2026-06-15T22:47:50Z,23,"Stack from ghstack (oldest at bottom): #185339 -> #182631 #185530 #183270 [aten] Add aten._foreach_mm operator (CUTLASS grouped GEMM) Summary Batched matrix multiply for lists of same-shape tensors. One CUTLASS grouped GEMM launch replaces N independent cuBLAS calls, cutting kernel launch overhead and unlocking casc..."
180597,[nocommit] symmem custreamwaitvalue32,https://github.com/pytorch/pytorch/pull/180597,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,180407,,,0,1,0,,,"Stale, release notes: distributed (c10d), ciflow/h100-symm-mem",2026-04-16T19:21:14Z,2026-06-15T22:18:01Z,23,Stack from ghstack (oldest at bottom): -> #180597 #180407
180407,[nocommit] symmem nccl ib,https://github.com/pytorch/pytorch/pull/180407,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,180597,,,0,3,0,,,"Stale, release notes: distributed (c10d), ciflow/h100-symm-mem",2026-04-15T00:21:12Z,2026-06-15T22:18:00Z,23,Stack from ghstack (oldest at bottom): #180597 -> #180407
187106,[inductor] decompose broadcast-bias baddbmm into bmm + pointwise add,https://github.com/pytorch/pytorch/pull/187106,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,187093,187093,187093,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-11T23:37:27Z,2026-06-15T22:15:26Z,23,"Issue Fixes #187093 Summary For baddbmm with a bias broadcast along dim 1 (M) of the [B, M, N] output, register a conditional Inductor decomposition to bmm(batch1, batch2) + bias, following the existing conditional addmm/mm/bmm decomps in torch/_inductor/decomposition.py (returns NotImplemented when the condition do..."
180129,[scaled_grouped_mm] add support for out argument,https://github.com/pytorch/pytorch/pull/180129,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, Stale, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-04-11T15:54:49Z,2026-06-15T22:07:32Z,23,"Summary Add support for optional out argument in torch.scaled_grouped_mm, that the caller can use to pass a preallocated output buffer to. Tests Added test cases for 2d/2d and 2d/3d in test/test_scaled_matmul_cuda.py"
179528,[Codemod][PSS3] Upgrade fbcode/mobile-vision to Python Scientific Stack 3 (#179528),https://github.com/pytorch/pytorch/pull/179528,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,179528,,,,,0,1,0,,,"fb-exported, Stale, ciflow/trunk, topic: not user facing, meta-exported",2026-04-06T23:58:49Z,2026-06-15T18:54:22Z,23,Summary: Pull Request resolved: #179528 Differential Revision: D90838435
173894,[nn] Fix softplus numerical instability with large beta values,https://github.com/pytorch/pytorch/pull/173894,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,171249,171249,171249,,,,0,7,0,,,"module: nn, module: cpu, triaged, module: NaNs and Infs, open source, Stale, release notes: nn, topic: not user facing, ciflow/mps",2026-01-30T05:55:25Z,2026-06-15T18:54:17Z,23,"Summary Fixes #171249 When torch.nn.functional.softplus is called with very large beta values (e.g., 1e+30), the current implementation can produce inf values for positive inputs because exp(beta * x) overflows. Before fix: >>> import torch.nn.functional as F >>> input = torch.tensor([0.5, -1.0, 2.0]) >>> F.softplus..."
185977,Fix index_add out-of-bounds parity between eager and inductor,https://github.com/pytorch/pytorch/pull/185977,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/mps, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185885,185885,185885,,,,0,22,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, ciflow/mps, module: inductor, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-06-02T18:58:54Z,2026-06-15T18:47:12Z,23,"Fixes #185885 Eager index_add bounds-checks the index, but the inductor decomp in _index_add routes through index_put, which follows advanced-indexing semantics and silently accepts negative/out-of-range indices. Compiled index_add therefore returned a tensor where eager raises. This adds a side-effectful aten._asse..."
179941,[inductor] Add reduction loop peeling for Triton codegen,https://github.com/pytorch/pytorch/pull/179941,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,122840,122840,"122840, 148402",,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-04-10T17:46:50Z,2026-06-15T16:51:54Z,23,"Add loop peeling optimization for non-persistent reductions that splits the reduction loop into an unmasked main loop (vectorizable) and a masked tail loop, gated behind config.triton.loop_peeling (default off). During the single masked codegen pass, load() and reduction() register masked_line→unmasked_line mappings..."
187192,Protect GraphPassRegistry map with shared_mutex against concurrent access (#187192),https://github.com/pytorch/pytorch/pull/187192,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"topic: not user facing, meta-exported",2026-06-12T17:11:22Z,2026-06-15T16:49:09Z,23,Summary: Making GraphPassRegistry thread-safe Test Plan: CI Differential Revision: D108432093
184654,Short-circuit pytree.tree_map_only when no leaves match,https://github.com/pytorch/pytorch/pull/184654,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-21T04:12:54Z,2026-06-15T16:34:26Z,23,"tree_map_only always went through tree_map → flatten + per-leaf map_only(...) closure + unflatten, invoking the pred + func Python lambda once per leaf even when nothing matched. Fold the dispatch into a single pass over the flattened leaves: No match: skip the per-leaf calls; still treespec.unflatten(leaves) so the..."
187002,Use C++20 <bit> instead of c10::llvm bit helpers in c10,https://github.com/pytorch/pytorch/pull/187002,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/vllm,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: cpu, triaged, open source, release notes: quantization, topic: not user facing, ciflow/vllm",2026-06-11T01:39:08Z,2026-06-15T15:31:05Z,23,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
175476,[inductor] Hoist output buffer allocations into prior CUDAGraph partition,https://github.com/pytorch/pytorch/pull/175476,ready for maintainer decision,45,"CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,175450,,,0,17,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-02-21T06:26:08Z,2026-06-15T14:56:19Z,23,"Stack from ghstack (oldest at bottom): #175797 -> #175476 #175450 #175449 #174856 When a non-cudagraph partition (fallback region) contains ExternKernelOut ops (e.g., DeviceCopy for cpu<->cuda transfers), their output buffer allocations are hoisted into the prior cudagraph partition so they are captured once during..."
186642,[inductor][rocm] make AMD MM matrix_instr_nonkdim configurable,https://github.com/pytorch/pytorch/pull/186642,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-08T20:03:47Z,2026-06-15T13:09:55Z,23,"Motivation The AMD MM Triton autotune sweep hard-codes matrix_instr_nonkdim to [0, 16], so shapes where the mfma_32x32 path wins cannot be reached without patching inductor. What changed Add torch._inductor.config.rocm.mfma_nonkdim, read from the env var TORCHINDUCTOR_MFMA_NONKDIM, and drive the AMD MM autotune swee..."
169937,Fix vectorized float32 to uint8 conversion on AArch64,https://github.com/pytorch/pytorch/pull/169937,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/linux-aarch64, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,168972,168972,,,,0,18,0,,,"module: cpu, open source, module: arm, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/linux-aarch64, arm priority",2025-12-09T10:30:07Z,2026-06-15T12:50:01Z,23,"Stack from ghstack (oldest at bottom): -> #169937 Fixes #168972 This fixes numerical inconsistency problems in converting float32 to uint8 ( seen non AArch64 ). In particular converting negative floats results in 0 due to static_cast giving UB. This has been reported ( #156788 ) and fixed ( #157267 ) before, but onl..."
176746,Added 'gelsd' to CUDA torch.linalg.lstsq driver so that underdetermined least square systems can be solved on CUDA,https://github.com/pytorch/pytorch/pull/176746,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,117122,117122,117122,,,,0,7,0,,,"triaged, open source, release notes: linalg_frontend",2026-03-06T20:46:45Z,2026-06-15T12:42:30Z,23,"Fixes #117122 Updated torch.linalg.lstsq to be able to use driver='gelsd' and solve the system of equation using the SVD of the matrix $A$ for $||Ax-b||_2^2$. This allows a user to be able to solve any system, not just the systems where $A$ has full rank. In the case where there are multiple solutions, the 'gelsd' s..."
177866,Sparse semi structured setup/teardown latency,https://github.com/pytorch/pytorch/pull/177866,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,153825,153825,153825,,,,0,13,0,,,"triaged, open source, release notes: sparse",2026-03-19T16:09:23Z,2026-06-15T12:41:56Z,23,"Fixes #153825 Updated the sparse semi-structures mm implementation (_cslt_sparse_mm_impl() in pytorch/aten/src/ATen/native/sparse/cuda/cuSPARSELtOps.cpp) to cache matmul descriptors for matrices with specific parameters (size, dtyp, etc.). This helps reduce the runtime of subsequent matmul of the same kind of matrix..."
178797,Fix meta conv shape checks for invalid groups vs out_channels,https://github.com/pytorch/pytorch/pull/178797,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,177237,177237,177237,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-03-30T18:47:51Z,2026-06-15T12:33:52Z,23,"Grouped convolution requires both input and output channel counts to be compatible with groups. This change tightens the meta implementation so invalid combinations fail with the same ""Invalid channel dimensions"" error as on real devices, updates the conv1d / conv3d docs to state that both in_channels and out_channe..."
173182,Updated foreach_norm to accept dim and keepdim arguments.,https://github.com/pytorch/pytorch/pull/173182,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,133367,,,,0,17,0,,,"triaged, open source, release notes: foreach_frontend",2026-01-23T16:56:50Z,2026-06-15T12:32:54Z,23,"Related: #133367 Updated foreach_norm to accept dim and keepdim arguments. This implementation: row_norm = torch._foreach_norm(tensor_list, dim=-1, keepdim=True) and col_norm = torch._foreach_norm(tensor_list, dim=-2, keepdim=True) produce the same output as row_norm = [torch.linalg.vector_norm(t, dim=-1, keepdim=tr..."
187168,Fix accuracy degradation for baddbmm on CPU for float tensors,https://github.com/pytorch/pytorch/pull/187168,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, release notes: linalg_frontend",2026-06-12T11:14:22Z,2026-06-15T11:37:59Z,23,Use acc_type instead of opmath_type to avoid loosing precision leading to different results depending on tensor strides and compiler optimizations. The basic loop over ks gets conditionally vectorized by -O3 depending on the stride (Option -fversion-loops-for-strides). This leads to slightly different values for non...
171219,[DTensor] Support Partial to _StridedShard,https://github.com/pytorch/pytorch/pull/171219,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"171023, 171178, 171338",,,0,23,0,,,"Stale, topic: not user facing, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2025-12-24T00:05:30Z,2026-06-15T10:36:36Z,23,"As title. After refactoring to the padding/unpadding logic in the previous PR, it's much easier to support reduce-scatter for _StridedShard now. Stack from ghstack (oldest at bottom): #171338 -> #171219 #171178 #171023"
171338,[DTensor] Fully support uneven _StridedShard redistribution with device order,https://github.com/pytorch/pytorch/pull/171338,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"171023, 171178, 171219",,,0,23,0,,,"Stale, topic: not user facing, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2025-12-26T20:00:57Z,2026-06-15T10:36:34Z,23,Stack from ghstack (oldest at bottom): -> #171338 #171219 #171178 #171023
171023,[DTensor] Support Shard to _StridedShard through all-to-all,https://github.com/pytorch/pytorch/pull/171023,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"171178, 171219, 171338",,,0,24,0,,,"Stale, topic: not user facing, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2025-12-22T02:15:54Z,2026-06-15T10:36:33Z,23,Workflow: Stack from ghstack (oldest at bottom): #171338 #171219 #171178 -> #171023
171178,[DTensor] Refactor padding/unpadding with CollectivePaddingContext,https://github.com/pytorch/pytorch/pull/171178,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"171023, 171219, 171338",,,0,27,0,,,"Stale, topic: not user facing, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2025-12-23T08:59:01Z,2026-06-15T10:36:28Z,23,"This PR maintains the original padding/unpadding logic and introduces CollectivePaddingContext, a fluent API for managing padding and unpadding around collective operations in DTensor redistribution. Motivation The padding/unpadding logic for collective operations (all_gather, reduce_scatter, alltoall) was scattered..."
186762,Fix typo in quantization documentation,https://github.com/pytorch/pytorch/pull/186762,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-09T03:56:04Z,2026-06-15T10:35:42Z,23,Approved on branch. New pull request as this is only visible on release/2.12.
175964,Add validation for embedding_bag offsets to prevent illegal memory access,https://github.com/pytorch/pytorch/pull/175964,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,175368,175368,175368,,,,0,3,0,,,"triaged, open source, Stale",2026-02-27T09:52:50Z,2026-06-15T10:35:27Z,23,"This commit fixes a bug where embedding_bag would cause illegal memory access (segfaults on CPU, CUDA errors on GPU) when given invalid offsets that point beyond the indices length, instead of raising a proper error. The fix enhances the check_arguments() function to validate that all offsets are: Within bounds [0,..."
179160,Fix uint32 overflow in CUDA scan _with_indices kernels,https://github.com/pytorch/pytorch/pull/179160,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,167086,167086,167086,,,,0,3,0,,,"triaged, open source",2026-04-02T18:10:28Z,2026-06-15T10:35:26Z,23,"Fixes #167086 #167326 added 64-bit index support to scan_outer_dim (cumsum/cumprod). This PR does the same for the _with_indices variants (cummax/cummin), which still overflow when num_orows * row_size * num_irows > UINT_MAX. Both kernels (tensor_kernel_scan_outer_dim_with_indices, tensor_kernel_scan_innermost_dim_w..."
182034,Add no-tree-loop-optimize pragma for apply_grad_input on GCC/AArch64,https://github.com/pytorch/pytorch/pull/182034,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,177250,177250,,,,0,4,0,,,"open source, ciflow/trunk, release notes: nn",2026-04-30T15:19:56Z,2026-06-15T08:47:20Z,23,"Summary Fixes an AArch64 GCC optimization issue in CPU upsample backward for reduced-precision gradients. The reduced-precision apply_grad_input path accumulates a temporary float buffer into BFloat16/Half gradients. On AArch64 GCC builds, tree vectorization of this function can produce incorrect results or illegal..."
185157,[Inductor] Fix CPU bool argmin/argmax incorrect indices,https://github.com/pytorch/pytorch/pull/185157,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,184893,184893,,,,,0,3,0,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-25T19:05:02Z,2026-06-15T08:42:07Z,23,"Fixes #184893 Summary: When using torch.compile with the Inductor backend on CPU, max(dim)[1] (argmax) and min(dim)[1] (argmin) could return incorrect indices for boolean tensors. Root cause: CPU C++ codegen represents vectorized boolean values as VecMask<float, N>. In that representation, true lanes are all-bits-se..."
179127,Skip FBGEMM on Arm64,https://github.com/pytorch/pytorch/pull/179127,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-04-02T06:46:33Z,2026-06-15T04:36:48Z,23,"as per Quantized operations require FBGEMM. FBGEMM is only optimized for CPUs with instruction set support AVX2 or newer. pytorch/torch/testing/_internal/common_quantization.py Line 351 in e899f15 reason = ""Quantized operations require FBGEMM. FBGEMM is only optimized for CPUs with instruction set support AVX2 or ne..."
176393,Make avgpool3d support fp16/bf16,https://github.com/pytorch/pytorch/pull/176393,draft/noise,13,PR is draft; CI lane labels: ciflow/trunk,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,10,0,,,"module: cpu, open source, ciflow/trunk, topic: not user facing",2026-03-04T06:24:11Z,2026-06-15T03:33:18Z,24,Make avg_pool3d_out_cpu use avg_pool3d_kernel. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
177147,[FSDP2] Remove redundant event wait for all_gather_stream,https://github.com/pytorch/pytorch/pull/177147,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, release notes: distributed (fsdp)",2026-03-11T17:20:42Z,2026-06-15T03:33:33Z,24,"For implicit prefetch, making the all-gather stream waiting for the copy-out event is redundant because the all-gather stream already waits for the copy-in stream and copy-in stream waits for the copy-out event. From the memory perspective, the peak memory bound of parameter storage is 3x with or without this wait...."
170538,[XPU] Migrate test_testing for xpu,https://github.com/pytorch/pytorch/pull/170538,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,114850,114850,,,,,0,8,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, keep-going, ciflow/xpu",2025-12-16T09:26:16Z,2026-06-15T02:42:17Z,24,"Description Fixes #114850, we will port dynamo, fsdp tests to Intel GPU We could enable Intel GPU with following methods and try the best to keep the original code styles: Changes Get device type with from accelerator and get device type helper method Add dtypesIfXPU for dtypes test case. Use common utils replace th..."
180411,stateless dtensor,https://github.com/pytorch/pytorch/pull/180411,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180409, 180410",,,0,2,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-04-15T01:08:18Z,2026-06-15T01:36:39Z,24,Stack from ghstack (oldest at bottom): -> #180411 #180410 #180409
180409,stateless ghstack,https://github.com/pytorch/pytorch/pull/180409,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180410, 180411",,,0,1,0,,,"module: cpu, Stale",2026-04-15T01:08:12Z,2026-06-14T21:56:10Z,24,Changes based on #178160 Stack from ghstack (oldest at bottom): #180411 #180410 -> #180409 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
180410,debug,https://github.com/pytorch/pytorch/pull/180410,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180409, 180411",,,0,1,0,,,Stale,2026-04-15T01:08:15Z,2026-06-14T21:56:09Z,24,Fix the bugs in #180409 Stack from ghstack (oldest at bottom): #180411 -> #180410 #180409
184986,Handle unbatched input for non-zeros padding modes in Conv1d/2d/3d,https://github.com/pytorch/pytorch/pull/184986,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,104860,104860,104860,,,,0,1,0,,,"triaged, open source, release notes: nn, topic: not user facing",2026-05-23T12:44:19Z,2026-06-27T08:37:46Z,11,"Fixes #104860 When padding_mode is not ""zeros"", Conv1d, Conv2d, and Conv3d call F.pad on the input before performing the convolution. However, F.pad (particularly with circular mode) requires a batch dimension. This means unbatched inputs (e.g. a 3D tensor for Conv2d) fail with a cryptic error, even though the same..."
176936,Add driver api guard to cuda utils,https://github.com/pytorch/pytorch/pull/176936,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, Stale, release notes: distributed (c10d)",2026-03-09T21:43:56Z,2026-06-14T20:03:45Z,24,This change fixes the build when PYTORCH_C10_DRIVER_API_SUPPORTED is not defined. The change was implemented with #151125 but it was mark as stalled.
176321,test: replaces deprecated numpy.linalg.linalg module,https://github.com/pytorch/pytorch/pull/176321,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source, Stale, topic: not user facing",2026-03-03T18:31:44Z,2026-06-14T20:03:43Z,24,"Replaces numpy.linalg.linalg with numpy.linalg._linalg because the previous module was deprecated in numpy 2.0 but it was recently removed (commit b72ff69). Which makes test/torch_np/numpy_tests/linalg/test_linalg.py to fail with: File ""/third_party/py/torch/test/torch_np/numpy_tests/linalg/test_linalg.py"", line 15,..."
174605,make Store timeout_ atomic for thread safety.,https://github.com/pytorch/pytorch/pull/174605,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, Stale, release notes: distributed (c10d)",2026-02-09T18:57:27Z,2026-06-14T20:03:40Z,24,"There was a data race that ocurred when getTimeout() read timeout_ and setTimeout() wrote to timeout_ without synchronization, and multiple threads could call setTimeout or construct StoreTimeoutGuard on the same Store object concurrently. The timeout_ member in c10d::Store is now an std::atomic<std::chrono::millise..."
179642,Cache cuSPARSELt matmul plan/descriptors in _cslt_sparse_mm (#179642),https://github.com/pytorch/pytorch/pull/179642,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, ciflow/trunk, release notes: sparse, meta-exported",2026-04-07T23:05:35Z,2026-06-14T18:55:58Z,24,"Summary: Previously, every call to _cslt_sparse_mm recreated all cuSPARSELt descriptors, matmul plan, and workspace from scratch — even when the matrix shapes and parameters were identical. This caused significant CPU-side overhead (~50-200μs per call for plan creation). This diff adds a thread-local plan cache keye..."
177438,Infer device_type in fork_rng from devices argument,https://github.com/pytorch/pytorch/pull/177438,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,Stale,2026-03-13T23:35:25Z,2026-06-14T18:55:52Z,24,"Stack from ghstack (oldest at bottom): -> #177438 fork_rng defaulted device_type to ""cuda"", so callers like test_randperm that pass devices=[""openreg:0""] without an explicit device_type would incorrectly try to initialize CUDA. When devices is provided but device_type is not, infer it from the first device in the li..."
178014,add missing to() operator which is called in benchmark (#178014),https://github.com/pytorch/pytorch/pull/178014,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Merged, Reverted, Stale, ciflow/trunk, release notes: sparse, ci-no-td, meta-exported",2026-03-20T22:00:14Z,2026-06-14T17:59:57Z,24,"Summary: While running MTS benchmark, I notice a .cpu() operator is missing, this will be a blocking error in some workflows. Note: only the fallback logic is implemented here. Optimization for cuda -> cuda is not in the scope of this change. Test Plan: OSS CI: https://hud.pytorch.org/pr/178014 Reviewed By: jcaip Di..."
180460,Decompose all_gather+matmul into P2P+tiled mm for fine-grained overlap,https://github.com/pytorch/pytorch/pull/180460,draft/noise,13,"PR is draft; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, Stale, release notes: distributed (pipeline), module: inductor, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-04-15T13:32:25Z,2026-06-14T16:59:39Z,24,cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
180183,autograd: handle kMkldnn and kJagged layouts in record_stream_any_impl,https://github.com/pytorch/pytorch/pull/180183,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,60306,,,,0,1,0,,,"triaged, open source, Stale",2026-04-12T18:54:36Z,2026-06-14T16:59:38Z,24,"Previously, tensors with kMkldnn or kJagged layouts would hit the default TORCH_INTERNAL_ASSERT(false) in record_stream_any_impl, causing a hard crash if such tensors ever reached this code path. kMkldnn tensors do not support stream recording via storage() and are not used in CUDA stream contexts, so we skip silent..."
179627,Fix all masked Transformer eager/compile inconsistency,https://github.com/pytorch/pytorch/pull/179627,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,178677,178677,178677,,,,0,1,0,,,"triaged, open source, Stale, release notes: nn",2026-04-07T20:26:47Z,2026-06-14T16:59:38Z,24,Fixes #178677 eager/compiled inconsistency when src_key_padding_mask is all True. Compiled mode skips nested tensor optimization (which breaks in torch compile see the 481-482 comment) so it bypasses the error check that eager hits. Adds upfront validation so both modes reject invalid input
180167,Clarify that spectral_norm does not compute the true operator norm for conv layers,https://github.com/pytorch/pytorch/pull/180167,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,"13505, 99149",,,,0,1,0,,,"triaged, open source, Stale",2026-04-12T05:10:05Z,2026-06-14T16:59:36Z,24,"Summary Add a warning to the spectral_norm docstrings in both torch.nn.utils.parametrizations and torch.nn.utils.spectral_norm clarifying that for convolutional layers, the spectral norm of the reshaped weight matrix is not the operator norm (true Lipschitz constant) of the convolutional layer Fix misleading ""reduci..."
179594,[CUDA] Use #pragma unroll in Normaization.cuh,https://github.com/pytorch/pytorch/pull/179594,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, Stale, release notes: cuda",2026-04-07T16:08:12Z,2026-06-14T16:59:33Z,24,"I think this was just forgotten because I don't see a reason not to unroll these, but not tested Contributed by Benedikt Johannes"
179399,Improve GELU documentation formula,https://github.com/pytorch/pytorch/pull/179399,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-05T06:48:07Z,2026-06-14T16:59:33Z,24,"What does this PR do? Adds the explicit error function (erf) formulation for GELU when approximate='none' in the nn.GELU documentation. Previously, the documentation only described GELU as: GELU(x) = x * Φ(x) where Φ(x) is the Gaussian cumulative distribution function. This PR keeps that definition and additionally..."
179584,fix inference_mode getitem compile error,https://github.com/pytorch/pytorch/pull/179584,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,169477,169477,169477,,,,0,4,0,,,"triaged, open source, Stale, release notes: fx, module: dynamo",2026-04-07T14:42:37Z,2026-06-14T16:59:32Z,24,"Fixes #169477 In Interpreter.call_function, when the target is operator.getitem and inference mode is enabled on this thread, run the getitem under torch.inference_mode(False) so FX replay does not produce inference views. A regression test is added in test_aot_autograd.py. AOT runs the captured graph through the FX..."
179404,[DTensor] Replace __module__ hacks with @exposed_in decorators,https://github.com/pytorch/pytorch/pull/179404,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,171905,171905,171905,,,,0,1,0,,,"triaged, open source, Stale, release notes: distributed (dtensor)",2026-04-05T12:24:13Z,2026-06-14T16:59:31Z,24,"Summary Replace manual __module__ attribute assignments in torch.distributed.tensor with @exposed_in() decorators applied at definition sites Covers DTensor, distribute_tensor, distribute_module, factory functions (ones, empty, full, rand, randn, zeros), CommDebugMode, visualize_sharding, context_parallel, local_map..."
179379,[DTensor] Convert type aliases to TypeAliasType in experimental/,https://github.com/pytorch/pytorch/pull/179379,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,171905,171905,171905,,,,0,3,0,,,"triaged, open source, Stale, release notes: distributed (dtensor)",2026-04-04T12:24:48Z,2026-06-14T16:59:28Z,24,"Summary Convert bare type alias assignments (PlacementType, InputPlacements, OutputPlacements) to typing_extensions.TypeAliasType for proper type checker support Motivation Fixes #171905 Bare alias assignments like PlacementType = Sequence[Placement] | None aren't recognized as type aliases by type checkers. TypeAli..."
179000,Add type conversion test for IValue: implements TODO(gmagogsfm),https://github.com/pytorch/pytorch/pull/179000,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, Stale, release notes: vulkan, topic: not user facing",2026-04-01T14:00:41Z,2026-06-14T16:59:26Z,24,
187069,[torch][autograd] Fix Python refcount leaks in autograd C++ glue,https://github.com/pytorch/pytorch/pull/187069,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"ciflow/trunk, topic: not user facing, meta-exported",2026-06-11T18:39:57Z,2026-06-14T16:30:46Z,24,"Summary: Autograd's Python bindings were leaking references in several places, causing memory issues. This diff fixes these leaks by properly adopting owned references in the autograd C++ glue code, ensuring they are released correctly. Test Plan: arc lint + buck build of affected targets (see stack). Differential R..."
187101,Add missing dispatch for bf16 in tensor_cpu,https://github.com/pytorch/pytorch/pull/187101,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187089,187089,187089,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-11T23:06:00Z,2026-06-14T16:22:04Z,24,"The ctc_loss implementation currently has a bug when backend is PrivateUse1 and dtype is bf16. Specifically, when get_clamped_target_length is called for the bf16 dtype, it crashes because tensor_cpu is missing dispatch support for that type. This change adds support for that as well as the half dtype which was also..."
187068,[torch][profiler] Fix refcount leaks in _TensorMetadata layout/dtype getters,https://github.com/pytorch/pytorch/pull/187068,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, meta-exported",2026-06-11T18:38:31Z,2026-06-14T16:17:54Z,24,"Summary: wrap() returns a new owned reference (Py_NewRef), but the layout and dtype property getters wrapped it with reinterpret_borrow, which incref's again and leaks one reference per access. Use reinterpret_steal to adopt the new ref. Test Plan: arc lint + buck build of affected target (see stack). Differential R..."
179372,[Inductor] Fix view_as_complex stride requirement in Inductor backward by forcing contiguous layout,https://github.com/pytorch/pytorch/pull/179372,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,179368,,,,0,1,0,,,"triaged, open source, Stale, module: inductor",2026-04-04T07:49:24Z,2026-06-14T16:03:23Z,24,Summary This PR fixes a RuntimeError: Tensor must have a last dimension with stride 1 that occurs when using torch.compile on models involving complex number views and convolutions (Issue #179368). The Root Cause The view_as_complex operation (and the backward of view_as_real) has a strict hardware and mathematical...
180412,stateless spmd,https://github.com/pytorch/pytorch/pull/180412,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180409, 180410, 180411",,,0,1,0,,,Stale,2026-04-15T01:08:21Z,2026-06-14T06:06:23Z,24,Stack from ghstack (oldest at bottom): -> #180412 #180411 #180410 #180409
180379,[Dynamo] Inline root module forward hooks when compiling modules,https://github.com/pytorch/pytorch/pull/180379,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,117584,117584,117584,,,,0,3,0,,,"Stale, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-14T20:36:03Z,2026-06-14T03:48:23Z,24,"Fix #117584 Root cause The root nn.Module hook path still compiles in two frames by default. torch.compile(mod) only wrapped the top module call when wrap_top_frame=True or when Dynamo already knew it needed an inline wrapper, so per-module forward hooks on the root module were still traced separately from the modul..."
185958,torch/headeronly/util/TypeList.h: simplify contains and find_if with C++20 fold expressions and if constexpr,https://github.com/pytorch/pytorch/pull/185958,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, ciflow/vllm, meta-exported",2026-06-02T16:46:23Z,2026-06-14T01:33:13Z,25,"Summary: Two modernizations to the typelist utilities: contains: Replace the three-specialization recursive SFINAE pattern (primary + empty-list + two enable_if-discriminated non-empty cases) with a single partial specialization using a C++20 fold expression: (std::is_same_v<Types, Type> || ...). The result is non-r..."
185859,ATen/core/TensorBase.h: replace symint enable_if aliases with C++20 requires clauses,https://github.com/pytorch/pytorch/pull/185859,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, ciflow/vllm, meta-exported",2026-06-01T22:08:36Z,2026-06-14T01:33:09Z,25,"Summary: Remove the enable_if_symint and enable_if_int type aliases from the at::symint namespace and replace each overload pair with requires std::is_same_v<T, c10::SymInt> / requires std::is_same_v<T, int64_t> constraints. Affected function pairs: sizes, size, strides, numel. No behavior change. Test Plan: buck2 t..."
174409,[inductor] Add logging to autotune output showing cache hit,https://github.com/pytorch/pytorch/pull/174409,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-02-05T19:46:06Z,2026-06-13T21:52:52Z,25,"Example output, observe that we have AUTOTUNE mm(8192x64, 64x256) [CACHE HIT] when there is a cache hit. This helps when debugging long autotune times and identify if cache hit happens as expected. Autotune Choices Stats: {""num_choices"": 21, ""num_triton_choices"": 20, ""best_kernel"": ""triton_mm_11"", ""best_kernel_desc""..."
185875,c10/core/DispatchKeySet.h: add noexcept to non-constexpr query methods,https://github.com/pytorch/pytorch/pull/185875,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, release notes: cpp, topic: improvements, ciflow/vllm, meta-exported",2026-06-02T01:59:16Z,2026-06-13T20:33:14Z,25,"Summary: Add noexcept to the non-constexpr query methods in c10::DispatchKeySet that perform only bitwise arithmetic and integer/enum casts. The constexpr methods already can't throw (any exception would make them ill-formed), but the non-constexpr ones were missing the annotation. Affected methods: isSupersetOf, em..."
187142,Fully type-annotate torch/utils/data/graph_settings.py,https://github.com/pytorch/pytorch/pull/187142,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: dataloader",2026-06-12T03:09:41Z,2026-06-13T18:40:59Z,25,"Fixing an Issue Before submitting, please review: The Ultimate Guide to PyTorch Contributions AI-Assisted Development policy Issue Fixes # Summary Checklist Passes lint (spin fixlint) Added/updated tests Updated documentation (if applicable) Included benchmark results (for PRs impacting perf) BC-breaking?"
187249,perf(dynamo): add guard partial memo,https://github.com/pytorch/pytorch/pull/187249,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-13T15:15:24Z,2026-06-13T16:44:55Z,25,背景 客户模型里 TorchDynamo cache lookup 的开销主要集中在 guard evaluation，尤其是从 L['self'] 开始反复遍历整个 nn.Module 子树。实测里单次 lookup 大量时间花在稳定 module tree 的 accessor traversal 上，而不是 backend match、cache entry 排序，或者单个 dict/getattr 操作本身。 这类模型在 warmup 后 self module 结构通常稳定，但现有 guard 每次仍会完整递归检查 L['self'] 下的大量 module、parameter、buffer 路径，导致每 step...
177401,Allow non-CL images to go through NEON path in `F.interpolate()` for bilinear and bicubic modes when antialias=True.,https://github.com/pytorch/pytorch/pull/177401,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,17,0,,,"module: cpu, Stale, topic: not user facing",2026-03-13T17:41:34Z,2026-06-13T15:10:10Z,25,"Currently in main, the fast NEON path of F.interpolate is only used if both the input and the output are CL: pytorch/aten/src/ATen/native/cpu/UpSampleKernel.cpp Lines 1813 to 1816 in 381d938 if (input.size(1) == 3 && input.is_contiguous(at::MemoryFormat::ChannelsLast) && output.is_contiguous(at::MemoryFormat::Channe..."
184143,Bound Inductor op count for high-order diff,https://github.com/pytorch/pytorch/pull/184143,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,147075,147075,,,,0,10,0,,,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ci-no-td, ciflow/torchtitan, agentic",2026-05-18T01:52:24Z,2026-06-13T14:18:46Z,25,Stack from ghstack (oldest at bottom): -> #184143 Limit scalar op counting during Inductor realization heuristics so high-order torch.diff expression trees are realized before exponential traversal. Add a regression test covering n=30 diff compilation. Fixes #147075 Generated by my agent cc @voznesenskym @penguinwu...
180274,[WIP][Triton] update triton pin,https://github.com/pytorch/pytorch/pull/180274,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/trunk, topic: not user facing, ciflow/inductor",2026-04-13T20:20:04Z,2026-06-13T13:03:15Z,25,Test new pin
178922,Add CUDA support for Lanczos interpolation mode,https://github.com/pytorch/pytorch/pull/178922,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"Stale, release notes: nn, module: inductor, ciflow/inductor, ciflow/dtensor",2026-03-31T20:02:42Z,2026-06-13T13:03:11Z,25,#177320 added Lanczos interpolation mode for CPU images. This PR adds CUDA support. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
187056,[TESTING] Separated ROCm triton commit pin,https://github.com/pytorch/pytorch/pull/187056,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor, ciflow/inductor-rocm, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, ciflow/trunk, release notes: releng, ciflow/inductor, ciflow/inductor-rocm, ciflow/rocm-mi300, ciflow/docker",2026-06-11T15:05:19Z,2026-06-13T09:03:51Z,25,Reverted the logic that we had for separate triton commit pin for ROCm because we would like to bump it more frequently. Currently holds the same commit SHA. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
180207,[TEST] Refactor test_structured_trace.py to be device-generic,https://github.com/pytorch/pytorch/pull/180207,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,176188,,,,0,1,0,,,"open source, Stale, topic: not user facing, module: dynamo",2026-04-13T08:05:24Z,2026-06-13T07:47:52Z,25,"Refactor test/dynamo/test_structured_trace.py to be device-generic by replacing hardcoded ""cuda"" references with torch.accelerator API to enable tests to run on all accelerators including out-of-tree devices via PrivateUse1. Changes in test/dynamo/test_structured_trace.py: Remove import: requires_cuda_and_triton fro..."
187187,[typing] Replace TypeAlias with TypeAliasType in torch/types.py,https://github.com/pytorch/pytorch/pull/187187,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,171905,171905,171905,,,,0,3,0,,,"triaged, open source",2026-06-12T16:13:12Z,2026-06-13T07:15:22Z,25,"Fixes #171905 Description There is a common paradigm in PyTorch of modifying an imported module's __module__ attribute to re-expose it in a new location. However, this dynamic reassignment breaks static type checkers (like MyPy and Pyright) and IDE linters. [cite_start]This PR addresses the issue by migrating core t..."
187126,[c10] Remove deprecated CUDAAllocatorConfig forwarders,https://github.com/pytorch/pytorch/pull/187126,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-12T01:42:54Z,2026-06-13T02:18:03Z,26,
187116,Add stable tie handling to topk,https://github.com/pytorch/pytorch/pull/187116,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: cpu, release notes: quantization, release notes: mps, ciflow/mps, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-06-12T01:26:17Z,2026-06-13T02:08:04Z,26,"Stack from ghstack (oldest at bottom): -> #187116 The reverted deterministic-topk change made topk pick stable tie indices when global deterministic algorithms were enabled. That changed default user-visible tie behavior for existing deterministic-mode callers, including internal model checks that had recorded the p..."
185261,add meta registration for scaled_grouped_mm_v2,https://github.com/pytorch/pytorch/pull/185261,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-26T18:30:26Z,2026-06-13T01:01:30Z,26,Summary Add meta registrations for scaled grouped mm v2 for compile composability Tests pytest test/test_scaled_matmul_cuda.py
179658,Add OpaqueGenerator to allow make_fx trace Generator,https://github.com/pytorch/pytorch/pull/179658,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,179661,,,,,0,1,0,,,"Stale, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-08T03:04:13Z,2026-06-12T23:00:10Z,26,"Stack from ghstack (oldest at bottom): #179661 -> #179658 #179657 Following the context from #179657. Repro scripts call make_fx to retrace the module into a GraphModule, but make_fx couldn't handle torch._C.Generator arguments — FX's proxy infrastructure requires weakref support and create_arg/codegen support that..."
186190,[spmd_types] typecheck FlexAttention,https://github.com/pytorch/pytorch/pull/186190,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,181398,,,0,12,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-04T06:10:46Z,2026-06-12T22:48:39Z,26,"Simple spmd_types handling for FlexAttention, placing no_typecheck over internals, and typechecking inputs/outputs. Follows DTensor rule logic for Q/K/V: mesh axes must shard same tensor dims, only batch/head dims are allowed, no Partials. Output typing follows Q, and non-Q/K/V inputs aren't typechecked. will clean..."
186653,Make topk deterministic under deterministic algorithms,https://github.com/pytorch/pytorch/pull/186653,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Merged, Reverted, ciflow/trunk, release notes: cuda, ci-no-td, module: core aten, ciflow/torchtitan",2026-06-08T21:30:01Z,2026-06-12T22:15:52Z,26,"Stack from ghstack (oldest at bottom): -> #186653 When torch.use_deterministic_algorithms(True) is enabled, make topk resolve tied values with stable index tie-breaking. CPU topk now compares by value and then index in deterministic mode. CUDA keeps the existing top-k selection path and uses stable sorting for the s..."
177619,more ops,https://github.com/pytorch/pytorch/pull/177619,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"177471, 177618",,,0,8,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-03-17T04:33:23Z,2026-06-12T22:03:57Z,26,"Stack from ghstack (oldest at bottom): #177471 -> #177619 #177618 #176992 Register sharding strategies for many additional ops: kron, searchsorted, pdist, embedding_bag, histogram, histogramdd, quantile, addbmm, isin, cdist, cartesian_prod, block_diag, index_fill, index_add, scatter_reduce, and more. Also adds point..."
177618,replicate-only ops,https://github.com/pytorch/pytorch/pull/177618,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"177471, 177619",,,0,8,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-03-17T04:33:18Z,2026-06-12T22:03:55Z,26,"Stack from ghstack (oldest at bottom): #177471 #177619 -> #177618 #176992 Register replicate-only strategies for ops with no useful sharding (repeat_interleave, masked_scatter, put, take, etc.) and add sharding support for allclose via register_op_strategy with all_reduce(MIN) dispatch, following the same pattern as..."
177471,tag on ops,https://github.com/pytorch/pytorch/pull/177471,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"177618, 177619",,,0,13,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-03-15T05:43:19Z,2026-06-12T22:03:54Z,26,Stack from ghstack (oldest at bottom): -> #177471 #177619 #177618 #176992
186635,c10/util/ArrayRef.h: specialize std::ranges::enable_borrowed_range,https://github.com/pytorch/pytorch/pull/186635,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, release notes: cpp, meta-exported",2026-06-08T19:24:01Z,2026-06-12T21:53:32Z,26,"Summary: c10::ArrayRef<T> already satisfies std::ranges::contiguous_range structurally — its iterator type is a raw const T* (which models std::contiguous_iterator), data() returns const T*, and size() returns size_t. The one missing piece for full C++20 ranges integration is std::ranges::enable_borrowed_range. With..."
187213,Issue-106571: Lint Fixes - Rule B007 - gb_registry_linter - Replace …,https://github.com/pytorch/pytorch/pull/187213,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-06-12T21:13:48Z,2026-06-12T21:17:36Z,26,Description This is part of the ongoing bugbear-flak8 linting fixes that are tracked under this issue issue-106571 This PR deals with B007 lint violations in the gb_registry_linter.py file Testing To test for B007 lint violations: Remove B007 ignore flag from pyproject.toml (line 171) Run lint on the file gb_registr...
186848,Use C++20 concepts (1 of N),https://github.com/pytorch/pytorch/pull/186848,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,176662,179286,,,0,1,0,,,"triaged, open source, NNC, ciflow/trunk, release notes: jit, topic: not user facing, ciflow/vllm",2026-06-09T22:29:32Z,2026-06-12T21:00:38Z,26,C++20 introduces concepts which can simplify some template code. This PR is an attempt to bring it into the PyTorch codebase. #176662 cc @EikanWang @jgong5
183617,[fsdp2] fix mem accumulation of all reduce input buffers for hsdp with mixed precision,https://github.com/pytorch/pytorch/pull/183617,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,179128,179443,,,0,5,0,,,"ciflow/trunk, release notes: distributed (fsdp), ciflow/inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-05-13T20:15:28Z,2026-06-12T20:07:44Z,26,"Summary: Based on the conversation and implementation in #179443 and solves #179128. The basic idea is that the only condition where memory accumulates with all_reduce is when we have mixed_precision. In the previous implementation, we aggressively drop memory at the potential cost of perf due to increased stream sy..."
187189,[DO NOT MERGE][ROCm] Bump default trunk ROCm test shards to 20,https://github.com/pytorch/pytorch/pull/187189,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, triaged, open source, ciflow/trunk, topic: not user facing",2026-06-12T16:20:02Z,2026-06-12T18:47:15Z,26,[DO NOT MERGE] Summary Increase the number of shards for the default config of the linux-jammy-rocm-py3.10-mi350 test job in the trunk workflow from 8 to 20. This spreads the default ROCm test suite across more gfx950 runners to reduce per-shard wall-clock time. The inductor (2 shards) and distributed (3 shards) con...
187050,[TESTING] [Inductor] Add regression tests for large-batch BMM grid Y/Z split,https://github.com/pytorch/pytorch/pull/187050,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-rocm, ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/inductor-rocm, ciflow/rocm-mi300",2026-06-11T14:28:21Z,2026-06-12T18:01:04Z,26,"Summary Adds test/inductor/test_bmm_batch_grid_yz.py to lock in the behavior from #178617: when BMM batch size exceeds the CUDA gridDim.y cap (~65535), the launch splits batch across grid_y / grid_z, and the Triton BMM template must rebuild the batch index from tl.program_id(1) and tl.program_id(2) (via idx_q). The..."
186815,Remove deprecated torch.qr,https://github.com/pytorch/pytorch/pull/186815,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, module: amp (automated mixed precision), ciflow/trunk, release notes: mobile, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-09T18:10:15Z,2026-06-12T18:00:49Z,26,The time has come to remove deprecated linear algebra related functions. This PR removes torch.qr. Fresh reland of #70989 after the old 2022 PR was reverted and kept legacy Meta CodeSync check baggage. cc @mcarilli @ptrblck @leslie-fang-intel @jgong5 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @XiaobingSuper @...
187090,Fix nn.Module.to() failure on subclass parameters held by Python weakrefs,https://github.com/pytorch/pytorch/pull/187090,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,187008,187008,187008,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-11T22:09:53Z,2026-06-12T17:59:41Z,26,"Fixes #187008 Summary This PR resolves an issue where calling nn.Module.to(), device/dtype conversions, or load_state_dict on modules with tensor subclass parameters fails if any Python-level weak references are held to them (e.g. by EMA trackers, memory profilers, or Dynamo guards). Action & Rationale We introduce..."
187169,Support weights_only and safetensors format in torch.export.save and …,https://github.com/pytorch/pytorch/pull/187169,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,153410,153410,153410,,,,0,1,0,,,"triaged, open source, release notes: export",2026-06-12T11:22:12Z,2026-06-12T17:55:13Z,26,"Summary This PR adds support for weights_only=True and safetensors format in torch.export.save and torch.export.load, addressing security concerns with pickle-based deserialization of custom objects in exported programs (issue #153410). Fixes #153410 Proposed Changes Updated load and save signatures in torch/export/..."
185122,Validate parameter/buffer storage bounds at export time,https://github.com/pytorch/pytorch/pull/185122,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,184996,184996,184996,,,,0,1,0,,,"triaged, open source, release notes: export",2026-05-25T08:33:36Z,2026-06-12T17:39:33Z,26,"Fixes: #184996 Summary : Adds early validation in torch.export.export() to reject parameters/buffers whose underlying storage is too small for their declared shape/stride/offset metadata. Previously, export would silently succeed and produce an unusable ExportedProgram that failed later during ep.module() or torch.e..."
177287,"Default backward_pass_autocast to ""off"" to match eager autocast semantics",https://github.com/pytorch/pytorch/pull/177287,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,153044,"153044, 174929",,,,0,2,0,,,"module: bc-breaking, Stale, release notes: autograd, topic: bc breaking, ciflow/inductor",2026-03-12T18:41:46Z,2026-06-12T17:27:50Z,26,"*** DON'T LAND UNTIL #177286 HAS BEEN RELEASED (not just landed) (so this goes out in the next release) *** The backward pass of torch.compile'd regions no longer inherits the forward pass's autocast context by default. This matches eager PyTorch behavior, where autocast conventionally wraps only the forward pass. T..."
187097,[BE] Fix B007 unused loop variables,https://github.com/pytorch/pytorch/pull/187097,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,106571,,,,0,1,0,,,"triaged, open source, release notes: releng",2026-06-11T22:49:39Z,2026-06-12T17:12:17Z,26,Fixes a small set of B007 lint violations by renaming intentionally unused loop variables with a leading underscore. This addresses the B007 slice I mentioned in #106571. Test: ./venv/bin/python -m ruff check --no-cache --select B007 .github/scripts .ci/libtorch
186573,[Inductor] Stabilize checkpointed reduction autotuning,https://github.com/pytorch/pytorch/pull/186573,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,186572,186572,,,,0,5,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-07T23:37:55Z,2026-06-12T17:00:15Z,26,"Stack from ghstack (oldest at bottom): -> #186573 Activation checkpointing wants the original forward AC region and its recompute to use the same numerics-affecting compiled strategy. The ideal long-term design is compile-once/reuse for AC subgraphs in AOTAutograd or partitioning, but today those copies are compiled..."
185366,[ROCm][Distributed] Add synthesized test for sac_milp to ensure numerical correctness,https://github.com/pytorch/pytorch/pull/185366,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/rocm-mi200, ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: rocm, triaged, open source, ciflow/trunk, topic: not user facing, ciflow/rocm-mi300, ciflow/rocm-mi200",2026-05-27T13:04:00Z,2026-06-12T12:56:54Z,26,"Summary Adds TestSACILP.test_sac_ilp_case1_with_pinned_hw, a synthesized variant of test_sac_ilp_case1 that restores the exact numerical assertions which #182670 had to drop in order to make test_sac_ilp_case1 portable across A100/H100/MI300/MI350. Background #182670 (review) test_sac_ilp_case1 feeds the SAC MILP wi..."
186343,Enable FA build for SDPA on Windows+CUDA,https://github.com/pytorch/pytorch/pull/186343,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, ciflow/binaries_wheel",2026-06-05T07:58:13Z,2026-06-12T11:45:28Z,26,"Summary Description Enable building the FlashAttention backend for scaled_dot_product_attention on Windows + CUDA (MSVC). Today, the flag is hard-disabled on MSVC, so every Windows CUDA build ships without the flash SDPA backend and silently falls back to slower, memory-heavy backends. With this PR, the FlashAttenti..."
179675,[WIP]port dynamo cases to Intel GPU,https://github.com/pytorch/pytorch/pull/179675,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,8,0,,,"open source, Stale, topic: not user facing, module: dynamo, ciflow/inductor",2026-04-08T06:07:15Z,2026-06-12T08:05:06Z,26,Stack from ghstack (oldest at bottom): -> #179675 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo @azahed98
187146,test all named dims test with dynamic sepcs using automatic conversion,https://github.com/pytorch/pytorch/pull/187146,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,2,0,,,topic: not user facing,2026-06-12T03:30:51Z,2026-06-12T07:28:34Z,26,Stack from ghstack (oldest at bottom): #187014 #187010 #186751 #186167 -> #187146 #187145 #187144 #187143 #185982
187135,[xpu] Remove deprecated at::xpu CachingHostAllocator forwarders,https://github.com/pytorch/pytorch/pull/187135,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source",2026-06-12T02:37:48Z,2026-06-12T05:51:32Z,26,
186046,[MPS] Remove dead xfail/tolerance keys from MPS test infra,https://github.com/pytorch/pytorch/pull/186046,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/mps",2026-06-03T06:26:44Z,2026-06-12T05:51:16Z,26,
180192,[DO NOT REVIEW] fix torchtitan CI error,https://github.com/pytorch/pytorch/pull/180192,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"Stale, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-04-13T00:03:40Z,2026-06-12T03:48:16Z,26,Stack from ghstack (oldest at bottom): -> #180192
178371,[Profiler] Fix profile_memory not working with profile_all_threads,https://github.com/pytorch/pytorch/pull/178371,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,166121,166121,166121,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-03-25T08:01:55Z,2026-06-12T03:48:16Z,26,"Fix #166121 Summary Fix profile_memory=True having no effect when profile_all_threads=True. Both features are useful in our use case and should work together. Root cause profile_all_threads=True stores profiler state in a global manager at the torch/csrc layer, but the memory allocator in c10 only checks ThreadLocal..."
184249,Fix extern autotune tensor kwargs,https://github.com/pytorch/pytorch/pull/184249,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,CLEAN,False,0,0,,,,139077,139077,,,,0,2,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan, agentic, needs-review",2026-05-18T18:41:38Z,2026-06-11T23:28:09Z,27,"Stack from ghstack (oldest at bottom): -> #184249 Materialize tensor keyword arguments for extern max-autotune benchmarking and keep generated extern calls using realized kwargs, with regression coverage for optional tensor kwargs.\n\nFixes #139077\nGenerated by my agent cc @voznesenskym @penguinwu @EikanWang @jgong..."
186940,[ShardedTensor] use device-agnostic fallback in tensor_device (fixes #186938),https://github.com/pytorch/pytorch/pull/186940,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/torchtitan,review when ready,,UNKNOWN,False,0,0,,,186938,186938,186938,,,,0,2,0,,,"triaged, open source, release notes: distributed (sharded), ciflow/torchtitan",2026-06-10T15:55:53Z,2026-06-11T20:01:15Z,27,"Fixes #186938. Summary ShardedTensor.device's no-local-shards fallback in torch/distributed/_shard/sharded_tensor/_ops/tensor_ops.py:54 hardcodes torch.device(torch.cuda.current_device()). On any non-CUDA build of PyTorch (XPU, MPS, HPU, …) reading .device on a ShardedTensor with empty _local_shards raises Assertion..."
175509,Add fast path for min/max dim reductions with size 2 on CPU,https://github.com/pytorch/pytorch/pull/175509,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,140000,140000,140000,,,,0,2,0,,,"module: cpu, triaged, open source, Stale, release notes: cpu (x86)",2026-02-23T00:14:41Z,2026-06-11T19:24:58Z,27,"Summary When reducing a dimension of size 2, torch.min(tensor, dim=...) and torch.max(tensor, dim=...) are significantly slower than NumPy because compare_base_kernel strides along the reduction dimension for each output element. For dim-0 on large contiguous tensors, this causes cache misses on every comparison (e...."
173239,Test openreg inference hooks,https://github.com/pytorch/pytorch/pull/173239,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,172321,172321,172321,,,,0,8,0,,,"triaged, open source, Stale, topic: not user facing, module: openreg",2026-01-24T07:39:52Z,2026-06-11T19:24:28Z,27,Fixes #172321 A lot of what was required in the issue was already solved in b76165a. This PR just adds the two missing tests from #172321: test_autograd_inference_mode test_backward_hook cc mansiag05 @parsshar-RH
185236,Detect fp32 precision flag leaks in TestCase and fix offenders,https://github.com/pytorch/pytorch/pull/185236,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"triaged, open source, release notes: distributed (c10d), module: inductor, module: dynamo, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-26T15:32:01Z,2026-06-11T18:53:33Z,27,"Follow-up to #180926. In review there, @albanD asked that we not spread defensive with_ieee_matmul_precision decorators across matmul-using tests, and instead fix the root cause: tests that leave a global fp32 precision flag in a non-default state, plus the base TestCase not catching such leaks. This makes any test..."
186619,[pytorch] Fix Inductor missing wait on AsyncCollectiveTensor nested inside DTensor,https://github.com/pytorch/pytorch/pull/186619,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180614,180614,,,,,0,1,0,,,"topic: not user facing, ciflow/inductor, meta-exported",2026-06-08T17:11:05Z,2026-06-11T18:11:55Z,27,Summary: AOT autograd's process_inputs scans flat_args for top-level AsyncCollectiveTensor (ACT) and records act_input_indices so the compiled graph's inner_fn can emit args[i].trigger_wait() before the kernel runs (introduced for the top-level case in #179849). That scan is non-recursive. When a DTensor carries an...
179257,[c10] Fix off-by-one issue in ApproximateClock.cpp,https://github.com/pytorch/pytorch/pull/179257,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-03T15:36:11Z,2026-06-11T17:43:17Z,27,"Problem When replicates is odd (e.g., 1001), the correct 0‑based median index is replicates / 2 = 500. The code incorrectly added +1, using index 501 instead. This selects the element right of the true median, skewing the scale factor and t0 correction. Example replicates = 1001 Wrong: scale_factors[1001/2 + 1] = sc..."
187006,Avoid duplicate sm_103 family-specific CUDA codegen,https://github.com/pytorch/pytorch/pull/187006,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: build",2026-06-11T02:49:43Z,2026-06-11T17:40:48Z,27,"Summary This fixes a CUDA 13.1 codegen conflict when the build already requests compute_103/sm_103 globally. Codegen.cmake adds family-specific 103a and 103f per-file flags for a few CUDA sources when compute_100 is present. If the global arch list also contains compute_103, nvcc sees both non-family-specific sm_103..."
186950,Fixed issue with pickling of custom decomp dict,https://github.com/pytorch/pytorch/pull/186950,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-10T17:07:36Z,2026-06-17T19:18:49Z,21,"This is a follow-up PR for #175954. Although the fix implemented there was working, it can cause some issues with FX graph caching. In particular, when supplying a custom decomposition table, it would be wrapped in compile_fx.<locals>.get_decomp_fn, which would eventually cause issues with pickling further downstrea..."
183669,[inductor] Make experimental benchmarker parameters configurable via env vars,https://github.com/pytorch/pytorch/pull/183669,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-14T07:30:20Z,2026-06-11T17:02:53Z,27,💡 Summary The experimental benchmarker currently hardcodes parameters like estimation_iters and memory_warmup_iters. This PR makes them configurable via environment variables to allow easier tuning for different hardware and workloads without modifying source code. Added environment variables: TORCHINDUCTOR_EXPERIME...
183556,[inductor] Fix MemoryTracker assertion with shared storages in reordered schedules,https://github.com/pytorch/pytorch/pull/183556,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-13T14:00:52Z,2026-06-11T16:59:49Z,27,"Stack from ghstack (oldest at bottom): -> #183556 When selective activation checkpointing (SAC) duplicates forward nodes into the backward region via graph.node_copy, the recomputed node shares the same FakeTensor storage (_cdata) as the original forward node. GraphAliasTracker attributes each storage to the first n..."
182964,[inductor][overlap] Wire _low_contention_all_reduce into LC FX pass,https://github.com/pytorch/pytorch/pull/182964,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,182963,,,0,3,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/torchtitan",2026-05-08T16:29:39Z,2026-06-11T16:54:07Z,27,"Stack from ghstack (oldest at bottom): -> #182964 #182963 #182955 Extend the low-contention collectives pass to also replace _c10d_functional.all_reduce with symm_mem._low_contention_all_reduce. Refactors the internal type from is_ag bool to a coll_type string (""ag"", ""rs"", ""ar"") to support the three collective types..."
182963,[symm_mem] Add _low_contention_all_reduce to symmetric memory,https://github.com/pytorch/pytorch/pull/182963,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,182964,,,0,3,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/h100-symm-mem, ciflow/torchtitan",2026-05-08T16:29:34Z,2026-06-11T16:53:59Z,27,"Stack from ghstack (oldest at bottom): #182964 -> #182963 #182955 One-shot pull-and-accumulate all-reduce using copy engine P2P: Each rank copies input to symm_mem workspace (if needed) Barrier to ensure all inputs visible Pull local rank's data, then for each remote rank: P2P copy into temp buffer + accumulate into..."
181748,[Inductor][Triton] Enable decompose_k and use Triton for inner BMM on HIP (#181748) (#181748),https://github.com/pytorch/pytorch/pull/181748,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/slow, ciflow/rocm-mi300, release notes: inductor (aoti), ciflow/h100, meta-exported, ciflow/b200, ciflow/torchtitan",2026-04-28T16:03:16Z,2026-06-11T16:51:30Z,27,"Summary: Enable decompose_k on AMD/ROCm by removing the conditional that disabled it (previously defaulted num_decompose_k_splits to 0 on HIP). Both NVIDIA and AMD now default to 10 splits. On HIP, the inner BMM produced by decompose_k is routed through a Triton template (max_autotune_gemm_backends=""TRITON"") since i..."
178300,[Inductor] Retry Triton compilation without pointer_range_32 on failure,https://github.com/pytorch/pytorch/pull/178300,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,1,0,,,"fb-exported, Stale, topic: not user facing, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-03-24T17:47:31Z,2026-06-11T16:51:23Z,27,"Summary: #176675 added tt.pointer_range=32 annotations for AMD/HIP to enable buffer load/store ops via 32-bit pointer offsets. However, some kernels (e.g. fused_group_contiguous_nan_clamp from ads_mkl) have complex strided pointer patterns that the CanonicalizePointers pass cannot decompose, causing PassManager::run..."
175450,[inductor] CUDAGraph P2P pool handling for symm_mem,https://github.com/pytorch/pytorch/pull/175450,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,175476,,,0,18,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-02-20T21:25:23Z,2026-06-11T16:50:40Z,27,Stack from ghstack (oldest at bottom): #175797 #175476 -> #175450 #175449 #174856 Summary CUDAGraph tree assumes every live CUDA tensor belongs to the caching allocator's private pool. P2P symmetric memory buffers don't — they are backed by cuMemCreate/cuMemMap and live outside the pool entirely. Without special han...
171780,Add support shared mempool for accelerator Graph,https://github.com/pytorch/pytorch/pull/171780,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,171445,,,0,38,0,,,"open source, ciflow/trunk, release notes: python_frontend",2026-01-06T12:17:23Z,2026-06-11T16:50:34Z,27,"Stack from ghstack (oldest at bottom): #171445 -> #171780 Motivation This PR introduces a new API torch.accelerator.generate_graph_pool_handle, as a unified replacement for torch.cuda.graph_pool_handle. This API returns an opaque and unique memory pool handle that can be shared across multiple graphs to explicitly c..."
181007,[inductor][comms] Optional use of process_group alloc_tensor for collectives,https://github.com/pytorch/pytorch/pull/181007,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181008, 181009, 181010",,,0,1,0,,,"ciflow/trunk, release notes: distributed (c10d), topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/torchtitan",2026-04-21T18:29:32Z,2026-06-11T16:48:36Z,27,"Stack from ghstack (oldest at bottom): #181010 #181009 #181008 -> #181007 C++ backend (Backend.hpp, ProcessGroupNCCL.hpp, init.cpp) Adds isInitialized() virtual method to Backend base class (default: true). ProcessGroupNCCL overrides it (already had the method, now marked override) since NCCL initializes lazily. Exp..."
181292,[inductor] Autotune cache stability and perf improvements (#181292),https://github.com/pytorch/pytorch/pull/181292,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-23T21:42:35Z,2026-06-11T16:46:20Z,27,"Summary: Fixes two bugs in the autotune cache that caused unnecessary re-autotuning, and makes cache keys stable across kernel renumbering. Bug 1: found_by_coordesc lost during pickle. Config.found_by_coordesc is a dynamic attribute that gets dropped by Config.__setstate__ during the async compile worker-parent pick..."
179548,[inductor] Add dedicated thread pool for nogil Triton compilation (#179548) (#179548),https://github.com/pytorch/pytorch/pull/179548,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-07T04:33:23Z,2026-06-11T16:40:50Z,27,Summary: Create AsyncCompile.thread_pool() as parallel implementation to process_pool() for use with free-threaded Python. This enables zero-overhead kernel compilation when GIL is disabled. This is Commit 3 of the nogil thread-mode compilation series. Motivation: Fork overhead is 10-50ms and holds GIL Pickle/unpick...
179526,inductor: handle matmuls gracefully in foreach_map,https://github.com/pytorch/pytorch/pull/179526,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,158975,158975,158975,,,,0,2,0,,,"triaged, open source, ciflow/trunk, release notes: foreach_frontend, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-06T23:52:31Z,2026-06-11T16:40:41Z,27,"Add FallbackSubgraphLowering to handle non-pointwise ops like mm and matmul in foreach_map. Previously, SubgraphLoweringException would propagate when PointwiseSubgraphLowering encountered ops that require buffer creation. Now _foreach_map catches this exception and falls back to individual op lowerings via Fallback..."
179142,[Inductor] Add runtime assertion support to FX wrapper backend (#179142),https://github.com/pytorch/pytorch/pull/179142,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, release notes: fx, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-02T13:50:25Z,2026-06-11T16:35:11Z,27,Summary: Add WrapperLine subclasses and FX converter methods so that backends using FX IR (like MTIA) get the same runtime assertion support as the Python and C++ wrapper backends. New WrapperLine types and their FX IR generation: SizeAssertLine: size/stride validation via sym_size/sym_stride → eq → torch._check Ali...
178924,[inductor] Fix FP8 dtype detection in _get_estimated_runtime,https://github.com/pytorch/pytorch/pull/178924,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, topic: not user facing, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-03-31T20:09:52Z,2026-06-11T16:34:57Z,27,"Summary: _get_estimated_runtime() uses buf.node.maybe_get_dtype() to determine the dtype for peak TFLOPS lookup. For FP8 matmul ops (aten._scaled_mm), this returns the output dtype (FP16/BF16 accumulator precision), not the compute dtype (FP8). This causes compute-bound roofline times to be ~2x too high for FP8 kern..."
178083,Update inductor expected accuracy files,https://github.com/pytorch/pytorch/pull/178083,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing, module: dynamo, ciflow/inductor",2026-03-22T03:51:46Z,2026-06-11T16:27:40Z,27,Summary This PR updates the expected accuracy CSV files for inductor benchmarks based on CI results from PyTorch commit 27c9a35. These files serve as reference points for dynamo/inductor CI to track: Graph breaks Model accuracy Changes Updated CUDA expected accuracy files in benchmarks/dynamo/ci_expected_accuracy/ U...
178048,"[inductor] Add custom_pass_context for scoped, composable custom pass registration",https://github.com/pytorch/pytorch/pull/178048,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,1,0,,,"triaged, open source, Stale, module: inductor, ciflow/inductor",2026-03-21T08:33:33Z,2026-06-11T16:26:19Z,27,"Introduces custom_pass_context, a context manager based mechanism for registering custom Inductor FX passes. This replaces the existing pattern of assigning stateful objects directly to Inductor config fields, which caused config serialization failures and broken FX graph caching. The existing custom pass registrati..."
177854,[Inductor] Run first cudagraph warmup in eager pool (#177854),https://github.com/pytorch/pytorch/pull/177854,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-03-19T15:09:21Z,2026-06-11T16:26:02Z,27,"Summary: Models that lazily initialize persistent state (e.g. KV caches) during cudagraph warmup allocate those tensors in the cudagraph private pool. On generation changes, dealloc_current_path_weakrefs() frees and poisons them, causing ""tensor output of CUDAGraphs has been overwritten"" errors. Fix: when cudagraph_..."
175584,fix aten.remainder in inductor,https://github.com/pytorch/pytorch/pull/175584,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-02-23T23:37:15Z,2026-06-11T16:20:54Z,27,cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
171445,Intorduce torch.accelerator.is_graph_available to check if Graph is available,https://github.com/pytorch/pytorch/pull/171445,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,171780,,,0,51,0,,,"open source, ciflow/trunk, release notes: python_frontend",2025-12-29T06:40:14Z,2026-06-11T16:17:54Z,27,Stack from ghstack (oldest at bottom): -> #171445 #171780 Motivation This PR aims to introduce a new API torch.accelerator.is_graph_available to detect if the current accelerator supports graph capture and replay feature. This will help the user do something such as: if torch.accelerator.is_graph_available(): graph...
176295,[inductor] Fix cudagraph output->input feedback crashes,https://github.com/pytorch/pytorch/pull/176295,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,176294,176294,,,,,0,4,0,,,"triaged, open source, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-03-03T15:45:27Z,2026-06-11T15:45:18Z,27,Closes #176294 Fixes #176294 Fixes crashes when CUDA graph outputs are fed back as inputs across successive calls — the standard RL/simulation loop pattern state = compiled_step(state). The root cause is dealloc_current_path_weakrefs: on a generation transition it frees all pool storage from the previous path and ma...
170609,Add auto-generated cuDNN depthwise conv heuristic,https://github.com/pytorch/pytorch/pull/170609,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: cudnn, module: cuda, module: cpu, triaged, open source",2025-12-16T22:32:36Z,2026-06-11T14:58:57Z,27,"This PR revisits the depthwise convolution heuristic introduced in #58749, to automate the heuristic generation process and to improve behavior with newer devices and cuDNN versions. The updated heuristic performs about 5-8% better on average than the existing heuristic, and the entire decision tree can be replicate..."
185053,[Test]Refactor test_custom_ops.py to be device agnostic,https://github.com/pytorch/pytorch/pull/185053,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, ciflow/trunk, topic: not user facing",2026-05-24T07:28:48Z,2026-06-11T14:54:37Z,27,"Make 5 tests in TestCustomOp and TestCustomOpAPI classes device-agnostic by replacing hardcoded CUDA specific code with accelerator generic patterns. Changes: Replace @unittest.skipIf(not TEST_CUDA/TEST_XPU) with @unittest.skipIf(not TEST_ACCELERATOR) Replace .cuda() with .to(device_type) Replace hardcoded ""cuda"" wi..."
187004,perf(dynamo): add guard lookup fast paths,https://github.com/pytorch/pytorch/pull/187004,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,35,0,,,"open source, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-11T02:35:38Z,2026-06-11T14:53:09Z,27,背景 在大模型/推荐模型的 torch.compile 推理场景中，单次 step 进入 compiled region 前会经过 TorchDynamo cache lookup。这个 lookup 需要执行 guard 检查，用来判断当前输入、模块状态、全局状态是否仍然匹配已编译 graph。 现场 profile 发现，部分模型的耗时并不只在算子下发和 NPU kernel 上，step 边界处的 TorchDynamo Cache Lookup 也会成为明显瓶颈。进一步统计显示，主要开销集中在 guard tree 遍历，尤其是从 L['self'] 开始递归检查大型 nn.Module 树： L['self'] L[...
171617,[FSDP] Preserve container object identity in _recursive_to function,https://github.com/pytorch/pytorch/pull/171617,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,157832,157832,157832,,,,0,3,0,,,"triaged, open source, ciflow/trunk, release notes: distributed (fsdp), ciflow/torchtitan",2026-01-02T10:07:58Z,2026-06-11T13:06:41Z,27,"Fix a bug where _recursive_to in torch/distributed/utils.py was creating new container objects (list, tuple, dict, namedtuple) even when all elements were unchanged. This broke code relying on object identity being preserved through FSDP-wrapped modules. The fix adds identity checks for the single-device case, if al..."
186061,Fixing torch.pow crash with integral dtypes on CUDA by gating reciprocal/sqrt/rsqrt shortcuts,https://github.com/pytorch/pytorch/pull/186061,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,137440,137440,137440,,,,0,1,0,,,"triaged, open source",2026-06-03T09:53:03Z,2026-06-11T11:41:37Z,27,"Gated the reciprocal/sqrt/rsqrt fast-path shortcuts in pow_tensor_scalar_kernel on !isIntegralType() so integral dtypes skip the shortcuts and fall through to powi(), matching CPU behavior. (Fixes #137440). Added a corresponding test for int8/int16/int32/int64 on CUDA."
182233,Fix torch.combinations dynamic shape support by replacing numel with sym_numel and propagating SymInt,https://github.com/pytorch/pytorch/pull/182233,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,11,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/dtensor",2026-05-02T20:54:21Z,2026-06-11T10:59:22Z,27,"Fix torch.combinations dynamic shape support Problem torch.combinations fails when used with dynamic shape compilation: torch.compile(model, dynamic=True) This results in: RuntimeError: Cannot call numel() on tensor with symbolic sizes/strides The issue is caused by the use of: self.numel() numel() requires a concre..."
186924,Also search $MKLROOT/lib in Linux for mkl wheel and install_mkl.sh,https://github.com/pytorch/pytorch/pull/186924,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cpu, triaged, open source, module: intel",2026-06-10T10:12:25Z,2026-06-11T10:08:12Z,27,Since current FindMKL.cmake only searchs $MKLROOT/lib/intel64 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01 @frank-wei
186232,Add opt-in cross-stream segment reclaim to CUDA allocator,https://github.com/pytorch/pytorch/pull/186232,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-04T13:30:21Z,2026-06-11T08:32:39Z,27,"This PR allows the allocator to reclaim a whole idle segment cached under another stream, cutting cudaMalloc/OOM in multi-stream workloads."
184190,[fx]: Add SymNode helper to get raw expr,https://github.com/pytorch/pytorch/pull/184190,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: fx, ciflow/inductor",2026-05-18T11:29:15Z,2026-06-11T08:23:21Z,27,Adds helper to access raw sympy expression without triggering replacement logic. This is useful for lowering of dynamic models in ExecuTorch where symbolic expressions are to be materialized into shape operations. Also adds a test case which highlights the differences between expr() and raw_expr().
178258,[submodule] Upgrade to protobuf v33.6,https://github.com/pytorch/pytorch/pull/178258,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/linux-aarch64, ciflow/trunk",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,6,0,,,"module: mkldnn, module: third_party, topic: build, open source, NNC, ciflow/trunk, release notes: build, module: bazel, ciflow/linux-aarch64",2026-03-24T10:15:02Z,2026-06-11T07:36:22Z,27,"This PR upgrade Protobuf to v33.6 with the help of Claude Code. It also contains the following changes: Protobuf v3.13 → v33.5 for CMake builds Remove ~150 lines of legacy protobuf patching (GetEmptyStringAlreadyInited, move-constructor rewriting) Fix aarch64 builds by building protobuf as shared to avoid relocation..."
185932,[MPS] Add `lcm` Metal kernel,https://github.com/pytorch/pytorch/pull/185932,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: cpu, triaged, open source, Merged, Reverted, ciflow/trunk, release notes: mps, ciflow/mps, ci-no-td",2026-06-02T14:01:15Z,2026-06-11T06:58:02Z,27,This PR adds lcm Metal kernel and switches gcd to binary GCD. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
180162,Add CPU tensor validation to FBGEMM operators to prevent segfault,https://github.com/pytorch/pytorch/pull/180162,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,173495,173495,173495,,,,0,1,0,,,"open source, Stale, release notes: quantization",2026-04-12T02:44:04Z,2026-06-11T06:06:57Z,27,"Fixes #173495 FBGEMM operators segfault when passed CUDA tensors because they read GPU memory pointers as CPU memory. Added TORCH_CHECK(tensor.is_cpu(), ...) to all FBGEMM functions that directly access tensor data, and a test verifying the error is raised. cc @jerryzh168 @malfet"
184895,[Inductor] Index expression codegen for user Triton kernel fusion,https://github.com/pytorch/pytorch/pull/184895,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,DIRTY,False,0,0,,,,,,181138,,,0,39,0,,,"triaged, open source, module: inductor, module: dynamo, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-22T14:58:11Z,2026-06-11T05:29:28Z,27,"Enables epilogue fusion for epilogues that require additional index expressions outside the user kernel's scope. This is done by parameterising FusedUserTritonKernel with the output tile/block dimension, which the user annotates using wrap_triton: wrap_triton(some_kernel, output_tile=(""BLOCK_SIZE"",))[GRID](...) Exam..."
182526,[DTensor] register missing pooling backward sharding strategies,https://github.com/pytorch/pytorch/pull/182526,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-05T16:53:37Z,2026-06-11T05:16:27Z,27,"Eight pooling backward aten ops lacked sharding strategies, causing hard crashes (RuntimeError: no sharding strategy registered) when backpropagating through sharded pooling layers. Add them to existing strategy functions: interp_pool_1out_2in_strategy: avg_pool2d/3d_backward, _adaptive_avg_pool3d_backward pool_back..."
185000,debug_mode: avoid hashing async collectives before wait,https://github.com/pytorch/pytorch/pull/185000,ready for maintainer decision,67,needs-review label present; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor,review diff and decide acceptance path,,CLEAN,False,0,0,,,,169150,169150,,,,0,3,0,,,"topic: not user facing, ciflow/dtensor, agentic, needs-review",2026-05-23T18:01:15Z,2026-06-11T03:18:46Z,28,"Stack from ghstack (oldest at bottom): -> #185000 DebugMode.log_tensor_hashes installed generic dispatch hooks that hashed every tensor input and output. Eager functional collectives can expose asynchronous results through _c10d_functional op outputs and AsyncCollectiveTensor wrappers before wait_tensor runs, so deb..."
184825,[DTensor] add sharding strategy for group norm backward,https://github.com/pytorch/pytorch/pull/184825,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-21T23:48:52Z,2026-06-11T02:25:56Z,28,"Register a single-dim strategy for aten.native_group_norm_backward.default that enables batch-dim (Shard(0)) sharding. Each sample is independent for grad_input; grad_weight and grad_bias reduce over batch via Partial(""sum""). Also register the scalar shape adjuster for the backward op so N/C/HxW arguments are rewrit..."
180971,"[inductor] Using is_gpu() instead of hardcoded ""cuda"" check",https://github.com/pytorch/pytorch/pull/180971,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180951,180951,,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-21T10:04:55Z,2026-06-11T01:24:30Z,28,"Resolves: #180951 Issue: In the pytorch/torch/_inductor/output_code.py file, the logging of cudagraph skip reasons happens only when the device_types has ""cuda"". This check occurs in three locations in the file, that are fixed with this pr This hard coded ""cuda"" check is inconsistent with that of done in pytorch/tor..."
186776,Fix __qualname__ for OpOverload and OpOverloadPacket,https://github.com/pytorch/pytorch/pull/186776,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,186140,186140,186140,,,,0,1,0,,,"triaged, open source",2026-06-09T09:25:51Z,2026-06-11T01:17:39Z,28,"Summary Fix two bugs in the qualname PR that caused CI failures: Namespace doubling in opoverloadpacket.module: In _OpNamespace.getattr, opoverloadpacket.module was set to self.module + ""."" + namespace_name, which doubled the namespace (e.g., ""torch.ops.aten.aten"" instead of ""torch.ops.aten""). Fixed to just self.mod..."
186949,Fix for scan autograd issue,https://github.com/pytorch/pytorch/pull/186949,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,open source,2026-06-10T17:05:53Z,2026-06-11T00:25:57Z,28,"This PR is an alternative approach to #184529. In particular, it forces the init to have requires_grad=True during the tracing phase, in order to not have the wrong zero propagation. This should only affect the tracing phase for the backward and not the operation itself. In addition, I expanded the existing testcase..."
174196,Add validation for lstm.data hidden state shapes to prevent segfaults,https://github.com/pytorch/pytorch/pull/174196,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,173946,173946,173946,,,,0,1,0,,,"triaged, open source, Stale, release notes: nn",2026-02-03T13:59:26Z,2026-06-10T23:56:04Z,28,"Add shape validation checks in both C++ native implementation and Python decomposition path for lstm.data operation. The validation ensures that hx[0] and hx[1] have correct dimensions matching num_layers * num_directions, consistent batch sizes, and compatibility with batch_sizes tensor. This prevents segfaults whe..."
176670,[CUDAGraph] Enable `graph_capture_record_stream_reuse` with conditional nodes,https://github.com/pytorch/pytorch/pull/176670,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175001,175001,176657,,,0,3,0,,,"module: cuda, open source, module: cuda graphs, topic: not user facing, module: CUDACachingAllocator",2026-03-06T00:40:50Z,2026-06-10T23:50:28Z,28,"See #175001. Stack from ghstack (oldest at bottom): -> #176670 #176657 Conditional nodes introduce three memory-reuse cases: parent alloc, child free: disallowed; PR #176657 detects the capture-ID mismatch in free() child alloc, child free/reuse: already works via normal get_free_block() matching child alloc, parent..."
186608,[user-streams][aoti] Add minimal AOTI user stream support (#186608),https://github.com/pytorch/pytorch/pull/186608,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), meta-exported, ciflow/torchtitan",2026-06-08T14:51:21Z,2026-06-10T21:53:51Z,28,"Summary: This implements the MVP AOTI path for explicit user streams by inlining record_event and wait_event in the generated C++ wrapper, allocating per-thread cached side streams/events in a generated helper header, preserving scheduler stream metadata across grouped/fused nodes, and rejecting unsupported stream s..."
185765,[Inductor] Fix correctness discrepancy in torch.minimum and torch.maximum with signed zeros (-0.0 vs 0.0),https://github.com/pytorch/pytorch/pull/185765,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185610,185610,,,,,0,6,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-31T18:03:12Z,2026-06-10T19:47:04Z,28,"PR Description Fixes #185610 Summary This PR fixes a correctness discrepancy between eager mode and Inductor for torch.minimum, torch.maximum, and their _with_index (argmin/argmax) variants when handling signed zeros (-0.0 vs +0.0). Because standard GPU Floating Point Units (FPUs) enforce 0.0 == -0.0, standard < or..."
186753,[fx] Promote raw-SymInt arg warning to RuntimeError,https://github.com/pytorch/pytorch/pull/186753,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-09T02:13:42Z,2026-06-10T19:35:55Z,28,Summary: part of three of re- lnand plan for #186272 due to executorch dependency Differential Revision: D107938875 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @a...
175468,[DO NOT MERGE] Test AMD Capacity.,https://github.com/pytorch/pytorch/pull/175468,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor-rocm-mi300, ciflow/periodic-rocm-mi300, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,31,0,,,"triaged, open source, Stale, topic: not user facing, keep-going, ci-no-td, ciflow/rocm-mi300, ciflow/periodic-rocm-mi300, ciflow/inductor-rocm-mi300",2026-02-21T00:53:09Z,2026-06-10T19:23:01Z,28,"Switch MI300 ROCm workflows to linux.rocm.gpu.gfx942.{1,4}.stg and extend actionlint labels accordingly. Authored with Cursor."
186863,[MPS] Register unsigned binary Metal kernels,https://github.com/pytorch/pytorch/pull/186863,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,176296,176296,176296,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-06-10T01:26:35Z,2026-06-10T19:20:14Z,28,"Summary: Register MPS binary Metal kernels for uint16, uint32, and uint64 across integer, int-to-float, comparison castout, and add/sub/lerp alpha families. Add unsigned MPS regression coverage for tensor, scalar, strided, comparison out=, shift, alpha, and lerp scalar-weight paths. Fixes #176296 After #176343 fixed..."
183343,dynamo: trace torch.cuda.Event.record / Event.wait under fx proxy mode,https://github.com/pytorch/pytorch/pull/183343,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/torchtitan; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: dynamo, ciflow/torchtitan",2026-05-11T18:00:58Z,2026-06-10T19:15:57Z,28,"When Event.record(stream) and Event.wait(stream) are called from regular Python bytecode that dynamo symbolically traces, the calls are picked up by EventVariable.call_method and lowered to torch.ops.streams.{record,wait}_event FX nodes. When the same calls happen during eager execution under an active fx proxy trac..."
186887,Update test_mm_sparse_first_NT/NN input generation,https://github.com/pytorch/pytorch/pull/186887,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: sparse, triaged, open source, topic: not user facing",2026-06-10T03:56:23Z,2026-06-10T18:05:18Z,28,"In B = torch.rand(dense_input_shape, device=A_sparse.device).to(dtype), torch.rand() generates values in [0, 1), converting to int8 produces an all-zero tensor. As a result, both dense and sparse matmul paths operate on a zero matrix and produce all-zero outputs, allowing the test to pass without validating INT8 num..."
113258,[pytree] support PyStructSequence types for Python pytree,https://github.com/pytorch/pytorch/pull/113258,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,165,0,,,"open source, Stale, ciflow/trunk, topic: not user facing, module: pytree, module: dynamo, ciflow/inductor, no-stale, ci-test-showlocals, ciflow/torchtitan",2023-11-08T12:36:46Z,2026-06-10T16:06:12Z,28,Stack from ghstack (oldest at bottom): -> #113258 cc @zou3519
181188,[AOTInductor] Outline per-kernel-call teardown into AOTI_NOINLINE helpers,https://github.com/pytorch/pytorch/pull/181188,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-04-22T23:17:09Z,2026-06-10T16:02:28Z,28,"Stack from ghstack (oldest at bottom): -> #181188 For each kernel call, gather the contiguous trailing FreeLine/FreeIfNotReusedLine/NullLine entries into an AOTI_NOINLINE static void free_after_<kernel>_<id>(...) helper that holds the input-buffer .reset() calls, and replace them in run_impl with a single call to th..."
182980,Clear exception tracebacks to avoid reference cycles,https://github.com/pytorch/pytorch/pull/182980,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/docker, ciflow/dtensor, ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,167906,167906,167906,,,,0,13,0,,,"triaged, open source, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), release notes: distributed (checkpoint), release notes: export, ciflow/torchtitan, ciflow/dtensor, ciflow/docker, jansel-agent-skip",2026-05-08T18:27:28Z,2026-06-10T14:48:54Z,28,"Summary Fixes #167906 When an exception is caught with except Exception as e and stored in a variable that outlives the except block, the traceback object holds references to the frame's locals, creating a reference cycle. This can delay garbage collection of large objects like tensors, causing unexpected memory ret..."
186808,[inductor] Re-key range_tree_nodes on IterationRangesEntry rename,https://github.com/pytorch/pytorch/pull/186808,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-09T17:21:59Z,2026-06-10T14:44:27Z,28,"Stack from ghstack (oldest at bottom): -> #186808 set_name() updated an entry's name and codegen but left it in V.kernel.range_tree_nodes under its old symbol (the dict is keyed by symbol(name)). External-template epilogue fusion then looked up the renamed symbol, missed, and failed codegen with: InductorError: Asse..."
178485,[DO NOT MERGE][Inductor] Migrate oneDNN Inner Product to oneDNN MatMul for MKLDNN linear ops,https://github.com/pytorch/pytorch/pull/178485,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,13,0,,,"module: cpu, module: mkldnn, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-03-26T06:53:08Z,2026-06-10T12:04:13Z,28,This PR is rebased based on the following two PRs: #147855 #147360 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @gujinghui @PenghuiCheng @jianyuh @min-jean-cho @yanbing-j @Guobing-Chen @Xia-Weiwen @snadampal @voznesenskym @penguinwu @EikanWang @zhuhaozhe @blzheng @wenzh...
178545,[MPS] Add native ops fast path for SDPA on long sequences (~2x speedup),https://github.com/pytorch/pytorch/pull/178545,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, Stale, release notes: mps",2026-03-26T21:54:36Z,2026-06-10T11:45:28Z,28,"Summary F.scaled_dot_product_attention on MPS is ~2x slower than equivalent manual at::matmul + at::softmax for sequences longer than 8 tokens. This PR adds a fast path that bypasses the MPSGraph-based sdpa_general_mps and calls native MPS ops directly, eliminating graph compilation/dispatch overhead. Root Cause For..."
178100,docs: mention torch.cuda.memory.mem_get_info in CUDA memory management notes,https://github.com/pytorch/pytorch/pull/178100,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,174625,174625,174625,,,,0,2,0,,,"triaged, open source, release notes: cuda, topic: docs, topic: not user facing",2026-03-22T17:46:11Z,2026-06-10T08:17:44Z,28,"Fixes #174625 Adds a brief description of torch.cuda.memory.mem_get_info to the CUDA memory management section of the notes, placed alongside the other memory monitoring functions. The new paragraph: Explains that mem_get_info queries free/total GPU memory from the driver (outside PyTorch's caching allocator) Notes..."
186590,[Draft][xpu] Add NVFP4 support on Intel GPU,https://github.com/pytorch/pytorch/pull/186590,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: cpu, open source, module: inductor, ciflow/inductor, ciflow/xpu, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-06-08T08:47:36Z,2026-06-10T08:10:27Z,28,"This PR adds NVFP4 (Float4_e2m1fn_x2) scaled_mm support for XPU, following the pattern established by the Float8 blockwise and mirroring the CUDA NVFP4 implementation. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng..."
183852,ci: declare workflow-level `contents: read` on 8 workflows,https://github.com/pytorch/pytorch/pull/183852,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-15T05:41:52Z,2026-06-10T08:09:37Z,28,Pins the default GITHUB_TOKEN to contents: read on 8 workflows in .github/workflows/ that don't call a GitHub API beyond the initial checkout. The following files were left implicit because they reference GITHUB_TOKEN / use a write-scope action / trigger on pull_request_target. Those scopes are best declared by main...
180087,[DTensor] Error on random ops with device mismatch,https://github.com/pytorch/pytorch/pull/180087,draft/noise,13,"PR is draft; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,180088,,,,0,1,0,,,"Stale, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-04-10T23:51:04Z,2026-06-10T07:55:25Z,28,"Stack from ghstack (oldest at bottom): -> #180087 error on buggy case: #180088 we can have DTensor with cuda mesh and cpu local tensor, for example mesh = init_device_mesh(""cuda"", (world_size,), mesh_dim_names=(""dp"",)) with torch.device(""meta""): t = torch.empty(256, 256) dt = DTensor.from_local(t, mesh, [Shard(0)])..."
178805,[openreg] Add comprehensive test documentation for OpenReg backend test suite,https://github.com/pytorch/pytorch/pull/178805,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,172324,172324,172324,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing",2026-03-30T20:02:49Z,2026-06-10T07:55:21Z,28,"Documents the standardized testing patterns, skip/expectedFailure conventions, assertion idioms, serialization and profiler patterns, device-generic test instantiation, and contributor workflow for adding new tests to the OpenReg (PrivateUse1) extension test suite. Fixes #172324"
186882,RFC Change error order,https://github.com/pytorch/pytorch/pull/186882,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186880,,,0,2,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-10T03:36:56Z,2026-06-10T06:44:40Z,28,Stack from ghstack (oldest at bottom): -> #186882 #186880 #186876 Before: After:
184334,[CUDA] Enable bf16 grouped_mm CUTLASS fast path on SM12x and Windows,https://github.com/pytorch/pytorch/pull/184334,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,170955,170955,170955,,,,0,2,0,,,"triaged, open source, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-19T08:05:59Z,2026-06-10T06:44:23Z,28,"Summary Fixes silent CPU grouping fallback for torch.nn.functional.grouped_mm on consumer Blackwell (SM12.x) and enables the CUTLASS bf16 grouped GEMM fast path on Windows. _grouped_mm_cuda now allows the CUTLASS fast path when the CUDA device CC major is 9–12 (Hopper through SM12.x), instead of only majors 9 and 10..."
186010,Surface ROCm GPU device faults in TorchElastic signal-failure messages (#186010),https://github.com/pytorch/pytorch/pull/186010,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, fb-exported, release notes: distributed (torchelastic), meta-exported, ciflow/torchtitan",2026-06-02T23:30:50Z,2026-06-10T06:43:26Z,28,"Summary: A worker killed by a signal (e.g. SIGABRT from a GPU fault) produces no Python traceback, so the propagated ChildFailedError shows only ""Signal N (SIG...)"". On AMD the real cause is hidden: ROCr aborts on the fault (""Memory access fault by GPU"" / HSA_STATUS_ERROR_*) before any catchable error is raised, whe..."
183389,Port 7 tests under test/distributed folder to XPU,https://github.com/pytorch/pytorch/pull/183389,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,12,0,,,"triaged, open source, topic: not user facing, ciflow/inductor",2026-05-12T00:39:16Z,2026-06-10T06:08:23Z,28,"For #114850, we will port distributed tests to Intel GPU. We will enable Intel GPU with following methods and keep the original code styles: Example: ""torch.accelerator.current_accelerator()"" to determine the accelerator backend enabled XPU for the following files: test/distributed/test_dynamo_distributed.py test/di..."
186209,Unify triton-xpu CD nightly build into 'triton' package with version suffix,https://github.com/pytorch/pytorch/pull/186209,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, release notes: releng",2026-06-04T08:54:08Z,2026-06-10T06:01:31Z,28,Summary Unifies triton XPU wheel build under the single package name triton (previously triton-xpu) Differentiates XPU backend via PEP 440 local version identifier: +xpu suffix Uses triton's TRITON_WHEEL_VERSION_SUFFIX env var to inject the backend suffix into the wheel version CUDA and ROCm builds are unchanged (tr...
178960,profiler: fix USE_KINETO=OFF build failure (unconditional ActivityType.h include),https://github.com/pytorch/pytorch/pull/178960,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,178939,,,,0,3,0,,,"triaged, open source, Stale, topic: not user facing",2026-04-01T02:39:53Z,2026-06-10T05:03:49Z,28,"When building with USE_KINETO=OFF, kineto_shim.h was unconditionally including <ActivityType.h>, a Kineto-only header, causing a fatal compile error on systems where Kineto is not installed (e.g. Gentoo with system libraries instead of bundled third-party). The fix wraps the include in #ifdef USE_KINETO and provides..."
179400,Use std::in_range for integer range checking,https://github.com/pytorch/pytorch/pull/179400,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries, ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,"176662, 186868",,,,0,2,0,,,"triaged, open source, Merged, Reverted, ciflow/binaries, ciflow/trunk, topic: not user facing, ci-no-td, ciflow/vllm",2026-04-05T07:21:27Z,2026-06-10T05:02:37Z,28,"C++20 provides std::in_range which allows checking if an integer value can fit in a given integer type. This PR replaces such checks with in_range. std::in_range is safer than direct comparisons because it avoids implicit casts. The call sites to update were obtained by grepping for ""numeric_limits.*min"". PyTorch su..."
186852,[profiler][cupti] ProfilerObserver + record_function annotation routing,https://github.com/pytorch/pytorch/pull/186852,draft/noise,13,PR is draft,wait for author to mark ready,,CLEAN,True,0,0,,,,,,186655,,,0,1,0,,,topic: not user facing,2026-06-09T23:55:38Z,2026-06-10T03:28:56Z,29,Stack from ghstack (oldest at bottom): #186655 #186802 #186812 -> #186852 #186439 #186811 #186438 #186437 #186436 Splits the consumer half out of the CUPTI multiplexer engine (the parent PR). The engine is one shared CUPTI subscription that demuxes decoded columns to registered observers; this PR adds the observer a...
186851,[profiler][cupti] ProfilerObserver + record_function annotation routing,https://github.com/pytorch/pytorch/pull/186851,draft/noise,13,PR is draft,wait for author to mark ready,,CLEAN,True,0,0,,,,,,186655,,,0,1,0,,,,2026-06-09T23:35:35Z,2026-06-10T02:16:15Z,29,Stack from ghstack (oldest at bottom): #186655 #186802 #186812 -> #186851 #186439 #186811 #186438 #186437 #186436 Splits the consumer half out of the CUPTI multiplexer engine (the parent PR). The engine is one shared CUPTI subscription that demuxes decoded columns to registered observers; this PR adds the observer a...
186835,[RFC]: Auto-select and run Operator Performance Benchmarks from PR Diffs,https://github.com/pytorch/pytorch/pull/186835,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, topic: not user facing",2026-06-09T20:38:25Z,2026-06-10T02:01:33Z,29,"RFC: Auto-select and run Operator Performance Benchmarks from PR Diffs Summary This RFC explores automatically mapping code changes in a PR to the operator performance benchmarks that should run against those changes. Today this mapping does not exist. The benchmarks exist, the comparison workflow exists, but nothin..."
180081,[DTensor] Fix GIL-unsafe thread-local py::object in DTensor dispatch logging,https://github.com/pytorch/pytorch/pull/180081,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Stale, meta-exported",2026-04-10T22:41:21Z,2026-06-10T01:27:12Z,29,"Summary: D92459325 introduced thread_local py::object variables in python_variable.cpp to cache Python logger objects for DTensor shard prop cache logging. When threads exit (e.g., in tests using spawn_threads_and_init_comms), C++ destroys these thread-locals, but the GIL is not held at thread teardown time. py::obj..."
184459,Decompose supported aten._trilinear in Inductor,https://github.com/pytorch/pytorch/pull/184459,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,105560,105560,,,,0,1,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan, agentic, needs-review",2026-05-20T00:50:01Z,2026-06-10T00:11:38Z,29,"Stack from ghstack (oldest at bottom): -> #184459 Add an Inductor-only conditional decomposition for aten._trilinear so broadcast-compatible floating and complex cases can fuse, while unsupported dtype and shape cases continue through the native fallback path. Fixes #105560 Generated by my agent cc @voznesenskym @pe..."
186831,Improve cudagraph stack trace coverage and assertions,https://github.com/pytorch/pytorch/pull/186831,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186828, 186829, 186830",,,0,2,0,,,"module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-09T19:42:54Z,2026-06-10T00:10:48Z,29,"Stack from ghstack (oldest at bottom): -> #186831 #186830 #186829 #186828 Cherry-pick the cudagraph stack trace coverage work from PR 177890 onto the CUDA graph annotation stack. The patch snapshots output stack traces before graph passes can strip them, uses first-user stack traces for placeholder outputs saved for..."
182712,[MPS] Enhance Linear tensor ops to avoid unnecessary .contiguous() calls on weight tensor,https://github.com/pytorch/pytorch/pull/182712,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,181946,183759,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-06T20:59:58Z,2026-06-09T23:57:34Z,29,Summary: Removing .contiguous() call on weight tensor as it goes through Placeholder which natively makes tensor contiguous on any device running macOS < 15.0 without strided tensor support. Performance Analysis: Benchmarked using following script: import torch import torch.nn.functional as F import torch.utils.benc...
185989,Bump onnx to 1.22.0rc1 to integrate ONNX 1.22.0 release branch,https://github.com/pytorch/pytorch/pull/185989,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/docker,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,185955,185955,,,,186969,0,5,0,,,"open source, topic: not user facing, ciflow/docker",2026-06-02T20:26:42Z,2026-06-09T23:44:57Z,29,"Integrate the upcoming ONNX 1.22.0 release by bumping the pinned dependency to the latest available release candidate, so CI can surface any incompatibilities ahead of the final ONNX 1.22.0 release. Summary Bump onnx==1.21.0 → onnx==1.22.0rc1 in .ci/docker/requirements-ci.txt. Merged latest main to integrate recent..."
186829,Auto-promote CUDA graph static input copies,https://github.com/pytorch/pytorch/pull/186829,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186828, 186830, 186831",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-09T19:42:43Z,2026-06-09T23:44:06Z,29,"Stack from ghstack (oldest at bottom): #186831 #186830 -> #186829 #186828 Add a runtime option that promotes repeatedly changing static input addresses to copy semantics for a CUDA graph function. The tree tracks static-input address churn by function and parent node, records promoted indices into the wrapped functi..."
186830,Add CUDA graph disable regions,https://github.com/pytorch/pytorch/pull/186830,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186828, 186829, 186831",,,0,1,0,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-09T19:42:48Z,2026-06-09T23:44:04Z,29,"Stack from ghstack (oldest at bottom): #186831 -> #186830 #186829 #186828 Add torch.compiler.cudagraph_disable() as a metadata-based region annotation. The context manager emits begin/end marker ops during tracing, the proxy-tensor metadata pass marks traced nodes inside the region as cudagraph-unsafe and erases the..."
186828,Add CUDA graph output and input copy annotations,https://github.com/pytorch/pytorch/pull/186828,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186829, 186830, 186831",,,0,1,0,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-09T19:42:37Z,2026-06-09T23:44:00Z,29,"Stack from ghstack (oldest at bottom): #186831 #186830 #186829 -> #186828 Add explicit torch.compiler annotations for CUDA graph output cloning and input copy semantics, then thread the metadata through AOTAutograd, FX cache hashing, Inductor output code, and CUDA graph trees. Output clone preserves output alias gro..."
186636,c10/core/impl/SizesAndStrides.h: add [[nodiscard]] to size() and isInline(),https://github.com/pytorch/pytorch/pull/186636,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, meta-exported",2026-06-08T19:24:51Z,2026-06-09T23:43:28Z,29,Summary: Add [[nodiscard]] to two methods in c10::impl::SizesAndStrides: size() — the public dimensionality accessor; discarding it is always a logic error isInline() — private helper that gates the entire inline vs. out-of-line storage path; discarding it silently would corrupt subsequent pointer arithmetic Test Pl...
184155,Fix Inductor bool vector mask casts,https://github.com/pytorch/pytorch/pull/184155,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,146306,146306,,,,0,4,0,,,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ci-no-td, ciflow/torchtitan, agentic",2026-05-18T04:04:11Z,2026-06-09T23:14:06Z,29,"Stack from ghstack (oldest at bottom): -> #184155 Teach C++ codegen to route bool vector mask casts through an Inductor helper that supports both VecMask and raw Vectorized values, and add focused compile/codegen coverage for the regression. Fixes #146306 Generated by my agent cc @voznesenskym @penguinwu @EikanWang..."
184114,Add cpp wrapper fallback alignment assertions,https://github.com/pytorch/pytorch/pull/184114,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,151027,151027,,,,0,5,0,,,"Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, ci-no-td, release notes: inductor (aoti), ciflow/torchtitan, agentic",2026-05-17T19:31:59Z,2026-06-09T23:13:58Z,29,Stack from ghstack (oldest at bottom): -> #184114 Generate fallback output alignment assertions through the wrapper abstraction so cpp_wrapper emits native AOTI runtime checks alongside existing size/stride checks. Add CPU cpp_wrapper coverage for generated checks and incorrect-meta runtime failures. Fixes #151027 G...
178277,Skip fake autocast tests for `linalg.pinv` and `pinverse` on XPU,https://github.com/pytorch/pytorch/pull/178277,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing, ciflow/xpu",2026-03-24T15:05:10Z,2026-06-09T23:09:15Z,29,"Add fake_autocast_device_skips entry for XPU, aligned with the existing CUDA and CPU behavior. FakeTensor cannot correctly model autocast dtype promotion through at::matmul (called internally by linalg_pinv and pinverse), producing float32 output while real autocast produces float16. CUDA already skips these tests (..."
186138,Support data-dependent 0-dim tensor indexing under tracing,https://github.com/pytorch/pytorch/pull/186138,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186139,,,0,4,0,,,topic: not user facing,2026-06-03T21:31:55Z,2026-06-09T23:08:41Z,29,"Stack from ghstack (oldest at bottom): #186139 -> #186138 Indexing a fixed-size dim with a 0-dim integer tensor whose value is data-dependent (e.g. table[:, t] where t's value is unbacked) raised a GuardOnDataDependentSymNode under torch.compile, even though the output shape does not depend on the index value. Two c..."
186798,[TESTING ONLY] [ROCm] testing heuristic config approach,https://github.com/pytorch/pytorch/pull/186798,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/inductor-rocm-mi355, ciflow/rocm-mi355","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, module: inductor, ciflow/inductor, ciflow/rocm-mi355, ciflow/inductor-perf-test-nightly-rocm-mi350, ciflow/inductor-rocm-mi355, ciflow/torchtitan",2026-06-09T15:17:17Z,2026-06-09T23:08:29Z,29,Generating CI results for experimental method cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchule...
186139,Support closed-over data-dependent indexing in scan (forward and autograd),https://github.com/pytorch/pytorch/pull/186139,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186138,,,0,7,0,,,,2026-06-03T21:32:00Z,2026-06-10T15:56:56Z,28,"Stack from ghstack (oldest at bottom): -> #186139 #186138 A scan combine_fn that closes over a tensor and indexes it with the per-step integer xs element (e.g. table[:, t], a common JAX idiom in HMMs and decoding loops) previously failed, while jax.lax.scan handles it. This builds on data-dependent 0-dim tensor inde..."
182546,Fix sparse compressed set_data device metadata,https://github.com/pytorch/pytorch/pull/182546,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,136258,136258,136258,,,,0,3,0,,,"triaged, open source",2026-05-05T18:59:52Z,2026-06-09T22:15:55Z,29,"Fixes #136258 This adds a SparseCsrTensorImpl::shallow_copy_from override so sparse compressed tensors copy their internal component tensors during compatible shallow-copy assignment. nn.Module.to() can route parameter updates through param.data = param_applied. For sparse compressed tensors, the base TensorImpl::sh..."
185189,[Test] Make binary_ufuncs file device agnostic for out-of-tree backends,https://github.com/pytorch/pytorch/pull/185189,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,8,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-05-26T04:19:44Z,2026-06-09T21:20:23Z,29,Remove device restrictions that prevented out-of-tree backends from running the binary_ufuncs tests: @onlyNativeDeviceTypes decorator is removed from the below tests: test_out_resize_warning test_inplace_dunders test_complex_scalar_pow_tensor test_pow_scalar_type_promotion test_div_and_floordiv_script_vs_python test...
186541,[MPS] [WIP] Port all MPSGraph matmuls to Metal,https://github.com/pytorch/pytorch/pull/186541,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,16,0,,,"open source, topic: improvements, release notes: mps, ciflow/mps",2026-06-07T12:18:34Z,2026-06-09T19:15:22Z,29,"Pretty big PR, putting up a draft here to track the progress. Advantages of this are listed as speedups below. Apart from this, we will be able to, in further PRs, to merge epilogue activation functions in matmuls which should bring more speedups to MPS torch compile. Currently this is the result on M5 Pro: speedup..."
174559,[lint] Enable Ruff PYI056 by replacing __all__.append/extend usage,https://github.com/pytorch/pytorch/pull/174559,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,110950,110950,,,,,0,2,0,,,"module: lint, triaged, open source, Stale, release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-02-08T18:33:54Z,2026-06-09T15:34:39Z,29,Summary remove PYI056 from the Ruff ignore list in pyproject.toml replace __all__.append(...) / __all__.extend(...) with slice-based list updates in affected modules keep behavior unchanged while making __all__ updates compatible with flake8-pyi's PYI056 Resolves #110950 Validation python3 -m ruff check --config pyp...
186177,[autograd] Defer device thread creation when multithreading is disabled,https://github.com/pytorch/pytorch/pull/186177,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184783,184783,184783,,,,0,1,0,,,"module: autograd, triaged, open source, topic: not user facing",2026-06-04T03:14:57Z,2026-06-09T15:31:17Z,29,"Summary Fixes #184783 Root cause: Engine::initialize_device_threads_pool() used a function-local static bool lambda that unconditionally called start_device_threads() on the very first backward — without checking get_multithreading_enabled(). Additionally, start_device_threads() bundled two unrelated concerns: alloc..."
185507,Pre-compute igates for XPU LSTM to reduce per-step kernel launches,https://github.com/pytorch/pytorch/pull/185507,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cpu, triaged, open source",2026-05-28T12:05:35Z,2026-06-09T14:02:36Z,29,"Summary On XPU, batch-compute input gates (linear_ih) for the entire sequence upfront, then pass pre-computed results to _thnn_fused_lstm_cell per time step. This reduces kernel launch overhead from O(T) GEMM launches to a single large GEMM. Correctness Standard LSTM formula per time step: gates_t = W_ih @ x_t + b_i..."
178242,[XPU] Add grouped_mm dispatch for XPU via sycl-tla,https://github.com/pytorch/pytorch/pull/178242,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, triaged, open source, ciflow/trunk, topic: not user facing, ciflow/xpu",2026-03-24T05:56:34Z,2026-06-09T13:01:20Z,29,"Add XPU dispatch key for _grouped_mm in native_functions.yaml and implement _grouped_mm_xpu dispatch function that routes BF16 grouped GEMM to the sycl-tla kernel in torch-xpu-ops, with fallback for other dtypes. Add XPU: _grouped_mm_xpu dispatch in native_functions.yaml Add GroupedBlas.cpp with XPU dispatch impleme..."
182013,Rename vec256_tests to vec_tests and fix executable search,https://github.com/pytorch/pytorch/pull/182013,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, ciflow/trunk, topic: not user facing, ci-no-td",2026-04-30T11:27:08Z,2026-06-09T10:11:09Z,29,"It looks like #58438 renamed vec265 to vec, which appears that it would cause vec256_tests=$(find . -maxdepth 1 -executable -name 'vec256_test*') to return empty in test.sh. So it looks like we may not have been running any vectorized c++ tests in CI for around 5 years. Stack from ghstack (oldest at bottom): -> #182..."
176600,Fix 1C1W tensor's shape and stride,https://github.com/pytorch/pytorch/pull/176600,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, Stale, ciflow/trunk, release notes: fx",2026-03-05T13:19:12Z,2026-06-09T07:32:32Z,29,The comment mistakenly wrote H instead of W. This commit fixs it.
186634,c10/util/intrusive_ptr.h: replace hand-written comparison operators with operator<=>,https://github.com/pytorch/pytorch/pull/186634,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, ciflow/vllm, meta-exported",2026-06-08T19:23:05Z,2026-06-09T06:20:33Z,29,"Summary: Replace the 9 hand-written free-function comparison operators for intrusive_ptr and weak_intrusive_ptr (operator<, operator==, operator!=, and the nullptr overloads) with C++20 operator<=> + operator==. The spaceship operator synthesizes <, >, <=, >= automatically, and the retained operator== synthesizes !=..."
179882,[Devmate Tasks][Butterfly] Fix flaky fbcode//caffe2/test/distributed/tensor:dtensor_ops - test_single_dim_strategy_i0_cpu_float (#179882),https://github.com/pytorch/pytorch/pull/179882,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,179882,,,,,0,1,0,,,"fb-exported, Stale, ciflow/trunk, topic: not user facing, meta-exported, ciflow/dtensor",2026-04-10T01:08:32Z,2026-06-09T05:53:40Z,29,Summary: Pull Request resolved: #179882 Reviewed By: weifengpy Differential Revision: D99766883
185101,Drop redundant tuple/list/int wrappers around tensor shape and numel,https://github.com/pytorch/pytorch/pull/185101,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: quantization, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-25T03:13:40Z,2026-06-09T05:23:49Z,29,Drop redundant tuple(t.shape) / list(t.shape) wrappers and int(t.numel()) / int(t.size()) / int(t.dim()) casts across torch/. Both are no-ops on concrete tensors: torch.Size is already an immutable tuple subclass that compares equal to plain tuples and is SymInt-aware. Wrapping it in tuple() / list() discards that a...
185668,Fix typos in comments and docstrings across torch,https://github.com/pytorch/pytorch/pull/185668,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: amp (automated mixed precision), ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-29T21:20:58Z,2026-06-09T04:14:09Z,29,"Stack from ghstack (oldest at bottom): -> #185668 Correct spelling mistakes in comments and docstrings across dynamo, inductor, autograd, distributed, JIT, nn, and utility modules. These are comment-only changes with no functional impact. Authored with Claude (typo_terminator2). cc @mcarilli @ptrblck @leslie-fang-in..."
186607,Raise a clear error for malformed allocator config instead of INTERNAL_ASSERT,https://github.com/pytorch/pytorch/pull/186607,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,186453,186453,186453,,,,0,1,0,,,"triaged, open source",2026-06-08T14:43:48Z,2026-06-09T03:28:06Z,30,"Fixes #186453 Summary import torch crashes with INTERNAL ASSERT FAILED ... Index out of bounds in ConfigTokenizer ... please report a bug to PyTorch when a malformed allocator config is set, for example PYTORCH_CUDA_ALLOC_CONF=expandable_segments (a key with no value, which the reporter in #186453 hit). ConfigTokeni..."
184585,[MPS] Wire SDPBackend dispatch for scaled_dot_product_attention,https://github.com/pytorch/pytorch/pull/184585,ready for maintainer decision,45,"CI lane labels: ciflow/inductor, ciflow/mps; open non-draft PR has no obvious automated blocker",review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,9,0,,,"triaged, open source, topic: improvements, release notes: mps, ciflow/mps, module: inductor, ciflow/inductor, release notes: inductor (aoti)",2026-05-20T18:08:21Z,2026-06-09T01:51:15Z,30,Wire SDPBackend dispatch for MPS SDPA: existing prefill and vector decode kernels are now reachable via flash_attention and efficient_attention (previously always math). This is to lay base for future work of migrating attention from MPS Graph. Matches CUDA (where applicable) cc @voznesenskym @penguinwu @EikanWang @...
186251,Wait for Docker images before starting builds,https://github.com/pytorch/pytorch/pull/186251,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"ciflow/trunk, topic: not user facing",2026-06-04T18:39:08Z,2026-06-09T00:18:59Z,30,"Add reusable _wait-for-docker-image.yml workflow that polls ECR for a container image manifest, with configurable timeout and poll interval Add wait-for-image-* jobs in trunk.yml for each distinct Docker image used by Linux build jobs (CUDA, ROCm, inductor, clang18, aarch64) Wire build jobs to depend on their corres..."
186609,[user-streams][aoti] Add minimal AOTI user stream support,https://github.com/pytorch/pytorch/pull/186609,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-06-08T14:51:34Z,2026-06-08T18:33:12Z,30,"Summary: This implements the MVP AOTI path for explicit user streams by inlining record_event and wait_event in the generated C++ wrapper, allocating per-thread cached side streams/events in a generated helper header, preserving scheduler stream metadata across grouped/fused nodes, and rejecting unsupported stream s..."
176234,[autograd] Explicitly error when layer_norm computes 3rd order derivatives,https://github.com/pytorch/pytorch/pull/176234,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,175754,,,,,0,35,0,,,"Merged, Reverted, ciflow/trunk, release notes: autograd, topic: bug fixes, no-stale, ci-no-td",2026-03-02T23:00:28Z,2026-06-08T18:20:19Z,30,"Stack from ghstack (oldest at bottom): #177806 #177070 #174079 -> #176234 The saved mean and invstd tensors passed to layer_norm_double_backward have no autograd history connecting them to the input (they come from the forward kernel). The double backward formula analytically accounts for their dependency on input,..."
184559,[Test]Refactor test/test_reductions.py to be device generic,https://github.com/pytorch/pytorch/pull/184559,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,6,0,,,"triaged, open source, topic: not user facing",2026-05-20T15:27:13Z,2026-06-08T17:31:19Z,30,"Replace @OnlyOn([""cuda"", ""xpu""]) with @onlyAccelerator for 12 tests to enable execution on all accelerators including out of tree devices via PrivateUse1. Changes in test/test_reductions.py: Add import: onlyAccelerator from common_device_type Replace @OnlyOn([""cuda"", ""xpu""]) with @onlyAccelerator for 12 tests: test_..."
186207,Fix segfault and invalid cast in UntypedStorage,https://github.com/pytorch/pytorch/pull/186207,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,169210,169210,169210,,,,0,1,0,,,"triaged, open source",2026-06-04T08:49:56Z,2026-06-08T14:19:24Z,30,"Summary Fixes two error-checking gaps reported in #169210 where invalid usage of UntypedStorage caused a segfault rather than a clear Python error. Fix 1 — Segfault on invalid allocator= argument (c10/, torch/csrc/) Passing tensor.data_ptr() as allocator= to UntypedStorage() caused a segfault because the integer was..."
186458,Fix SDPA backend detection for autograd cache,https://github.com/pytorch/pytorch/pull/186458,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor",2026-06-06T09:50:54Z,2026-06-08T14:10:03Z,30,"The choice of backend for scaled_dot_product_attention is currently ignored by the Autograd compile cache. This means that after compiling once and then switching backends (e.g. FLASH_ATTENTION to MATH), torch will continue to use the originally selected kernel until the on-disk cache is cleared. This PR adds the va..."
186568,Add expected-failure test for ConstantLR factor=0,https://github.com/pytorch/pytorch/pull/186568,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,168044,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-07T20:32:01Z,2026-06-08T14:08:52Z,30,"This PR adds an expected-failure test documenting the current behavior of ConstantLR when factor=0. ConstantLR uses factor to scale the learning rate during the constant phase and later restores the original learning rate by applying the inverse scaling. When factor=0 is provided, the scheduler can be constructed su..."
186307,Add contains() to ska::flat_hash_map/set,https://github.com/pytorch/pytorch/pull/186307,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-05T02:17:19Z,2026-06-08T14:06:30Z,30,We need contains to eliminate clang-tidy warnings.
186389,[BE][Ez]: Update string_util APIs to use std::string_view more,https://github.com/pytorch/pytorch/pull/186389,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source",2026-06-05T17:32:14Z,2026-06-08T14:04:31Z,30,
186519,c10/util/intrusive_ptr.h: add [[nodiscard]] to query methods,https://github.com/pytorch/pytorch/pull/186519,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"ciflow/trunk, meta-exported",2026-06-07T05:50:03Z,2026-06-08T13:55:27Z,30,"Summary: Add [[nodiscard]] to the pure-query methods on both intrusive_ptr and weak_intrusive_ptr. Silently discarding the result of get(), defined(), use_count(), weak_use_count(), unique(), is_uniquely_owned(), or expired() is always a logic error. The get() annotation is particularly important — calling ptr.get()..."
186565,Support parametrized modules in functional_call,https://github.com/pytorch/pytorch/pull/186565,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-07T19:37:37Z,2026-06-08T13:51:20Z,30,"Support parametrized modules in functional_call Previously, when a module had active parametrizations registered on a parameter or buffer, calling functional_call with the regular name of the parameter (e.g. 'weight') did not disable the parametrization. This contradicts the documentation which states that passing t..."
186559,Remove deprecated pytree_to_str and str_to_pytree functions,https://github.com/pytorch/pytorch/pull/186559,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-07T18:34:10Z,2026-06-08T13:41:45Z,30,Remove deprecated pytree_to_str and str_to_pytree functions These functions were deprecated in favor of treespec_dumps and treespec_loads respectively. There are no usages of these deprecated functions anywhere in the PyTorch repository. Test Plan: Run the pytree unit tests: python test/test_pytree.py Disclosed: Thi...
186473,[accelerator] Support multiple private backends,https://github.com/pytorch/pytorch/pull/186473,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,"133288, 180345",,,,0,1,0,,,"triaged, open source, module: amp (automated mixed precision), release notes: python_frontend",2026-06-06T15:50:24Z,2026-06-08T13:36:52Z,30,"Issue Addresses #180345 Related to #133288 Summary This Draft PR implements an incremental approach for supporting multiple out-of-tree accelerator backends. Instead of supporting an arbitrary number of private backends, it exposes PrivateUse2 and PrivateUse3 alongside PrivateUse1. Supporting an arbitrary number wou..."
186239,[ROCm][CI] Add linux.rocm.mi350.docker-cache to docker-cache-rocm,https://github.com/pytorch/pytorch/pull/186239,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"module: rocm, triaged, open source, topic: not user facing",2026-06-04T15:58:23Z,2026-06-08T13:34:18Z,30,Summary Add linux.rocm.mi350.docker-cache to the docker-cache-rocm runner matrix alongside MI210. Test plan workflow_dispatch docker-cache-rocm on this branch with branch=main and a recent docker-builds run_id Authored w/ assistance from cursor cc @jeffdaily @sunway513 @pruthvistony @ROCmSupport @jataylo @hongxiayan...
186205,[dynamo] Add as_strided_ handler to fix metadata propagation (#185888),https://github.com/pytorch/pytorch/pull/186205,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185888,185888,185888,,,,0,10,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-04T08:30:54Z,2026-06-08T13:32:56Z,30,"Fixes #185888 Problem x.as_strided_() under torch.compile failed to propagate mutated size/stride metadata because var_getattr returned a DelayGraphBreakVariable (since as_strided_ is tagged inplace_view in ATen), blocking the call before call_method could dispatch it. Fix var_getattr: Bypass the inplace_view → Dela..."
182147,Support XPU and other accelerators in RNG state handling,https://github.com/pytorch/pytorch/pull/182147,draft/noise,23,"PR is draft; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk",wait for author to mark ready,,UNKNOWN,True,0,0,,,1970,1970,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor",2026-05-01T17:35:58Z,2026-06-08T11:40:17Z,30,"Fixes intel/torch-xpu-ops#1970 Refactors hardcoded CUDA-specific RNG state helpers to dynamically resolve the active accelerator (CUDA, XPU, etc.) via torch.accelerator. This fixes test_dropout failures on XPU when using AOTAutograd/Inductor by replacing torch.cuda calls with device-agnostic equivalents. Changes: In..."
186586,Propagate override_cudagraphs annotation across CM-lifecycle gaps,https://github.com/pytorch/pytorch/pull/186586,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-06-08T07:29:48Z,2026-06-08T10:29:44Z,30,"Summary: torch._dynamo.override_cudagraphs stores its annotation on the active OutputGraph during symbolic execution. The annotation only takes effect if that OG reaches compile_subgraph's finalize line, where it is copied into gm.meta[""cudagraph_annotation""] and consumed downstream by Inductor's create_compiler_con..."
186575,Update FindMKLDNN CMake configuration,https://github.com/pytorch/pytorch/pull/186575,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: mkldnn, open source",2026-06-08T02:37:53Z,2026-06-08T03:07:22Z,31,"This pull request makes a small but important update to the MKLDNN (oneDNN) CMake configuration. The change ensures that the CMAKE_BUILD_TYPE is explicitly passed when configuring the MKLDNN build, which helps maintain consistent build types (such as Debug or Release) across the project and its dependencies. CMake c..."
186334,Fix torch.kron crash on non-contiguous inputs,https://github.com/pytorch/pytorch/pull/186334,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend",2026-06-05T06:18:45Z,2026-06-06T19:08:10Z,32,"This commit fixes a RuntimeError where torch.kron crashes on non-contiguous inputs with ""view size is not compatible with input tensor's size and stride"". Root Cause: The implementation was using at::_unsafe_view which requires the tensor to be contiguous or viewable without copying. When passed a non-contiguous ten..."
172829,Fixes registration of custom sharding strategies when using register_sharding with TensorList,https://github.com/pytorch/pytorch/pull/172829,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,167435,167435,,,,,0,1,0,,,"triaged, open source, Merged, Reverted, release notes: distributed (dtensor), ci-no-td",2026-01-20T15:47:40Z,2026-06-06T15:09:53Z,32,"Fixes #167435 Problem DTensor dispatch fails with ""Cannot find device mesh"" when using register_sharding with operators that have TensorList kwargs like out=[tensor1, tensor2]. Lists of DTensors in kwargs were stored directly instead of being converted to DTensorSpecs, causing nested dispatch failures during shape p..."
185187,Validate forward-mode autograd.functional vectorize usage,https://github.com/pytorch/pytorch/pull/185187,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184842,184842,184842,,,,0,1,0,,,"triaged, open source, release notes: autograd",2026-05-26T03:10:57Z,2026-06-05T21:45:21Z,33,"Issue Fixes #184842 Summary Reject torch.autograd.functional.jacobian(..., strategy=""forward-mode"", vectorize=False) with a ValueError before evaluating the user function. This also adds the same public validation for torch.autograd.functional.hessian(..., outer_jacobian_strategy=""forward-mode"", vectorize=False) and..."
186038,add distortion benchmark,https://github.com/pytorch/pytorch/pull/186038,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,19,0,,,topic: not user facing,2026-06-03T04:02:15Z,2026-06-05T21:12:04Z,33,Stack from ghstack (oldest at bottom): -> #186038 #186295 Add an in-tree benchmark harness for the experimental CUPTI monitor backend. The harness exercises the same multistream mixed workloads used during prototype development and can compare always-on raw dumping and profiled window distortion against the stock pr...
186378,[activation_checkpointing] Support AutoNamingMode + policy name keys under torch.compile,https://github.com/pytorch/pytorch/pull/186378,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"186277, 186382",,,0,4,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-05T14:48:02Z,2026-06-05T19:51:42Z,33,"Stack from ghstack (oldest at bottom): -> #186378 #186382 #186277 Follow-up to the previous PR, which added AutoNamingMode and name / module-FQN keys for checkpoint's policy= in eager. This PR makes both work under torch.compile. The names come from different sources in the two worlds, normalized to a common root-re..."
186382,[activation_checkpointing] AutoNamingMode + name/module-FQN keys in checkpoint policy= (eager),https://github.com/pytorch/pytorch/pull/186382,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"186277, 186378",,,0,3,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-05T15:19:19Z,2026-06-05T19:51:35Z,33,"Stack from ghstack (oldest at bottom): #186378 -> #186382 #186277 Introduces AutoNamingMode, a public TorchDispatchMode in torch.utils.checkpoint that auto-names each op output {fqn}_{op}_{count} (module-qualified, via ModuleTracker forward hooks) and exposes a names registry (tensor -> name). AutoNamingMode stays o..."
185821,[BE][Ez]: Add c10::SmallVector reserve calls and simplify log func calls,https://github.com/pytorch/pytorch/pull/185821,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/mps,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"module: cpu, triaged, open source, release notes: quantization, release notes: linalg_frontend, ciflow/mps",2026-06-01T15:32:19Z,2026-06-05T19:51:27Z,33,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @aditew01
185649,[Test] Make test_multihead_attention device generic for out-of-tree accelerators,https://github.com/pytorch/pytorch/pull/185649,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: tests/core",2026-05-29T17:51:18Z,2026-06-05T19:09:58Z,33,"Summary Refactors test/nn/test_multihead_attention.py to break the tight coupling between test cases and specific hardware backends, enabling new accelerator backends to reuse these tests automatically. Remove dead TEST_CUDA guard (_do_cuda_memory_leak_check, _do_cuda_non_default_stream) from TestMultiheadAttentionN..."
186277,[activation_checkpointing] Add policy= kwarg and name() to torch.utils.checkpoint,https://github.com/pytorch/pytorch/pull/186277,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"186378, 186382",,,0,3,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-06-04T21:53:54Z,2026-06-05T19:09:33Z,33,"Stack from ghstack (oldest at bottom): #186378 #186382 -> #186277 Adds two dict-driven controls for selective activation checkpointing (AC1, use_reentrant=False) that work in both eager and torch.compile: checkpoint(..., policy={op_or_name: CheckpointPolicy}): a convenient form of selective activation checkpointing...."
185864,Fix segfault in native_batch_norm with empty running_mean/running_var,https://github.com/pytorch/pytorch/pull/185864,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,169208,169208,169208,,,,0,2,0,,,"triaged, open source, release notes: nn, topic: bug fixes",2026-06-01T23:09:47Z,2026-06-05T18:37:05Z,33,"Fixes #169208 torch.native_batch_norm (the batch_norm_cpu path) segfaults when running_mean or running_var is a defined but empty tensor. The size of the running stats is never checked on this path, so the stats update loop indexes a zero length buffer out of bounds. F.batch_norm does not hit this because it routes..."
185901,[Test] Refactor test_schema_check.py to be device-agnostic,https://github.com/pytorch/pytorch/pull/185901,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, topic: not user facing, module: tests/core",2026-06-02T06:56:31Z,2026-06-05T12:51:21Z,33,"Summary This PR refactors test/test_schema_check.py to decouple it from specific hardware backends, enabling new accelerator backends to reuse PyTorch's schema-checking test suite. Changes New TestSchemaCheckFunctionality(TestCase): Extracted 13 functionality tests that verify SchemaCheckMode preserves op correctnes..."
184409,Fix local_map deferred inlining to functionalize mutable operations,https://github.com/pytorch/pytorch/pull/184409,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,182403,182403,182403,,,,0,2,0,,,"triaged, open source, module: functionalization, module: higher order operators",2026-05-19T17:23:03Z,2026-06-05T12:26:59Z,33,"Fix local_map deferred inlining to functionalize mutable operations Fixes #182403 Problem When local_map uses deferred inlining with activation checkpointing, mutable operations like index_put_ were not functionalized before partitioning. This caused the default partitioner to reject the graph and fall back to min-c..."
185415,[ROCm] Fix LayerNorm backward kernel for AMD Strix Halo GPUs,https://github.com/pytorch/pytorch/pull/185415,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,False,0,0,,,,183861,,,,,0,1,0,,,"module: rocm, open source, release notes: cuda",2026-05-27T17:58:06Z,2026-06-05T12:18:38Z,33,"FIXES #183861 LayerNorm/RMSNorm's gamma beta backward kernel produced wrong results for AMD Strix Halo and other RDNA GPUs. It was caused by a hardcoded warp size (incorrect for RDNA/Strix Halo), and mixed usage of thread block and warp coordinates. This caused half of the values (along the x-axis/features) being dr..."
186079,Add test cases for 0-sized tensor padding,https://github.com/pytorch/pytorch/pull/186079,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,152750,152750,152750,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-06-03T13:27:17Z,2026-06-05T11:10:05Z,33,"Summary Closes #152750. PR #153037 corrected the C++ check in constant_pad_nd from new_dim > 0 to new_dim >= 0, but went stale before the matching regression tests landed, leaving the fix in C++ but uncovered. This PR adds the missing coverage by extending the constant-mode cases in sample_inputs_nn_pad (which flow..."
177263,Eliminate the unnecessary logic,https://github.com/pytorch/pytorch/pull/177263,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-03-12T14:10:22Z,2026-06-05T10:44:50Z,33,Stack from ghstack (oldest at bottom): -> #177263 As the title stated.
185190,Make constant SymInts hashable,https://github.com/pytorch/pytorch/pull/185190,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/dynamo, ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,"139111, 157417","139111, 157417",,,,,0,4,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, release notes: fx, test-config/distributed, ciflow/inductor, release notes: distributed (dtensor), ci-no-td, ciflow/vllm, ciflow/dynamo, ciflow/torchtitan, ciflow/dtensor",2026-05-26T05:34:13Z,2026-06-05T08:06:05Z,33,"Makes constant SymInts hashable and fixes DTensorSpec hashing under dynamic shapes, re-enabling test_ddp_tp. Fixes #139111 Closes #157417 SymInt.__hash__ hashes by value when expr.is_number; symbolic SymInts keep raising TypeError (einops relies on it to bypass its lru_cache). DTensorSpec._hash_key recursed into sym..."
185648,[Test] Make sigmoid_add extension test device-agnostic in test_cpp_extensions_aot.py,https://github.com/pytorch/pytorch/pull/185648,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: tests/graph",2026-05-29T17:19:55Z,2026-06-05T07:47:34Z,33,Refactor test/test_cpp_extensions_aot.py to remove hardcoded device strings from the sigmoid_add kernel tests and run them on every available accelerator backend. Changes Add import importlib and module-level _SIGMOID_ADD_BACKENDS constant mapping each device type to its pre-built extension module - Merge test_cuda_...
186304,"c10/core/DispatchKeySet: promote Full, FullAfter, Raw to enum class",https://github.com/pytorch/pytorch/pull/186304,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: dynamo, ciflow/inductor, release notes: inductor (aoti), meta-exported",2026-06-05T01:39:04Z,2026-06-05T06:35:30Z,33,"Summary: Change the three constructor-tag enums inside c10::DispatchKeySet from unscoped to scoped: enum Full { FULL } → enum class Full { FULL } enum FullAfter{ FULL_AFTER} → enum class FullAfter{ FULL_AFTER} enum Raw { RAW } → enum class Raw { RAW } With unscoped enums, the enumerators FULL, FULL_AFTER, and RAW we..."
185555,Add torch._rust Rust extension and build plumbing (#185555),https://github.com/pytorch/pytorch/pull/185555,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, release notes: releng, ciflow/inductor, meta-exported, ciflow/docker",2026-05-28T19:43:34Z,2026-06-05T06:35:13Z,33,Summary: Adds initial Rust/PyO3 support to PyTorch by introducing Rust into the build ecosystem. This is foundational infrastructure; the included tensor helpers exist to exercise the path end-to-end and are not yet a stable public API. The crate lives at torch/_rust/. The C++/Rust boundary uses the cxx crate. Becau...
186073,Support weights_only=True in torch.export.load,https://github.com/pytorch/pytorch/pull/186073,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/inductor,review when ready,,UNKNOWN,False,0,0,,,,186073,,,,,0,4,0,,,"triaged, open source, ciflow/inductor, release notes: export",2026-06-03T11:56:36Z,2026-06-05T05:57:30Z,33,Fixes #186073 would that be cool ? hehe
186218,Limit expandable IPC receiver VA reservation,https://github.com/pytorch/pytorch/pull/186218,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,186213,,,,0,1,0,,,"triaged, open source",2026-06-04T09:32:45Z,2026-06-05T03:16:09Z,34,"When CUDA expandable segments are shared through IPC, the receiver rebuild path constructs an ExpandableSegment from the producer's share header. This path only imports and maps the handles exported by the producer, but it previously used the normal expandable segment reservation policy and reserved enough VA space..."
185956,torch/headeronly/core/TensorAccessor.h: replace typedef with using,https://github.com/pytorch/pytorch/pull/185956,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, release notes: cpp, topic: improvements, meta-exported",2026-06-02T16:44:49Z,2026-06-05T01:46:11Z,34,"Summary: Mechanically replace all typedef T* PtrType and typedef typename PtrTraits<T>::PtrType PtrType declarations in TensorAccessorBase, TensorAccessor, GenericPackedTensorAccessorBase, and the DefaultPtrTraits / RestrictPtrTraits structs with equivalent using declarations. The CUDA __restrict__ qualifier is pres..."
186268,[TEST][CUDA] Add some additional cleanup to `TestCudaallocator`and re-enable `test_memory_plots`,https://github.com/pytorch/pytorch/pull/186268,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cuda, module: tests, open source, topic: not user facing",2026-06-04T21:14:44Z,2026-06-05T00:21:48Z,34,authored with codex cc @ptrblck @msaroufim @tinglvv @nWEIdia @mruberry
186253,[fsdp2] Add SPMD no-typecheck scopes to hooks,https://github.com/pytorch/pytorch/pull/186253,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-06-04T18:50:45Z,2026-06-04T22:28:07Z,34,"Stack from ghstack (oldest at bottom): (to be filled) Wrap the existing post-forward, pre-backward, and post-backward hook bodies in dist.spmd_no_typecheck(). This is a mechanical prep change for later FSDP and spmd_types integration; it keeps the existing Dynamo-disabled hook entrypoints and does not change the hoo..."
186128,Reject dtype conversion of quantized TypedStorage with a clear error,https://github.com/pytorch/pytorch/pull/186128,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,169210,169210,169210,,,,0,1,0,,,"triaged, open source",2026-06-03T20:59:03Z,2026-06-04T22:18:57Z,34,"Fixes #169210. The dtype-conversion methods on TypedStorage (bfloat16, float, double, half, int, ...) call into the C++ storage-cast path with a quantized dtype. The cast has no well-defined meaning — a quantized dtype is not self-contained and needs the tensor's scale and zero_point to be interpreted — and previous..."
177546,[inductor] Fix pad_mm leaking padded strides to user-visible outputs,https://github.com/pytorch/pytorch/pull/177546,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,"152059, 168207, 176109, 176110","152059, 168207, 176109, 176110",,,,,0,7,0,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-03-16T17:03:23Z,2026-06-04T18:58:21Z,34,"Stack from ghstack (oldest at bottom): -> #177546 pad_mm pads mm operands to alignment boundaries, then slices the output back to the original size. This slice is a view that inherits the padded base strides (e.g., stride (4, 1) instead of (2, 1) for a 3x2 output). Two bugs allowed these padded strides to leak throu..."
185830,"Fix typos in comments, docstrings, and error messages",https://github.com/pytorch/pytorch/pull/185830,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: cpu, module: amp (automated mixed precision), ciflow/trunk, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), release notes: export, ciflow/inductor-pallas, ciflow/torchtitan, ciflow/dtensor",2026-06-01T17:37:48Z,2026-06-04T18:40:31Z,34,"Stack from ghstack (oldest at bottom): -> #185830 Correct spelling errors, missing words, punctuation issues, and extra whitespace across torch source files. These are comment-only and string-only changes with no functional impact. Authored with Claude (typo_terminator2)."
183847,[distributed][docs] Fix group_src docstring in irecv,https://github.com/pytorch/pytorch/pull/183847,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, release notes: distributed (c10d), ciflow/torchtitan",2026-05-15T04:34:49Z,2026-06-04T18:45:08Z,34,"The group_src parameter in irecv is described as ""Destination rank"" but it specifies the source rank to receive from. This is consistent with the src parameter description (""Source rank"") in the same function."
179769,[cuda graphs] Integrate kernel annotations into inductor cudagraph trees,https://github.com/pytorch/pytorch/pull/179769,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,8,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-08T23:30:36Z,2026-06-04T17:45:42Z,34,"Stack from ghstack (oldest at bottom): -> #179769 #179867 #179768 Add config.triton.cudagraph_kernel_annotations (default False) and wire up clear/resolve/remap calls in CUDAGraphNode._record() so that mark_kernels() scopes firing during cudagraph capture are automatically processed. WIP: current progress: it works,..."
182344,Inductor: Fix Searchsorted/Bucketize Sliced View Boundaries,https://github.com/pytorch/pytorch/pull/182344,ready for maintainer decision,80,"approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,182343,182343,,,,,0,3,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-04T20:35:45Z,2026-06-04T15:34:22Z,34,"Materialize view-backed boundary tensors before bucketize/searchsorted boundary metadata extraction, and add focused Inductor and AOTI regression tests for sliced boundaries. Fixing #182343 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy..."
186159,Restore Any return type for Module.__getattr__ to fix typing false positives,https://github.com/pytorch/pytorch/pull/186159,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,185023,185023,185023,,,,0,1,0,,,"triaged, open source",2026-06-04T00:06:57Z,2026-06-04T13:43:08Z,34,"Fixes #185023. Problem Module.__getattr__ is annotated as Union[Tensor, ""Module""]. This produces false-positive type errors on common, well-typed code: class MyModel(nn.Module): def __init__(self) -> None: super().__init__() self.register_buffer('running_mean', torch.zeros(4)) self.weight = nn.Parameter(torch.randn(..."
184315,[Test] Make autograd/test_functional.py device-agnostic for out-of-tree backends.,https://github.com/pytorch/pytorch/pull/184315,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, topic: not user facing",2026-05-19T05:26:53Z,2026-06-04T11:47:37Z,34,Replace hardcoded CUDA references with device-agnostic equivalents so out-of-tree accelerator backends can run these tests. Replace TEST_CUDA with TEST_ACCELERATOR Replace device=cuda and .cuda() with torch.accelerator API Remove common_cuda import (no longer needed) Verified all 124 tests pass on both CPU and CUDA...
186048,Dedupe namespaced codegen and annotation parsing in torchgen,https://github.com/pytorch/pytorch/pull/186048,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, suppress-bc-linter",2026-06-03T06:29:15Z,2026-06-04T09:37:25Z,34,"This PR removes three duplicated pieces of logic in torchgen, with no change to generated code: register_dispatch_key.py: collapse the near-identical namespaced declaration/definition codegen paths into a single gen_namespaced_target helper. model.py: share the Tensor(a!) alias-annotation parsing between Argument.pa..."
183691,[Test] Add tolerance and precision override mechanisms for parameteried tests,https://github.com/pytorch/pytorch/pull/183691,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,177968,177968,177968,,,,0,1,0,,,"triaged, open source, module: testing, ciflow/trunk, topic: not user facing, module: PrivateUse1, module: openreg",2026-05-14T09:21:17Z,2026-06-04T08:32:40Z,34,"Fixes #177968 Enhancements to test configuration system: Added precision_overrides and tolerance_overrides class variables to DeviceTypeTestBase, allowing out-of-tree backends to set per-dtype (and per-test) precision (atol) and tolerance (atol+rtol) overrides via set_test_configs(). Implemented _bake_class_level_ov..."
185184,[Test][distributed] Refactor some distributed test helpers to be device-agnostic,https://github.com/pytorch/pytorch/pull/185184,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, module: testing, ciflow/trunk, topic: not user facing",2026-05-26T02:26:53Z,2026-06-04T06:55:13Z,34,Use torch.accelerator instead of device specific APIs in some helpers skip_if_no_gpu at_least_x_gpu skip_if_lt_x_gpu init_multigpu_helper
174241,[MPS] Fix conv backward to preserve input memory format,https://github.com/pytorch/pytorch/pull/174241,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,174269,174269,,,,,0,2,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, release notes: mps, ciflow/mps, module: inductor, ciflow/inductor, ci-no-td",2026-02-03T22:09:00Z,2026-06-04T01:30:31Z,35,"Summary Fixes convolution_backward on MPS to preserve the original input tensor's memory format, rather than deriving it from grad_output. When convolution_backward is called with a channels_last input but contiguous grad_output, the grad_input should preserve the input's memory format (channels_last), not adopt gra..."
186130,Add _single ProcessGroup methods and deprecate the _base aliases,https://github.com/pytorch/pytorch/pull/186130,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186131,,,0,2,0,,,"release notes: distributed (c10d), ciflow/h100-symm-mem, ciflow/torchtitan",2026-06-03T21:01:42Z,2026-06-04T00:09:31Z,35,"Stack from ghstack (oldest at bottom): #186131 -> #186130 #186125 #186124 #186123 Add all_gather_single / reduce_scatter_single / all_to_all_single to the C++ ProcessGroup, matching the torchcomms backend naming scheme. These now hold the canonical dispatch implementations; the old _allgather_base / _reduce_scatter_..."
170644,[dynamo] Support dict assignment with lazy constant keys,https://github.com/pytorch/pytorch/pull/170644,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"170092, 170567",,,0,104,0,,,"Merged, Reverted, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, ci-no-td, ciflow/vllm, ciflow/torchtitan, jansel-agent-skip",2025-12-17T05:43:12Z,2026-06-03T23:37:38Z,35,Stack from ghstack (oldest at bottom): -> #170644 #170567 #170092 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98 @Lucaskabela @mlazos
170567,[dynamo] Support dict mutation without a guard,https://github.com/pytorch/pytorch/pull/170567,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"170092, 170644",,,0,80,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan, jansel-agent-skip",2025-12-16T16:26:00Z,2026-06-03T23:37:29Z,35,Stack from ghstack (oldest at bottom): #170644 -> #170567 #170092 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98 @Lucaskabela
170092,[dynamo] Support constant ops on unrealized LazyConstantVariables,https://github.com/pytorch/pytorch/pull/170092,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"170567, 170644",,,0,107,0,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, skip-pr-sanity-checks, module: dynamo, ciflow/inductor, keep-going, ci-no-td, ciflow/vllm, ciflow/dynamo, ciflow/torchtitan, jansel-agent-skip",2025-12-10T17:32:26Z,2026-06-03T23:37:16Z,35,Stack from ghstack (oldest at bottom): #170644 #170567 -> #170092 cc @albanD @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo @azahed98 Differential Revision: D105076819
186131,Migrate ProcessGroup method usages to the _single names,https://github.com/pytorch/pytorch/pull/186131,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,186130,,,0,2,0,,,release notes: distributed (c10d),2026-06-03T21:01:44Z,2026-06-03T23:30:37Z,35,Stack from ghstack (oldest at bottom): -> #186131 #186130 #186125 #186124 #186123 Update the remaining first-party usages of the deprecated ProcessGroup methods to the new names: the Python ProcessGroup subclass used by the multi-threaded test backend now implements all_gather_single / reduce_scatter_single / all_to...
186133,[ROCm] remove redundant @skipIfRocm from test_cudnn_attention_broken_166211,https://github.com/pytorch/pytorch/pull/186133,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,168872,168872,168872,,,,0,1,0,,,"module: rocm, open source, topic: not user facing",2026-06-03T21:09:31Z,2026-06-03T22:56:03Z,35,"Investigative summary FIXES #168872 Test environment -- PyTorch version: 2.13.0a0+gitf6f4e17 -- Hip version: 7.2.53211 -- GPU 0 name: AMD Instinct MI300X Command to reproduce $ PYTORCH_TEST_WITH_ROCM=1 pytest test/test_transformers.py -k ""test_cudnn_attention_broken_166211"" -v Default behaviour test/test_transformer..."
185883,Simplify ostream formatting with fmt::join and fmt::print,https://github.com/pytorch/pytorch/pull/185883,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, release notes: cpp",2026-06-02T03:42:34Z,2026-06-03T22:45:38Z,35,"Simplifies ostream string building in a few spots: replaces manual separator loops with fmt::join, and uses fmt::print to format directly into the stream instead of building a temporary std::string via fmt::format. Uses fmt::ptr for descriptor handles. No behavior change."
185993,Fix SM121 compute capability corruption in CMake GPU arch detection,https://github.com/pytorch/pytorch/pull/185993,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-06-02T20:49:15Z,2026-06-03T22:38:53Z,35,"Problem The string(REPLACE) on line 138 of select_compute_arch.cmake uses literal substring matching to rewrite ""2.1"" to ""2.1(2.0)"" for legacy Fermi GPU handling. However, this also matches the ""2.1"" substring within SM121 (compute capability 12.1), corrupting it into ""12.1(2.0)"" — an invalid arch that causes nvcc c..."
185959,Fix SIGFPE in torch.lcm with the integer type minimum,https://github.com/pytorch/pytorch/pull/185959,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,121343,121343,121343,,,,0,1,0,,,"module: cpu, triaged, open source",2026-06-02T16:48:15Z,2026-06-03T22:23:51Z,35,"Fixes #121343 torch.lcm raises SIGFPE for the integer type minimum, for example torch.lcm(torch.tensor([-9223372036854775808]), torch.tensor(215)). The CPU lcm_kernel computes std::abs(a / g * b) where g = calc_gcd(a, b). calc_gcd takes abs of its inputs, but abs(INT64_MIN) overflows and stays negative, so calc_gcd(..."
185906,add type hint to seed argument in torch.random.manual_seed,https://github.com/pytorch/pytorch/pull/185906,ready for maintainer decision,45,CI lane labels: ciflow/torchtitan; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: devx, ciflow/torchtitan",2026-06-02T08:50:43Z,2026-06-03T22:18:32Z,35,"The type hint for the seed argument is not specified in the function signature, which makes the function ""partially unknown"". Especially in strict mode, type checkers do not like that: from typing import reveal_type import torch.random torch.random.manual_seed(42) # Pylance(reportUnknownMemberType): Type of ""manual_..."
185868,"Update clamp, clamp_min, clamp_max to use the minimum norm subgradient",https://github.com/pytorch/pytorch/pull/185868,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184572,184572,184572,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-06-02T00:28:00Z,2026-06-03T22:01:26Z,35,"Fixes #184572 torch.clamp, torch.clamp_min, torch.clamp_max subgradients violate PyTorch's documented minimum norm convention, returning 1.0 instead of 0.0. Changes made: tools/autograd/derivatives.yaml: Update boundary conditions for clamp_min and clamp_max torch/csrc/autograd/FunctionsManual.cpp: Update boundary c..."
185927,Fix clamp boundary subgradients,https://github.com/pytorch/pytorch/pull/185927,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,184572,184572,184572,,,,0,2,0,,,"triaged, open source",2026-06-02T12:51:14Z,2026-06-03T22:01:22Z,35,"summary Use minimum-norm subgradients for clamp boundaries. Scalar bounds now give zero self-gradient at the boundary, while one-sided Tensor bounds split ties evenly between self and the bound. Leave two-Tensor-bound clamp unchanged. This PR updates clamp boundary gradients to follow the minimum-norm subgradient co..."
186045,Merge identical elif branches and inline single-use locals in torchgen,https://github.com/pytorch/pytorch/pull/186045,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: inductor (aoti)",2026-06-03T06:02:17Z,2026-06-03T22:01:20Z,35,Authored with Claude.
185973,[Accelerator] Enable Default Stream API,https://github.com/pytorch/pytorch/pull/185973,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185975,185975,185975,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, module: accelerator, ciflow/torchtitan",2026-06-02T18:24:25Z,2026-06-03T22:01:12Z,35,PrivateUse1 already exposes default stream API in DeviceGuardImplInterface::getDefaultStream. It seems straightforward to integrate it into torch.accelerator: torch.accelerator.default_stream() Fix #185975. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @j...
186110,Remove test_transformers from windows blocklist,https://github.com/pytorch/pytorch/pull/186110,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-06-03T18:01:18Z,2026-06-03T22:01:09Z,35,"Allow math SDPA in CUDA MHA fastpath and enable transformers tests on Windows The native MHA fastpath previously skipped fused SDPA and fell through to legacy bmm/softmax, which can produce NaNs in fp16. Accept SDPBackend::math in the fastpath so SDPA runs instead. Add xfailIfNoAcceleratorTriton on test_transformer_..."
182208,Add missing stacklevel to 57 warnings.warn() calls in torch/,https://github.com/pytorch/pytorch/pull/182208,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: fx, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-02T05:41:22Z,2026-06-03T21:42:10Z,35,"Summary 57 warnings.warn() calls across 38 files in torch/ (excluding tests) are missing stacklevel, causing warning messages to point to internal lines inside PyTorch rather than the caller's code. Problem Without stacklevel, warnings reference internal PyTorch lines: torch/cuda/__init__.py:1350: UserWarning: Canno..."
186095,[inductor] Cache per-node tiling in fusion candidate evaluation,https://github.com/pytorch/pytorch/pull/186095,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186096, 186097, 186101, 186102, 186106",,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-03T16:03:01Z,2026-06-03T21:13:17Z,35,"Stack from ghstack (oldest at bottom): #186102 #186106 #186101 #186097 #186096 -> #186095 select_tiling is called for each node in every fusion pair. With N pairs, this means O(N) redundant calls for the same node. Cache results per node id, and share across unfused nodes with the same group (device + iteration doma..."
186086,[PyTorch] Add missing meta kernel for aten::_sample_dirichlet,https://github.com/pytorch/pytorch/pull/186086,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, module: dynamo, ciflow/inductor, meta-exported",2026-06-03T14:31:28Z,2026-06-03T21:13:12Z,35,"Summary: Fixes torch.compile(fullgraph=True) for models that use torch.distributions.Dirichlet (e.g. OBA). The op previously had no meta kernel, causing graph breaks under Dynamo. The implementation mirrors _standard_gamma — output shape equals input shape. Differential Revision: D107307274 cc @voznesenskym @penguin..."
186106,[inductor] Fast paths in sizevars for concrete integer comparisons,https://github.com/pytorch/pytorch/pull/186106,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"186095, 186096, 186097, 186101, 186102",,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-03T17:35:20Z,2026-06-03T21:13:11Z,35,"Stack from ghstack (oldest at bottom): #186102 -> #186106 #186101 #186097 #186096 #186095 statically_known_equals and check_leq create sympy.Eq/sympy.Le objects on every call. For concrete integers, use direct Python comparisons. Profiled on 20K-node graph (10K params, H100), on top of previous fix: Total before: 10..."
177069,[xpu] Increase tolerance for tests that fail due to non-deterministic operators behavior.,https://github.com/pytorch/pytorch/pull/177069,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,"2238, 2751, 3103","2238, 2751, 3103",,,,,0,3,0,,,"open source, ciflow/trunk, release notes: python_frontend",2026-03-10T20:41:20Z,2026-06-18T20:52:27Z,20,Is fix for: intel/torch-xpu-ops#3103 Is fix for: intel/torch-xpu-ops#2751 Summary This PR relaxes test tolerances for selected XPU tests that compare numerically sensitive computations across different execution paths and/or repeated runs. The root issue is broader than a single operator: some XPU kernels and reduct...
185934,[NVSHMEM] Fix link_whole and -fPIC for nvshmem_device_link_target,https://github.com/pytorch/pytorch/pull/185934,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/h100-symm-mem, ciflow/trunk",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"fb-exported, ciflow/trunk, release notes: distributed (c10d), ciflow/h100-symm-mem, meta-exported",2026-06-02T14:11:58Z,2026-06-03T18:44:54Z,35,"Summary: Four build fixes for NVSHMEM integration with python_binary targets that use Buck link groups: Add -fPIC to compiler_flags in the nvshmem obj library. Objects go through the nvcc -dlink pipeline and end up in the dlink archive, which may be placed in shared libraries (link groups) by python_binary. Without..."
185851,[autograd] Fall back to tensor backward for symbolic pow exponent,https://github.com/pytorch/pytorch/pull/185851,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185715,185715,,,,0,1,0,,,"topic: not user facing, module: inductor, ciflow/inductor",2026-06-01T21:26:06Z,2026-06-03T18:37:37Z,35,"Stack from ghstack (oldest at bottom): (to be filled) Under dynamic=True a Python-int exponent to torch.pow becomes a symbolic Scalar. pow_backward short-circuited on exponent.equal(0.0), but Scalar::equal asserts ""NYI SymInt equality"" for symbolic scalars, so compiling pow + autograd.grad with dynamic shapes crashe..."
184965,[Inductor] Cap combo-kernel seed autotune configs by size,https://github.com/pytorch/pytorch/pull/184965,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"183626, 184226, 184955, 185183",,,0,10,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T05:35:43Z,2026-06-03T18:27:25Z,35,Stack from ghstack (oldest at bottom): #184226 #185183 #183626 -> #184965 #184955 #182901 Differential Revision: D106878653 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @ja...
186020,[WIP] Reintroduce ptxas bundling for CUDA 13.0 binaries (revert #174716),https://github.com/pytorch/pytorch/pull/186020,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-06-03T00:50:12Z,2026-06-03T16:34:21Z,35,"Reintroduces the ptxas bundle for CUDA 13.0 binaries -- effectively a revert of #174716, scoped to 13.0. The bundle was originally added (#163988) to work around #163801, where Triton's ptxas could miscompile kernels. #174716 removed it once that issue was believed fixed upstream; we are reinstating it for CUDA 13.0..."
172945,Bf16 optimized heuristics,https://github.com/pytorch/pytorch/pull/172945,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cpu, triaged, open source, release notes: releng",2026-01-21T14:05:38Z,2026-06-03T15:44:33Z,35,This PR introduces: BGEMM backend integration for BF16 GEMM A data-driven decision-tree heuristic for selecting BGEMM vs oneDNN Significant CPU inference improvements OpenBLAS update OpenBLAS version updated to 0.3.31.dev (this build contains BGEMM kernels for BF16). BGEMM vs SBGEMM: BGEMM is a BFloat16-native GEMM...
185475,[pytorch] Nested custom Python process groups (experimental) (#185475),https://github.com/pytorch/pytorch/pull/185475,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/torchtitan,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, release notes: distributed (c10d), meta-exported, ciflow/torchtitan",2026-05-28T06:46:15Z,2026-06-03T15:25:40Z,35,"Summary: Add experimental infrastructure for custom Python process groups that can intercept dist.* collective calls with the original Python arguments. See CUSTOM_PG.md for full documentation. Two kinds of custom PGs are supported: Terminal -- extends ProcessGroup directly, implements collectives with the dist.* AP..."
185117,nn: add missing memory_format overload to Module.to(),https://github.com/pytorch/pytorch/pull/185117,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184642,184642,184642,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-25T07:47:18Z,2026-06-03T10:48:12Z,35,"Summary Module.to(memory_format=torch.channels_last) is documented and works at runtime, but there was no matching @overload for type checkers. This causes no-matching-overload errors in tools like ty (Astral). Add the missing overload and import memory_format from torch. Fixes #184642."
184898,autograd: raise ValueError for forward-mode without vectorize in jacobian/hessian,https://github.com/pytorch/pytorch/pull/184898,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184842,184842,184842,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-05-22T15:09:35Z,2026-06-03T10:48:10Z,35,Summary torch.autograd.functional.jacobian with strategy='forward-mode' and vectorize=False (the default) silently accepts the call then crashes with NotImplementedError deep inside _jacfwd. The same issue affects hessian with outer_jacobian_strategy='forward-mode'. This PR adds an early ValueError at argument-valid...
186068,Add skipIfXpu decorator to tests for semi-structured sparsity support,https://github.com/pytorch/pytorch/pull/186068,draft/noise,23,PR is draft; PR closes or references open issue numbers,wait for author to mark ready,,UNKNOWN,True,0,0,,,2663,2663,,,,,0,1,0,,,"open source, topic: not user facing",2026-06-03T10:43:35Z,2026-06-03T10:46:23Z,35,"Add skipIfXpu to semi-structured sparsity tests Semi-structured sparsity (SparseSemiStructuredTensor and its backends) is not yet supported on XPU. Without the skipIfXpu decorator, running the test suite on an XPU device causes all semi-structured tests to fail unconditionally rather than being skipped. This PR adds..."
183272,[DTensor] Switch test_math_ops and test_matrix_ops to DTensorOpTestBase,https://github.com/pytorch/pytorch/pull/183272,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/dtensor,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,108744,183269,,,0,4,0,,,"triaged, open source, topic: not user facing, ciflow/dtensor",2026-05-11T12:34:34Z,2026-07-05T17:51:54Z,3,"Summary Converts DistMathOpsTest and DistMatrixOpsTest from DTensorTestBase (MultiProcessTestCase) to DTensorOpTestBase (MultiThreadedTestCase), following the pattern established in test_pointwise_ops.py. test/distributed/tensor/test_math_ops.py: DistMathOpsTest → DTensorOpTestBase test/distributed/tensor/test_matri..."
184548,[Test]Refactor test/test_sort_and_select.py to be device generic,https://github.com/pytorch/pytorch/pull/184548,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-20T14:27:59Z,2026-06-03T09:35:44Z,35,"Remove @onlyCUDA decorators from 5 tests to enable execution on all accelerators including out of tree devices via PrivateUse1. Changes in test/test_sort_and_select.py: Add import: onlyAccelerator from common_device_type Remove @onlyCUDA from 4 device-agnostic tests: test_sort_large, test_sort_discontiguous, test_to..."
185747,Use at::mps::OptionalMPSGuard for MPS dispatch keys in codegen,https://github.com/pytorch/pytorch/pull/185747,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,77170,77170,77170,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/mps",2026-05-31T07:20:01Z,2026-06-03T06:03:40Z,35,Use at::mps::OptionalMPSGuard for MPS dispatch keys in codegen. Forward-declares MPSEvent in MPSGuardImpl.h so the generated plain-C++ RegisterMPS_*.cpp can include the guard without pulling in Objective-C Metal types. Closes #77170 as obsolete: MPSGuardImpl now has real event support and MPS-specific device semanti...
183855,[c10d] Surface operational metadata in WorkNCCL operator<<,https://github.com/pytorch/pytorch/pull/183855,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: distributed (c10d)",2026-05-15T06:02:13Z,2026-06-03T04:34:32Z,35,This change enriches the WorkNCCL operator<< representation by surfacing IsP2P and IsBarrier metadata. These fields are passive CPU-side state and do not invoke CUDA queries or synchronization-sensitive runtime probing. Including this metadata improves debugging context for ProcessGroupNCCL failures by making it eas...
175542,[DTensor] Register DTensorSpec as pytree constant at module level,https://github.com/pytorch/pytorch/pull/175542,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,175467,175467,175467,,,,0,2,0,,,"triaged, open source, release notes: distributed (dtensor)",2026-02-23T16:24:36Z,2026-06-03T03:46:21Z,36,"Summary Fix torch.export.export() failing with RuntimeError: DTensorSpec not registered as pytree constant when exporting tensor-parallel models that use DTensor (#175467). Commit 1: During export, temporarily register non-graphable hashable leaf types (e.g. DTensorSpec) as pytree constants so tree_flatten absorbs t..."
186002,"Temporary Commit at 6/2/2026, 2:55:48 PM",https://github.com/pytorch/pytorch/pull/186002,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,,2026-06-02T22:02:58Z,2026-06-03T03:45:31Z,36,Stack from ghstack (oldest at bottom): -> #186002
185834,[inductor] Preserve signed zero in Triton neg codegen,https://github.com/pytorch/pytorch/pull/185834,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185696,185696,,,,0,2,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-01T17:58:30Z,2026-06-03T03:44:46Z,36,"Stack from ghstack (oldest at bottom): -> #185834 Triton lowers unary minus (-x) to 0 - x (see semantic.minus), which yields +0.0 for a +0.0 input and so loses the IEEE sign bit of zero. As a result torch.neg(+0.0) returned +0.0 under inductor, and atan2(x, -x) for x=+0.0 evaluated libdevice.atan2(0.0, +0.0)=0.0 ins..."
185308,Align subclass output SymInt unwrapping with metadata,https://github.com/pytorch/pytorch/pull/185308,draft/noise,61,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/dtensor, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,185296,185296,185296,,,,0,2,0,,,"ciflow/inductor, ciflow/dtensor",2026-05-27T01:39:03Z,2026-06-03T00:16:59Z,36,"Agent Report Summary torch.compile failed for a DTensor-returning module because AOTAutograd's metadata pass and joint graph capture disagreed on the flattened output arity for traceable wrapper subclass outputs. The generated two-rank repro failed before the fix with: AssertionError: Node view_3 was invalid, but is..."
179547,[inductor] Add thread-based compilation support for free-threaded Python (#179547) (#179547),https://github.com/pytorch/pytorch/pull/179547,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-04-07T04:32:08Z,2026-06-03T00:14:38Z,36,"Summary: This change adds support for thread-based Triton compilation workers when running with free-threaded Python (PEP 703). When the GIL is disabled, we can use threads instead of processes for compilation, avoiding expensive process spawning overhead. Key changes: Add should_use_thread_workers() to detect free-..."
185846,[decomp] Match eager's device-dependent threshold comparison dtype,https://github.com/pytorch/pytorch/pull/185846,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185470,185470,,,,0,2,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor",2026-06-01T20:04:44Z,2026-06-03T00:04:20Z,36,"Stack from ghstack (oldest at bottom): -> #185846 torch.threshold's boundary comparison uses a different dtype across eager backends: the CUDA kernel casts the scalar threshold to the input dtype and compares in the input dtype, while the CPU kernel compares in fp32. The decomposition used elementwise_type_promotion..."
185844,[CuTeDSL] Just use tensor strides directly in `_to_fake_cute_tensor`,https://github.com/pytorch/pytorch/pull/185844,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/b200, ciflow/torchtitan",2026-06-01T19:28:31Z,2026-06-02T22:56:35Z,36,Fixes a few outstanding cases (due to padding?) authored with codex cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
181730,"Prevent DCE eliminate ""triton_kernel_wrapper_mutation""",https://github.com/pytorch/pytorch/pull/181730,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, release notes: fx",2026-04-28T08:59:43Z,2026-06-02T22:47:07Z,36,"triton_kernel_wrapper_mutation has [num_users=0] (returns None) but mutates its input buffers in-place, soNode.is_impure()incorrectly returned False and DCE eliminated it, corrupting downstream consumers of the mutated buffers. This pr add TritonKernelWrapperMutation to torch.fx.node._side_effectful_functions."
185842,[decomp] Fix multilabel_margin_loss for padded targets,https://github.com/pytorch/pytorch/pull/185842,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185464,185464,,,,0,2,0,,,"module: inductor, ciflow/inductor",2026-06-01T18:58:22Z,2026-06-02T21:59:02Z,36,"Stack from ghstack (oldest at bottom): -> #185842 multilabel_margin_loss_forward builds its loss tensor z with shape [target_position, batch, input_class]. Target positions at or after the first -1 are invalid, but their gathered index was clamped to 0, so those rows summed in input[0]'s hinge terms. The decompositi..."
185592,Remove android directory,https://github.com/pytorch/pytorch/pull/185592,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, release notes: vulkan, topic: bc breaking, topic: not user facing, skip-pr-sanity-checks, suppress-bc-linter",2026-05-29T04:14:38Z,2026-06-02T20:51:03Z,36,Is has been replaced by ExecuTorch cc @albanD
169694,Increase precision for golden solution in transformer tests.,https://github.com/pytorch/pytorch/pull/169694,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/rocm-mi300, ciflow/rocm-mi355",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,25,0,,,"triaged, open source, Merged, Reverted, release notes: python_frontend, topic: not user facing, module: inductor, module: dynamo, ci-no-td, ciflow/rocm-mi300, ciflow/rocm-mi355",2025-12-05T19:07:35Z,2026-06-02T20:35:18Z,36,"Overview Force FP32 precision for ""golden"" solution computation when TF32 is set for compute test solution. Rationale The test/test_transformers.py testing suite calculates the numerical tolerance by comparing output tensors from the same precision (""reference"") and higher precision (""golden""), both calculated by SD..."
179732,Fix torch.compile istft failing when length exceeds reconstructed sig…,https://github.com/pytorch/pytorch/pull/179732,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,179593,179593,,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-04-08T18:14:01Z,2026-06-02T20:34:32Z,36,"Fix torch.compile istft failing when length exceeds reconstructed signal size The Python ref for istft used narrow which hard-checks that start + length <= dim_size, causing torch.compile to error when the length parameter exceeds the reconstructed signal size. Eager mode succeeds because C++ slice silently clamps t..."
185935,[DO NOT MERGE] ci: add manywheel build telemetry (May 6f7bb01c),https://github.com/pytorch/pytorch/pull/185935,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2451,0,,,open source,2026-06-02T14:20:15Z,2026-06-02T18:31:35Z,36,Telemetry-only branch for CI artifact collection. Do not merge. Base commit: 6f7bb01 (May current) Adds build-analysis instrumentation to .ci/manywheel/build.sh to collect ninja_log.txt and cache stats from linux-binary-manywheel CI runs. Goal is to diagnose 6+ hour ROCm nightly build times.
185838,[inductor] Propagate subgraph tensor constants to the root graph,https://github.com/pytorch/pytorch/pull/185838,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,185482,185482,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-06-01T18:26:16Z,2026-06-02T17:39:02Z,36,Stack from ghstack (oldest at bottom): -> #185838 A tensor constant materialized only inside a torch.cond branch (e.g. an index tensor for scatter_add) was registered in the SubgraphLowering's own constants dict but never in the root graph's. At load time PyCodeCache attaches constants to the module only from the ro...
185822,[DO NOT MERGE] ci: add manywheel build telemetry (Feb baseline 0cd9ecaf),https://github.com/pytorch/pytorch/pull/185822,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/binaries_wheel, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2371,0,,,"open source, ciflow/trunk, ciflow/binaries_wheel",2026-06-01T15:36:21Z,2026-06-02T14:54:15Z,36,Telemetry-only branch for CI artifact collection. Do not merge. Base commit: 0cd9eca (Feb baseline) Adds build-analysis instrumentation to .ci/manywheel/build.sh to collect ninja_log.txt and cache stats from linux-binary-manywheel CI runs. Goal is to diagnose 6+ hour ROCm nightly build times.
185936,[DO NOT MERGE] ci: add manywheel build telemetry (Apr midpoint 977c5623),https://github.com/pytorch/pytorch/pull/185936,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2426,0,,,open source,2026-06-02T14:22:17Z,2026-06-02T14:31:34Z,36,Telemetry-only branch for CI artifact collection. Do not merge. Base commit: 977c562 (Apr midpoint) Adds build-analysis instrumentation to .ci/manywheel/build.sh to collect ninja_log.txt and cache stats from linux-binary-manywheel CI runs. Goal is to diagnose 6+ hour ROCm nightly build times.
185706,Fix typos across comments and docstrings in torch,https://github.com/pytorch/pytorch/pull/185706,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, release notes: quantization, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), release notes: export, ciflow/h100-symm-mem, ciflow/torchtitan, ciflow/dtensor",2026-05-30T13:38:34Z,2026-06-02T14:01:32Z,36,"Stack from ghstack (oldest at bottom): -> #185706 Fix spelling and grammar errors in comments and docstrings across 62 files spanning dynamo, inductor, distributed, functorch, JIT, nativert, ONNX, and other subsystems. Corrections include misspelled words (instasnces, inviolated, unsolbed, refacotred, arugument), du..."
181578,[Draft] [04/27/26] Dashboard result update for ROCm,https://github.com/pytorch/pytorch/pull/181578,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-periodic, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"module: rocm, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/inductor-periodic, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355",2026-04-27T13:28:22Z,2026-06-02T11:45:26Z,36,Trying new scriptable approach to update expected results cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @cha...
178590,Correct type hints for ConvNd subclasses,https://github.com/pytorch/pytorch/pull/178590,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,6,0,,,"module: nn, triaged, open source, topic: not user facing",2026-03-27T12:23:07Z,2026-06-02T09:19:15Z,36,"When trying to use one Conv2d to create another, I was seeing errors with ty like so: error[invalid-argument-type]: Argument to bound method `__init__` is incorrect --> test.py:180:9 | 178 | stride=updated_stride, 179 | padding=updated_padding, 180 | dilation=layer.dilation, | ^^^^^^^^^^^^^^^^^^^^^^^ Expected `int |..."
184226,[Inductor] Allow mixed pointwise/reduction combo kernels,https://github.com/pytorch/pytorch/pull/184226,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"183626, 184955, 184965, 185183",,,0,26,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-18T16:36:40Z,2026-06-02T05:19:36Z,36,Stack from ghstack (oldest at bottom): -> #184226 #185183 #183626 #184965 #184955 #182901 Differential Revision: D106878655 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @ja...
185183,[Inductor] Fix combo-kernel num_warps regression on mixed reductions,https://github.com/pytorch/pytorch/pull/185183,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"183626, 184226, 184955, 184965",,,0,9,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-26T02:14:19Z,2026-06-02T05:19:20Z,36,Stack from ghstack (oldest at bottom): #184226 -> #185183 #183626 #184965 #184955 #182901 Differential Revision: D106878654 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @ja...
185606,Enable TF32 fpmath mode for XPU deconvolution,https://github.com/pytorch/pytorch/pull/185606,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, open source, topic: not user facing, ciflow/xpu",2026-05-29T07:31:15Z,2026-06-02T05:19:06Z,36,"Summary The convolution path already calls apply_tf32_if_allowed() to set oneDNN fpmath_mode::tf32 when the user enables it via torch.backends.mkldnn.allow_tf32 = True, but the deconvolution path has not been enabled yet. This caused ConvTranspose ops to always run in strict FP32 mode regardless of the user setting...."
183626,[Inductor] Seed combo-kernel autotune from standalone subkernels,https://github.com/pytorch/pytorch/pull/183626,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"184226, 184955, 184965, 185183",,,0,22,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-05-13T22:17:36Z,2026-06-02T05:18:51Z,36,Stack from ghstack (oldest at bottom): #184226 #185183 -> #183626 #184965 #184955 #182901 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo Differential Revision: D1068...
181915,Support returning None from torch.cond() true_fn and false_fn,https://github.com/pytorch/pytorch/pull/181915,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,181891,181891,181891,,,,0,3,0,,,"triaged, open source, release notes: cuda, release notes: fx, module: dynamo, ciflow/inductor, module: higher order operators, ciflow/torchtitan",2026-04-29T20:01:12Z,2026-06-02T03:46:13Z,37,"This allows us to run conditional code that only mutates inputs or captured variables, but does not return anything. While making this change, I realized as well that it should be disallowed for true_fn and false_fn to return different non-tensor outputs in the same output slot when using torch.cond() during cuda gr..."
185758,Tidy MPSGuardImpl: name capability bitmask and fix clang-tidy nits,https://github.com/pytorch/pytorch/pull/185758,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-31T12:44:42Z,2026-06-02T03:10:57Z,37,"Name the device-capability bitmask via a local constexpr, default the OptionalMPSGuard default ctor and give it a destructor, and name the unused getNewStream parameters (dropping the (void)priority cast). Authored with Claude."
171313,Support torch.accelerator.Graph on CUDA,https://github.com/pytorch/pytorch/pull/171313,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"171445, 171780",,,0,41,0,,,"open source, Stale, ciflow/trunk, release notes: python_frontend, no-stale, module: accelerator",2025-12-26T06:56:54Z,2026-06-02T02:02:32Z,37,Stack from ghstack (oldest at bottom): #171445 #171780 #171443 -> #171313 Motivation This PR adds support for at::accelerator::Graph and torch.accelerator.Graph on the CUDA backend. It serves as a concrete example of how to support the unified Graph interface by reusing an existing device-specific graph implementati...
182618,[Profiler] Forward declare nlohmann::json in public header,https://github.com/pytorch/pytorch/pull/182618,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,181565,181565,181565,,,,0,1,0,,,"triaged, open source",2026-05-06T06:11:10Z,2026-06-02T01:31:40Z,37,"Avoid including nlohmann/json.hpp from the profiler traceback public header by forward declaring nlohmann::json and keeping the full include in the .cpp file. This follows the suggested fwd-declaration approach for torch/csrc/profiler/python/combined_traceback.h. The header only needs to declare json_symbolize(), so..."
184533,Defer AOTAutograd version-check error for forward-only graph,https://github.com/pytorch/pytorch/pull/184533,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,158561,,,,0,2,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: aot autograd",2026-05-20T13:10:09Z,2026-06-02T00:24:56Z,37,"Summary: Add a regression test for the AOTAutograd timing mismatch in #158561. Defer a narrow autograd version-check failure from joint capture to backward for a single-output, non-alias forward graph. Add guardrail tests for parameter gradients, input gradients, inference controls, and true forward-time errors. Val..."
183390,[SymmMem] Add NCCL implementation of all_to_all_vdev,https://github.com/pytorch/pytorch/pull/183390,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/h100-symm-mem, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, release notes: distributed (c10d), ci-no-td, ciflow/h100-symm-mem, module: symm_mem, ciflow/torchtitan",2026-05-12T00:41:27Z,2026-06-01T23:47:50Z,37,"Mirrors the existing NVSHMEM all_to_all_vdev (with on-device split sizes) on top of the NCCL symmetric memory backend, so applications already using NCCL symm memory can dispatch the same torch.ops.symm_mem.all_to_all_vdev schema without pulling in NVSHMEM. Implementation pattern follows nccl_reduce_scatter_offset.c..."
185727,"Document bernoulli(input, p) overload in torch docs",https://github.com/pytorch/pytorch/pull/185727,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,152095,152095,152095,,,,0,1,0,,,"triaged, open source, release notes: python_frontend",2026-05-30T20:36:01Z,2026-06-01T23:46:24Z,37,"Fixes #152095 Summary torch.bernoulli accepts two distinct call forms: torch.bernoulli(input, *, generator=None, out=None) draws from per-element probabilities held in input. torch.bernoulli(input, p, *, generator=None) draws using a single scalar probability p for every element of input. Only the first form was doc..."
179205,[functorch] Fix vmap/functionalize under inference_mode,https://github.com/pytorch/pytorch/pull/179205,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,177750,177750,177750,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-04-02T23:38:58Z,2026-06-01T22:47:09Z,37,"Fixes #177750 This PR completes the inference_mode handling for the functorch transforms begun in #177596. cc @aorenste Under torch.inference_mode(), vmap and functionalize outputs became permanently marked as inference tensors because these transforms didn't save/restore the inference_mode flag. #177596 fixed this..."
185792,test/distributed/pipelining: add end-to-end tests with HuggingFace GPT-2 and BERT,https://github.com/pytorch/pytorch/pull/185792,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,185780,185780,185780,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-06-01T04:53:33Z,2026-06-01T22:37:41Z,37,"Closes #185780. Adds test/distributed/pipelining/test_hf_transformers.py with 8 CPU tests that exercise torch.distributed.pipelining against real HuggingFace Transformer submodules (GPT-2 and BERT), complementing the existing handwritten-model tests in test_transformer.py. What is tested test_pipeline_ir_num_stages..."
185835,[c10d] Add delegate backend to FakeProcessGroup for numerics testing,https://github.com/pytorch/pytorch/pull/185835,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,release notes: distributed (c10d),2026-06-01T18:19:36Z,2026-06-01T21:41:43Z,37,"Stack from ghstack (oldest at bottom): -> #185835 FakeProcessGroup.set_delegate(backend) forwards all collectives to a real backend instead of FakePG's no-op implementations. This enables multi-process numerics testing with deterministic collectives -- 8 processes init with FakePG, attach a gloo delegate, and get bi..."
185689,Upload docs preview directly to S3 from the OSDC pod when authenticated,https://github.com/pytorch/pytorch/pull/185689,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"Merged, Reverted, topic: not user facing, ci-no-td",2026-05-30T00:43:35Z,2026-06-01T20:23:56Z,37,"Summary The OSDC docs build (build-docs-osdc in _docs.yml) always staged the built docs as a GHA artifact and relied on the upload-docs-preview workflow_run job to do the S3 sync — even when the pod had successfully assumed arn:aws:iam::308535385114:role/arc (same-repo PRs). See e.g. this run where ""Configure AWS cr..."
185646,[AI Codemod][AsyncioGetEventLoopMigration] [asyncio-codemod] xplat/caffe2/test,https://github.com/pytorch/pytorch/pull/185646,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, topic: not user facing, meta-exported",2026-05-29T17:08:34Z,2026-06-01T18:51:45Z,37,Differential Revision: D106774158
181833,Add boolean evaluation for LazyVal wrapping str,https://github.com/pytorch/pytorch/pull/181833,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"180828, 180954",,,0,2,0,,,"open source, ciflow/trunk, topic: not user facing",2026-04-29T03:22:52Z,2026-06-01T17:59:36Z,37,"Stack from ghstack (oldest at bottom): #180828 -> #181833 #180954 #180827 Motivation This change adds __bool__ support for LazyVal to enable its use in boolean contexts, such as: if (ACCELERATOR_TYPE): # do something ACCELERATOR_TYPE is a LazyVal wrapping a string value. In python, str does not define __bool__ expli..."
180954,Add value property to LazyVal,https://github.com/pytorch/pytorch/pull/180954,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,"180828, 181833",,,0,8,0,,,"open source, ciflow/trunk, topic: not user facing",2026-04-21T03:28:26Z,2026-06-01T17:59:35Z,37,"Stack from ghstack (oldest at bottom): #180828 #181833 -> #180954 #180827 Motivation Add a value property to LazyVal to make it easier to use, especially when' LazyVal' is not a bool."
185743,Mark CUDA kernel helpers static,https://github.com/pytorch/pytorch/pull/185743,ready for maintainer decision,45,CI lane labels: ciflow/trunk; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, release notes: foreach_frontend",2026-05-31T04:35:45Z,2026-06-01T16:44:09Z,37,"Adds static to free functions in aten/src/ATen/native/cuda/ whose body does not contain GPU_LAMBDA / device lambdas, are not declared in any header, and are not referenced from any other TU. Authored with Claude."
185432,Upgrade pyrefly 1.0.0,https://github.com/pytorch/pytorch/pull/185432,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: mkldnn, release notes: quantization, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), release notes: distributed (checkpoint), release notes: export, ciflow/rocm-mi300, release notes: inductor (aoti), ciflow/h100, ciflow/b200, ciflow/torchtitan, ciflow/dtensor",2026-05-27T20:23:16Z,2026-06-01T15:07:27Z,37,Upgrading PyTorch to use our latest release v1.0.0. Adds suppressions and fixes formatting where needed. Will follow up with additional PR's to remove any unused suppression comments after this lands. cc @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfeima @sanchitintel @ashokei @jingxu10 @min-jean-ch...
183056,[lint] Fix PYI036 violations and re-enable the rule,https://github.com/pytorch/pytorch/pull/183056,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/inductor, ciflow/torchtitan",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,6,0,,,"module: cpu, triaged, open source, module: amp (automated mixed precision), Merged, Reverted, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-05-09T18:47:18Z,2026-06-01T15:07:15Z,37,"Summary Partial progress on #110950: this PR re-enables ruff's PYI036 rule only. It fixes the __exit__ / __aexit__ annotations across torch/, tools/, and test/ that previously used Any for their exception parameters. With these annotations corrected, PYI036 is removed from the ignore list in pyproject.toml so the ru..."
181087,[MPS] RDMA for distributed MPS,https://github.com/pytorch/pytorch/pull/181087,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,37,0,,,"triaged, open source, release notes: distributed (c10d), release notes: mps, ciflow/mps, ciflow/torchtitan",2026-04-22T08:45:54Z,2026-06-01T14:59:43Z,37,"RDMA-only c10d backend for Apple Silicon Macs connected via Thunderbolt 5. Construction requires ibv_alloc_pd to succeed on every rank, otherwise the constructor throws and users should select the gloo backend. Implemented Op Transport Notes allreduce RDMA mesh SUM / PRODUCT / MIN / MAX over float32 / int32 / int64..."
185286,Add version counter check for gradient tensors in the autograd engine,https://github.com/pytorch/pytorch/pull/185286,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"185284, 185285, 185287",,,0,1,0,,,,2026-05-26T21:40:20Z,2026-06-01T14:40:50Z,37,"Stack from ghstack (oldest at bottom): #185287 -> #185286 #185285 #185284 #185283 #185282 The engine now records the version counter of each gradient tensor when it is placed into an InputBuffer (via add()), checks it before accumulation, and verifies it again when the next node consumes the buffer. This detects in-..."
185217,Fix Debug/RelWithDebInfo PyTorch+XPU on Windows,https://github.com/pytorch/pytorch/pull/185217,draft/noise,13,PR is draft,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,185215,,,,0,6,0,,,"module: mkldnn, open source",2026-05-26T12:19:37Z,2026-06-01T12:10:10Z,37,"Fix for #185215. Building and using PyTorch with Windows Debug and RelWithDebInfo configuration with XPU support, exposed several Windows-specific issues: debug Python ABI/runtime mismatches (/MD vs /MDd, Py_DEBUG propagation), oneDNN CRT mismatch in debug builds, incorrect SYCL toolkit import library selection for..."
185518,[inductor] Add probability checks for Bernoulli lowering,https://github.com/pytorch/pytorch/pull/185518,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185246,185246,185246,,,,0,3,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-28T14:14:44Z,2026-06-01T07:24:06Z,37,"Summary This updates Inductor's handling of aten.bernoulli.default to emit an explicit lowering instead of relying on the generic decomposition. The new lowering preserves the previous decomposition semantics (rand < p, including non-floating probability inputs), while adding the missing runtime probability range ch..."
174198,Adjust test_linalg_lstsq_input_check to be compatible with other devices,https://github.com/pytorch/pytorch/pull/174198,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers; CI lane labels: ciflow/xpu,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,2817,2817,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/xpu",2026-02-03T14:58:50Z,2026-06-01T07:16:27Z,37,Fixes: intel/torch-xpu-ops#2817 Make test_linalg_lstsq_input_checks more universal as the CUDA specific check should be under if device == 'cuda' and not under if device != 'cpu' as it makes more sense and resolves xpu specific issues with this UT.
185787,[WIP] Migrate XPU test runners to linux.xpu.test and disable legacy jobs,https://github.com/pytorch/pytorch/pull/185787,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/xpu,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing, ciflow/xpu",2026-06-01T03:29:49Z,2026-06-01T07:12:16Z,37,"Summary Disable legacy XPU CI jobs (linux-jammy-xpu-n-1, client build/test, windows XPU builds) Migrate linux-noble-xpu-n test shards from linux.idc.xpu to linux.xpu.test runners Test Plan CI validation via this PR. Co-authored-by: AI assistant"
180496,feat: Add OpenTelemetry trace export support for torch.profiler,https://github.com/pytorch/pytorch/pull/180496,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,153614,153614,153614,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-04-15T19:47:31Z,2026-06-01T01:53:04Z,38,"Summary Adds opentelemetry_trace_handler() to torch.profiler that converts PyTorch profiler events into OpenTelemetry spans, enabling seamless integration with any OTel-compatible observability backend (Jaeger, Zipkin, OTLP, Datadog, etc.). Fixes #153614 Motivation As discussed in #153614, PyTorch's profiler produce..."
185361,[pooling] Reject negative divisor_override in avg_pool2d/3d,https://github.com/pytorch/pytorch/pull/185361,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185247,185247,185247,,,,0,1,0,,,"module: cpu, triaged, open source, release notes: quantization, ciflow/mps, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-27T12:47:44Z,2026-06-01T01:37:44Z,38,"What does this PR do? Fix segmentation fault in torch.linalg.eigvals triggered by invalid avg_pool3d divisor. Root Cause F.avg_pool3d creates Negative Values: When passing divisor_override=-1, PyTorch's pooling kernel sums the values in the window and divides them by -1. Because there is no check preventing negative..."
182924,[Fix] DLPack toDLPackNonOwning: handle MPS storage_offset,https://github.com/pytorch/pytorch/pull/182924,ready for maintainer decision,70,approved with no detected failing/pending checks,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,open source,2026-05-08T02:08:58Z,2026-05-31T15:02:25Z,38,"Summary toDLPackNonOwning writes src.data_ptr() into DLTensor.data and sets byte_offset = 0. For MPS tensors this is wrong: PyTorch's MPS allocator encodes id<MTLBuffer> directly into c10::DataPtr.data in aten/src/ATen/mps/MPSAllocator.mm:768-770, so for any sliced/viewed MPS tensor data_ptr() returns id<MTLBuffer>..."
185263,Fix torch.compile normal tensor std validation,https://github.com/pytorch/pytorch/pull/185263,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185248,185248,185248,,,,0,1,0,,,"triaged, open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-26T18:47:27Z,2026-05-31T05:25:25Z,38,Fixes #185248 Summary Add tensor std validation in the torch.normal decomposition. Preserve eager behavior for negative tensor standard deviations under torch.compile. Add an Inductor CPU regression test. Test plan python3 -m py_compile torch/_refs/__init__.py test/inductor/test_cpu_repro.py git diff HEAD~1 HEAD --c...
185712,"Fix typos in comments, docstrings, and strings across the codebase",https://github.com/pytorch/pytorch/pull/185712,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: quantization, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-30T16:06:06Z,2026-05-30T20:18:47Z,39,"Stack from ghstack (oldest at bottom): -> #185712 Correct a variety of spelling, grammar, and punctuation errors in comments, docstrings, error messages, and documentation strings. These range from simple misspellings (e.g. ""encloding"" -> ""enclosing"", ""iteam"" -> ""item"", ""versuib"" -> ""version"") to grammatical fixes (..."
185713,"Fix typos in comments, docstrings, and error messages",https://github.com/pytorch/pytorch/pull/185713,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/h100-symm-mem, ciflow/inductor, ciflow/mps","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: quantization, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (dtensor), release notes: inductor (aoti), ciflow/h100-symm-mem, ciflow/torchtitan, ciflow/dtensor",2026-05-30T16:12:41Z,2026-05-30T19:27:43Z,39,"Stack from ghstack (oldest at bottom): (to be filled) Fix spelling and grammar errors across torch/_dynamo, torch/_inductor, torch/_ops, torch/_utils, torch/ao, torch/backends, torch/cuda, torch/distributed, torch/export, and torch/storage. Notably fixes a variable name typo (""supgraph"" -> ""subgraph"") in graph_dedup..."
185705,"Fix typos in comments, docstrings, and error messages",https://github.com/pytorch/pytorch/pull/185705,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: amp (automated mixed precision), release notes: distributed (rpc), module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-30T13:32:11Z,2026-05-30T16:46:33Z,39,"Stack from ghstack (oldest at bottom): (to be filled) Fix spelling and grammar issues across torch, including misspelled words (reconstrtuction, regione, Jacboian, oppurtunies, etc.), incorrect articles (a/an), missing/extra words in error messages, and broken RST cross-references. Authored with Claude (typo_termina..."
185048,[Bugfix] fix `torch.library._register_fake` crashes when any unrelated `LazyLoader` module in `sys.modules` fails to import,https://github.com/pytorch/pytorch/pull/185048,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,185046,185046,185046,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-05-24T05:16:08Z,2026-05-30T16:17:35Z,39,"fix #185046 Root cause of #185046 torch/_library/utils.py: def get_source(stacklevel: int) -> str: frame = inspect.getframeinfo(sys._getframe(stacklevel)) source = f""{frame.filename}:{frame.lineno}"" return source Only frame.filename and frame.lineno are used. But inspect.getframeinfo does much more than needed: it c..."
185565,upstream torch.comms,https://github.com/pytorch/pytorch/pull/185565,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/docker, ciflow/h100-distributed, ciflow/h100-symm-mem, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,11,0,,,"release notes: distributed (c10d), ciflow/h100-distributed, ciflow/h100-symm-mem, ciflow/torchtitan, ciflow/docker",2026-05-28T22:13:03Z,2026-05-30T09:28:11Z,39,"Upstream torchcomms (gloo + nccl) into torch.comms Adds the torchcomms communication library to the PyTorch tree as torch.comms, limited to the gloo and nccl backends. C++ lives in torch/csrc/comms, the Python package in torch/comms, and tests in test/comms. A new USE_TORCH_COMMS CMake option (default ON when USE_DI..."
185666,Add ROCm CI workflows for mi210,https://github.com/pytorch/pytorch/pull/185666,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, triaged, open source, topic: not user facing",2026-05-29T20:57:51Z,2026-05-30T06:58:45Z,39,"Summary Add ROCm CI workflows for mi210 Workflows rocm-mi210 — 1 shard(s), runner linux.rocm.gpu.mi210.1 (testing mode) Ciflow tags ciflow/rocm-mi210/* Test plan Verify workflows parse and actionlint passes cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @prag..."
184859,Deprecate c10::ssize as alias for std::ssize,https://github.com/pytorch/pytorch/pull/184859,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-22T07:07:16Z,2026-05-30T03:13:49Z,40,"Summary c10::ssize was kept in-tree as a C++17-friendly stand-in for C++20's std::ssize. This PR keeps c10::ssize as a [[deprecated]] alias that forwards to std::ssize. New code is steered to std::ssize, out-of-tree code keeps compiling, and the unit test is dropped since the symbol no longer has its own implementat..."
185125,Remove unused caffe2/utils/string_utils,https://github.com/pytorch/pytorch/pull/185125,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-25T09:11:54Z,2026-05-30T01:29:45Z,40,The header was included by three binaries/*_torch.cc files that define their own local split() and never call any caffe2:: symbol from it. No other callers exist in-tree. Authored by Claude.
185339,[not-for-land][aten] Add cuBLAS/cublasLt grouped GEMM backends for _foreach_mm,https://github.com/pytorch/pytorch/pull/185339,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182631, 183270",,,0,5,0,,,"ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-27T08:22:09Z,2026-05-29T22:43:48Z,40,Stack from ghstack (oldest at bottom): -> #185339 #182631 #185530 #183270 Experimental alternative backends for _foreach_mm benchmarking: cublasGemmGroupedBatchedEx (CUDA 12.5+): toggled via TORCH_FOREACH_MM_CUBLAS=1. Uses device pointer arrays with pinned-memory async H2D copy. 1.2-2.4x slower than CUTLASS. cublasL...
177823,[Inductor] Cap num_warps for non-1x1 Triton conv kernels to fix incorrect results,https://github.com/pytorch/pytorch/pull/177823,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,173068,173068,174929,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-03-19T08:00:57Z,2026-05-29T21:44:31Z,40,"Summary Fixes #173068 Non-1x1 convolution kernels use a flattened single-loop pattern (UNROLL=False) that triggers a Triton compiler bug (triton-lang/triton#1254) when num_warps >= 8, producing incorrect results (max_diff=132+ with 84.7% mismatched elements). The root cause is in Triton's code generation for the fla..."
185358,[PrivateUse1] Add dedicated LSTM dispatch branch for PU1 backends.,https://github.com/pytorch/pytorch/pull/185358,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,170737,170737,170737,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: PrivateUse1",2026-05-27T12:28:00Z,2026-05-29T21:41:39Z,40,"Previously, PrivateUse1 tensors in the top-level lstm() functions had no explicit dispatch branch, they fell through all backend-specific checks (cuDNN, MIOpen, oneDNN, MPS) before reaching the generic fallback path. Additionally, LSTMCell and GRUCell unconditionally routed PrivateUse1 tensors to _thnn_fused_lstm_ce..."
185579,Improve error message in linalg.eigh for large matrices,https://github.com/pytorch/pytorch/pull/185579,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,92141,92141,92141,,,,0,2,0,,,"triaged, open source",2026-05-29T01:51:07Z,2026-05-29T18:28:37Z,40,"When calling linalg.eigh on a large matrix (e.g., 53000 x 53000), the LAPACK workspace size computation 1 + 6*n + 2*n*n overflows int32, causing a cryptic MKL error: Intel MKL ERROR: Parameter 8 was incorrect on entry to SSYEVD. RuntimeError: false INTERNAL ASSERT FAILED at .../BatchLinearAlgebra.cpp:... linalg.eigh..."
184335,[spmd_types] pipeline parallel support for spmd_types-annotated tensors,https://github.com/pytorch/pytorch/pull/184335,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,181398,,,0,4,0,,,"release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-05-19T08:09:21Z,2026-05-29T18:14:50Z,40,"Send spmd_types annotations over PP ranks with _SPMDTensorMeta, re-annotating on recv side. no_typechecks PP internals, and also re-applies metadata after input splits. Stack from ghstack (oldest at bottom): -> #184335 #181519 #181398 #180880"
184142,Fix LPPool signed fractional inputs,https://github.com/pytorch/pytorch/pull/184142,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: cpp",2026-05-18T01:17:43Z,2026-05-29T17:19:31Z,40,"Fixes LPPool1d, LPPool2d, and LPPool3d for signed inputs with non-integer norm_type. Lp pooling should compute over abs(input) ** p; using input ** p produces NaNs for negative values when p is fractional. This updates the Python and C++ frontend implementations to use absolute values before exponentiation. This als..."
183535,[MPS] fix older gpu gen mac matmul failures,https://github.com/pytorch/pytorch/pull/183535,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/mps,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,177116,177116,177116,,,,0,3,0,,,"triaged, open source, topic: bug fixes, release notes: mps, ciflow/mps",2026-05-13T09:17:45Z,2026-05-29T16:12:23Z,40,Fixes #177116 Another MPS Graph issue...
185256,Fix PendingUnbackedSymbolNotFound for tensor-derived slice bounds,https://github.com/pytorch/pytorch/pull/185256,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185185,185185,185185,,,,0,1,0,,,"module: cpu, triaged, open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-26T18:03:00Z,2026-05-29T15:38:10Z,40,"Fixes #185185 _compute_slice_index returns None for unbacked symbolic integers because no guard_or_false check can resolve statically. This forces slice_forward to create fresh unbacked symbols for size and storage offset. In chained slices (e.g. x[:, :, h0:h1, w0:w1]), the intermediate storage offset symbol gets su..."
185454,"fix clamp and clamp_{min,max} gradient at scalar bound to be 0",https://github.com/pytorch/pytorch/pull/185454,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184574,184574,,,,,0,1,0,,,"triaged, open source",2026-05-28T01:20:35Z,2026-05-29T15:35:41Z,40,"Issue Fixes #184574 Summary Scalar clamp / clamp_min / clamp_max returned subgradient 1 at their bound, while relu / hardtanh return 0. The backward used the closed interval (self >= min, self <= max): switching to strict inequalities (>, <) puts the boundary on the saturated side, so clamp(x, lo, hi) now matches ha..."
185231,[C++] Ensure wrapped tensors can be printed,https://github.com/pytorch/pytorch/pull/185231,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185200,185200,185200,,,,0,1,0,,,"triaged, open source",2026-05-26T15:00:50Z,2026-05-29T14:55:50Z,40,E.g. a functorch::TensorWrapper cannot be printed as it errors out with Cannot access data pointer of Tensor that doesn't have storage Instead the underlying tensor needs to be printed. This adds a recursive unwrapper and uses it for printing. Note that e.g. a tensor.to(...).contiguous() on an unwrapped tensor wraps...
182289,[MPS] Pinned memory - Add pinned memory allocator,https://github.com/pytorch/pytorch/pull/182289,ready for maintainer decision,45,CI lane labels: ciflow/mps; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-04T14:54:29Z,2026-05-29T12:36:07Z,40,"It enables CPU pinned tensors for MPS, updates non-blocking CPU<->MPS copies to alias shared storage when possible, and removes the DataLoader warning that previously disabled pin_memory for MPS. Tested locally with: python test/test_accelerator.py -k pin_memory_on_non_blocking_copy python test/test_dataloader.py -k..."
184698,Materialize lazy modules fully on load_state_dict instead of leaving them in a partially initialized state.,https://github.com/pytorch/pytorch/pull/184698,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,73009,73009,73009,,,,0,2,0,,,"triaged, open source",2026-05-21T11:39:18Z,2026-05-29T09:45:11Z,40,"Materialize lazy modules fully on load_state_dict instead of leaving them in a partially initialized state, covering all 13 built-in lazy module types.(Fixes #73009) Factoring finalization into reusable _materialize() on LazyModuleMixin. Added _lazy_load_pre_materialize_hook() for subclasses to set metadata (in_feat..."
180742,[autograd] lazily start worker threads in the autograd engine,https://github.com/pytorch/pytorch/pull/180742,ready for maintainer decision,80,approved with no detected failing/pending checks; PR closes or references open issue numbers,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,91898,91898,91898,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-04-18T15:04:10Z,2026-05-29T09:36:24Z,40,"Before this change, the first call to backward() would spawn one OS thread per GPU device, even if only one GPU was ever used. On an 8-GPU machine, a single-GPU job created 8 device threads with 7 sitting idle for the entire run, wasting resources and adding latency to the first backward pass. This PR fixes that by..."
184164,Fix while_loop compile with empty carry and mutation-only state,https://github.com/pytorch/pytorch/pull/184164,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,1,1,statusCheckRollup,statusCheckRollup,182036,182036,182036,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-18T06:47:30Z,2026-05-29T08:25:29Z,40,Fixes #182036 To allow zero schema returns for while_loop_op only in auto_functionalize. Wiring MutationOutput for buffer mutations when body_fn returns (). Fixing wrapper codegen when there are no outer carried inputs. Adding test_while_loop_compile_empty_carry_mutation (cpu/cuda). cc @voznesenskym @penguinwu @Eika...
185178,[FSDP][AC] Bypass SAC tracking for all-gather internals,https://github.com/pytorch/pytorch/pull/185178,draft/noise,13,"PR is draft; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-05-26T01:09:20Z,2026-05-29T07:01:24Z,40,"Fixes a selective activation checkpointing replay mismatch seen in TorchTitan DeepSeek V3 with FSDP+EP. TorchTitan tracking issue: pytorch/torchtitan#3427 During non-reentrant checkpoint recompute, SAC installs a TorchDispatchMode to track ops for replay. FSDP pre-forward hooks may run during that recompute and unsh..."
185548,[test] Migrate in-test OpInfo skip to OpInfo skip DB,https://github.com/pytorch/pytorch/pull/185548,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/inductor",2026-05-28T17:55:52Z,2026-05-29T06:28:59Z,40,Stack from ghstack (oldest at bottom): -> #185548 Follows-up #182999
183269,[DTensor] Switch test_embedding_ops to DTensorOpTestBase (MultiThreadedTestCase),https://github.com/pytorch/pytorch/pull/183269,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/dtensor,review when ready,,UNKNOWN,False,0,0,,,108744,108744,108744,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/dtensor",2026-05-11T12:19:12Z,2026-05-29T05:58:52Z,40,"Summary Converts TestEmbeddingOp in test/distributed/tensor/test_embedding_ops.py from DTensorTestBase (MultiProcessTestCase) to DTensorOpTestBase (MultiThreadedTestCase), following the pattern already established in test_pointwise_ops.py. Changes: Replace DTensorTestBase with DTensorOpTestBase as base class Add Loc..."
184577,nn: validate num_groups > 0 in GroupNorm.__init__,https://github.com/pytorch/pytorch/pull/184577,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/dtensor,review when ready,,UNKNOWN,False,0,0,,,184280,184280,184280,,,,0,2,0,,,"triaged, open source, release notes: nn, ciflow/dtensor",2026-05-20T17:37:28Z,2026-06-03T10:48:25Z,35,"Description GroupNorm(num_groups=0, ...) raises a cryptic ZeroDivisionError from the divisibility check: ZeroDivisionError: integer division or modulo by zero Negative values also silently pass the check and produce wrong results. Fix Add an explicit num_groups <= 0 guard before the modulo operation so the error is..."
183870,[stats] Handle alternate test-reports path in backfill_test_jsons_while_running,https://github.com/pytorch/pytorch/pull/183870,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-15T08:50:11Z,2026-05-29T05:58:43Z,40,"Summary backfill_test_jsons_while_running hardcodes test/test-reports as the base path when computing relative paths for XML → JSON conversion. Some jobs (e.g. test-osdc-crossref) emit artifacts directly under test-reports/ (no test/ prefix), causing: ValueError: '...test-osdc-crossref-.../test-reports/...' is not i..."
185570,tmp,https://github.com/pytorch/pytorch/pull/185570,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-29T00:37:03Z,2026-05-29T04:59:41Z,40,Stack from ghstack (oldest at bottom): -> #185570 #185292
184253,[DO NOT MERGE][ROCm][CI] Test k8s dpx,https://github.com/pytorch/pytorch/pull/184253,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"module: rocm, triaged, open source, ciflow/trunk, topic: not user facing, keep-going, ci-no-td",2026-05-18T19:00:11Z,2026-05-29T04:59:10Z,40,Authored with assistance from Cursor. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
177445,Add SymInt and SymFloat to Number,https://github.com/pytorch/pytorch/pull/177445,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"module: typing, Merged, Reverted, ciflow/trunk, release notes: quantization, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-03-14T04:38:43Z,2026-05-29T02:21:36Z,41,Stack from ghstack (oldest at bottom): #177447 #177446 -> #177445 cc @lolpack @maggiemoss @ndmitchell @kinto0 @samwgoldman @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
182655,Update ACL/OpenBLAS/manywheel build scripts,https://github.com/pytorch/pytorch/pull/182655,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, release notes: releng, topic: not user facing, no-runner-experiments",2026-05-06T14:40:48Z,2026-05-29T02:18:09Z,41,Update: this essentially just replicates #167829 now. Previously it also tried to add ccache support to the ACL and OpenBLAS install scripts (from #169569) but that has now been removed. cc: @robert-hardwick @aditew01
182068,Make custom op aliasing check warn (not error) in CI,https://github.com/pytorch/pytorch/pull/182068,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,182006,182006,,,,,0,1,0,,,"ciflow/trunk, topic: not user facing",2026-04-30T19:08:36Z,2026-05-29T02:03:32Z,41,"Set error_on_custom_op_aliasing to False unconditionally so CI runs and local runs produce the same UserWarning instead of CI hard-erroring. Previously, error_on_custom_op_aliasing=bool(os.getenv(""CI"")) caused the check to raise RuntimeError under CI=1 while only warning elsewhere, which surfaced surprising failures..."
177170,[FSDP] Fix alias annotation for fsdp::all_gather_copy_in,https://github.com/pytorch/pytorch/pull/177170,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,170986,,,,,0,7,0,,,"module: bc-breaking, Merged, Reverted, ciflow/trunk, release notes: distributed (fsdp), topic: bc breaking, ciflow/inductor, suppress-bc-linter, ci-no-td, ciflow/torchtitan",2026-03-11T19:29:13Z,2026-05-29T01:57:56Z,41,"Stack from ghstack (oldest at bottom): -> #177170 The fsdp::all_gather_copy_in op mutates all_gather_output (via foreach_copy into views of it) and returns two tensors that alias it, but the schema declared no mutation or aliasing. This could lead to silent incorrectness under torch.compile. Fix the schema to use Te..."
178314,Separate tagging path for AC when non-strict tracing,https://github.com/pytorch/pytorch/pull/178314,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,13,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-03-24T19:04:42Z,2026-05-29T01:57:49Z,41,Stack from ghstack (oldest at bottom): -> #178314 Add SAC ac_graph_id tagging on top of the HOP tagging (tag_activation_checkpoint_impl) into _CachingTorchDispatchMode so both the HOP path (torch.compile) and the make_fx path get ac_graph_id through the same code path. Share _ac_graph_id_counter between wrap.py and...
178416,[dynamo] Migrate allow_in_graph → nonstrict_trace in torch/ production code,https://github.com/pytorch/pytorch/pull/178416,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,2,0,,,"release notes: sparse, module: dynamo, ciflow/inductor, module: compiled autograd, ciflow/torchtitan",2026-03-25T20:01:46Z,2026-05-29T01:57:47Z,41,"Stack from ghstack (oldest at bottom): #180616 -> #178416 #178340 We would like to deprecate allow_in_graph in favor of nonstrict_trace because they have the same functionality from user's perspecitve i.e. skip dynamo tracing. They have different unlying implmentations though, nonstrict_trace uses flat_apply HOP ins..."
177283,RFC : WIP warmup compile with fake tensor with dynamic dims then run with real,https://github.com/pytorch/pytorch/pull/177283,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,2,0,,,"module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-03-12T18:02:19Z,2026-05-29T01:57:45Z,41,"Stack from ghstack (oldest at bottom): -> #177283 one possible way to address: I want to mark batch size dynamic but get a single compilation even when there are graph breaks and/or eager code in between compile regions. is the following: Warm up the model first using fake tensors that have dynamic-size inputs, and..."
181526,Replace obsolete TORCHDYNAMO_DYNAMIC_SHAPES env var in torch.compile FAQ,https://github.com/pytorch/pytorch/pull/181526,waiting on contributor,53,review decision requests changes; PR closes or references open issue numbers,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,180467,180467,180467,,,,0,1,0,,,"triaged, open source, topic: not user facing, release notes: dynamo",2026-04-26T22:15:58Z,2026-05-29T01:46:26Z,41,"Summary The torch.compile FAQ tells users to enable/disable dynamic shapes via the TORCHDYNAMO_DYNAMIC_SHAPES environment variable, which no longer exists. Replace both occurrences with the supported torch.compile(..., dynamic=...) API form. Fixes #180467. Test plan Docs-only change; no code paths touched. CI doc bu..."
180828,Reuse ACCELERATOR_TYPE in test UT files,https://github.com/pytorch/pytorch/pull/180828,ready for maintainer decision,70,"approved with no detected failing/pending checks; CI lane labels: ciflow/dtensor, ciflow/h100-symm-mem, ciflow/inductor, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"180954, 181833",,,0,11,0,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), ciflow/h100-symm-mem, ciflow/dtensor",2026-04-20T09:27:28Z,2026-05-29T01:10:22Z,41,Stack from ghstack (oldest at bottom): -> #180828 #181833 #180954 #180827 Motivation This PR reuse ACCELERATOR_TYPE in UTs to avoid duplicated code. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aa...
183023,[Inductor] Filter triton configs by launch geometry in triton heuristics,https://github.com/pytorch/pytorch/pull/183023,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/xpu",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,182928,182928,182928,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/xpu, ciflow/torchtitan",2026-05-09T02:39:31Z,2026-05-28T19:17:24Z,41,"Implements the idea proposed in #182928: Implement the function to filter out extremely bad configs based on the grid size, cta work and the number of ctas. Keep the smallest grid size if all the configs are filtered. Use the function to filter configs at the end of pointwise, reduction, persistent_reduction, cooper..."
182414,fix(decomp): cast padding to int in replication/reflection pad decomposition,https://github.com/pytorch/pytorch/pull/182414,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,182339,182339,,,,,0,1,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, ciflow/inductor, ci-no-td",2026-05-05T05:47:09Z,2026-05-28T17:01:06Z,41,"Summary Fixes #182339 When use_deterministic_algorithms(True) routes F.pad(mode='replicate') through the decomposition path, @pw_cast_for_opmath promotes all tensor arguments -- including 0-d int64 padding elements in the padding tuple -- to float via increase_prec. The resulting float index tensor from torch.arange..."
184163,[Test]Refactor test/nn/test_dropout.py to be device-agnostic,https://github.com/pytorch/pytorch/pull/184163,ready for maintainer decision,70,approved with no detected failing/pending checks; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-05-18T06:36:34Z,2026-05-28T16:46:02Z,41,Moves test_native_dropout_corner_case from TestDropoutNN to TestDropoutNNDeviceType to enable device-generic testing across all accelerators. Changes: Moved test_native_dropout_corner_case to TestDropoutNNDeviceType class Added device parameter and runtime skip for CPU variant (test compares accelerator vs CPU) Remo...
176989,Add supports for itertools.batched and itertools.combinations_with_re…,https://github.com/pytorch/pytorch/pull/176989,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,175869,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: dynamo",2026-03-10T04:45:25Z,2026-05-28T16:30:16Z,41,Adds Dynamo polyfill support for itertools.batched and itertools.combinations_with_replacement in torch/_dynamo/polyfills/itertools.py #175869 .venv/bin/python test/dynamo/test_functions.py FunctionTests.test_itertools_batched FunctionTests.test_itertools_combinations_with_replacement .venv/bin/python test/dynamo/cp...
184426,Support getting CUDA patch version,https://github.com/pytorch/pytorch/pull/184426,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: python_frontend, topic: bc breaking",2026-05-19T20:22:22Z,2026-05-28T15:24:58Z,41,This adds a torch.version.cuda_build command accessible via Python and a TORCH_CUDA_VERSION_BUILD macro accessible via C++. This functionality already existed via CUDAToolkit_VERSION_PATCH in the CMake files but was not exposed anywhere. Useful for checking the CUDA version used at build time and for guarding edge c...
185407,[RFC][dynamo] Narrow GB6297 to only graph-break on differentiable duplicate tensor inputs,https://github.com/pytorch/pytorch/pull/185407,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-27T17:12:00Z,2026-05-28T15:19:54Z,41,"Summary: D105712843 added GB6297 which graph-breaks whenever the same tensor is passed to multiple forward inputs of autograd.Function.apply(). This was a correctness fix for a real gradient-loss bug where Dynamo's HOP deduplication silently drops one backward gradient contribution. However, the check is overly broa..."
183909,changing Sleef_asinhf8_u10 to std impl,https://github.com/pytorch/pytorch/pull/183909,draft/noise,61,PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,183767,183767,,,,,0,2,0,,,"module: cpu, open source, topic: not user facing, module: inductor, ciflow/inductor",2026-05-15T15:25:06Z,2026-05-28T14:56:18Z,41,fixes #183767 I am posting this as a draft to make it clear what the problem is and serve as a place holder. This is a numerical library problem not an inductor problem IMO. Reverting to std asinh has obvious performance implications . this will match eager behavior since it calls std::asinh. static void asinh_kerne...
148023,torch.utils.checkpoint preserves torch function mode stack during recompute,https://github.com/pytorch/pytorch/pull/148023,has linked issue,35,PR closes or references open issue numbers,review when ready,,CLEAN,False,0,0,,,,147995,147995,,,,0,7,0,,,"release notes: autograd, topic: bug fixes, no-stale",2025-02-26T23:04:34Z,2026-05-28T12:47:26Z,41,"Stack from ghstack (oldest at bottom): -> #148023 #146633 Fixes #147995 TorchFunctionModeTLS is part of the autograd tls, but because .backward() itself is a leaf for TorchFunctionMode, the mode is disabled before we enter into the engine. Conversely, since TorchDispatchMode traces through the .backward() python cal..."
152159,Add dynamo config to HOP-ify context managers,https://github.com/pytorch/pytorch/pull/152159,ready for maintainer decision,45,CI lane labels: ciflow/inductor; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,,,,0,21,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, no-stale",2025-04-25T03:17:09Z,2026-05-28T12:46:52Z,41,"Stack from ghstack (oldest at bottom): -> #152159 # Note [Hopifying Context Managers] # # If the context manager class has been added to a opt-in dynamo config, # we will convert it into a generic context manager HOP. When the # HOP is later called in AOTAutograd, it will run the captured # graph under the ctx. # #..."
185364,[not-for-land][native_dsl] Add _grouped_mm_from_ptrs pointer-packing path,https://github.com/pytorch/pytorch/pull/185364,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182631, 183270, 185339, 185340, 185363",,,0,1,0,,,"release notes: foreach_frontend, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-27T12:54:58Z,2026-05-28T08:58:03Z,41,"Stack from ghstack (oldest at bottom): #183271 #185340 #185339 -> #185364 #185363 #183270 #182631 Add aten::_grouped_mm_from_ptrs — takes CPU/CUDA int64 tensors of device data pointers, wraps them as from_blob tensors in C++, stacks into 3D, and delegates to _grouped_mm. This reduces Python-to-C++ roundtrips from 3..."
185172,Add support to custom devices to operator_benchmark.,https://github.com/pytorch/pytorch/pull/185172,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/op-benchmark,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, topic: not user facing, ciflow/op-benchmark",2026-05-25T22:28:23Z,2026-05-28T08:57:01Z,41,This PR modifies operator_benchmark scripts to support custom/arbitrary devices. Key Changes: Custom Device Support: Removed the hardcoded list of supported devices (_supported_devices in benchmark_utils.py) Using torch.get_device_module() for device-specific functions Repoting unimplemented operators: Detecting uni...
185363,[not-for-land][native_dsl] Python _foreach_mm override via stack + _grouped_mm,https://github.com/pytorch/pytorch/pull/185363,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182631, 183270, 185339, 185340, 185364",,,0,1,0,,,,2026-05-27T12:54:51Z,2026-05-28T07:42:14Z,41,"Stack from ghstack (oldest at bottom): #183271 #185340 #185339 #185364 -> #185363 #183270 #182631 Native DSL override for aten::_foreach_mm that stacks the TensorList inputs into contiguous 3D tensors, calls _grouped_mm in 3D x 3D mode, and unbinds the output back into a list (zero-copy views). Includes a zero-copy..."
185457,[Inductor][xpu] Enable grouped_mm Triton template on XPU with TMA disabled,https://github.com/pytorch/pytorch/pull/185457,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-05-28T02:09:49Z,2026-05-28T07:41:19Z,41,Stack from ghstack (oldest at bottom): -> #185457 Enable the persistent grouped_mm Triton template (triton_mm_grouped) on Intel XPU. The template uses pointer-based loads instead of TMA descriptors on XPU due to a known bug in triton-xpu (intel/intel-xpu-backend-for-triton#7030) where 3D tl.make_tensor_descriptor wi...
184141,Remove pointer_to_optional_list,https://github.com/pytorch/pytorch/pull/184141,waiting on contributor,43,review decision requests changes; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, ciflow/inductor, release notes: inductor (aoti)",2026-05-18T01:13:36Z,2026-05-28T03:00:31Z,42,pointer_to_optional_list returned an ArrayRef into a temporary vector when T != U. This PR removes it and adjusts torchgen/gen_aoti_c_shim.py
182256,[MPS] enable metal performance primitives attention,https://github.com/pytorch/pytorch/pull/182256,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-03T22:07:47Z,2026-05-28T03:00:04Z,42,"adds metal performance primitives matmuls in attention., batch size 1, bf16 perf gains: num_heads=16 head_dim qL MPP (ms) non-MPP (ms) speedup 64 1024 0.458 1.015 2.22x 64 2048 0.777 3.304 4.25x 64 16000 44.261 183.754 4.15x 64 32000 193.606 742.045 3.83x 64 64000 930.675 2999.505 3.22x 96 1024 0.319 1.428 4.48x 96..."
177568,[MTIA] Add dilation support for max_pool2d_with_indices_backward lowering (#177568),https://github.com/pytorch/pytorch/pull/177568,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, module: inductor, ciflow/inductor, meta-exported, ciflow/inductor-pallas, ciflow/torchtitan",2026-03-16T20:12:07Z,2026-05-28T01:35:09Z,42,"Summary: Re-land of D95566134 (reverted in D96091876, D96273115) with a fix for the 3D tensor regression. The Inductor lowering for aten.max_pool2d_with_indices_backward fell back to FallbackKernel when dilation != 1 (marked as ""NYI""). This caused compilation failures on the target backend since the fallback op is u..."
185418,[coor] Represent get_rank as a SymInt,https://github.com/pytorch/pytorch/pull/185418,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"release notes: distributed (c10d), module: dynamo, ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-27T18:17:30Z,2026-05-27T23:42:33Z,42,"Stack from ghstack (oldest at bottom): -> #185418 compile_on_one_rank assumes rank 0 can produce a graph that is valid across the group. Dynamo normally constant folds torch.distributed.get_rank(), so rank-dependent Python control flow can be baked into the graph before the artifact is reused on other ranks. Instead..."
185376,[not-for-land][native_dsl] Add _foreach_mm_from_ptrs pointer-bridge and py_ptrs override,https://github.com/pytorch/pytorch/pull/185376,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182631, 183270, 185339, 185340",,,0,1,0,,,"release notes: foreach_frontend, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-27T15:01:28Z,2026-05-27T23:06:36Z,42,"Stack from ghstack (oldest at bottom): #183271 -> #185376 #185340 #185339 #183270 #182631 Add aten::_foreach_mm_from_ptrs -- a thin C++ bridge that receives device data pointer values as a CPU int64 tensor from Python, wraps them as from_blob tensors (metadata only, no data copy), stacks into 3D, and delegates to at..."
184302,Add Stream-K MM backend support to Inductor,https://github.com/pytorch/pytorch/pull/184302,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/inductor-rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, module: inductor, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/inductor-rocm-mi300, ciflow/torchtitan",2026-05-19T01:07:57Z,2026-05-27T21:13:51Z,42,Add Stream-K GEMM candidates to Inductor MM lowering. Use Origami to select the base stream-K config. Add optional Stream-K autotune variants. Support force-Stream-K mode through env flags. Port Stream-K stores to current template output hooks. Add chiplet PID transform support for ROCm. Track autotune best kernel f...
184064,Fix silent int64 precision loss in F.pad constant mode,https://github.com/pytorch/pytorch/pull/184064,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,170798,170798,170798,,,,0,2,0,,,"triaged, open source, release notes: nn",2026-05-17T02:41:21Z,2026-05-27T20:32:14Z,42,"Issue Fixes #170798 Summary F.pad declares value as float?, so a Python int whose magnitude exceeds 2**53 is silently truncated when it crosses into the C++ binding. For integer tensors with an integer fill value, dispatch directly to constant_pad_nd, which takes a Scalar and preserves the value. native_functions.ya..."
185340,[not-for-land][native_dsl] Add Python _foreach_mm override via _grouped_mm,https://github.com/pytorch/pytorch/pull/185340,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"182631, 183270, 185339, 185376",,,0,1,0,,,"release notes: foreach_frontend, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-27T08:22:14Z,2026-05-27T15:01:38Z,42,"Stack from ghstack (oldest at bottom): #183271 #185376 -> #185340 #185339 #183270 #182631 Native DSL override for aten::_foreach_mm that delegates to aten::_grouped_mm. Stacks TensorList into 3D, calls _grouped_mm, unbinds output. Includes zero-copy fast path for pre-stacked inputs. 1.1-1.8x slower than C++ CUTLASS..."
184192,[Test]Refactor test/nn/test_lazy_modules.py to be device- generic by replacing hardcoded device selection with torch.accelerator API,https://github.com/pytorch/pytorch/pull/184192,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-18T11:54:38Z,2026-05-27T14:51:08Z,42,"Replace hardcoded device selection with torch.accelerator API in test_materialize_device to enable testing on all accelerators (CUDA, XPU, HPU, MPS, PrivateUse1). Changes Replace if TEST_CUDA: device = ""cuda"" elif TEST_PRIVATEUSE1: device = ... with device = torch.accelerator.current_accelerator() Update device comp..."
175746,RISC-V Vector Extension (RVV) support for ATen,https://github.com/pytorch/pytorch/pull/175746,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, triaged, open source, topic: not user facing, module: risc-v",2026-02-25T14:29:51Z,2026-05-27T12:59:11Z,42,"This is a revised and updated version of zhangfeiv0's RISC-V Vector Extension support for ATen: #135570 https://github.com/zhangfeiv0/pytorch/tree/vec-rvv (original author's feature branch) The main improvement of this version is an adjustment of the Vectorized<> type internal representation, which now allows to com..."
184531,[Inductor] Tolerance increase for test_bmm_large_batch_dynamic_cuda,https://github.com/pytorch/pytorch/pull/184531,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,179958,179958,,,,,0,4,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ci-no-td, ciflow/rocm-mi300, ciflow/rocm-mi355, ciflow/rocm-navi31, ciflow/rocm-mi200, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355",2026-05-20T13:03:08Z,2026-05-27T12:58:53Z,42,"FIXES #179958 Loosen accuracy tolerance for test_bmm_large_batch_dynamic (fp16 BMM on ROCm). Other callers of check_model_with_multiple_inputs are unaffected (default remains 1e-4, bit-equivalent to same()'s built-in default). Test logic is still preserved as any regression that happens with batch handling will go o..."
178106,Support Optional[list[torch.Tensor]] in infer_schema,https://github.com/pytorch/pytorch/pull/178106,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,160094,160094,160094,,,,0,1,0,,,,2026-03-23T01:02:00Z,2026-05-27T12:58:49Z,42,Fixes #160094 Testing python3 -m pytest test/test_custom_ops.py
183738,Multi-warp TMA scatter_add via PerWarpTmaPipeline,https://github.com/pytorch/pytorch/pull/183738,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,183739,,,0,5,0,,,topic: not user facing,2026-05-14T16:57:55Z,2026-05-27T12:57:44Z,42,"Stack from ghstack (oldest at bottom): #183739 -> #183738 Adds 8 warps per CTA, each running an independent producer/consumer pipeline over its own row stream. Stock cutlass.pipeline.PipelineTmaAsync gates mbarrier_init on warp_idx == 0 only, so multiple pipes per CTA only get warp 0's barriers initialized; quack an..."
185227,Error out when getting symmetric memory device pointers for NCCL < 2.28,https://github.com/pytorch/pytorch/pull/185227,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/h100-symm-mem,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,185216,185216,185216,,,,0,1,0,,,"triaged, open source, release notes: distributed (c10d), ciflow/h100-symm-mem",2026-05-26T14:36:34Z,2026-05-27T08:38:20Z,42,Those pointers are never set and hence stay NULL. Using them in CUDA kernels causes CUDA_ERROR_ILLEGAL_ADDRESS. Instead throw a descriptive error at the place of the issue. Skip all tests that made use of this. Fixes #185216
175758,Create a DeviceModule protocol,https://github.com/pytorch/pytorch/pull/175758,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,110080,,,,0,4,0,,,"module: cpu, triaged, open source",2026-02-25T17:29:25Z,2026-05-27T08:01:58Z,42,Proof of concept following the conversation on #110080 The main point of this new protocol is to make it easier for users to write device agnostic code by providing a strongly typed common interface for all device types. I've tried to keep changes to the backends to a minimum by only adding functions to the protocol...
183501,[PyTorch][Pinned Memory] Fix stuck event blocking in CachingHostAllocator,https://github.com/pytorch/pytorch/pull/183501,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, meta-exported",2026-05-13T03:07:16Z,2026-05-27T06:42:14Z,42,"Summary: CONTEXT: In CachingHostAllocator, the process_events() function stops processing the entire event queue when it encounters a single incomplete CUDA event. This means one stuck or slow event can block ALL other completed blocks from being returned to the free list, causing massive pinned memory accumulation..."
178031,[MPS] Add adaptive_avg_pool3d support for MPS,https://github.com/pytorch/pytorch/pull/178031,has linked issue,35,"PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/mps",review when ready,,UNKNOWN,False,0,0,,,141287,141287,141287,,,,0,4,0,,,"triaged, open source, release notes: mps, ciflow/mps, ciflow/inductor, release notes: inductor (aoti)",2026-03-21T01:11:17Z,2026-05-27T05:44:35Z,42,Add MPS support for adaptive_avg_pool3d and its backward path. Fixes #141287 (comment)
184607,"Speed up inductor test infrastructure (~4x collection, ~1.7x execution)",https://github.com/pytorch/pytorch/pull/184607,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,181617,,,0,2,0,,,"fb-exported, ciflow/trunk, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-20T20:31:40Z,2026-05-27T04:01:30Z,42,Stack from ghstack (oldest at bottom): -> #184607 Several independent optimizations to reduce inductor test overhead: ISA subprocess caching (cpu_vec_isa.py): Cache the result of CPU vector ISA load-check subprocesses using .load_ok marker files. Each check_build() call spawns a Python subprocess that imports torch...
185284,Add checkpoint_name() API for explicit tensor naming in SAC,https://github.com/pytorch/pytorch/pull/185284,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"185285, 185286, 185287",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-26T21:40:05Z,2026-05-27T01:25:54Z,43,Stack from ghstack (oldest at bottom): #185287 #185286 #185285 -> #185284 #185283 #185282 Authored with Claude.
185287,Add memory_budget() context manager for per-region activation memory budgets,https://github.com/pytorch/pytorch/pull/185287,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"185284, 185285, 185286",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-26T21:40:27Z,2026-05-27T01:25:20Z,43,"Stack from ghstack (oldest at bottom): -> #185287 #185286 #185285 #185284 #185283 #185282 Adds a memory_budget(budget) context manager that controls the recomputation vs. memory trade-off in the min-cut partitioner on a per-region basis under torch.compile. The implementation delegates to fx_traceback.annotate, whic..."
185285,Allow debug mode with context_fn in checkpoint(),https://github.com/pytorch/pytorch/pull/185285,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"185284, 185286, 185287",,,0,1,0,,,ciflow/torchtitan,2026-05-26T21:40:13Z,2026-05-27T00:21:48Z,43,"Stack from ghstack (oldest at bottom): #185287 #185286 -> #185285 #185284 #185283 #185282 Previously, passing both debug=True and a custom context_fn to torch.utils.checkpoint raised a ValueError. This composes the two by stacking the debug logging contexts on top of the user's contexts via a new _compose_context_fn..."
184420,"Fix pow(-0.0, +/-0.5) sign on sqrt/rsqrt shortcut",https://github.com/pytorch/pytorch/pull/184420,PR blocked,73,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,127163,127163,127163,,,,0,1,0,,,"module: cpu, triaged, open source, release notes: mps, ciflow/mps, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-19T19:14:19Z,2026-05-26T23:49:52Z,43,"Fixes #127163 pow(x, +/-0.5) routed through the sqrt / rsqrt fast path preserved the sign of -0.0, so pow(-0.0, 0.5) returned -0.0 and pow(-0.0, -0.5) returned -inf. According to IEEE-754 / Array API both should be positive. Fix: kernel(x + 0.0). -0.0 + 0.0 == +0.0 canonicalizes the sign; every other finite input is..."
178204,[PyTorch][OSS][AOTAutograd] Allow None values in saved tensors for vc_check split (#178204),https://github.com/pytorch/pytorch/pull/178204,waiting on contributor,43,"review decision requests changes; CI lane labels: ciflow/inductor, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, module: dynamo, ciflow/inductor, meta-exported",2026-03-23T22:16:55Z,2026-05-26T23:23:31Z,43,"Summary: When a compiled forward function (e.g., via inductor) returns None for certain saved activations — such as optional parameters like bias=None — the AOT autograd runtime previously crashed with: AssertionError: expected all tensors_saved_with_vc_check to be Tensors This happened because the assertion in Comp..."
184991,[inductor] Add memory-aware graph-level CSE pass,https://github.com/pytorch/pytorch/pull/184991,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,1,0,,,"release notes: releng, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T15:03:17Z,2026-05-26T22:23:37Z,43,"Summary Adds a Common Subexpression Elimination pass at the inductor post-grad level that identifies and folds identical computations (same op, same input nodes, same kwargs) into a single node. Motivation Many models produce identical computations that inductor doesn't deduplicate: LLaVA: 32 identical where(mask, 0..."
185097,Migrate deprecated torch.norm calls to torch.linalg.*,https://github.com/pytorch/pytorch/pull/185097,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: quantization, ciflow/torchtitan",2026-05-25T02:38:21Z,2026-05-26T20:05:15Z,43,torch.norm is deprecated. Replace 22 call sites with the new torch.linalg.vector_norm API. Authored by Claude.
182384,[autograd] Refresh ctx.needs_input_grad for partial backward,https://github.com/pytorch/pytorch/pull/182384,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,174017,174017,"166577, 174017",,,,0,1,0,,,"triaged, open source",2026-05-04T22:48:58Z,2026-05-26T19:44:04Z,43,"Fixes #174017. Alternative to #177070 (which makes ctx.needs_input_grad a dynamic getter); this PR refreshes the tuple once at backward entry instead. When torch.autograd.backward(tensors=loss, inputs=[a]) trims unreachable edges, native ops correctly skip the unneeded gradient viaNode::task_should_compute_output()...."
185253,[c10d] Make FakeProcessGroup reduce collectives numerically consistent,https://github.com/pytorch/pytorch/pull/185253,draft/noise,51,PR is draft; failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,release notes: distributed (c10d),2026-05-26T17:59:14Z,2026-05-26T19:37:58Z,43,Stack from ghstack (oldest at bottom): -> #185253 FakeProcessGroup previously produced wrong numerical results for reduce collectives: allreduce was a no-op and reduce_scatter just copied a chunk without any reduction. This made FakePG unsuitable for single-process numerics verification of distributed training graph...
179830,Fix upsample_nearest3d stride mismatch under torch.compile (#179272),https://github.com/pytorch/pytorch/pull/179830,waiting on contributor,53,"review decision requests changes; PR closes or references open issue numbers; CI lane labels: ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,179272,179272,179272,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, keep-going, release notes: inductor, ciflow/inductor-pallas, ciflow/torchtitan",2026-04-09T15:59:51Z,2026-05-26T17:22:28Z,43,"When F.interpolate(mode=""nearest"") is applied to a non-contiguous (channels_last_3d) 5D tensor, torch.compile produces a contiguous output while eager mode preserves the input's stride order. Root cause: the _upsample_nearest decomposition in torch/_decomp/decompositions.py decomposes all upsample_nearest variants i..."
183739,TMA scatter_add: 2D-tile multi-row path for small-N shapes,https://github.com/pytorch/pytorch/pull/183739,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,CLEAN,False,0,0,,,,,,183738,,,0,5,0,,,topic: not user facing,2026-05-14T16:57:59Z,2026-05-26T16:55:12Z,43,"Stack from ghstack (oldest at bottom): -> #183739 #183738 When the entire row fits in one TMA chunk (num_chunks == 1, i.e. row_bytes <= 512), each pipeline stage now holds K=4 rows loaded with a single 2D TMA descriptor (cta_tiler=(K, N)) instead of one row per stage. The consumer issues K bulk-reduces back-to-back..."
184822,Add Sphinx redirects for renamed autosummary Stream/stream pages,https://github.com/pytorch/pytorch/pull/184822,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,173318,173318,173318,,,,0,1,0,,,"open source, topic: not user facing",2026-05-21T23:28:07Z,2026-05-26T16:10:55Z,43,"Summary Fixes #173318. autosummary_filename_map in docs/source/conf.py (introduced by #164901) renames torch.{cuda,cpu,mtia,xpu}.Stream and the lowercase stream context manager pages to *_class.html / *_function.html to disambiguate case-only collisions. The rename leaves the original generated/*.html URLs as 404s o..."
185137,[fx][_library] Honor wrapped op's schema.is_mutable in is_impure's auto_functionalized HOP branch (#185137),https://github.com/pytorch/pytorch/pull/185137,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, meta-exported",2026-05-25T13:52:25Z,2026-05-26T15:09:04Z,43,"Summary: torch._library.utils.is_impure (consumed by torch.fx.Node.is_impure and by every FX pass that asks ""can I optimize across this node?"") has two parallel branches that should answer the same question for the same underlying op: OpOverload branch (lines 653-656): correctly consults schema.is_mutable, which cat..."
185091,Prefer .view() and drop redundant .contiguous() in tensor-creator chains,https://github.com/pytorch/pytorch/pull/185091,PR blocked,63,"failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, release notes: quantization, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-25T00:49:07Z,2026-05-26T13:51:40Z,43,"Replace .reshape() with .view() on 11 expression chains where the receiver is contiguous by construction, and drop one redundant .contiguous() on a torch.cat(...) result. The receivers in each case are produced by ops that always return contiguous tensors: torch.tensor(...) / torch.arange(...) / torch.cat(...) / tor..."
185026,[typing] Add Iterable base class to ModuleList,https://github.com/pytorch/pytorch/pull/185026,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,185022,185022,185022,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-05-23T23:31:51Z,2026-05-26T13:51:35Z,43,"🐛 Describe the bug Type checkers (astral/ty, pyright) report not-iterable when iterating over nn.ModuleList, despite __iter__ being properly defined. Root cause The @_copy_to_script_wrapper decorator on __iter__ obscures the return type from static analysis, preventing structural recognition of the Iterable protocol..."
184631,[MPS] Skip strided NDArray path in _mps_linear_nograph when inputs are contiguous,https://github.com/pytorch/pytorch/pull/184631,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,181725,181725,181725,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-21T00:37:42Z,2026-05-26T13:45:10Z,43,"Summary Fixes #181725. _mps_linear only takes the no-graph path when input, weight, and bias are all contiguous (Linear.mm:117). But _mps_linear_nograph still forwards t.strides() to getMPSNDArray for all three, pushing every F.linear down the strided arrayViewWithShape:strides: + permuteNDArray path. That's an extr..."
184885,docs: correct torch.addcdiv prose to match implementation order,https://github.com/pytorch/pytorch/pull/184885,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,170869,170869,170869,,,,0,1,0,,,"triaged, open source, release notes: python_frontend",2026-05-22T12:46:05Z,2026-05-26T13:44:56Z,43,"Fixes #170869. The first sentence of the torch.addcdiv docstring describes the operation as dividing tensor1 by tensor2 first and then multiplying the quotient by value, which suggests the evaluation order (tensor1 / tensor2) * value. The CPU and CUDA pointwise kernels in aten/src/ATen/native/cpu/PointwiseOpsKernel...."
184857,"Raise ValueError early when strategy=""forward-mode"" used without vectorize=True",https://github.com/pytorch/pytorch/pull/184857,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184842,184842,184842,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-22T06:33:33Z,2026-05-26T13:44:24Z,43,"jacobian() and hessian() accept strategy=""forward-mode"" with vectorize=False (the default) at call time but then raise NotImplementedError deep inside _jacfwd at execution time. This moves the check to argument validation so users get a clear ValueError with actionable guidance immediately. Also fixes docstring typo..."
184866,[distributed] Handle empty inputs and kwargs in _to_kwargs,https://github.com/pytorch/pytorch/pull/184866,ready for maintainer decision,45,CI lane labels: ciflow/torchtitan; open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: distributed (fsdp), ciflow/torchtitan",2026-05-22T08:26:26Z,2026-05-26T13:37:56Z,43,"Summary: Fix _to_kwargs so modules with no positional inputs and no keyword inputs still get a single empty call entry instead of empty input/kwarg lists. Review: @rohan-varma @weifengpy, would you mind taking a look when you have a chance? Your guidance on the distributed/FSDP input handling here would be greatly a..."
185193,Validate forward-mode functional autograd arguments,https://github.com/pytorch/pytorch/pull/185193,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184842,184842,184842,,,,0,1,0,,,"triaged, open source",2026-05-26T06:55:37Z,2026-05-26T13:36:34Z,43,"Fixes #184842 This PR adds early validation for forward-mode autograd functional APIs that require vectorize=True. Changes: Raises ValueError in torch.autograd.functional.jacobian when strategy=""forward-mode"" is used with vectorize=False. Raises ValueError in torch.autograd.functional.hessian when outer_jacobian_str..."
184840,Add Module.to overload for memory_format,https://github.com/pytorch/pytorch/pull/184840,has linked issue,35,PR closes or references open issue numbers,review when ready,,UNKNOWN,False,0,0,,,184642,184642,184642,,,,0,2,0,,,"module: nn, module: typing, triaged, enhancement, open source",2026-05-22T03:30:18Z,2026-05-26T13:35:02Z,43,Summary Add the missing nn.Module.to(memory_format=...) overload. Add typing coverage for documented Module.to call forms. Test plan python -m mypy --config-file mypy.ini test/typing/pass/nn_module.py` fixes #184642 cc @albanD @mruberry @jbschlosser @walterddr @mikaylagawarecki @lolpack @maggiemoss @ndmitchell @kint...
184509,[torch/csrc] Add bulk storage helpers; remove per-byte loops,https://github.com/pytorch/pytorch/pull/184509,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-20T08:05:41Z,2026-05-26T13:31:38Z,43,This PR add bulk operations for storage.
184657,[MPS] Tolerate lenient mixed-dtype batch_norm in InstanceNorm test,https://github.com/pytorch/pytorch/pull/184657,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, release notes: mps, ciflow/mps",2026-05-21T05:12:08Z,2026-05-26T13:29:59Z,43,
165837,[OpenReg][Random] Enrich OpenReg random number generator module implementation add focused documentation,https://github.com/pytorch/pytorch/pull/165837,PR blocked,63,failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, ciflow/trunk, topic: not user facing, module: openreg",2025-10-18T14:19:07Z,2026-05-26T13:28:08Z,43,"Summary This PR enriches the OpenReg (PRIVATEUSE1) random number generation (RNG) module with: Generator-aware random kernels for PrivateUse1 (rand, randn, and partial randint support) Per-device generator plumbing and convenience Python APIs under torch.openreg Focused documentation and targeted tests validating de..."
185219,[cherry-pick] no topological sort in manual overlaps,https://github.com/pytorch/pytorch/pull/185219,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,185218,,,0,2,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-26T12:23:18Z,2026-05-26T12:36:13Z,43,Stack from ghstack (oldest at bottom): -> #185219 #185218 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
185218,[not-for-land][squash] Regional inductor cache fix,https://github.com/pytorch/pytorch/pull/185218,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,185219,,,0,2,0,,,"module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-05-26T12:23:13Z,2026-05-26T12:36:12Z,43,Stack from ghstack (oldest at bottom): #185219 -> #185218 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98
159718,[Don't Review] Test XPU CI,https://github.com/pytorch/pytorch/pull/159718,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,43,0,,,"open source, Stale, ciflow/trunk, topic: not user facing, no-stale, ciflow/xpu",2025-08-02T15:31:23Z,2026-05-26T10:08:57Z,43,Stack from ghstack (oldest at bottom): -> #159718 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben
181993,[MPS] fix device binary shaders,https://github.com/pytorch/pytorch/pull/181993,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,181650,181650,181650,,,,0,3,0,,,"triaged, open source, topic: bug fixes, release notes: mps, ciflow/mps",2026-04-30T06:56:07Z,2026-05-26T10:06:18Z,43,Fixes #181650 I'm still confused on how the user ran into that error
183752,[BE]: Revert back to upstream uv action to solve rate-limiting,https://github.com/pytorch/pytorch/pull/183752,ready for maintainer decision,45,open non-draft PR has no obvious automated blocker,review diff and decide acceptance path,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ci-no-td",2026-05-14T18:42:31Z,2026-05-26T04:42:21Z,43,Reverts to upstream UV to properly mitigate the rate limiting as an alternative to #183643 . It now properly uses GH Actions API logged into to make the call.
184909,[Inductor][Draft] A Helion backend for Inductor,https://github.com/pytorch/pytorch/pull/184909,draft/noise,51,"PR is draft; failed checks or blocked/unstable merge state; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"release notes: releng, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T17:10:14Z,2026-05-25T23:36:49Z,44,"Stack from ghstack (oldest at bottom): -> #184909 #184908 Motivation Extend Inductor's Pallas coverage. The inductor -> pallas path is sparsely supported; the helion -> pallas path is much more mature. Routing Inductor through Helion gives us Pallas (TPU + CPU interpret) ""for free"" on the cases Helion already handle..."
182576,Improved static Type Hints for dataloader,https://github.com/pytorch/pytorch/pull/182576,waiting on contributor,43,review decision requests changes,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,179707,,,,0,4,0,,,"triaged, open source, release notes: dataloader",2026-05-05T22:06:29Z,2026-05-25T15:01:16Z,44,"General description Added a second covariant type var _T_Collated_co Added a second typevar to the DataLoader Added two overloads of DataLoader.__init__ (one with known collate_fn, one without) Annotated the Iterators as yielding the _T_Collated_co instance. This PR is linked to #179707 (Improve static type hints fo..."
184254,"Fix `replication_pad{2d,3d}` meta accepting negative output dims",https://github.com/pytorch/pytorch/pull/184254,PR blocked,63,failed checks or blocked/unstable merge state,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source, release notes: nn, topic: bug fixes",2026-05-18T19:02:12Z,2026-05-25T12:12:57Z,44,"Summary The shape-meta for replication_pad2d validates that negative padding does not shrink the output to zero or below with TORCH_CHECK(owidth >= 1 || oheight >= 1, ...); The || should be &&. The check only catches the case where both dims are too small, not the one where a single dim went negative. replication_pa..."
179223,[Test] Refactor test_higher_order_ops.py to be device-generic,https://github.com/pytorch/pytorch/pull/179223,draft/noise,13,"PR is draft; CI lane labels: ciflow/dynamo, ciflow/inductor",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/dynamo",2026-04-03T06:00:07Z,2026-05-25T06:33:10Z,44,Refactor test/dynamo/test_higher_order_ops.py to be device-generic by replacing hardcoded CUDA/GPU references with torch.accelerator API to enable tests to run on all accelerators including out-of-tree devices via PrivateUse1. Changes in test/dynamo/test_higher_order_ops.py: HigherOrderOpTests (2 methods): test_wrap...
184358,Add XPU to externalId,https://github.com/pytorch/pytorch/pull/184358,draft/noise,15,PR is draft; PR closes or references open issue numbers; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,3479,3479,,,,,0,1,0,,,open source,2026-05-19T11:58:37Z,2026-05-25T00:03:21Z,45,Fixes: intel/torch-xpu-ops#3479 Updates the mapping between libkineto::ChromeTraceLogger::handleActivity() and KinetoEvent::externalId() after following change: pytorch/kineto#1390
178513,Fix interpolate nearest mode with non-integer scale_factor,https://github.com/pytorch/pytorch/pull/178513,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,UNKNOWN,False,0,0,,,175154,175154,175154,,,,0,2,0,,,"triaged, open source, ciflow/inductor",2026-03-26T15:31:07Z,2026-05-24T17:45:53Z,45,"When using torch.nn.functional.interpolate with mode='nearest' and a non-integer scale_factor (e.g., 1.3), the decomposition was producing incorrect results. The issue was that the code was using truncation (.to(torch.int64)) instead of floor() to compute the input indices. This fix: Uses torch.floor() to properly c..."
185044,[AI SLOP] Support SymInt list custom op returns,https://github.com/pytorch/pytorch/pull/185044,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-05-24T03:55:15Z,2026-05-24T07:25:53Z,45,"Stack from ghstack (oldest at bottom): -> #185044 Allow custom operators annotated as list[int] to infer SymInt[] returns and carry unbacked SymInt bindings through export and Inductor. The missing pieces were schema inference for list[int], cpp-wrapper naming/materialization for scalar-list return slots, and mixed..."
185011,[inductor] Pair maxpool value and offset lowering,https://github.com/pytorch/pytorch/pull/185011,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T21:03:27Z,2026-05-24T06:14:38Z,45,"Summary: Add a small-window maxpool-with-offsets lowering path that carries pooled value and offset as one paired comparator state instead of lowering them as separate value/argmax work. Keep the path conservative: static kernel/window, supported padding/ceil cases, and fallback to the existing lowering otherwise. A..."
185036,[inductor] Prototype channel-resident reduction suffix,https://github.com/pytorch/pytorch/pull/185036,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,2,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-24T01:36:23Z,2026-05-24T06:14:35Z,45,"Submitted by my agent. Summary This is a draft prototype for better-benchmark issue #45. It adds a default-off channel-resident reduction suffix path for a narrow Triton reduction shape where the generated kernel stores a full suffix value to the output buffer, reloads it later in the same kernel, and then overwrite..."
185037,[inductor] Use persistent path for small fused online softmax,https://github.com/pytorch/pytorch/pull/185037,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,2,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-24T01:55:54Z,2026-05-24T06:14:34Z,45,"Submitted by my agent. Summary Use the persistent reduction path for small fused online_softmax_reduce kernels that are marked ReductionHint.DEFAULT but contain only online-softmax reductions. Some fused masked softmax rows are logically inner reductions, but the fused pointwise/multiple-output shape demotes the red..."
185027,[inductor] Fold CE backward one-hot row sums,https://github.com/pytorch/pytorch/pull/185027,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,3,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T23:37:59Z,2026-05-24T06:14:30Z,45,"Summary: Add a narrow post-grad fold for CE-backward-style one-hot row sums. Fold sum(where(iota(vocab_dim) == target, row_scalar, 0), vocab_dim) to a validity-guarded row scalar when the value is independent of the vocab dimension. Preserve guards for single sum dim, integer target, iota over the reduced dim, zero..."
185028,[inductor] Reassociate cat under surviving reductions,https://github.com/pytorch/pytorch/pull/185028,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,4,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T23:44:19Z,2026-05-24T06:14:26Z,45,"Submitted by my agent. Summary Reassociate channel cat under reductions when the cat dimension survives the reduction. This currently handles: sum(cat(xs, dim=C), reduce_dims) -> cat(sum(x, adjusted_reduce_dims), dim=C) mean(cat(xs, dim=C), reduce_dims) -> cat(mean(x, adjusted_reduce_dims), dim=C) The rewrite avoids..."
184941,[inductor] Enable reduction epilogue fusion by deferring split decision,https://github.com/pytorch/pytorch/pull/184941,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,6,0,,,"release notes: releng, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T22:50:53Z,2026-05-24T04:53:20Z,45,"Summary When a split reduction's output feeds a broadcast pointwise consumer (e.g., batch norm normalize after var_mean), suppress the split so the consumer can fuse into the reduction via fits_in_main_body. The codegen then picks persistent or cooperative reduction internally based on rnumel. This eliminates the re..."
185030,[inductor] Fold masked invariant expand reductions,https://github.com/pytorch/pytorch/pull/185030,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,3,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-24T00:07:16Z,2026-05-24T04:53:15Z,45,"Submitted by my agent. This adds a narrow post-grad rewrite for the pattern: sum(where(mask, 0, expand(base) / K), dims) when the reduced dimensions include large zero-stride expanded lanes and K equals the counted expanded size. The replacement counts unmasked lanes first, then scales the invariant base value, avoi..."
185032,[inductor] Reuse expensive broadcast factors in pointwise tiling,https://github.com/pytorch/pytorch/pull/185032,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-24T00:45:53Z,2026-05-24T04:53:14Z,45,"Submitted by my agent. Draft for better-benchmark issue 51. This adds a narrow pointwise tiling path for RoPE-like [outer, reuse, inner] shapes where expensive broadcast expressions depend on the coalesced inner dimension but not the interior reuse/head dimension. The scheduler selects 3D tiling so the expensive fac..."
184930,[inductor] Score broadcast reuse for expensive pointwise tiling,https://github.com/pytorch/pytorch/pull/184930,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T21:40:34Z,2026-05-24T03:28:29Z,46,"Summary: Score pointwise tiling splits when a split exposes reuse of expensive broadcast expressions. Require actual expensive work depending on broadcast-indexed reads, nontrivial reuse, coalesced suffix traffic, and enough total work. Harden against predicated cat-origin schedules and low-precision small-work regr..."
184932,[inductor] Use non-online softmax for huge inner reductions,https://github.com/pytorch/pytorch/pull/184932,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T21:44:45Z,2026-05-24T03:27:57Z,46,"Summary: Use regular max/sum softmax lowering for huge CUDA inner softmax rows where online softmax is instruction-bound. Keep the guard narrow: CUDA, innermost single reduction dim, xnumel >= 1024, rnumel >= max(8192, persistent threshold + 1), and bounded total input numel. Add focused codegen tests for huge softm..."
184936,[inductor] Avoid masks for zero-stride cat source loads,https://github.com/pytorch/pytorch/pull/184936,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T21:52:45Z,2026-05-24T03:27:52Z,46,"Summary: In pointwise_cat, load zero-stride cat sources directly instead of wrapping them in cat-region masks. Guard narrowly on stride[cat_dim] == 0, because out-of-region pointwise-cat indices only vary in the cat dimension. Add focused coverage for an expanded class-token cat feeding var_mean. Context: better-ben..."
184940,[inductor] Enable reduction+transpose fusion via loop reordering,https://github.com/pytorch/pytorch/pull/184940,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,5,0,,,"release notes: releng, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T22:50:35Z,2026-05-24T02:15:28Z,46,"Summary Enable fusion of a strict reduction + transposed pointwise/shared-input pattern by adding a narrow loop-reorder fallback in _try_reindex_pointwise_for_reduction. Pattern # Input: x [M, N] row-major transpose_out = x.permute(1, 0).contiguous() # [N, M] sum_out = x.sum(dim=0) # [N] The pointwise/transpose cons..."
184929,[inductor] Sink single-use dropout randoms into fused epilogues,https://github.com/pytorch/pytorch/pull/184929,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T21:40:34Z,2026-05-24T01:27:26Z,46,"Summary: Avoid materializing CUDA uniform random tensors when they are only used by a scalar threshold comparison feeding a fused dropout-style epilogue. Keep sinking disabled under align_random_eager, and keep multi-use and non-dropout threshold patterns materialized so RNG is not duplicated or semantically changed..."
184999,[inductor] Avoid pointwise cat for complex output branches,https://github.com/pytorch/pytorch/pull/184999,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T17:52:53Z,2026-05-24T01:26:01Z,46,"Stack from better-benchmark issue 27. This changes cat lowering so complex final-output cats use ConcatKernel instead of pointwise_cat, while keeping pointwise_cat for downstream pointwise consumers. The motivating case is branchy BN-style cats where pointwise_cat evaluates mutually exclusive branch bodies under mas..."
184998,[inductor] Use one warp for tiny inner persistent reductions,https://github.com/pytorch/pytorch/pull/184998,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T17:52:53Z,2026-05-24T01:25:53Z,46,"Stack from better-benchmark issue 27. This routes CUDA persistent INNER reductions with tiny rnumel through the existing one-warp persistent inner config instead of the generic persistent config list. For r=49, this changes the default from two-warps to XBLOCK=8,num_warps=1. Validation: python -m py_compile torch/_i..."
185003,[inductor] Tune small multi-output mix-order reductions,https://github.com/pytorch/pytorch/pull/185003,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T18:53:43Z,2026-05-24T01:25:46Z,46,"Summary: tune small multi-output mix-order reductions by using fewer split chunks for the validated rnumel=512 case postprocess the resulting RSPLIT_SIZE=32 config to use smaller XBLOCK, one stage, and one warp add a heuristic unit test covering the target case and a nearby rnumel=128 case that should keep generic b..."
185012,[inductor] Add FP64-heavy pointwise autotune candidate,https://github.com/pytorch/pytorch/pull/185012,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-23T21:29:14Z,2026-05-24T01:25:36Z,46,"Summary: Count FP64 compute-heavy CSE assignments in Triton pointwise codegen metadata. Add a guarded CUDA 1D pointwise autotune candidate XBLOCK=256,num_warps=8,num_stages=1 only when the generated kernel has enough FP64 compute work. Add guard tests so FP32 and low-op FP64 pointwise kernels do not get the candidat..."
184931,[inductor] Use persistent reduction for R=2048 online softmax,https://github.com/pytorch/pytorch/pull/184931,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T21:40:34Z,2026-05-23T23:33:20Z,46,Summary: raises the persistent-reduction threshold to 2048 only for non-cooperative inner online_softmax_reduce kernels keeps generic inner reductions unchanged adds tests for R=2048 online softmax and for generic R=2048 sum staying non-persistent Context: better-benchmark issue: eellison/better-benchmark#46 clean v...
184935,[inductor] Canonicalize duplicate mean and var reductions,https://github.com/pytorch/pytorch/pull/184935,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T21:49:39Z,2026-05-23T23:02:53Z,46,Summary: Canonicalize exact same-input aten.mean.dim + aten.var.correction pairs to aten.var_mean.correction in post-grad FX. Reuse the shared first reduction for LayerNorm-style graphs that currently emit duplicate mean and variance reductions. Add a focused FX pass test for the rewrite. Related better-benchmark is...
184949,Remove allow-untyped-defs from pyi files,https://github.com/pytorch/pytorch/pull/184949,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: distributed (c10d), ciflow/torchtitan",2026-05-23T00:53:51Z,2026-05-23T17:18:49Z,46,Stack from ghstack (oldest at bottom): -> #184949
184933,[inductor] Fold sparse one-hot sum reductions,https://github.com/pytorch/pytorch/pull/184933,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T21:44:45Z,2026-05-23T16:46:11Z,46,Summary: folds sparse CE-style one-hot reductions in Inductor post-grad FX handles vocab-invariant values as guarded scalar values handles vocab-dependent values as guarded gathers with safe indexing for invalid targets now gates the rewrite to cases that broadcast the target over the reduced vocab dim and have stat...
184777,Refactor test_dlpack for out of tree backend,https://github.com/pytorch/pytorch/pull/184777,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,2,0,,,"open source, topic: not user facing",2026-05-21T18:39:15Z,2026-05-23T14:55:29Z,46,Refactors test_dlpack.py to use @onlyAccelerator decorator and device-agnostic torch.Stream(device=device) calls
184962,[ROCm] Drop version cap on bf16/fp16 mm/bmm accuracy skip,https://github.com/pytorch/pytorch/pull/184962,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/b200, ciflow/h100, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, topic: not user facing, ciflow/rocm-mi300, ciflow/h100, ciflow/b200",2026-05-23T04:51:57Z,2026-05-23T09:13:26Z,46,"test_mm_bmm_dtype_overload skipped affected shapes only for ROCm < 7.2.1, but the hipblas/hipblaslt accuracy regression persists past 7.2.1 on MI200. Drop the version cap so the skip fires unconditionally on ROCm. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77am..."
184937,[inductor] Prototype channel-resident reduction suffix codegen,https://github.com/pytorch/pytorch/pull/184937,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-22T21:56:06Z,2026-05-23T02:58:55Z,47,"Summary: Prototype a guarded Triton path for channel-resident reduction suffixes in BN-backward-shaped mixed reductions. When the reduction can run with a full RBLOCK, keep full-tile values resident across reduction finalization, elide the intermediate scratch store, and emit the dependent full-output suffix without..."
184042,Tma scatter add remove software pipelining,https://github.com/pytorch/pytorch/pull/184042,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,183479,,,0,2,0,,,"triaged, open source, release notes: cuda",2026-05-16T18:44:59Z,2026-05-22T22:08:36Z,47,"Following #182675 Issue: In tma_scatter_add_kernel, both TMA load and reduce are async ops. We need at least 3 smem buffers to fully pipeline the ops to avoid back-to-back wait, e.g. one for load, one for reduce, the last one for issuing a new load without waiting for the completion of the immediately issued load or..."
184756,Fix Dynamo dict equality semantics,https://github.com/pytorch/pytorch/pull/184756,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:04Z,2026-05-22T20:46:16Z,47,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 -> #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Use stored dict key hashes and CPython-style rich c...
184873,Scope third-party install rules out of the wheel via CMake components,https://github.com/pytorch/pytorch/pull/184873,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180247, 180248, 180250",,,0,1,0,,,"module: mkldnn, open source",2026-05-22T09:20:35Z,2026-05-22T18:52:43Z,47,"Stack from ghstack (oldest at bottom): -> #184873 #180250 #180249 #180248 #180247 #180243 Move all install() rules onto named components so scikit-build-core can restrict the wheel to the subset that actually belongs there. The default component is ""libtorch"" (shared libraries, headers, runtime data); explicit ""torc..."
184758,Fix Dynamo defaultdict shallow copy semantics,https://github.com/pytorch/pytorch/pull/184758,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:14Z,2026-05-22T17:43:40Z,47,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 -> #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Teach Dynamo to route defaultdict.copy through the...
184813,Use custom Min Max in Inductor symbolic indexing,https://github.com/pytorch/pytorch/pull/184813,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/inductor-pallas, ciflow/mps, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,184812,,,0,1,0,,,"ciflow/mps, module: inductor, ciflow/inductor, ciflow/inductor-pallas, ciflow/torchtitan",2026-05-21T22:54:05Z,2026-05-22T16:34:35Z,47,Stack from ghstack (oldest at bottom): -> #184813 #184812 Route focused Inductor lowering and index-propagation paths through torch's custom symbolic Min and Max so existing simplification support applies. Update Pallas and MPS printing for the resulting expressions.\n\nAuthored by Codex. cc @voznesenskym @penguinwu...
184489,Forward CXX compiler to nvcc as host compiler for GCC too,https://github.com/pytorch/pytorch/pull/184489,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-05-20T05:59:04Z,2026-05-22T15:43:28Z,47,"cuda.cmake sets CMAKE_CUDA_HOST_COMPILER to CMAKE_CXX_COMPILER only when the C++ compiler is Clang. With GCC, nvcc was left to its default and picks up bare g++, which in my case was GCC 10 even though CXX=g++-14 was set to override it. CUDA 13's CCCL headers (cuda::std::optional and friends) break on GCC 10 with ""i..."
181077,Enable fbgemm on MacOS,https://github.com/pytorch/pytorch/pull/181077,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, triaged, open source, ciflow/trunk, release notes: quantization",2026-04-22T06:35:40Z,2026-05-22T15:42:12Z,47,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
180501,Fix clip_grad_norm_ failure with mixed DeviceMesh DTensor norms,https://github.com/pytorch/pytorch/pull/180501,PR blocked,65,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/dtensor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,180346,180346,180346,,,,0,2,0,,,"triaged, open source, ciflow/dtensor",2026-04-15T20:25:57Z,2026-05-22T14:26:14Z,47,"Summary Fixes #180346 This PR fixes two related failures in clip_grad_norm_ when gradients are DTensors from different DeviceMesh objects: 1. torch.stack on mixed-mesh norms in _get_total_norm() Per-parameter norms are stacked via torch.stack(). When the norms are DTensors from different meshes, torch.stack cannot r..."
182781,[1/N] Use const pointers and references,https://github.com/pytorch/pytorch/pull/182781,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: cpu, triaged, open source, release notes: quantization, release notes: linalg_frontend, module: dynamo, ciflow/inductor, module: compiled autograd, ciflow/rocm-mi300, release notes: inductor (aoti), ciflow/h100, ciflow/b200",2026-05-07T05:35:40Z,2026-05-22T13:34:55Z,47,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98 @xmfan @aditvenk
179397,Modernize map usage in ATen,https://github.com/pytorch/pytorch/pull/179397,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/mps, ciflow/trunk, ciflow/vllm","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,176662,,,,0,3,0,,,"module: cpu, open source, ciflow/trunk, release notes: quantization, topic: not user facing, ciflow/mps, ciflow/vllm",2026-04-05T05:24:20Z,2026-05-22T11:55:18Z,47,"C++20 provides a contains member function for all associative containers. Prior to C++20, the common pattern to check if an element is a member of a container was to either container.find(key) != container.end() or container.count(key) > 0. This PR replaces those instances with contains. PyTorch now supports C++20 #..."
184587,Fix torch.compile addmm failure with beta=0 nonbroadcastable bias,https://github.com/pytorch/pytorch/pull/184587,PR blocked,65,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,183899,183899,183899,,,,0,1,0,,,"open source, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-05-20T18:32:32Z,2026-05-22T07:46:30Z,47,Fixes #183899 Fixes torch compile addmm failure when beta = 0 and bias shape is not broadcastable to output. When beta = 0 bias is ignored (multiplied by 0). So my fix skips bias shape validation in both pattern matching and the lowering paths. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSu...
184724,[XPU][SDPA] Add XPU flash_attention_forward dispatch with dropout fallback,https://github.com/pytorch/pytorch/pull/184724,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,open source,2026-05-21T15:04:29Z,2026-05-22T05:18:57Z,47,Fixes: intel/torch-xpu-ops#3698 This PR enables XPU support for aten._flash_attention_forward and adds a safe fallback path for dropout > 0. Adds XPU dispatch for aten._flash_attention_forward in native_functions.yaml. Implements _flash_attention_forward_xpu in the XPU attention backend. Keeps the existing fast SYCL...
184812,Improve Inductor index simplification,https://github.com/pytorch/pytorch/pull/184812,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,184813,,,0,1,0,,,"release notes: fx, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-21T22:54:01Z,2026-05-22T03:31:02Z,48,"Stack from ghstack (oldest at bottom): #184813 -> #184812 Teach SizeVar simplification a few guarded modular arithmetic identities used by dynamic-shape indexing, including mod/floordiv joins and symbolic-coefficient floor-div expansion through SIMD codegen. Also tighten evaluate_min's GCD fallback so divisibility a..."
184801,Gate baddbmm expectedFailureDynamic on TEST_Z3,https://github.com/pytorch/pytorch/pull/184801,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,162287,,,,0,1,0,,,"topic: not user facing, module: dynamo, ciflow/inductor",2026-05-21T21:40:54Z,2026-05-22T02:16:57Z,48,"Summary The baddbmm subtest of test_scalar_arg_0d_tensor was added in #182660 marked expectedFailureDynamic to track a Z3 translation-validation gap (#162287). However, the dynamic_shapes test class only enables translation_validation when Z3 is importable (test/dynamo/test_dynamic_shapes.py:57): (fx_config, ""transl..."
178176,[DO NOT MERGE]Test New Landing Pg,https://github.com/pytorch/pytorch/pull/178176,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,13,0,,,"triaged, open source, topic: not user facing",2026-03-23T18:09:25Z,2026-05-21T23:38:45Z,48,Testing a new landing page. This is no where near final.
184617,Fix Dynamo sequence repeat identity semantics,https://github.com/pytorch/pytorch/pull/184617,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,4,0,,,"ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-20T22:05:23Z,2026-05-21T22:15:40Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 -> #184617 #184616 #184605 #184634 Validate sequence repeat counts before dispatching...
184755,Fix Dynamo sequence search semantics,https://github.com/pytorch/pytorch/pull/184755,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:00Z,2026-05-21T21:53:45Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 -> #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Dynamo's sequence index fallback now follows CPytho...
184754,Fix Dynamo reversed iterator consumption,https://github.com/pytorch/pytorch/pull/184754,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:31:55Z,2026-05-21T21:50:47Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 -> #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Make Dynamo lower reversed() over unpackable values...
184765,Fix Dynamo ordered contains semantics,https://github.com/pytorch/pytorch/pull/184765,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:45Z,2026-05-21T21:49:04Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 -> #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Teach Dynamo to expose descriptor-backed method nam...
184766,Fix Dynamo set discard and remove semantics,https://github.com/pytorch/pytorch/pull/184766,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:49Z,2026-05-21T21:49:02Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 -> #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Update Dynamo's set discard and remove handling to...
184635,Fix Dynamo cmp_to_key list sort semantics,https://github.com/pytorch/pytorch/pull/184635,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,3,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T01:26:41Z,2026-05-21T21:48:59Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 -> #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Support functools.cmp_to_key in Dynamo list sorting...
184761,Fix Dynamo mutable set dict-key hashing semantics,https://github.com/pytorch/pytorch/pull/184761,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:27Z,2026-05-21T21:48:03Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 -> #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Handle tracked mutable set operations without mater...
184759,Fix Dynamo frozenset dict-key hashing semantics,https://github.com/pytorch/pytorch/pull/184759,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:18Z,2026-05-21T21:48:02Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 -> #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Preserve tracked dict-key and set hash entries thro...
184763,Fix Dynamo dict copy GC tracking semantics,https://github.com/pytorch/pytorch/pull/184763,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:36Z,2026-05-21T21:47:57Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 -> #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Add narrow Dynamo support for gc.is_tracked on fres...
184762,Fix Dynamo list extend iterator semantics,https://github.com/pytorch/pytorch/pull/184762,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:31Z,2026-05-21T21:47:56Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 -> #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Handle user-defined iterables in list.extend with C...
184757,Fix Dynamo dict fromkeys semantics,https://github.com/pytorch/pytorch/pull/184757,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:09Z,2026-05-21T21:47:49Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 -> #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Teach Dynamo to trace dict and dict subclass fromke...
184616,Fix Dynamo list constructor iteration protocol,https://github.com/pytorch/pytorch/pull/184616,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,4,0,,,"ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-20T22:01:23Z,2026-05-21T21:47:36Z,48,"Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 -> #184616 #184605 #184634 Route list() through the generic iterator protocol,..."
184622,Fix Dynamo tuple constructor semantics,https://github.com/pytorch/pytorch/pull/184622,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,3,0,,,"ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-20T22:42:49Z,2026-05-21T21:47:24Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 -> #184622 #184617 #184616 #184605 #184634 Route tuple construction through Dynamo's iterator...
184645,Fix Dynamo defaultdict inplace union semantics,https://github.com/pytorch/pytorch/pull/184645,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,3,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T02:57:00Z,2026-05-21T21:46:35Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 -> #184645 #184627 #184622 #184617 #184616 #184605 #184634 DefaultDictVariable now handles the nb_inplace_or s...
184753,Fix Dynamo frozenset copy identity semantics,https://github.com/pytorch/pytorch/pull/184753,stale/low urgency,17,"PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,CLEAN,False,0,0,,,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:31:50Z,2026-05-21T21:46:27Z,48,"Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 -> #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Exact frozenset.copy now preserves identity, while..."
184627,Fix Dynamo dict update semantics,https://github.com/pytorch/pytorch/pull/184627,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,3,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T00:09:39Z,2026-05-21T21:46:24Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 -> #184627 #184622 #184617 #184616 #184605 #184634 Teach ConstDictVariable.update to handle no-argumen...
184764,Fix Dynamo iterator reduce semantics,https://github.com/pytorch/pytorch/pull/184764,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:40Z,2026-05-21T21:11:34Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 -> #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Add Dynamo support for iterator reduce semantics ac...
184760,Fix Dynamo frozenset subclass dict-key hashing semantics,https://github.com/pytorch/pytorch/pull/184760,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184761, 184762, 184763, 184764, 184765, 184766",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T17:32:22Z,2026-05-21T21:10:39Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 -> #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 #184605 #184634 Delay user-defined class constant-arg checks until...
184605,Add CPython Dynamo agentic loop plan,https://github.com/pytorch/pytorch/pull/184605,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,CLEAN,False,0,0,,,,,,"184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673, 184674, 184753, 184754, 184755, 184756, 184757, 184758, 184759, 184760, 184761, 184762, 184763, 184764, 184765, 184766",,,0,5,0,,,"topic: not user facing, skip-pr-sanity-checks",2026-05-20T20:19:39Z,2026-05-21T19:51:43Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 #184669 #184635 #184766 #184765 #184764 #184763 #184762 #184761 #184760 #184759 #184758 #184757 #184756 #184755 #184754 #184753 #184645 #184627 #184622 #184617 #184616 -> #184605 #184634 Adds the CPython Dynamo expected-failure relevance...
184300,[xplat/caffe2] Add define_aten_lite macros for selective CUDA registration composed on top of pt_operator_registry,https://github.com/pytorch/pytorch/pull/184300,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, topic: not user facing, meta-exported",2026-05-19T00:50:45Z,2026-05-21T18:20:49Z,48,Summary: Adds xplat/caffe2:aten_lite_defs.bzl with two public macros — define_aten_lite_from_yaml and define_aten_lite — that build a link_whole=True selective ATen CUDA registration library composed on top of a CPU-side pt_operator_registry / get_pt_ops_deps library. Each macro emits a gen_aten_files genrule for th...
184536,Skip `TestFullyShardSymmMem` for NCCL < 2.28,https://github.com/pytorch/pytorch/pull/184536,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing",2026-05-20T13:26:03Z,2026-05-21T17:39:56Z,48,NCCL_CTA_POLICY_ZERO was introduced with NCCL 2.28 so skip that test similar to NCCLCopyEngineCollectives.
161044,Thread based dataloading workers,https://github.com/pytorch/pytorch/pull/161044,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,"101699, 135704",,,,0,22,0,,,"module: dataloader, module: data, ciflow/trunk, release notes: dataloader, skip-pr-sanity-checks",2025-08-20T06:43:03Z,2026-05-21T17:05:17Z,48,"Context The current PyTorch Dataloader relies on multiprocessing to create n workers. This works well in standard Python as multiprocessing gets us closest to true parallelism in CPU bound tasks because of the GIL. More recently, the dataloading space has been exploring ways to incorporate parallelism using threads..."
184673,Fix Dynamo dict fromkeys semantics,https://github.com/pytorch/pytorch/pull/184673,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184674",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T07:00:03Z,2026-05-21T16:02:36Z,48,"Stack from ghstack (oldest at bottom): #184674 -> #184673 #184672 #184671 #184670 #184669 #184645 #184635 #184627 #184622 #184617 #184616 #184605 #184634 Teach Dynamo to trace dict and dict subclass fromkeys, including closure-backed local subclass constructors, and add CPython/Dynamo coverage. Remove expected-failu..."
177961,[ROCm] Enable native AsyncTP,https://github.com/pytorch/pytorch/pull/177961,waiting on contributor,35,"review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/h100-symm-mem, ciflow/periodic-rocm-mi300, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,11,0,,,"module: rocm, triaged, open source, ciflow/trunk, release notes: distributed (c10d), ciflow/periodic-rocm-mi300, ciflow/h100-symm-mem, ciflow/torchtitan",2026-03-20T14:01:11Z,2026-05-21T14:38:28Z,48,"Enables the native asynchronous Tensor Parallel (fused_all_gather_matmul) path on ROCm, using AMD's Composable Kernel (CK) tile library as the GEMM backend. This mirrors the existing CUTLASS-based CUDA path. Kernel implementation (AsyncMM.cu): Adds async_input_mm_impl_ck_tile<LayoutB>() , a CK persistent GEMM kernel..."
182811,Reenable test_addcmul_alpha_one_fma_parity dtypes F32 and F64 on XPU.,https://github.com/pytorch/pytorch/pull/182811,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-05-07T11:19:26Z,2026-05-21T14:14:04Z,48,To be merged after intel/torch-xpu-ops#3275 is merged and reference updated.
184672,Fix Dynamo dict equality semantics,https://github.com/pytorch/pytorch/pull/184672,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184673, 184674",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T06:59:59Z,2026-05-21T11:24:17Z,48,"Stack from ghstack (oldest at bottom): #184674 #184673 -> #184672 #184671 #184670 #184669 #184645 #184635 #184627 #184622 #184617 #184616 #184605 #184634 Use stored dict key hashes and CPython-style rich comparison dispatch when Dynamo evaluates dict equality. This preserves lookup operand ordering, subtype-priority..."
184671,Fix Dynamo sequence search semantics,https://github.com/pytorch/pytorch/pull/184671,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184672, 184673, 184674",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T06:59:55Z,2026-05-21T11:24:14Z,48,"Stack from ghstack (oldest at bottom): #184674 #184673 #184672 -> #184671 #184670 #184669 #184645 #184635 #184627 #184622 #184617 #184616 #184605 #184634 Dynamo's sequence index fallback now follows CPython's identity-then-item-richcompare behavior, which fixes list.remove and related index semantics for custom equa..."
184670,Fix Dynamo reversed iterator consumption,https://github.com/pytorch/pytorch/pull/184670,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184671, 184672, 184673, 184674",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T06:59:51Z,2026-05-21T11:24:13Z,48,Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 -> #184670 #184669 #184645 #184635 #184627 #184622 #184617 #184616 #184605 #184634 Make Dynamo lower reversed() over unpackable values to a mutable iterator so list(reversed_dict) consumes the iterator state correctly. Add focused dict coverage a...
184669,Fix Dynamo frozenset copy identity semantics,https://github.com/pytorch/pytorch/pull/184669,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,174929,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184670, 184671, 184672, 184673, 184674",,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-21T06:59:47Z,2026-05-21T11:24:12Z,48,"Stack from ghstack (oldest at bottom): #184674 #184673 #184672 #184671 #184670 -> #184669 #184645 #184635 #184627 #184622 #184617 #184616 #184605 #184634 Exact frozenset.copy now preserves identity, while frozenset subclass copies still produce a distinct base frozenset. Internal set binary operations clone base sto..."
184674,fixup! Add CPython Dynamo agentic loop plan,https://github.com/pytorch/pytorch/pull/184674,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"184605, 184616, 184617, 184622, 184627, 184635, 184645, 184669, 184670, 184671, 184672, 184673",,,0,1,0,,,topic: not user facing,2026-05-21T07:00:08Z,2026-05-21T10:14:30Z,48,Stack from ghstack (oldest at bottom): -> #184674 #184673 #184672 #184671 #184670 #184669 #184645 #184635 #184627 #184622 #184617 #184616 #184605 #184634
177636,Refactor platform support utilities,https://github.com/pytorch/pytorch/pull/177636,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/b200, ciflow/dtensor, ciflow/h100, ciflow/inductor",wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,10,0,,,"open source, release notes: distributed (c10d), module: inductor, module: dynamo, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/dtensor",2026-03-17T10:06:26Z,2026-05-21T09:06:13Z,48,"This pull request refactors how platform capability constants (such as PLATFORM_SUPPORTS_FP8 and PLATFORM_SUPPORTS_FLASH_ATTENTION) are defined and imported in the test suite. The main change is moving the evaluation logic and constant definitions from common_cuda.py to a new file, common_gpu.py, and updating all te..."
183922,Unskip test_vector_norm on AArch64,https://github.com/pytorch/pytorch/pull/183922,PR blocked,65,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,125438,125438,,,,0,3,0,,,"open source, ciflow/trunk, topic: not user facing",2026-05-15T16:50:08Z,2026-05-21T08:20:29Z,48,Looks like these tests are passing now so unskip them Stack from ghstack (oldest at bottom): -> #183922 Fixes #125438
182269,[MPS] Fix channels_last batch_norm backward,https://github.com/pytorch/pytorch/pull/182269,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-05-04T03:30:12Z,2026-05-21T06:25:07Z,48,"Summary Extends the existing inputTensor NHWC→NCHW reshape+transpose dance in batch_norm_backward_mps to gradOutputTensor and gradInputTensor, so channels-last inputs route through the same graph path symmetrically. Also drops the now-unnecessary .contiguous() from the 5D→4D flatten added in #180335, since the chann..."
181060,[fft] Validate _fft_c2r last_dim_size to fix heap-buffer overflow,https://github.com/pytorch/pytorch/pull/181060,PR blocked,65,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/mps, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,141448,141448,,,,,0,1,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, release notes: mps, ciflow/mps, ci-no-td",2026-04-22T02:23:31Z,2026-05-21T05:51:18Z,48,"Fixes #141448. Summary The low-level _fft_c2r op didn't validate that last_dim_size was consistent with the input's last transformed dim. Calling it directly with a mismatch (e.g. input (3,1,3,1) and last_dim_size=536870912) caused a heap-buffer overflow in pocketfft on CPU, or tripped an INTERNAL_ASSERT deep inside..."
178082,Fix ninja workdir collisions in BuildExtension,https://github.com/pytorch/pytorch/pull/178082,waiting on contributor,45,review decision requests changes; PR closes or references open issue numbers; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,177265,177265,177265,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-03-22T03:48:14Z,2026-05-21T03:55:36Z,48,Summary isolate ninja manifests/workdirs for BuildExtension object compilation derive a per-compile ninja subdirectory from the object paths instead of writing directly into the shared setuptools output_dir keep object output paths unchanged so distutils/setuptools behavior stays intact Problem When multiple CppExte...
181073,Validate norm_type in lp_pool1d/2d/3d,https://github.com/pytorch/pytorch/pull/181073,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,134841,134841,134841,,,,0,2,0,,,"triaged, open source",2026-04-22T05:31:47Z,2026-05-20T21:44:52Z,49,Fixes #134841. Summary F.lp_pool1d/2d/3d compute (avg_pool(x**p))**(1/p); with norm_type=0 the 1/p divides by zero and the result is NaN/Inf instead of a useful error. The nn.LPPool* modules forward to these and inherit the bug. Raises ValueError at the top of each functional when norm_type == 0. Check is guarded by...
88221,Add torch.tensor replacement and int_tensor prim,https://github.com/pytorch/pytorch/pull/88221,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, no-stale",2022-11-01T18:22:59Z,2026-05-20T21:13:38Z,49,Stack from ghstack (oldest at bottom): -> #88221
184318,[opaque obj] Check opaque object schema types at dispatch,https://github.com/pytorch/pytorch/pull/184318,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,topic: not user facing,2026-05-19T06:26:26Z,2026-05-20T20:10:14Z,49,"Opaque object schemas now carry the registered Python class name instead of using an unqualified PyObject type. The schema parser rejects raw PyObject spelling, dispatch validates objects against the named schema type, and repeated parses of the same opaque type share the same type object so schema hashes stay consi..."
184151,[CUDA] Fix float32 std/var M2 overflow by using double accumulation,https://github.com/pytorch/pytorch/pull/184151,PR blocked,65,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,180156,180156,,,,0,1,0,,,"triaged, open source, release notes: cuda",2026-05-18T03:18:46Z,2026-05-20T15:57:32Z,49,"Summary torch.std and torch.var return inf on CUDA for float32 inputs with large magnitudes, while CPU returns the correct result. import torch torch.manual_seed(0) x = torch.randn(1000, dtype=torch.float32) * 1e19 + 1e20 print(torch.std(x.double())) # 1.0287e+19 (correct) print(torch.std(x)) # 1.0287e+19 (correct,..."
184257,[scatter_add] Reject misaligned data_ptr / row stride in CuTeDSL override cond,https://github.com/pytorch/pytorch/pull/184257,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"fb-exported, ciflow/trunk, meta-exported",2026-05-18T19:03:29Z,2026-05-20T15:21:15Z,49,"Summary: The CuTeDSL scatter_add override at torch/_native/ops/scatter_add/ (added in D105049595, reverted in D105339314, re-landed in D105407754) currently accepts inputs whose effective data_ptr() or outer row stride is not 16-byte aligned. Both override paths require 16B alignment of the global memory operand: TM..."
184498,[profiler] Propagate record_funtion_id from TorchOp ancestor to Kinet…,https://github.com/pytorch/pytorch/pull/184498,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, release notes: profiler",2026-05-20T06:59:15Z,2026-05-20T14:21:28Z,49,"When profiling GPU workloads, Kineto GPU kernel and CUDA runtime activities do not carry any record_function_id, making it impossible to correlate them back to the originating TorchOp in the profiler output. This matters for tools that consume profiler traces and need to reconstruct the execution graph: without the..."
176893,Add save_state_dict and load_state_dict functions (fixes #150857),https://github.com/pytorch/pytorch/pull/176893,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,150857,150857,150857,,,,0,3,0,,,"triaged, open source, Stale, release notes: cpp, topic: not user facing, no-stale, bot-triaged",2026-03-09T16:43:34Z,2026-05-20T13:58:06Z,49,"Summary This PR adds torch::save_state_dict() and torch::load_state_dict() template functions to enable C++ compatibility with Python's torch.save(model.state_dict()) format. Fixes Closes #150857 Problem Currently, there's no direct way in C++ PyTorch to save/load only the parameters and buffers of a module (state d..."
175537,Fix: Exclude third-party headers from PyTorch wheels,https://github.com/pytorch/pytorch/pull/175537,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,164883,164883,164883,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing, no-stale",2026-02-23T15:12:28Z,2026-05-20T13:56:32Z,49,"Replace wildcard patterns include//*.h, include//.hpp, include/**/.cuh with specific patterns for PyTorch core directories only. This prevents protobuf headers (google/), fmt/, pybind11/, fbgemm/, oneapi/ from being included in wheels, fixing diamond dependency conflicts. Fixes #164883"
174882,Fix torch.histc float16 precision inconsistency,https://github.com/pytorch/pytorch/pull/174882,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,174668,174668,174668,,,,0,2,0,,,"module: cpu, triaged, open source, Stale, topic: not user facing, no-stale",2026-02-12T14:59:58Z,2026-05-20T13:55:21Z,49,Fix torch.histc float16 precision inconsistency Summary Fixes numerical inconsistency in torch.histc with float16 inputs where the histogram additivity invariant histc(x) == histc(x[::2]) + histc(x[1::2]) was violated due to cumulative rounding errors. Problem The issue occurs when torch.histc processes large datase...
180904,Enable MyST Markdown docstrings in Sphinx docs,https://github.com/pytorch/pytorch/pull/180904,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-04-20T20:08:32Z,2026-05-20T12:46:39Z,49,
180888,Update requirements-docs.txt,https://github.com/pytorch/pytorch/pull/180888,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,2,0,,,"open source, topic: not user facing",2026-04-20T18:17:22Z,2026-05-20T12:46:35Z,49,
183925,Covert notes/cuda from .rst to .md,https://github.com/pytorch/pytorch/pull/183925,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, topic: not user facing",2026-05-15T17:00:12Z,2026-05-20T12:46:29Z,49,
178628,[DRAFT]Add TorchInductor documentation to User Guide,https://github.com/pytorch/pytorch/pull/178628,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,12,0,,,"open source, topic: not user facing",2026-03-27T18:55:26Z,2026-05-20T12:16:49Z,49,"Add comprehensive TorchInductor documentation covering the compilation pipeline internals. New pages include: Landing page with pipeline overview Architecture & starter example (relu+add fusion walkthrough) Operator decomposition mechanics FX graph passes (pre-grad, joint, post-grad) Inductor IR system (nodes, buffe..."
179538,Fix incorrect type annotation in Future.set_exception (#179538),https://github.com/pytorch/pytorch/pull/179538,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, fb-exported, meta-exported",2026-04-07T02:15:19Z,2026-05-20T12:16:47Z,49,"Summary: Future.set_exception() parameter result was typed as T (the generic Future value type), but it should be BaseException since this method only accepts exception objects. The docstring already correctly documents the type as BaseException, and the runtime check enforces isinstance(result, Exception). This fix..."
184389,[ROCm] Fix test_origami autotune comparison after PR #181617 cap,https://github.com/pytorch/pytorch/pull/184389,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,181617,,,0,2,0,,,"module: rocm, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/inductor-rocm-mi200, ciflow/inductor-rocm-mi300, ciflow/inductor-rocm-mi355",2026-05-19T15:41:53Z,2026-05-20T12:07:31Z,49,PR #181617 added a global test_configs.max_mm_configs=2 cap in InductorTestCase. test_origami_reduces_compile_work_vs_regular_max_autotune compares benchmark_gpu_calls between origami (topk-limited) and regular max_autotune; with both paths capped at 2 candidates the autotuner short-circuits in both branches and pro...
181617,"Speed up inductor test infrastructure (~4x collection, ~1.7x execution)",https://github.com/pytorch/pytorch/pull/181617,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-04-27T17:58:07Z,2026-05-20T12:07:30Z,49,Stack from ghstack (oldest at bottom): -> #181617 Several independent optimizations to reduce inductor test overhead: ISA subprocess caching (cpu_vec_isa.py): Cache the result of CPU vector ISA load-check subprocesses using .load_ok marker files. Each check_build() call spawns a Python subprocess that imports torch...
184464,Cache PyCompilerInterface method lookups during graph capture,https://github.com/pytorch/pytorch/pull/184464,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, module: dynamo, ciflow/inductor, module: compiled autograd",2026-05-20T01:29:18Z,2026-05-20T11:05:43Z,49,Authored by Claude. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo @azahed98 @xmfan @aditvenk
184483,Make DTensor.local_map pytree-aware on in/out_placements,https://github.com/pytorch/pytorch/pull/184483,draft/noise,53,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,182031,182031,182031,,,,0,1,0,,,"ciflow/inductor, release notes: distributed (dtensor), ciflow/torchtitan, ciflow/dtensor",2026-05-20T05:11:20Z,2026-05-20T08:45:10Z,49,"Agent Report Summary torch.distributed.tensor.experimental.local_map pytree-flattens its positional arguments via pytree.tree_flatten(args), but it then treats in_placements, in_grad_placements, and out_placements as flat sequences to be compared by length against the flattened args/output. This is the ""worst of bot..."
182055,[Inductor] Fix flaky test_allow_reuse_disable_if_exceed_peak by checking reuse marker,https://github.com/pytorch/pytorch/pull/182055,PR blocked,65,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,179776,179776,,,,,0,3,0,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor",2026-04-30T17:40:44Z,2026-05-20T07:44:18Z,49,"Stack from ghstack (oldest at bottom): -> #182055 The _is_triggering_buffer_reuse helper was comparing full generated code strings from two independent compilations (allow_buffer_reuse=True vs False). Any non-determinism between compilations (autotuning, global counters surviving dynamo.reset()) could cause spurious..."
181352,Re-enable nightly tests on s390x,https://github.com/pytorch/pytorch/pull/181352,waiting on contributor,35,"review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/s390",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,154074,181359,,,0,3,0,,,"triaged, open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/s390",2026-04-24T08:20:57Z,2026-05-20T06:57:57Z,49,"Revert ""[s390x CI] Disable scheduled s390x-periodic runs (#181005)"" This reverts commit 17b1573. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @Lucaskabela @azahed98"
184440,[dynamic_spec]Add optimizer to ease user's dynamic_spec adoption,https://github.com/pytorch/pytorch/pull/184440,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"release notes: fx, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-19T22:13:10Z,2026-05-20T03:01:16Z,50,Stack from ghstack (oldest at bottom): #184129 #184299 #184271 -> #184440 #182764 Summary Adds a library that observes torch.compile recompile reasons during a normal workload and emits a recommended ShapesSpec (or ParamsSpec) so the next run skips the static→dynamic transition and the chain of one-by-one auto-disco...
183875,Add claude skill for PrivateUse1 (PU1) third-party device integration,https://github.com/pytorch/pytorch/pull/183875,draft/noise,23,PR is draft; review decision requests changes; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,CLEAN,True,0,0,,,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-05-15T10:06:05Z,2026-05-20T02:07:32Z,50,Stack from ghstack (oldest at bottom): -> #183875
183386,Add tests for OneCycleLR zero divisor initialization behavior,https://github.com/pytorch/pytorch/pull/183386,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,168044,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-05-11T23:41:47Z,2026-05-19T22:50:52Z,50,"This PR adds tests documenting the current behavior of OneCycleLR when div_factor=0.0 or final_div_factor=0.0. Internally, these values are used during initialization to compute derived learning rates (initial_lr = max_lr / div_factor and min_lr = initial_lr / final_div_factor). When either value is zero, the schedu..."
184401,Revert D105603961,https://github.com/pytorch/pytorch/pull/184401,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, release notes: fx, module: dynamo, ciflow/inductor, ci-no-td, meta-exported, ciflow/torchtitan",2026-05-19T16:48:12Z,2026-05-19T22:13:51Z,50,"Summary: This diff reverts D105603961 (The context such as a Sandcastle job, Task, SEV, etc. was not provided.) Depends on D105603961 Test Plan: NA Differential Revision: D105706835 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhan..."
182921,[DO NOT MERGE][TESTING ONLY] Set MI300 ROCm build arch to SPIR-V,https://github.com/pytorch/pytorch/pull/182921,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/rocm-mi300,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, topic: not user facing, keep-going, ci-no-td, ciflow/rocm-mi300",2026-05-08T01:09:36Z,2026-05-19T20:45:28Z,50,Summary Add an optional pytorch-rocm-arch input to the reusable Linux build workflow. Pass PYTORCH_ROCM_ARCH into the build environment for both Linux build paths. Set PYTORCH_ROCM_ARCH=amdgcnspirv only for the regular ROCm MI300 workflow. Test plan Parsed .github/workflows/_linux-build.yml and .github/workflows/roc...
169107,Add failing test for PolynomialLR total_iters=0 closed-form crash,https://github.com/pytorch/pytorch/pull/169107,waiting on contributor,35,review decision requests changes; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,168044,,,,0,1,0,,,"triaged, open source, topic: not user facing",2025-11-26T04:23:21Z,2026-05-19T19:11:12Z,50,"This PR introduces a failing test that exposes a ZeroDivisionError when calling scheduler.step(epoch=...) on PolynomialLR with total_iters=0. Closed form LR computation performs: t = min(epoch, total_iters) / total_iters which produces a 0 / 0 division by zero crash when total_iters == 0. This PR is tests only, foll..."
184349,Improve complex input error for BatchNorm,https://github.com/pytorch/pytorch/pull/184349,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,184282,184282,184282,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-05-19T10:42:14Z,2026-05-19T17:21:33Z,50,"Fixes #184282 Summary This PR introduces an early complex data type check in _BatchNorm.forward. Currently, CUDA BatchNorm2d and BatchNorm3d, along with LazyBatchNorm2d and LazyBatchNorm3d, can reach cudnn_batch_norm with complex inputs. This leads to a lower-level input and weight data type mismatch. This update ca..."
184116,[distributed] Add NVFP4 scaled matmul + reduce-scatter fast path #182915,https://github.com/pytorch/pytorch/pull/184116,stale/low urgency,17,"PR has not updated for at least 45 days; CI lane labels: ciflow/h100-symm-mem, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,,,182915,,,,0,4,0,,,"triaged, open source, ciflow/h100-symm-mem, ciflow/torchtitan",2026-05-17T19:58:02Z,2026-05-19T17:20:45Z,50,The implementation uses: MSLK/CUTLASS NVFP4 grouped GEMM for partial matmul production global row offsets for NVFP4 activation scale addressing NCCL reduce_scatter_tensor for the reduction/communication path optional stripe-based overlap for large M symmetric-memory workspaces for Inductor/runtime buffer realization...
183592,[ROCm] op benchmark: use noble nightly image,https://github.com/pytorch/pytorch/pull/183592,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/nightly, ciflow/op-benchmark","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: rocm, triaged, open source, topic: not user facing, ciflow/nightly, no-runner-experiments, ciflow/op-benchmark",2026-05-13T18:06:28Z,2026-05-19T15:50:40Z,50,Summary Point ROCm operator_microbenchmark at ci-image:pytorch-linux-noble-rocm-nightly-py3 (TheRock nightly stack) instead of the Jammy benchmarks image. Set build-environment to linux-noble-rocm-nightly-py3.12-gfx942 to match existing nightly naming; job still runs on linux.rocm.gpu.gfx950.1 (see comment in YAML)....
177605,Default to fused implementation when available,https://github.com/pytorch/pytorch/pull/177605,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"no-stale, release notes: optim",2026-03-17T01:32:19Z,2026-05-19T15:43:47Z,50,"We want to default to the fused implementation whenever available. Well, when is it actually good to default to fused? The following conditions must be met: all params are floating point all params are on a fused supported device differentiable is not specified no params are complex (not yet checked in _default_to_f..."
178412,Move steps to CPU for fused optims,https://github.com/pytorch/pytorch/pull/178412,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/mps,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"release notes: mps, ciflow/mps, no-stale",2026-03-25T19:04:07Z,2026-05-19T15:37:43Z,50,Stack from ghstack (oldest at bottom): -> #178412
184236,[inductor][overlap] Enable simple_overlap by default,https://github.com/pytorch/pytorch/pull/184236,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, keep-going, ciflow/torchtitan",2026-05-18T17:39:59Z,2026-05-19T15:36:52Z,50,"Stack from ghstack (oldest at bottom): -> #184236 #184235 Summary: Flip enable_simple_overlap to True so that all inductor-compiled distributed workloads get comm/compute overlap out of the box. The pass has no collective reordering, no memory regression, and 0.06s overhead -- safe for default use. Authored by Claud..."
184067,[submodule] Drop unused third_party/gemmlowp submodule,https://github.com/pytorch/pytorch/pull/184067,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: cpu, triaged, open source, ciflow/trunk, release notes: quantization",2026-05-17T03:48:44Z,2026-05-19T15:21:11Z,50,cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
183749,Support auto_functionalized_v2 in can_fuse_into_auto_functionalized,https://github.com/pytorch/pytorch/pull/183749,waiting on contributor,35,"review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,3,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-14T18:20:12Z,2026-05-19T14:46:12Z,50,"Summary This PR updates can_fuse_into_auto_functionalized to recognize torch.ops.higher_order.auto_functionalized_v2. Previously, the check only handled torch.ops.higher_order.auto_functionalized. This caused the fusion logic to miss cases using the newer auto_functionalized_v2 higher-order operator. This change ext..."
183976,[CUDA] Correct inner scope Math.cuh int i -> int j,https://github.com/pytorch/pytorch/pull/183976,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: cuda",2026-05-15T23:15:44Z,2026-05-19T14:45:02Z,50,"We already have i defined before in scope, so we need j Contributed by Benedikt Johannes"
57772,Deprecate torch.svd and change svd -> linalg_svd,https://github.com/pytorch/pytorch/pull/57772,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"open source, module: linear algebra, cla signed, ciflow/trunk, release notes: linalg_frontend, no-stale",2021-05-06T21:13:29Z,2026-05-19T12:26:08Z,50,"This PR adds a warning that torch.svd is deprecated, with @lezcano's instructions from #57549. In addition, all usage of the old svd function is replaced with a new one from torch.linalg module. The PR is currently blocked by torch/xla. Once torch/xla is able to work with at::linalg_svd the XLA CI build will be gree..."
184308,Route privateuse1-overrideable ops through AOTI dispatcher (,https://github.com/pytorch/pytorch/pull/184308,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, release notes: inductor (aoti), ciflow/torchtitan",2026-05-19T02:42:10Z,2026-05-19T06:20:42Z,50,"The C-ABI shims aoti_torch___scaled_dot_product_fused_attention_overrideable[backward] were emitted in c_shim{cpu,cuda,mps}.cpp (and the backward in c_shim_xpu.cpp) with bodies that called at::compositeexplicitautograd::* directly, bypassing the dispatcher. The CompositeExplicitAutograd kernel for these ops is a TOR..."
184307,Update SDPA error diagnostics,https://github.com/pytorch/pytorch/pull/184307,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,module: cpu,2026-05-19T02:40:26Z,2026-05-19T06:20:24Z,50,Stack from ghstack (oldest at bottom): -> #184307 Human Note This is something I have wanted to do for a while but wasn't high enough pri for me to deal with all the c++ templates. This brings the older error message from ❯ python repro_sdpa_no_backend.py /home/drisspg/.ptq_workspace/jobs/20260311-adhoc-e9e0a9/repro...
183394,lint: ignore B904 for vendored code,https://github.com/pytorch/pytorch/pull/183394,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, topic: not user facing",2026-05-12T01:27:51Z,2026-05-19T01:17:01Z,51,"Summary Follow-up to reviewer feedback that PyTorch should not hand-edit vendored quack code. Added a B904 per-file ignore for torch/_vendor/**, alongside the existing vendored-code lint exception. Restored the vendored quack file to match the recorded upstream snapshot. Opened the upstream quack fix at Dao-AILab/qu..."
184058,nn.functional: raise ValueError for tau <= 0 in gumbel_softmax,https://github.com/pytorch/pytorch/pull/184058,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,184031,184031,184031,,,,0,1,0,,,"triaged, open source, release notes: nn",2026-05-17T01:17:29Z,2026-05-18T21:44:52Z,51,"Summary Fixes #184031. F.gumbel_softmax computes (logits + gumbels) / tau. When tau=0, this produces ±Inf, and softmax(Inf - Inf) yields NaN for every output element. The docstring described tau as ""non-negative"" but tau=0 has no valid interpretation — the mathematical limit as tau → 0 is an argmax one-hot, not NaN...."
183522,Fix using `torch.cuda.amp. autocast` with JIT on Python 3.13,https://github.com/pytorch/pytorch/pull/183522,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,183520,183520,,,,0,2,0,,,"triaged, open source, topic: not user facing",2026-05-13T07:25:23Z,2026-05-18T21:32:58Z,51,"The @deprecated wrapper causes a wrapper method __new__ to be injected. This causes it to no longer be treated as a builtin by inspect.isbuiltin but e.g inspect.getsource looks at the unwrapped function, which is still the object.__new__ builtin and fails with TypeError: module, class, method, function, traceback, f..."
183568,Fix timeout in `test_share_memory` on error,https://github.com/pytorch/pytorch/pull/183568,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/h100-symm-mem,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing, ciflow/h100-symm-mem",2026-05-13T15:09:05Z,2026-05-18T21:32:41Z,51,"When the RPC init fails, e.g. due to RuntimeError: In operator() at tensorpipe/common/ibv.h:172 """": Invalid argument (see pytorch/tensorpipe#413) then the rpc.shutdown will hang forever waiting for processes that didn't start. So do that only after successful rpc.init_rpc."
183798,[ROCm] Fix HIP stream pool aliasing when GPU tracing is active,https://github.com/pytorch/pytorch/pull/183798,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,DIRTY,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: rocm, open source, release notes: rocm",2026-05-14T21:57:27Z,2026-05-18T19:50:37Z,51,"On ROCm, getStreamFromPool can return the same underlying hipStream_t to two consecutive callers when a GPU trace hook is installed. PR #119996 made HIP streams lazy: initSingleStream now runs from a per-stream call_once inside CUDAStream::stream(), interleaved with normal getStreamFromPool traffic. The trace branch..."
183335,Internal,https://github.com/pytorch/pytorch/pull/183335,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"open source, fb-exported, ciflow/trunk, topic: not user facing, meta-exported",2026-05-11T17:03:18Z,2026-05-18T18:46:39Z,51,Summary: Internal Test Plan: Internal Differential Revision: D103705404
183651,[OSDC] Migrate h100-distributed.yml to OSDC (ARC) via dial-up pattern,https://github.com/pytorch/pytorch/pull/183651,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,8,0,,,topic: not user facing,2026-05-14T02:21:58Z,2026-05-18T17:35:59Z,51,"Stack from ghstack (oldest at bottom): -> #183651 #183650 #183649 Add check_experiments: arc,lf on the existing get-label-type job, plumb the OSDC inputs into the sm90 dist build and test jobs: build: ci-docker-hash, use-arc, python-version ""3.10"", compiler gcc11, cuda-version ""13.0"". test: add get-label-type to nee..."
180420,linalg.qr: reject overlapping or aliased Q and R in out=,https://github.com/pytorch/pytorch/pull/180420,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,180377,180377,180377,,,,0,4,0,,,"triaged, open source, release notes: linalg_frontend",2026-04-15T02:02:59Z,2026-05-18T17:34:49Z,51,"Previously, passing the same tensor for both out arguments of torch.linalg.qr could succeed on square matrices and produce wrong results. This PR adds a check at the start of TORCH_META_FUNC(linalg_qr) that checks at::get_overlap_status on the out tensors if they exist. That runs before set_output_strided, so it app..."
183857,Support float8 dtypes in index_copy on CPU and CUDA,https://github.com/pytorch/pytorch/pull/183857,ready for maintainer decision,72,approved with no detected failing/pending checks; PR closes or references open issue numbers; PR has not updated for at least 45 days,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,180855,180855,"113663, 180855",,,,0,1,0,,,"module: cpu, open source, release notes: indexing",2026-05-15T06:31:54Z,2026-05-18T16:41:02Z,51,"Fixes #180855. Summary Adds float8 dtype support to torch.index_copy on both CPU and CUDA. The issue's repro import torch input = torch.empty(1, dtype=torch.float8_e5m2) index = torch.tensor([0]) source = torch.empty(1, dtype=torch.float8_e5m2) torch.index_copy(input, 0, index, source) currently raises NotImplemente..."
184194,[inductor] Add simple_overlap pass for defensive comm/compute overlap,https://github.com/pytorch/pytorch/pull/184194,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-18T12:12:05Z,2026-05-18T15:44:01Z,51,"Stack from ghstack (oldest at bottom): -> #184194 Summary: Add a simple, predictable FX graph pass that overlaps NCCL collectives with compute by moving collective starts earlier and waits later. Designed to be safe as a default for any FSDP workload. The pass provides four guarantees: Hang-safe: collectives on the..."
182850,[inductor][symm_mem] Add low-contention allgather variants,https://github.com/pytorch/pytorch/pull/182850,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,4,0,,,"open source, release notes: distributed (c10d), module: inductor, ciflow/inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-05-07T17:16:20Z,2026-05-18T14:13:51Z,51,"Summary This PR adds other POC variants implementations of low-contention allgather, featuring direct-input and cached-output, for comparing non-multicast P2P, NVLS copy-engine multicast, and multimem kernel approaches. Variants: v3: closest continuation of existing v2; non-multicast P2P path using direct input, cac..."
176857,[XPU] test_indexing test_reductions: Remove obsolete XPU skip guards after verification,https://github.com/pytorch/pytorch/pull/176857,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"open source, ciflow/trunk, topic: not user facing, ciflow/xpu",2026-03-09T06:08:48Z,2026-05-18T09:17:54Z,51,Fixes: intel/torch-xpu-ops#1973 Remove obsolete XPU skip guards after verification in: test_indexing test_index_put_non_accumulate_deterministic_xpu test_index_copy_deterministic_xpu Verified on XPU: both tests pass. test_reductions test_reductions_large_half_tensors_xpu_float16 test_reductions_large_half_tensors_xp...
183903,[XPU] Relax Muon float32 tolerance in test_state_dict_with_cuda_params,https://github.com/pytorch/pytorch/pull/183903,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,open source,2026-05-15T13:55:33Z,2026-05-18T02:33:45Z,52,"Fixes: intel/torch-xpu-ops#1973 This PR adds an XPU-specific tolerance override for Muon in optimizer tests. The override is scoped to TestOptimRenewed.test_state_dict_with_cuda_params for torch.float32 on xpu only. This accounts for small, expected XPU versus CPU float32 accumulation differences observed in state_d..."
183507,einsum: fast path for small-K batched-dot to avoid cuBLAS launch overhead,https://github.com/pytorch/pytorch/pull/183507,draft/noise,15,PR is draft; PR closes or references open issue numbers; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,101249,101249,101249,,,,0,2,0,,,"open source, release notes: linalg_frontend",2026-05-13T03:54:33Z,2026-05-18T02:07:30Z,52,"Fixes #101249. cc @ruro For einsum('Nc,Nc->N', a, b) and similar diag-of-bmm patterns where both outer dims are 1, bmm degenerates to (B,1,K) @ (B,K,1). At small K, cuBLAS batched-GEMV launches one chunk per ~64K-batch slice and per-launch overhead dominates — 40x slower than (a*b).sum(-1). This adds an elementwise..."
183953,[DO NOT MERGE] Point CI workflows to test-infra IPv6 branch,https://github.com/pytorch/pytorch/pull/183953,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/binaries, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"ciflow/binaries, ciflow/trunk, topic: not user facing",2026-05-15T19:59:20Z,2026-05-18T01:16:18Z,52,Switch all pytorch/test-infra action refs from @main to @jeanschmidt/imds_ipv6 across 14 workflow files Covers setup-ssh action and linux_job_v2 reusable workflow Testing the IMDS IPv6-compatible test-infra changes against PyTorch CI before merging them to test-infra main. Authored by Claude.
183978,"[docs] Add .. deprecated:: directives to torch.{cuda,cpu}.amp public APIs",https://github.com/pytorch/pytorch/pull/183978,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, topic: not user facing",2026-05-16T00:01:44Z,2026-05-17T23:50:17Z,52,"Summary The deprecated AMP entry points in torch.cuda.amp and torch.cpu.amp (autocast, custom_fwd, custom_bwd, GradScaler) are wrapped with @deprecated(...) and emit FutureWarning at runtime, but their docstrings only mention the deprecation in plain prose. As a result the published pages on pytorch.org/docs render..."
183479,TMA scatter_add: add multi-row-per-warp path for small rows,https://github.com/pytorch/pytorch/pull/183479,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,open source,2026-05-12T22:22:43Z,2026-05-16T21:13:13Z,53,"Following #182675 Add tma_scatter_add_kernel_2d for the single-chunk case (D ≤ 512/sizeof). Each warp processes tiles of K=4 rows using cp.async.bulk.tensor.2d with a CUtensorMap descriptor. The 2D TMA instruction loads K rows in one shot even when src is non-contiguous (D < stride), and the hardware zero-fills OOB..."
183966,[CUDA] Change argmax to int64_t for better numerics,https://github.com/pytorch/pytorch/pull/183966,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, release notes: cuda",2026-05-15T22:12:34Z,2026-05-16T19:37:12Z,53,"We got int64_t for ptr_ind as well, so I'm pretty sure that (while not tested) this should be correct for better numerics Contributed by Benedikt Johannes"
183455,Port D104346887/PR 182675 for index large select (#183455),https://github.com/pytorch/pytorch/pull/183455,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td, meta-exported",2026-05-12T17:20:08Z,2026-05-16T10:01:13Z,53,"Summary: Adopts the vectorized scatter_add(dim=0) kernel from D104346887 / PR #182675 as a fast path for index_add_ on CUDA. index_add_(0, idx, src) is mathematically equivalent to self.scatter_add_(0, idx.unsqueeze(-1).expand_as(src), src) when alpha == 1, so we can call tma_scatter_add_kernel_launch directly. Elig..."
173459,Add dtype mismatch validation for mm/bmm in meta registrations (#171190),https://github.com/pytorch/pytorch/pull/173459,PR blocked,65,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/mps, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,171190,171190,,,,,0,11,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-01-27T00:51:04Z,2026-05-16T08:53:35Z,53,"Fixes #171190 Summary torch.compile with inductor backend silently succeeds when mm/bmm operations have mismatched input dtypes, while eager mode correctly raises RuntimeError. This violates the semantic equivalence guarantee between eager and compiled modes. Changes Add dtype validation for meta_mm and common_meta_..."
183993,"Back out ""Port D104346887/PR 182675 for index large select""",https://github.com/pytorch/pytorch/pull/183993,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, meta-exported",2026-05-16T03:07:24Z,2026-05-16T04:29:38Z,53,Summary: Original commit changeset: cc26aa4ac0a0 Original Phabricator Diff: D104669063 Test Plan: S664499 https://www.internalfb.com/chronos/job_instance/gp/99079202246114267/info Differential Revision: D105399771
180118,[dynamo] Skip pad_packed_sequence during tracing,https://github.com/pytorch/pytorch/pull/180118,PR blocked,65,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,178260,178260,,,,,0,2,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-04-11T05:42:12Z,2026-05-16T04:18:07Z,53,"Summary Fixes #178260. torch.compile was tracing through torch.nn.utils.rnn.pad_packed_sequence, which eventually reaches _VF._pad_packed_sequence and fails fake tensor propagation with: The tensor has a non-zero number of elements, but its data is not allocated yet. This change matches the existing handling for pac..."
179096,[Inductor] Add _scaled_mm_v2 lowering,https://github.com/pytorch/pytorch/pull/179096,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2026-04-02T00:19:18Z,2026-05-16T01:31:44Z,54,Stack from ghstack (oldest at bottom): -> #179096 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @Lucaskabela @azahed98
174893,[Inductor][NVGEMM] Enable Epilogue Fusions,https://github.com/pytorch/pytorch/pull/174893,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181622, 181902, 181903, 181909",,,0,17,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-02-12T17:55:04Z,2026-05-16T01:31:05Z,54,Stack from ghstack (oldest at bottom): #181909 #181903 #181902 #181622 -> #174893 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
181909,[Inductor][Triton] Forward hint_override in fused Triton compile,https://github.com/pytorch/pytorch/pull/181909,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"174893, 181622, 181902, 181903",,,0,2,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-29T19:15:06Z,2026-05-16T01:26:00Z,54,Stack from ghstack (oldest at bottom): -> #181909 #181903 #181902 #181622 #174893 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
181520,dataloader: add performance warning for in-memory datasets with num_workers>0,https://github.com/pytorch/pytorch/pull/181520,waiting on contributor,35,review decision requests changes; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,154318,,,,0,1,0,,,"triaged, open source, release notes: dataloader",2026-04-26T18:58:21Z,2026-05-16T01:07:15Z,54,"Summary Adds a UserWarning when DataLoader is constructed with an in-memory dataset (torch.Tensor, TensorDataset, numpy.ndarray) and num_workers > 0, since this combination is consistently slower than direct indexing due to IPC and collation overhead. Motivation As documented in #154318, DataLoader with num_workers..."
181501,docs: add DataLoader performance guidance for in-memory datasets,https://github.com/pytorch/pytorch/pull/181501,waiting on contributor,35,review decision requests changes; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,154318,,,,0,2,0,,,"triaged, open source, release notes: dataloader",2026-04-26T00:20:41Z,2026-05-16T01:07:10Z,54,"Summary Adds a ""Performance Tips"" section to the DataLoader docstring and a new ""Choosing the right num_workers"" subsection to docs/source/data.md, explaining when num_workers > 0 helps vs hurts performance. Fixes the documentation gap identified in #154318, where users experience 3–124× slowdowns using DataLoader w..."
183203,perf(dataloader): zero-copy contiguous-slice fast path for TensorDataset,https://github.com/pytorch/pytorch/pull/183203,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,154318,"181501, 181520",,,0,1,0,,,"triaged, open source, release notes: dataloader",2026-05-10T23:24:22Z,2026-05-16T01:06:58Z,54,"Summary Relates to #154318 (related: #181501, #181520) When iterating a TensorDataset with shuffle=False (the default for validation/inference), the DataLoader currently: Calls dataset[i] in a Python loop — one call per sample per batch Builds a temporary list of per-sample tuples in Python memory Calls default_coll..."
177705,[Do Not Land] Lookup Table Recorder,https://github.com/pytorch/pytorch/pull/177705,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-03-18T00:08:44Z,2026-05-16T01:01:56Z,54,Stack from ghstack (oldest at bottom): -> #177705 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
179011,[Inductor][NVGEMM] Add fast path for GEMM dispatch,https://github.com/pytorch/pytorch/pull/179011,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"179012, 179013, 179014",,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-01T16:29:10Z,2026-05-16T01:01:55Z,54,Stack from ghstack (oldest at bottom): #179014 #179013 #179012 -> #179011 #176859 #176847 #176845 #176549 #176548 #176547 #176546 #176545 #176544 #176543 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhan...
179012,[Inductor][NVGEMM] Add supplement heuristic configs behind flag,https://github.com/pytorch/pytorch/pull/179012,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"179011, 179013, 179014",,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-01T16:29:13Z,2026-05-16T01:01:54Z,54,Stack from ghstack (oldest at bottom): #179014 #179013 -> #179012 #179011 #176859 #176847 #176845 #176549 #176548 #176547 #176546 #176545 #176544 #176543 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhan...
178840,Add parallel subprocess compilation and disk caching for NVGEMM backend,https://github.com/pytorch/pytorch/pull/178840,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-03-31T01:14:18Z,2026-05-16T01:01:54Z,54,Stack from ghstack (oldest at bottom): -> #178840 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
179167,[Inductor][NVGEMM] Fix FP4 autotuning tensor creation,https://github.com/pytorch/pytorch/pull/179167,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,1,0,,,"open source, topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-02T19:12:14Z,2026-05-16T01:01:51Z,54,Stack from ghstack (oldest at bottom): (to be filled) cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
179014,[Inductor][NVGEMM] Wrap GroupedGemmArguments in try/except,https://github.com/pytorch/pytorch/pull/179014,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"179013, 179049",,,0,4,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-01T16:29:20Z,2026-05-16T01:01:51Z,54,Stack from ghstack (oldest at bottom): #179049 -> #179014 #179013 #176859 #176847 #176845 #176549 #176548 #176547 #176546 #176545 #176544 #176543 GroupedGemmArguments construction can fail with edge-case tensor shapes during kernel enumeration. Catch the exception and skip NVGEMM for that op rather than crashing com...
179013,[Inductor][NVGEMM] Fix FP4 autotuning tensor creation,https://github.com/pytorch/pytorch/pull/179013,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"179014, 179049",,,0,4,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-01T16:29:17Z,2026-05-16T01:01:51Z,54,Stack from ghstack (oldest at bottom): #179049 #179014 -> #179013 #176859 #176847 #176845 #176549 #176548 #176547 #176546 #176545 #176544 #176543 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhu...
179015,[Inductor][NVGEMM] Add fast path for GEMM dispatch,https://github.com/pytorch/pytorch/pull/179015,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"179013, 179014, 179016",,,0,2,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-01T16:39:55Z,2026-05-16T01:01:50Z,54,"Stack from ghstack (oldest at bottom): #179014 #179013 #179016 -> #179015 #176859 #176847 #176845 #176549 #176548 #176547 #176546 #176545 #176544 #176543 Cache compiled_obj and call it directly for all three GEMM variants (dense, scaled, grouped), bypassing per-call argument object construction. The overhead came fr..."
179016,[Inductor][NVGEMM] Add supplement heuristic configs behind flag,https://github.com/pytorch/pytorch/pull/179016,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"179013, 179014, 179015",,,0,2,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-01T16:39:58Z,2026-05-16T01:01:49Z,54,Stack from ghstack (oldest at bottom): #179014 #179013 -> #179016 #179015 #176859 #176847 #176845 #176549 #176548 #176547 #176546 #176545 #176544 #176543 Add 28 kernel configs that nvMatmulHeuristics doesn't explore but empirically beat cuBLAS on B200 bf16 decode shapes (found via exhaustive search on Llama-70B and...
179049,[Inductor][NVGEMM] Add supplement heuristic configs behind flag,https://github.com/pytorch/pytorch/pull/179049,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,"179013, 179014",,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-01T20:14:58Z,2026-05-16T01:01:48Z,54,Stack from ghstack (oldest at bottom): -> #179049 #179014 #179013 #176859 #176847 #176845 #176549 #176548 #176547 #176546 #176545 #176544 #176543 Add 28 kernel configs that nvMatmulHeuristics doesn't explore but empirically beat cuBLAS on B200 bf16 decode shapes (found via exhaustive search on Llama-70B and Qwen3-32...
179175,[Inductor][NVGEMM] Fix FP4 autotuning tensor creation,https://github.com/pytorch/pytorch/pull/179175,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,174929,,,,0,1,0,,,"open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-04-02T19:45:48Z,2026-05-16T01:01:46Z,54,"Stack from ghstack (oldest at bottom): #179169 #179168 -> #179175 #176859 #176847 #176845 #176549 #176548 #176547 #176546 #176545 #176544 #176543 FP4 tensors in Inductor IR use uint8 storage with a ReinterpretView to float4_e2m1fn_x2. When creating dummy tensors for cutlass_api's supports() checks, the layout produc..."
182730,[DTensor] Preserve placement for aten.clone.default,https://github.com/pytorch/pytorch/pull/182730,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,15,0,,,"Merged, Reverted, ciflow/trunk, ciflow/inductor, release notes: distributed (dtensor), ci-no-td, ciflow/torchtitan, ciflow/dtensor",2026-05-06T22:08:39Z,2026-05-16T00:03:33Z,54,"Stack from ghstack (oldest at bottom): -> #182730 #183629 We registered clone op twice, one in _tensor_ops.py and another in _pointwise_ops.py. The former one is correct. This PR removes the later one. aten.clone.default is tagged torch.Tag.pointwise and was getting auto-registered with the default pointwise single-..."
183726,[poc] yet another nonstrict api,https://github.com/pytorch/pytorch/pull/183726,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,,2026-05-14T15:02:20Z,2026-05-15T21:54:44Z,54,"[no-ci] Prototype make_fx, jit, and compile_artifact APIs for a target-specific nonstrict flow. Authored with Codex."
183940,[Native DSL] Vendor kernelagent-oink library,https://github.com/pytorch/pytorch/pull/183940,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,183941,,,0,1,0,,,,2026-05-15T18:52:33Z,2026-05-15T21:47:45Z,54,Stack from ghstack (oldest at bottom): #183941 -> #183940 Summary: Vendor a subset of the kernelagent-oink library (https://github.com/meta-pytorch/KernelAgent) under torch._vendor.oink. Only the modules needed by the upcoming cuteDSL RMSNorm override are vendored (rmsnorm.py) -- softmax/cross-entropy/layernorm and...
183941,[Native DSL] Oink-based cuteDSL RMSNorm,https://github.com/pytorch/pytorch/pull/183941,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,183940,,,0,1,0,,,,2026-05-15T18:52:38Z,2026-05-15T18:53:58Z,54,"Stack from ghstack (oldest at bottom): -> #183941 #183940 Summary: Add fused rmsnorm fwd/bwd overrides backed by the vendored kernelagent-oink kernels (cuteDSL). The override gates on Blackwell (SM10.x); on Hopper or older, calls fall through to aten so existing behavior is preserved. torch/_native/ops/norm/oink_rms..."
180557,[docs] Improve documentation of normalization layers,https://github.com/pytorch/pytorch/pull/180557,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,"51455, 75862","51455, 75862","51455, 75862, 113750",,,,0,7,0,,,"triaged, open source, release notes: python_frontend",2026-04-16T13:36:36Z,2026-06-22T14:33:15Z,16,"Fixes #51455 (its documentation part, not the implementations; the impl part of that issue seems to be covered by #113750 ) Fixes #75862 This PR improves documentation for normalization layers. Changes: Clarify normalization layers' documentation by explicitly listing the remaining dimension indices of tensors after..."
179719,Skip div-to-mul-reciprocal when division_rounding is enabled,https://github.com/pytorch/pytorch/pull/179719,waiting on contributor,45,"review decision requests changes; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,179577,179577,,,,,0,1,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ci-no-td, ciflow/torchtitan",2026-04-08T15:52:17Z,2026-05-15T13:39:14Z,54,"(a) Added code comment explaining why Inductor still converts f/const to f * (1/constant) even when eager_numerics.division_rounding is enabled. (b) Added 3 tests to show current status: Inspired by #179577, note that this PR can't fix #179577 since it was caused by the design of the model which has x/1e-8 cc @vozne..."
182196,[dynamo][dicts] Fix delitem for dicts,https://github.com/pytorch/pytorch/pull/182196,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-01T23:31:29Z,2026-05-15T07:54:17Z,54,Stack from ghstack (oldest at bottom): -> #182196 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo @azahed98
182348,[SymmMem] Implement StoreExchange::broadcast for multicast handle exchange,https://github.com/pytorch/pytorch/pull/182348,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/h100-symm-mem,review when ready,,UNKNOWN,False,0,0,,,169099,169099,169099,,,,0,2,0,,,"triaged, open source, release notes: distributed (c10d), ciflow/h100-symm-mem",2026-05-04T20:42:38Z,2026-05-15T07:08:50Z,54,"Adds broadcast primitive to StoreExchange instead of all_gather for symmetric memory multicast handle exchange. All_gather was wasteful here because only the leader produces a real handle. Every other rank had to publish a meaningless value just to participate, and N-1 of the gathered values were discarded, so it wa..."
179565,[WIP][CI] Add RHEL 9.6 build and test support,https://github.com/pytorch/pytorch/pull/179565,draft/noise,23,"PR is draft; review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: cpu, open source, release notes: releng, ciflow/inductor, no-runner-experiments, ciflow/torchtitan",2026-04-07T12:20:41Z,2026-05-15T05:47:27Z,54,"Summary Adds RHEL 9.6 as a CI platform for PyTorch, running on self-hosted runners with NVIDIA H200 GPUs (Hopper, sm_90) and Podman as the container runtime. This has been tested end-to-end on a self-hosted RHEL runner — Docker image builds, PyTorch CUDA build completes, and test shards execute. The PR is intentiona..."
183797,fix internal tests for [BE][Ez]: Make more dataclass slots=True (#183449),https://github.com/pytorch/pytorch/pull/183797,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/inductor, meta-exported",2026-05-14T21:54:54Z,2026-05-15T04:28:17Z,54,"Summary: As title, addresses internal build/test issues introduced by #183449. Maintains backward compatibility for serialization. Test Plan: CI passes. Differential Revision: D105233897"
183487,Update XNNPACK submodule to 0b66c9f,https://github.com/pytorch/pytorch/pull/183487,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, ciflow/trunk, release notes: mobile, release notes: quantization, merging",2026-05-12T23:53:01Z,2026-05-15T04:27:59Z,54,"Update the XNNPACK submodule to commit 0b66c9f. Note that XNNPACK has made source-breaking changes to the unstable operator-level APIs. Specifically - the JIT code cache arg was removed. To handle this, I need to do a two stage update, as I want to land the OSS and Meta-internal dependency updates separately. To do..."
183086,Remove unused 'type: ignore' suppressions in dynamo/inductor/functorch,https://github.com/pytorch/pytorch/pull/183086,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/inductor-pallas","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, Merged, Reverted, release notes: fx, ciflow/mps, module: inductor, module: dynamo, ciflow/inductor, module: compiled autograd, ci-no-td, ciflow/rocm-mi300, release notes: inductor (aoti), ciflow/h100, ciflow/b200, ciflow/inductor-pallas, ciflow/torchtitan",2026-05-10T06:27:33Z,2026-05-15T03:01:29Z,55,Authored by Claude. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98 @xmfan @aditvenk
166063,Add Gluon based grouped MM kernel for Blackwell into Inductor,https://github.com/pytorch/pytorch/pull/166063,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/b200, ciflow/h100, ciflow/inductor, ciflow/rocm-mi300","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,46,0,,,"open source, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/rocm-mi300, ciflow/h100, ciflow/b200, ciflow/torchtitan",2025-10-22T11:34:05Z,2026-05-15T03:00:43Z,55,Stack from ghstack (oldest at bottom): -> #166063 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98 @Lucaskabela @chenyang78
180545,fix AOTInductor fails to compile dynamic-shape CUDA gather pattern with ValueError: The argument '((0)) + 48' is not comparable,https://github.com/pytorch/pytorch/pull/180545,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-04-16T08:55:24Z,2026-05-15T03:00:27Z,55,GitHub comment draft for follow-up PR on issue #179900 Below is a ready-to-paste GitHub comment for your PR conversation. Thanks for taking a look. This PR is a follow-up to issue #179900: #179900 That issue was closed after it was reported that the original repro worked on main. I re-checked the situation on our si...
183747,Adding new test_c10d_ops_privateuse1.py variant for distributed testing,https://github.com/pytorch/pytorch/pull/183747,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,5,0,,,"open source, release notes: distributed (ddp), ciflow/torchtitan",2026-05-14T18:11:07Z,2026-05-14T23:11:21Z,55,This PR is part of a series to refactor distributed tests to be completely device-agnostic. This test was created with respect to test_c10d_ops_nccl.py and then adapted for PrivateUse1 backends based on the tests from that file. All other changes are due to the PR being dependent upon #PR178336 and should be ignored...
183787,[PyTorch] Support selective TorchScript tensor data elision (#183787),https://github.com/pytorch/pytorch/pull/183787,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, release notes: jit, meta-exported",2026-05-14T21:02:17Z,2026-05-14T22:48:56Z,55,Summary: Allow ScriptModuleSerializer callers to omit only selected tensor data records from the standard TorchScript data archive while preserving all other tensors and graph constants. Add a standard-archive import overload that accepts a pre-populated DeserializationStorageContext so omitted storages can be subst...
177450,DataPipes: prevent failing properties from masking invalid-iterator errors,https://github.com/pytorch/pytorch/pull/177450,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: dataloader",2026-03-14T11:28:43Z,2026-05-14T21:34:59Z,55,Summary Make IterDataPipe diagnostic argument rendering robust when constructor-matching properties raise. Prevent unrelated property exceptions from masking the intended invalid-iterator RuntimeError. Add a regression test for this behavior. Root Cause _generate_input_args_string in torch/utils/data/datapipes/_hook...
172003,Add __getitems__ to TensorDataset for faster DataLoader iteration,https://github.com/pytorch/pytorch/pull/172003,waiting on contributor,45,review decision requests changes; PR closes or references open issue numbers; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,106704,106704,106704,,,,0,1,0,,,"triaged, open source, release notes: dataloader",2026-01-08T16:50:34Z,2026-05-14T21:34:37Z,55,"Fixes #106704 Summary Implements the __getitems__ protocol for TensorDataset, enabling batch fetching of samples instead of individual __getitem__ calls. Root Cause TensorDataset was the only standard dataset missing __getitems__ — both StackDataset and Subset already implement it. Without it, _MapDatasetFetcher fal..."
175320,[MPS] Ensure pool2d inputs are contiguous in target memory format,https://github.com/pytorch/pytorch/pull/175320,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,,,175190,,,,0,1,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-02-19T09:12:38Z,2026-05-14T19:40:35Z,55,"Summary When pool2d_template creates Placeholder objects with gatherTensorData=false for ChannelsLast inputs, the raw Metal buffer is passed directly to MPSNDArray. Non-contiguous tensors (e.g., expanded views from autograd) can have buffers that are too small or laid out incorrectly, causing SIGABRT or wrong result..."
182114,Add torch.utils.nan_guard for forward-pass NaN/Inf debugging,https://github.com/pytorch/pytorch/pull/182114,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source",2026-05-01T13:57:03Z,2026-05-14T17:51:21Z,55,"Summary Adds torch.utils.nan_guard.NaNGuard, a forward-pass context manager that names the first nn.Module whose output contains NaN/Inf, the path within its output structure (via pytree), and the tensor's shape/dtype/device. torch.autograd.detect_anomaly already covers the backward pass through the autograd engine,..."
183087,[profiler] Synchronize all active devices on profiler exit,https://github.com/pytorch/pytorch/pull/183087,PR blocked,65,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,181238,181238,181238,,,,0,2,0,,,"triaged, open source",2026-05-10T06:30:20Z,2026-05-14T17:08:37Z,55,"The profiler's exit previously called synchronize() without a device argument, which only waits on the current device (default: device 0). On multi-GPU workloads, this causes long-running kernels on non-zero devices to be missed in profiling results. Fix by synchronizing all devices with an active CUDA context in bo..."
182583,Skip RST<->MD source conversions in doc-redirects-check,https://github.com/pytorch/pytorch/pull/182583,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: docs, triaged, open source, topic: not user facing, easy, rst-to-md",2026-05-05T23:35:42Z,2026-05-14T16:37:32Z,55,"Summary Fixes a doc-redirects-check lint failure that blocks docathon RST→MD conversion PRs. When a contributor converts docs/source/foo.rst to docs/source/foo.md using MyST, the conversion typically rewrites enough of the file (RST directives → MyST {eval-rst} blocks, role syntax changes, code-fence reformat) that..."
181262,"Show node.meta[""custom""][""stream""] in Graph.__str__",https://github.com/pytorch/pytorch/pull/181262,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, ciflow/trunk, release notes: fx",2026-04-23T17:39:36Z,2026-05-14T16:23:10Z,55,"print(gm.graph) does not currently surface stream annotations -- that information lives in node.meta[""custom""][""stream""] and is only visible via gm.print_readable(), which emits it as a separate # Annotation: {'stream': N} comment line. Extend Node.format_node (which Graph.str delegates to) so that call_* nodes with..."
180891,[ROCm] Migrate from rocm_smi to amd_smi,https://github.com/pytorch/pytorch/pull/180891,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/rocm-mi300, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, ciflow/trunk, release notes: releng, ciflow/rocm-mi300",2026-04-20T18:32:04Z,2026-05-14T16:21:13Z,55,cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
170807,Adding Glossary,https://github.com/pytorch/pytorch/pull/170807,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,17,0,,,"module: docs, triaged, open source, release notes: releng",2025-12-18T20:52:37Z,2026-05-14T16:11:27Z,55,Fresh PR for glossary feature. cc @svekars @sekyondaMeta
183367,Avoid initializing CUDA in mem_get_info,https://github.com/pytorch/pytorch/pull/183367,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,103584,103584,103584,,,,0,1,0,,,"triaged, open source",2026-05-11T21:25:00Z,2026-05-14T16:06:37Z,55,"Fixes #103584 Summary return (0, 0) from torch.cuda.mem_get_info() when CUDA has not been initialized avoid resolving the current CUDA device or calling cudart() on the uninitialized path add a subprocess regression test covering the uninitialized behavior and the initialized cudaMemGetInfo path"
183262,[CUDA] Expose cudaHostGetDevicePointer in cudart bindings,https://github.com/pytorch/pytorch/pull/183262,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: cuda",2026-05-11T10:33:37Z,2026-05-14T16:06:33Z,55,"Summary: expose cudaHostGetDevicePointer through torch.cuda.cudart() return both the cudaError_t status and CUDA-visible device pointer extend the existing cudart host registration test to register mapped host memory and retrieve the device pointer Example: import torch from torch.cuda import check_error, cudart t =..."
183405,Fix unused groups in test_convs,https://github.com/pytorch/pytorch/pull/183405,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-12T04:32:11Z,2026-05-14T16:06:31Z,55,This PR enhance test_convs to use groups
183502,Remove outdated code for macOS < 12,https://github.com/pytorch/pytorch/pull/183502,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source, topic: not user facing",2026-05-13T03:12:49Z,2026-05-14T16:06:15Z,55,
183433,[inductor] Fix backslash escaping order in user-defined Triton kernel…,https://github.com/pytorch/pytorch/pull/183433,waiting on contributor,35,"review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-12T12:57:33Z,2026-05-14T16:04:45Z,55,"… embedding As seen on mikekg:fix-inductor-userdef-triton-backslash (PR #183421). define_user_defined_triton_kernel embeds kernel source in a triple-quoted string in the generated AOT module. Backslashes must be doubled BEFORE the triple-quote escaping steps, otherwise: (original bug #183420) \n in a tl.constexpr de..."
182788,Fix decorators of `ProcessGroupGlooTest` & `ProcessGroupNCCLTest`,https://github.com/pytorch/pytorch/pull/182788,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-07T06:55:52Z,2026-05-14T16:04:27Z,55,The requires_gloo/requires_nccl decorator cause the function to just return. In the way they are used this skips the initialization done by a helper function. So the test is not skipped and then fails due to missing variables. Decorate the class instead.
182803,Fix torch.trapz documentation signature,https://github.com/pytorch/pytorch/pull/182803,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,71392,71392,71392,,,,0,1,0,,,"triaged, open source, release notes: python_frontend",2026-05-07T09:06:32Z,2026-05-14T16:03:57Z,55,"Summary This PR updates the torch.trapz documentation signature so it matches torch.trapezoid, since torch.trapz is documented as an alias of torch.trapezoid. Specifically, it adds the missing dx=None keyword argument to the torch.trapz docstring signature. Verification This is a documentation-only change. I compare..."
182836,[ROCm][CI] Update mi200 workflows to use arc label,https://github.com/pytorch/pytorch/pull/182836,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, triaged, open source, topic: not user facing",2026-05-07T15:39:09Z,2026-05-14T16:03:52Z,55,cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
182746,[MPS] Memory-efficient SDPA backward (opt-in),https://github.com/pytorch/pytorch/pull/182746,stale/low urgency,17,"PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/mps",review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps, ciflow/inductor, release notes: inductor (aoti)",2026-05-07T00:20:07Z,2026-05-14T16:03:47Z,55,"Summary Adds an MPS implementation of _scaled_dot_product_efficient_attention and its backward, mirroring the CUDA op signature. The backward never materializes the [B, H, S, S] attention matrix that the existing decomposed path requires, it tiles the recomputation along the K/V sequence dimension. The forward saves..."
182845,[AI Codemod][RlAutotuneFbsourceWearablesMemory] PackedLinearWeightsQnnp::apply_dynamic (qlinear_dynamic.cpp),https://github.com/pytorch/pytorch/pull/182845,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, triaged, open source, fb-exported, release notes: quantization, meta-exported",2026-05-07T16:56:05Z,2026-05-14T16:03:43Z,55,Reviewed By: jankrom Differential Revision: D104223823 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
182846,fix: safely handle non-standard boolean values in the CUDA kernel,https://github.com/pytorch/pytorch/pull/182846,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: cuda",2026-05-07T17:03:31Z,2026-05-14T16:03:33Z,55,"I got the following error executing TestCommonCUDA.test_non_standard_bool_values: File ""/third_party/py/torch/test/test_ops.py"", line 1604, in test_non_standard_bool_values actual = op(transformed.input, *transformed.args, **transformed.kwargs) File ""/third_party/py/torch/testing/_internal/opinfo/core.py"", line 1237..."
182873,test: skip some test_non_standard_bool_values test cases,https://github.com/pytorch/pytorch/pull/182873,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: python_frontend",2026-05-07T19:35:09Z,2026-05-14T16:03:28Z,55,Skip test_non_standard_bool_values cases with cuda for: masked_scatter msort nn_functional_unfold nonzero_static scatter / scatter_add / scatter_reduce_sum tril / triu CUB produces incorrect boolean truth values whith non-standard bool values as shown in the logs. Logs FAILED test/test_ops.py::TestCommonCUDA::test_n...
182609,Fix inconsistent CPU/CUDA casting for fmax/fmin/maximum/minimum with narrower out dtype,https://github.com/pytorch/pytorch/pull/182609,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,181805,181805,181805,,,,0,1,0,,,"triaged, open source",2026-05-06T02:30:51Z,2026-05-14T16:02:23Z,55,"Summary Fixes #181805. On non-CPU devices (CUDA, XPU, etc.), TensorIterator does not create a temporary output tensor when cast_common_dtype_to_outputs is set and the output dtype differs from the compute (common) dtype. The device kernel's internal dynamic-casting path produces wrong results for fmax/fmin/maximum/m..."
183037,[Distributed] Add `ranks_only` to `StateDictOptions` for controlling which ranks receive the full state dict,https://github.com/pytorch/pytorch/pull/183037,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, release notes: distributed (fsdp), release notes: distributed (checkpoint), ciflow/torchtitan",2026-05-09T05:42:58Z,2026-05-14T16:02:20Z,55,"Background / Motivation When exporting a model trained with Pipeline Parallelism (PP) + FSDP/TP for inference, users typically need to: Call get_model_state_dict(stage_module, options=StateDictOptions(full_state_dict=True)) on each PP stage to gather the full (unsharded) parameters. Collect the per-stage state dicts..."
178354,[XPU] Add XPU dispatch for _scaled_grouped_mm,https://github.com/pytorch/pytorch/pull/178354,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"module: cpu, open source, ciflow/trunk, topic: not user facing, ciflow/inductor, ciflow/xpu, release notes: inductor (aoti)",2026-03-25T00:46:08Z,2026-05-14T12:16:43Z,55,"Register XPU dispatch key for _scaled_grouped_mm in native_functions.yaml and add the dispatch function in ScaledGroupedBlas.cpp`. The kernel validates inputs (FP8 dtypes, transposition, divisibility, rowwise float32 scales), creates the output tensor, and forwards to the sycl-tla kernel via the USE_SYCLTLA-guarded..."
183612,"Update NVPL versions: blas 0.3.0 → 0.5.0.1, lapack 0.2.3.1 → 0.3.2",https://github.com/pytorch/pytorch/pull/183612,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/binaries,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/binaries, topic: not user facing, no-runner-experiments",2026-05-13T19:42:44Z,2026-05-14T05:59:26Z,55,"Updating NVPL to latest versions, last update in upstream was from 2 years ago in #132811. blas 0.3.0 → 0.5.0.1, lapack 0.2.3.1 → 0.3.2 cc @Aidyn-A @ptrblck @nWEIdia @atalman @malfet"
182179,[CI] Floor known-slow file times when stale per-config data under-shards,https://github.com/pytorch/pytorch/pull/182179,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/rocm-mi200, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,167616,,,,0,4,0,,,"open source, ciflow/trunk, topic: not user facing, keep-going, ci-no-td, ciflow/rocm-mi200",2026-05-01T20:05:07Z,2026-05-14T05:59:22Z,55,"TODO: Revert temporary change in 0e9cabf Motivation Prevent stale/suspiciously-low test times data from causing timeouts in CI for any job/config. Floor each known-slow file's estimate at the max of the active config's value, the same job's default config value, and default[""default""]. Only applies to the default an..."
181281,[dynamo] Refactor IsInstanceBuiltinVariable,https://github.com/pytorch/pytorch/pull/181281,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181337, 181338, 182598",,,0,12,0,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, module: dynamo, ciflow/inductor, ci-no-td, ciflow/torchtitan",2026-04-23T20:16:23Z,2026-05-14T05:59:16Z,55,Stack from ghstack (oldest at bottom): #182598 #181338 #181337 -> #181281 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo
181338,[dynamo] Refactor TypeBuiltinVariable,https://github.com/pytorch/pytorch/pull/181338,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181281, 181337, 182598",,,0,13,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-24T06:02:49Z,2026-05-14T05:58:47Z,55,Stack from ghstack (oldest at bottom): #182598 -> #181338 #181337 #181281 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo
181337,[dynamo] Refactor StrBuiltinVariable,https://github.com/pytorch/pytorch/pull/181337,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181281, 181338, 182598",,,0,12,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-04-24T06:02:43Z,2026-05-14T05:58:46Z,55,Stack from ghstack (oldest at bottom): #182598 #181338 -> #181337 #181281 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @jataylo
182598,[dynamo] Refactor builtins init static type check using _fn,https://github.com/pytorch/pytorch/pull/182598,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"181281, 181337, 181338",,,0,12,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-06T01:07:52Z,2026-05-14T05:58:45Z,55,Stack from ghstack (oldest at bottom): -> #182598 #181338 #181337 #181281 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @chauhang @amjames @Lucaskabela @jataylo
183374,"[ROCm] [ TEST , PLEASE DON""T MERGE ] TheRock 7.13 nightly deb install path",https://github.com/pytorch/pytorch/pull/183374,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/op-benchmark,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,6,0,,,"module: rocm, open source, topic: not user facing, no-runner-experiments, ciflow/op-benchmark",2026-05-11T22:07:07Z,2026-05-14T05:58:28Z,55,Install ROCm 7.13 in the Docker image from TheRock nightly debs. cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang
183611,[inductor] Record output alias for mutates_and_returns_first_arg fallbacks,https://github.com/pytorch/pytorch/pull/183611,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-13T19:34:13Z,2026-05-13T23:28:38Z,56,"Stack from ghstack (oldest at bottom): -> #183611 FallbackKernel's shortcut for ops covered by torch._library.utils.mutates_and_returns_first_arg only recorded the mutation, not the (output aliases first arg) alias relationship. When such an op was fed to compile_fx_inner directly (bypassing AOT autograd functionali..."
178360,Add device/backend agnostic support for distributed c10d/common tests,https://github.com/pytorch/pytorch/pull/178360,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"triaged, open source, release notes: distributed (ddp), module: PrivateUse1, ciflow/torchtitan",2026-03-25T01:41:26Z,2026-05-13T23:26:57Z,56,"This PR is part of a series to refactor distributed tests to be completely device-agnostic. Changes Tests: test_c10d_common.py, test_c10d_functional_native.py, test_c10d_nccl.py, test_c10d_spawn.py, test_c10d_spawn_nccl.py, test_device_mesh.py, test_backends.py Use torch.accelerator.current_accelerator() to derive D..."
177690,"DO NOT MERGE Default pre_grad_pass_timing to ""late"" in fbcode",https://github.com/pytorch/pytorch/pull/177690,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,8,0,,,"Merged, Reverted, ciflow/trunk, topic: not user facing, module: aotdispatch, module: inductor, ciflow/inductor, ci-no-td, autorevert: disable, ciflow/torchtitan",2026-03-17T20:25:37Z,2026-05-13T22:28:58Z,56,"Stack from ghstack (oldest at bottom): -> #177690 #183576 This ensures pre-grad passes run after cache lookups in Meta-internal builds, avoiding unnecessary pass execution on cache hits. Authored with Claude."
183384,[inductor] preserve scan carry strides with shape padding,https://github.com/pytorch/pytorch/pull/183384,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-11T23:37:05Z,2026-05-13T21:58:12Z,56,"Stack from ghstack (oldest at bottom): -> #183384 #183383 Scan lowering decomposes scan into while_loop, whose carried values must preserve metadata across iterations. Shape padding can return same-sized matmul results with padded strides, so normalize next_carry back to the incoming carry stride and cover the force..."
181306,Integrate profiler documentation into torch.profiler docs,https://github.com/pytorch/pytorch/pull/181306,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, topic: not user facing",2026-04-23T23:36:05Z,2026-05-13T21:16:39Z,56,"Enrich the torch.profiler API reference page with conceptual overview, getting started guide, Perfetto visualization instructions, advanced profiling techniques, and an API quick reference table. Add two new developer notes covering the profiler architecture (Kineto/CUPTI data flow) and memory profiling with memory..."
183594,TMA scatter_reduce min/max for sm_90+ (CUDA 12.8+),https://github.com/pytorch/pytorch/pull/183594,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,183479,,,0,2,0,,,open source,2026-05-13T18:30:51Z,2026-05-13T20:07:40Z,56,Depends on #183479 Generalize to support TMA scatter_reduce min/max for sm_90+ (CUDA 12.8+) Supported type matrix for `cp.reduce.async.bulk`: | op | f32 | f64 | f16 | bf16 | s32 | s64 | |---------|-----|-----|-----|------|-----|-----| | add | ✓ | ✓ | ✓ | ✓ | | | | min/max | | | ✓ | ✓ | ✓ | ✓ | Benchmarks on GB200 (1...
180748,Fix NumPy 2 deprecation warning in __array_wrap__,https://github.com/pytorch/pytorch/pull/180748,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,180657,180657,,,,,0,3,0,,,"triaged, open source",2026-04-18T21:46:51Z,2026-05-13T18:32:12Z,56,"NumPy 2 introduces a deprecation warning for array_wrap requiring additional arguments (context and return_scalar). This PR updates the array_wrap signature to accept these arguments while preserving existing behavior, ensuring compatibility with NumPy 2. Verified locally by reproducing the warning with NumPy 2 and..."
183550,[AI Codemod][RlAutotuneFbsourceWearablesMemory] PackedLinearWeightsQnnp::apply_dynamic (qlinear_dynamic.cpp),https://github.com/pytorch/pytorch/pull/183550,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, open source, fb-exported, ciflow/trunk, release notes: quantization, meta-exported",2026-05-13T11:28:46Z,2026-05-13T18:32:00Z,56,Reviewed By: jankrom Differential Revision: D104973676 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
182406,Fix misleading error when caching_allocator_alloc size overflows ssize_t,https://github.com/pytorch/pytorch/pull/182406,PR blocked,65,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,BLOCKED,False,0,0,,,182277,182277,182277,,,,0,2,0,,,"triaged, open source",2026-05-05T02:54:19Z,2026-05-13T12:47:59Z,56,"When a Python integer larger than SSIZE_MAX (e.g. 2**63) is passed to caching_allocator_alloc, PyLong_AsSsize_t returns -1 and sets an OverflowError. Without checking for this, the code falls through to TORCH_CHECK_VALUE which reports ""Invalid memory size: -1"" — hiding the real issue (integer overflow) behind a conf..."
183456,[ROCm] [opbench] Temporary profiler workaround for rocprofiler crash (#182719),https://github.com/pytorch/pytorch/pull/183456,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/op-benchmark,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: rocm, open source, topic: not user facing, ciflow/op-benchmark",2026-05-12T17:24:13Z,2026-05-13T11:54:00Z,56,"Summary Workaround for the ROCm rocprofiler / Kineto segfault in #182719: wrap the operator-benchmark Timer paths in a CPU-only torch.profiler.profile context when torch.version.hip is not None, so the profiler is initialized and the crash on AMD is avoided. CUDA / NVIDIA builds are unchanged (nullcontext). This is..."
182398,"Revert ""[dynamo-wrapped] Skip all Dynamo wrapped tests for numpy (#182240)""",https://github.com/pytorch/pytorch/pull/182398,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,182134,,,0,2,0,,,"topic: not user facing, ci-no-td",2026-05-05T01:14:34Z,2026-05-13T07:22:42Z,56,Stack from ghstack (oldest at bottom): #182134 -> #182398 This reverts commit 6e529f3.
183360,Prototype lazy activation checkpoint compile,https://github.com/pytorch/pytorch/pull/183360,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,"183359, 183361",,,0,2,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-11T19:49:54Z,2026-05-13T03:28:06Z,57,Stack from ghstack (oldest at bottom): #183361 -> #183360 #183359 Lower activation checkpointing to a lazy HOP that uses the traced checkpointed forward for forward execution while recomputing through the original function under lazy Dynamo compile. This keeps lifted forward state visible to the outer graph and lets...
183361,Prototype lazy module backward hook compile,https://github.com/pytorch/pytorch/pull/183361,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"183359, 183360",,,0,2,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-11T19:49:58Z,2026-05-13T02:15:55Z,57,"Stack from ghstack (oldest at bottom): -> #183361 #183360 #183359 Install module backward hook boundary nodes in the forward graph while deferring the user hook bodies to lazy Dynamo compilation when autograd invokes them. This keeps eager BackwardHook packing semantics and covers full backward pre-hooks, multi-inpu..."
183359,Prototype lazy autograd.Function backward compile,https://github.com/pytorch/pytorch/pull/183359,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,"183360, 183361",,,0,2,0,,,"module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-11T19:49:50Z,2026-05-13T02:15:17Z,57,"Stack from ghstack (oldest at bottom): #183361 #183360 -> #183359 Trace autograd.Function forward through the existing Dynamo subgraph path so forward tensor state and ctx setup are visible to the outer graph, then defer backward tracing by compiling the original backward when autograd invokes it. The runtime wrappe..."
171686,Add failing test documenting implicit step during `LRScheduler` initialization,https://github.com/pytorch/pytorch/pull/171686,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,168044,,,,0,2,0,,,"triaged, open source, topic: not user facing, no-stale",2026-01-04T23:30:49Z,2026-05-12T23:39:52Z,57,"This PR adds a failing test that documents the current behavior where constructing an LRScheduler performs an implicit learning rate step during initialization. From a user perspective, scheduler construction is typically expected to be side-effect-free, with learning rate updates occurring only after an explicit ca..."
183452,[BE]: Update cutlass submodule to 4.5.0rc,https://github.com/pytorch/pytorch/pull/183452,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, topic: not user facing",2026-05-12T16:47:44Z,2026-05-12T21:53:53Z,57,Make sure nothing breaks
183418,[Avocado][MI350X] Non-overlay compatibility for new AITER/CK overlays (#183418),https://github.com/pytorch/pytorch/pull/183418,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, meta-exported",2026-05-12T08:08:07Z,2026-05-12T18:10:44Z,57,"Summary: Adds select()-guarded compatibility changes to shared paths (caffe2, xplat, flash_attention_v2, third-party/aiter) needed by the new AITER dde1703 + CK v1.2.0 overlays. All changes are guarded by select({""fbcode//ck:avocado_prod"": [...], ""DEFAULT"": []}) or #ifndef guards, so they are no-ops without the avoc..."
181110,Workaround for nvrtcCompileProgram changing locale,https://github.com/pytorch/pytorch/pull/181110,waiting on contributor,35,review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/binaries,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"triaged, open source, ciflow/binaries, release notes: cuda, topic: bug fixes",2026-04-22T14:27:35Z,2026-05-12T15:38:03Z,57,There is a bug in CUDA 11.7 until CUDA 12.6.2 which changes the current thread locale when calling nvrtcCompileProgram. See e.g. https://stackoverflow.com/questions/74044994 This also includes the encoding used by Python by default for e.g. subsequent invocations of subprocess calls. When the user environment is now...
182857,[inductor][mtia] Skip matvec bmm/addmm decomposition on MTIA; add MTIA matvec bmm lowering (#182857),https://github.com/pytorch/pytorch/pull/182857,stale/low urgency,17,"PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"fb-exported, topic: not user facing, module: inductor, ciflow/inductor, meta-exported, ciflow/torchtitan",2026-05-07T18:14:32Z,2026-05-12T15:37:03Z,57,"Summary: Two-part change so MTIA emits a real matmul for matvec torch.bmm (M=1 or N=1), instead of running the broadcast-mul-sum decomposition. Part 1 -- Inductor decomposition (PyTorch side): Add ""mtia"" to the device-type exclusion list in decomp_bmm and decomp_addmm so the coordinate_descent_tuning matvec rewrite..."
183404,test/dynamo: add CPython coverage planning docs,https://github.com/pytorch/pytorch/pull/183404,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, ciflow/torchtitan",2026-05-12T04:13:30Z,2026-05-12T14:57:19Z,57,"Stack from ghstack (oldest at bottom): #183377 -> #183404 Adds the CPython coverage manager and plan docs, plus Dynamo protocol mirroring guidance used by those agents. The import linter now ignores Dynamo Markdown files so the new protocol guidance is treated as documentation. Authored by Codex. cc @voznesenskym @p..."
182274,[docs] Generate doc for values in ScalingType and SwizzleType enum,https://github.com/pytorch/pytorch/pull/182274,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"triaged, open source",2026-05-04T09:58:58Z,2026-05-12T13:55:53Z,57,Current PR Docs: https://docs.pytorch.org/docs/2.12/generated/torch.nn.functional.ScalingType.html https://docs.pytorch.org/docs/2.12/generated/torch.nn.functional.SwizzleType.html PR Doc Link - https://docs-preview.pytorch.org/pytorch/pytorch/182274/nn.functional.html#torch.nn.functional.ScalingType PR Doc Link - h...
177726,Introduce test skip markers for Sandcastle,https://github.com/pytorch/pytorch/pull/177726,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: dataloader",2026-03-18T09:56:12Z,2026-05-12T11:27:16Z,57,"Simplify the markers a bit to make them more expressive It also makes it easier to skip those tests ""manually"" by changing the single definition of the skip marker. This is important to reduce potential false positives (of failed tests) in some environments, such as HPC clusters"
162706,Fix accelerator-related skip-decorators for tests,https://github.com/pytorch/pytorch/pull/162706,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,162179,,,,0,4,0,,,"oncall: distributed, triaged, open source, ciflow/trunk, release notes: distributed (fsdp)",2025-09-11T06:59:29Z,2026-05-12T11:27:07Z,57,Avoid failures caused by tests exiting via sys.exit instead of unittest.skip In particular it will not try to start the test (causing forks into subprocess) just to stop them (killing the subprocess) which is done in the test setup Using unittest.skip decorators avoids the starting of the test in the first place. Re...
166625,Fix skip condition in test_overrides,https://github.com/pytorch/pytorch/pull/166625,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2025-10-30T09:15:42Z,2026-05-12T10:35:17Z,57,"This fixes the condition introduced in in #150059 The reason for the failure is running the test on a non-CUDA machine: File ""/var/lib/jenkins/workspace/test/test_overrides.py"", line 786, in <module> generate_tensor_like_override_tests(TestTorchFunctionOverride) File ""/var/lib/jenkins/workspace/test/test_overrides.p..."
179139,Report ACL version info,https://github.com/pytorch/pytorch/pull/179139,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, release notes: build",2026-04-02T13:19:14Z,2026-05-12T09:36:29Z,57,Especially for reproducibility and bug reports it is important to know the version of the Arm Compute Library too. Show together with MKLDNN version info. Similar to #175236 so cc @Skylion007
183052,Fix OOM error suggesting expandable_segments when already enabled,https://github.com/pytorch/pytorch/pull/183052,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,173049,173049,173049,,,,0,2,0,,,"triaged, open source, release notes: cuda, topic: improvements",2026-05-09T16:13:59Z,2026-05-12T05:42:23Z,57,"Fixes #173049 Summary When a CUDA OOM occurs and expandable_segments is already enabled via PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, the error message still tells the user to try setting it. This is confusing — they already did. The fix checks AcceleratorAllocatorConfig::use_expandable_segments() before inc..."
182612,Ignore decoding errors when reading CMake cache files (errors='ignore'),https://github.com/pytorch/pytorch/pull/182612,waiting on contributor,35,review decision requests changes; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"triaged, open source, topic: not user facing",2026-05-06T03:51:42Z,2026-05-12T02:00:14Z,58,"When reading CMakeCache.txt, invalid byte sequences may appear，which cause build error: File ""tools/setup_helpers/env.py"", line 68, in __init__ cmake_cache_vars = get_cmake_cache_variables_from_file(f) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File ""tools/setup_helpers/cmake_utils.py"", line 57, in get_cmake_cache_varia..."
173072,s390x: rework ZVECTOR data into 16-byte-sized vectors,https://github.com/pytorch/pytorch/pull/173072,waiting on contributor,45,"review decision requests changes; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/s390, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,171870,171870,"154074, 171870",,,,0,10,0,,,"module: cpu, triaged, open source, release notes: releng, module: inductor, ciflow/inductor, ciflow/s390, ciflow/torchtitan",2026-01-22T12:21:53Z,2026-05-11T23:32:03Z,58,Previously 32-byte-sized vector was emulated via pair of 16-byte-sized ones. Now there's infrastructure in pytorch for 16-byte-sized vectors to use. Fixes: #171870 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe...
182934,[RFC][distributed] Add parallel_map with collective operation interception…,https://github.com/pytorch/pytorch/pull/182934,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,review when ready,,UNKNOWN,False,0,0,,,,,180550,,,,0,1,0,,,"triaged, open source, release notes: distributed (c10d), ciflow/torchtitan",2026-05-08T08:05:25Z,2026-05-11T22:15:00Z,58,"Relates to #180550 Motivation Under free-threaded Python (nogil), user code that runs independent iterations in parallel threads can easily hit collective ordering issues — when multiple threads call dist.all_reduce (or other collectives) concurrently, cross-rank pairing becomes non-deterministic, leading to silent..."
181823,Docs: update torch.multinomial num_samples requirement note for bette…,https://github.com/pytorch/pytorch/pull/181823,PR blocked,73,failed checks or blocked/unstable merge state; PR closes or references open issue numbers,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,181347,181347,,,,,0,1,0,,,"triaged, open source, release notes: python_frontend",2026-04-29T02:29:00Z,2026-05-25T11:25:22Z,44,"Fixed inconsistency in torch.multinomial. Previously the note said ""lower than"" but example allows ""equal to"" for num_samples without replacement. Evidence from Example: weights = torch.tensor([0, 10, 3, 0], dtype=torch.float) # Contains exactly 2 non-zero elements torch.multinomial(weights, 2) # Runs successfully,..."
183035,Fix arange int64 fractional-step zero handling on CPU and CUDA.,https://github.com/pytorch/pytorch/pull/183035,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source",2026-05-09T05:00:35Z,2026-05-11T21:51:59Z,58,"Prevent divide-by-zero crashes when a fractional step is used with int64 dtype by validating step magnitude before integer conversion in both CPU and CUDA arange paths, and add regression coverage in tensor creation tests."
181052,Add docs for torch version information,https://github.com/pytorch/pytorch/pull/181052,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,32864,32864,32864,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-04-22T01:25:17Z,2026-05-11T21:39:44Z,58,Fixes #32864 This PR adds documentation entries for torch.__version__ and the torch.version module fields. Testing: Not run; docs-only change.
182975,Document 5 public APIs in docs/source/utils.md,https://github.com/pytorch/pytorch/pull/182975,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,64185,,,,0,1,0,,,"triaged, open source",2026-05-08T17:46:43Z,2026-05-11T21:26:01Z,58,Add documentation for 5 undocumented public APIs in docs/source/utils.md: torch.utils.backend_registration.generate_methods_for_privateuse1_backend torch.utils.backend_registration.rename_privateuse1_backend torch.utils.cpp_backtrace.get_cpp_backtrace torch.utils.hooks.unserializable_hook torch.utils.throughput_benc...
181096,inductor/flex_attention: fix fake block-mask generator for autotuner,https://github.com/pytorch/pytorch/pull/181096,waiting on contributor,45,"review decision requests changes; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,180651,180651,180651,,,,0,2,0,,,"triaged, open source, release notes: distributed (c10d), module: inductor, ciflow/inductor, ciflow/h100-symm-mem, ciflow/torchtitan",2026-04-22T11:33:17Z,2026-05-11T21:25:32Z,58,"Fixes #180651 Summary When the Inductor autotuner benchmarks flex attention configs, it needs fake input tensors to drive the kernels. One of those inputs is the block-mask, which tells the kernel how many KV blocks each Q-row needs to process. The block-mask splits KV blocks into two disjoint sets per Q-row: Partia..."
181367,[CPU] Fix fp16 matmul ~100x slower than fp32 on CPUs without AVX512-FP16,https://github.com/pytorch/pytorch/pull/181367,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,146508,146508,146508,,,,0,1,0,,,"triaged, open source, release notes: linalg_frontend",2026-04-24T12:49:31Z,2026-05-11T20:54:39Z,58,"Fixes #146508. Summary On CPUs without AVX512-FP16 / AMX-FP16 (pre-Sapphire-Rapids Intel, most AMD, older Arm, Apple silicon at CPU-capability=DEFAULT), torch.matmul(fp16) falls through oneDNN's fp16 check and BLAS shgemm_ and lands in a naive scalar kernel in BlasKernel.cpp. That kernel already accumulates in fp32..."
171101,workaround for transpose_mxn test failure on AArch64,https://github.com/pytorch/pytorch/pull/171101,ready for maintainer decision,72,"approved with no detected failing/pending checks; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,170877,170877,,,,0,4,0,,,"module: cpu, open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, merging",2025-12-22T15:38:59Z,2026-05-11T20:08:35Z,58,Stack from ghstack (oldest at bottom): -> #171101 Fixes #170877 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutrub...
180874,Add CPU offload support for selective activation checkpointing,https://github.com/pytorch/pytorch/pull/180874,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,"180866, 180867",,,0,18,0,,,ciflow/torchtitan,2026-04-20T16:48:45Z,2026-05-11T20:06:23Z,58,"Stack from ghstack (oldest at bottom): -> #180874 #180867 #180866 #180919 #176455 When the policy function returns MUST_CPU_OFFLOAD or PREFER_CPU_OFFLOAD, SAC now automatically copies the op's output to pinned CPU memory during forward and restores it to the original device during backward. This avoids both recomput..."
183354,[CI] Match MI355 ROCm shards with trunk,https://github.com/pytorch/pytorch/pull/183354,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"module: rocm, open source, topic: not user facing",2026-05-11T19:02:34Z,2026-05-11T19:30:26Z,58,Summary Increase rocm-mi355 default test shards from 6 to 10. Increase periodic-rocm-mi355 distributed test shards from 3 to 4. Why The MI355 trunk workflow is already producing artifacts with 10 default shards and 4 distributed shards. Aligning the dedicated rocm-mi355 and periodic-rocm-mi355 workflows with those s...
181435,Add operator benchmark monitoring skill for daily regression detection,https://github.com/pytorch/pytorch/pull/181435,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,release notes: benchmark,2026-04-24T19:58:59Z,2026-05-11T19:05:46Z,58,This adds automated monitoring for PyTorch operator microbenchmarks via a Claude Code skill that: Monitors daily operator_microbenchmark workflow runs Detects performance regressions by comparing against baseline CSVs Creates GitHub issues for significant regressions (>20% slowdown) Uses context-aware thresholds bas...
183039,[c10d] Remove false-positive DeviceMesh pickle warning in async checkpoint,https://github.com/pytorch/pytorch/pull/183039,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,review when ready,,UNKNOWN,False,0,0,,,182102,182102,182102,,,,0,1,0,,,"triaged, open source, release notes: distributed (c10d), ciflow/torchtitan",2026-05-09T06:21:07Z,2026-05-11T18:56:36Z,58,"When using dcp.async_save with AsyncCheckpointerType.PROCESS, DTensors are pickled and sent to a checkpoint subprocess. On unpickle, DeviceMesh.__setstate__ tries to resolve the original process groups which don't exist in the subprocess, emitting a spurious warning per PG per DTensor. Fixes #182102"
183028,[Inductor] Generalize parallel Triton precompile to all multi-config kernels,https://github.com/pytorch/pytorch/pull/183028,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,182987,,,0,4,0,,,"module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-05-09T03:16:02Z,2026-05-11T18:22:26Z,58,Stack from ghstack (oldest at bottom): -> #183028 #182987 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @mlazos
183029,[make_fx] Fix wrong 2nd-order gradients by decomposing detach to alias,https://github.com/pytorch/pytorch/pull/183029,draft/noise,53,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,175477,175477,,,,0,4,0,,,"ciflow/trunk, release notes: fx, module: dynamo, ciflow/inductor",2026-05-09T04:11:45Z,2026-05-11T17:39:34Z,58,"Stack from ghstack (oldest at bottom): -> #183029 During make_fx tracing, autograd's SavedVariable::unpack() dispatches aten.detach through TensorImpl::shallow_copy_and_detach_core, which goes through the Python dispatch stack when ProxyTorchDispatchMode is active. These spurious detach nodes get recorded in the FX..."
177867,Add ASIMD_BF16 Vectorized class specialisation,https://github.com/pytorch/pytorch/pull/177867,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"module: cpu, triaged, open source, Merged, Reverted, ciflow/trunk, topic: not user facing, ci-no-td",2026-03-19T16:38:45Z,2026-05-11T16:41:56Z,58,"On AArch64 we compile PyTorch with march=armv8-a. This means the default Vectorized class instance does not support native FP32-to-BF16 conversions, and instead emulates them using several instructions per vector. For smaller operations such as broadcast mul or div, a large part of their runtime is just converting f..."
181359,S390x: fixes for docker image build,https://github.com/pytorch/pytorch/pull/181359,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,12,0,,,"triaged, open source, ciflow/trunk, topic: not user facing",2026-04-24T10:36:56Z,2026-05-11T15:18:31Z,58,"Update OpenSSL old downloads location. Install gcc-toolset-14-binutils-gold: it fixes linking issue on s390x when building nightly wheels. Add retries when building additional dependencies for s390x image, and move building those dependencies into a separate script file. Install fixed protobuf. Update cmake to versi..."
182900,test: relax tolerance of test_Embedding_discontiguous_cuda,https://github.com/pytorch/pytorch/pull/182900,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-05-07T21:14:57Z,2026-05-11T14:47:56Z,58,"Relaxed the tolerance for the test_Embedding_discontiguous_cuda test case. File ""/third_party/py/torch/test/test_nn.py"", line 7934, in with_tf32_off test.test_cuda(self, **kwargs) ~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^ File ""/third_party/py/torch/testing/_internal/common_nn.py"", line 3640, in test_cuda self.test_noncontig(t..."
183254,Add RISC-V native fp16 conversion paths,https://github.com/pytorch/pytorch/pull/183254,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"module: cpu, triaged, open source",2026-05-11T06:41:03Z,2026-05-11T14:42:55Z,58,"Summary This PR adds RISC-V native fp16 conversion support for c10::Half. Use native _Float16 conversion for scalar Half <-> float on RISC-V targets with zfh/zfhmin. Add RVV bulk float <-> Half conversion when v + zve32f + zvfh/zvfhmin are available. Keep existing x86, AArch64, and generic fallback paths unchanged...."
183264,[WIP][Dynamo] Add mark_batch_invariant API,https://github.com/pytorch/pytorch/pull/183264,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,181089,,,0,1,0,,,"module: inductor, module: dynamo, ciflow/inductor, release notes: export, ciflow/torchtitan",2026-05-11T11:18:09Z,2026-05-11T14:19:47Z,58,Stack from ghstack (oldest at bottom): -> #183264 #181089 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98
181089,[Inductor] Disable split-reduction lowering under batch_invariant,https://github.com/pytorch/pytorch/pull/181089,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,183264,,,0,5,0,,,"ciflow/trunk, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-04-22T08:51:15Z,2026-05-11T14:19:44Z,58,"Stack from ghstack (oldest at bottom): #183264 -> #181089 Problem Reduction.num_splits calls V.choices.reduction_split_factor(reduction_numel_hint, numel_hint, ...) where numel_hint is batch-scaled (product of output dims, typically including batch). At bs=N and bs=N/2 the heuristic picks different K → different bf1..."
176001,[test] Resolve undefined GRAPH_EXECUTOR in test_jit_profiling.py,https://github.com/pytorch/pytorch/pull/176001,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing",2026-02-27T18:59:39Z,2026-05-11T14:19:27Z,58,"test/test_jit_profiling.py works fine if launched using pytest since pytest has setup code to parse command arguments. But if launching this test from python it'll fail: > python /opt/pytorch/pytorch/test/test_jit_profiling.py Traceback (most recent call last): File ""/opt/pytorch/pytorch/test/test_jit_profiling.py"",..."
181875,[ROCm][release/2.11] Fix test_transformers.py test suite for ROCm,https://github.com/pytorch/pytorch/pull/181875,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",REVIEW_REQUIRED,BLOCKED,False,0,0,,,,,,,,,0,1,0,,,"module: rocm, open source, release notes: rocm",2026-04-29T14:56:24Z,2026-05-11T11:46:55Z,58,"test/test_transformers.py fails during test collection with: TypeError: 'NoneType' object is not subscriptable. Only release/2.11 is affected as main, release/2.10, and release/2.12 branch have this fix. The fix: Add torch.version.cuda is not None check to IS_THOR to fix test_transformers.py test suite on ROCm. cc @..."
176430,Don't use sharding for tests with a single shard,https://github.com/pytorch/pytorch/pull/176430,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-03-04T14:42:45Z,2026-05-11T11:21:08Z,58,"Passing --num-shards will initialize the (custom) pytest-shard plugin. This will then attempt to filter tests by calculating the SHA256 hashes. With num_shards == 1 no tests will be filtered so this work is not necessary, so ommit the parameter in this case. The only possible issue I see is that the message ""Running..."
159862,Fix skipIfXpu and skipIfHpu and similar skip decorators,https://github.com/pytorch/pytorch/pull/159862,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,150779,150779,150779,,,,0,1,0,,,"triaged, open source, topic: not user facing",2025-08-05T13:36:24Z,2026-05-11T11:21:05Z,58,"The decorators won't work when used on classes so use the unittest skip functions in their implementation. In many cases those can be simple aliases. Fixes #150779 This exposed failures in tests that are currently not run: def skipIfHpu(fn): @wraps(fn) def wrapper(*args, **kwargs): if TEST_HPU: raise unittest.SkipTe..."
174768,Normalize object names in TreeProfiler test,https://github.com/pytorch/pytorch/pull/174768,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-02-11T13:26:03Z,2026-05-11T11:20:52Z,58,"Based on a failure in python profiler/test_profiler_tree.py TestProfilerTree.test_profiler_experimental_tree_with_stack_and_modules Depending on versions of Python, pybind11, ... the object might be slightly different: - <built-in method _get_tracing_state of PyCapsule object at 0xXXXXXXXXXXXX> + <built-in method _g..."
165986,Skip tests requiring NCCL if no GPU is found,https://github.com/pytorch/pytorch/pull/165986,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2025-10-21T12:09:20Z,2026-05-11T11:20:51Z,58,If a test is restricted to NCCL it will very likely also require a CUDA device. Put a check directly into the decorators to skip the tests if no CUDA is present With this many test decorators additionally checking for CUDA/GPUs can be removed as is done for (most?) of the tests using the requires_nccl* decorators. W...
183253,[BE]: Improve torch utils checkpoint typing,https://github.com/pytorch/pytorch/pull/183253,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,1,0,,,"open source, ciflow/torchtitan",2026-05-11T06:02:45Z,2026-05-11T10:16:32Z,58,
174959,Make server creation in distributed/elastic/test_control_plane more reliable,https://github.com/pytorch/pytorch/pull/174959,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2026-02-13T10:21:36Z,2026-05-11T10:01:53Z,58,"Using """" may cause a ""connection refused"" error in some environments. Using the localhost IP is more reliable."
180225,Enable ACL operator fusions for aarch64,https://github.com/pytorch/pytorch/pull/180225,draft/noise,23,"PR is draft; review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/linux-aarch64",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/linux-aarch64",2026-04-13T16:43:23Z,2026-05-11T09:07:09Z,58,"The inductor mkldnn fusion pass previously disabled ALL fusions when ACL was detected as the backend. This was overly conservative -- ACL has supported fused post-ops (relu, tanh, sigmoid, elu, etc.) for a while now. Changes: Register unary, binary, binary+unary, and inplace fusions unconditionally (only quantizatio..."
177126,Fix search for OpenMP using e.g. GCC,https://github.com/pytorch/pytorch/pull/177126,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source",2026-03-11T10:10:52Z,2026-05-11T08:45:15Z,58,When Clang is installed a libomp.so will exist in addition to libgomp.so. GCC uses -fopenmp which will link libgomp.so. However in the MKL workaround path an explicit search for libomp.so is inserted taking precedence over libgomp.so. This is not required as the library search will be done further below based on com...
169936,Fix `TestSaveLoad.test_version_error`,https://github.com/pytorch/pytorch/pull/169936,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, topic: not user facing",2025-12-09T10:27:51Z,2026-05-11T08:45:03Z,58,"The test makes an implicit assumption that $TMP=/tmp which isn't neccessarily the case. The code tries to extract the file stem (name without extension) which can be done clearer and without hardcoding offsets. I suspect the skip condition: @unittest.skipIf( IS_FBCODE or IS_MACOS or IS_WINDOWS, ""The file path is dif..."
167306,Fix integer overflow in (fake) quantization,https://github.com/pytorch/pytorch/pull/167306,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,111471,"111471, 129127",111471,,,,0,3,0,,,"module: cpu, triaged, open source, release notes: quantization",2025-11-07T09:29:04Z,2026-05-11T08:45:02Z,58,The static_cast<int64_t> can overflow for large float values and/or a small scale (e.g. 9.2e14 & 1e-4) Fix a similar issue in the mask calculation where std::lrint is used which may convert to a 32 bit float returning an implementation defined value on overflow. Stay in float mode using std::round and fmin/fmax to a...
176709,Remove Eigen d_cnjg and r_cnjg source reference,https://github.com/pytorch/pytorch/pull/176709,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, open source",2026-03-06T11:22:31Z,2026-05-11T07:56:57Z,58,Eigen no more ships these two files since 5.x
175969,[inductor][xpu] Fix convolution_backward accuracy for mixed channels-last / contiguous inputs,https://github.com/pytorch/pytorch/pull/175969,draft/noise,53,"PR is draft; failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan, ciflow/trunk, ciflow/xpu","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,1877,1877,,,,,0,8,0,,,"open source, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, ciflow/xpu, ciflow/torchtitan",2026-02-27T13:24:19Z,2026-05-11T07:32:58Z,58,"Fixes: intel/torch-xpu-ops#1877 Inductor's layout optimization converts conv weights to channels-last in the forward graph, but the input tensor saved for backward retains its original contiguous strides. This creates a mixed-format scenario where convolution_backward receives a channels-last weight but contiguous i..."
182779,[WIP] Test playground,https://github.com/pytorch/pytorch/pull/182779,draft/noise,43,"PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/dtensor, ciflow/h100-symm-mem, ciflow/inductor, ciflow/inductor-pallas","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, release notes: distributed (checkpoint), ciflow/h100-symm-mem, ciflow/inductor-pallas, ciflow/dtensor",2026-05-07T04:50:35Z,2026-05-11T06:05:27Z,58,cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @azahed98
183211,[FSDP] Break cross-language ref cycle keeping FlatParameter alive,https://github.com/pytorch/pytorch/pull/183211,draft/noise,5,"PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,1,0,,,"release notes: distributed (fsdp), ciflow/inductor, ciflow/torchtitan",2026-05-10T23:49:20Z,2026-05-11T03:11:52Z,59,Stack from ghstack (oldest at bottom): -> #183211 FSDP v1 wraps a module's parameters in a FlatParameter and stores ParamInfo/SharedParamInfo entries on flat_param that strong-ref the wrapped nn.Modules. The wrapped modules in turn hold view tensors of flat_param whose grad_fn (ViewBackward) carries an Edge to flat_...
183009,[Inductor] Fall back native_layer_norm; route var_mean through inductor lowering,https://github.com/pytorch/pytorch/pull/183009,draft/noise,15,"PR is draft; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for author to mark ready,,CLEAN,True,0,0,,,,168126,168126,,,,0,3,0,,,"topic: not user facing, module: inductor, ciflow/inductor, ciflow/torchtitan",2026-05-08T23:25:46Z,2026-05-11T03:11:51Z,59,"Stack from ghstack (oldest at bottom): -> #183009 Under bf16 autocast, the native_layer_norm decomposition produces slightly different fp32 results than the native kernel (~7e-7 max diff from different reduction algorithms). The subsequent matmul rounds to bf16, and values near bf16 rounding boundaries tip different..."
178390,[inductor] fold half addmm+relu into _addmm_activation,https://github.com/pytorch/pytorch/pull/178390,waiting on contributor,35,"review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/torchtitan",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, module: inductor, ciflow/inductor, release notes: inductor, ciflow/torchtitan",2026-03-25T14:31:39Z,2026-05-10T20:49:42Z,59,"Summary fold GPU relu(addmm(...)) on torch.float16/torch.bfloat16 into aten._addmm_activation in the post-grad pattern matcher keep the change narrowly gated on keep_addmm_fused_for_half_dtypes, GPU tensor metadata, and single-use addmm outputs add pattern-matcher coverage for the enabled and disabled cases Validati..."
183105,Fix formatting in extensions.json,https://github.com/pytorch/pytorch/pull/183105,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, ciflow/trunk, topic: not user facing",2026-05-10T16:01:33Z,2026-05-10T20:02:14Z,59,Description Fixes invalid JSON syntax in .vscode/extensions.json by removing a trailing comma from the recommendations array. Type of Change Bug fix (non-breaking change which fixes an issue) Checklist JSON syntax is now valid VSCode extensions configuration will load correctly
152828,[MSVC] Remove redundant MEM_EFF_ATTENTION_CUDA_SOURCES definition as /Zc:lambda is set globally for C++20 by default,https://github.com/pytorch/pytorch/pull/152828,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,92600,92600,,,,0,1,0,,,"module: build, module: windows, module: cpu, triaged, open source, topic: not user facing, skip-url-lint",2025-05-05T15:20:29Z,2026-05-10T18:50:12Z,59,"Fixes: #92600 MEM_EFF_ATTENTION_CUDA_SOURCES is used to set /Zc:lambda compilation flag for memory-efficient attention sources. /Zc:lambda is automatically enabled by the /std:c++20. Since C++20 is used to compile the code, MEM_EFF_ATTENTION_CUDA_SOURCES is no longer needed. cc @malfet @seemethere @peterjc123 @mszha..."
183054,coverage-td plan,https://github.com/pytorch/pytorch/pull/183054,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,topic: not user facing,2026-05-09T16:37:10Z,2026-05-10T14:28:20Z,59,Stack from ghstack (oldest at bottom): -> #183054
177511,[export] preserve original traceback when augmenting export error messages,https://github.com/pytorch/pytorch/pull/177511,PR blocked,65,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,177510,177510,177510,,,,0,1,0,,,"triaged, open source, module: dynamo, ciflow/inductor, release notes: export",2026-03-16T09:05:25Z,2026-05-10T05:35:42Z,59,"Fixes #177510 Summary torch.export.export() mutates exception messages to append a helpful draft_export() hint, but then re-raises with raise e, which resets __traceback__ to the except block rather than the real failure site. This PR replaces raise e with bare raise so the original traceback is preserved while the..."
178245,Replace c10::call_once with static initialization,https://github.com/pytorch/pytorch/pull/178245,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, release notes: sparse, ciflow/inductor",2026-03-24T06:35:22Z,2026-05-10T02:45:08Z,60,This PR replace some callers of c10::call_once with static initialization to simplify the code.
182424,Warning oversized storage copy,https://github.com/pytorch/pytorch/pull/182424,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,"65604, 122099, 182105",,,,0,4,0,,,"triaged, open source",2026-05-05T10:09:12Z,2026-05-18T09:53:37Z,51,"Adresses #182105. Multiple users have reported confusion regarding tensor view serialization, where the full underlying storage is serialized instead of just the viewed data (#122099, #95278, #65604, #1995). While this behavior is intentional (to preserve storage sharing semantics), it is often surprising and can le..."
180602,Enable ruff rule PLW1510: require explicit check= in subprocess calls,https://github.com/pytorch/pytorch/pull/180602,PR blocked,65,"failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/dtensor, ciflow/inductor, ciflow/torchtitan","identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,115016,115016,115016,,,,0,3,0,,,"triaged, open source, release notes: releng, topic: not user facing, module: inductor, module: dynamo, ciflow/inductor, ciflow/torchtitan, ciflow/dtensor",2026-04-16T19:48:59Z,2026-05-20T15:55:18Z,49,Fixes #115016 Summary Adds PLW1510 to the ruff select list in pyproject.toml and fixes all 89 violations across 53 files. The rule requires an explicit check= argument on subprocess.run() and similar calls. Each call site was examined to determine the correct value: check=True (23 calls): Where subprocess failure sh...
179863,Teach CSAN to detect caching allocator memory reuse races,https://github.com/pytorch/pytorch/pull/179863,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,7,0,,,"triaged, open source, release notes: rocm, release notes: cuda",2026-04-09T21:09:38Z,2026-05-14T05:15:44Z,55,"CSAN (CUDA Sanitizer) tracks tensor-level accesses to detect cross-stream data races, but previously could not detect races caused by the caching allocator reusing memory. When a tensor was freed, CSAN discarded its access history. If the allocator later handed the same memory block to a new allocation on a differen..."
141842,Refactor NJT to hold metadata on nested int,https://github.com/pytorch/pytorch/pull/141842,stale/low urgency,17,"PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk",review when ready,,CLEAN,False,0,0,,,,,,"145922, 146101, 146172",,,0,61,0,,,"ciflow/trunk, release notes: nested tensor, fx, module: dynamo, ciflow/inductor, no-stale, keep-going, suppress-bc-linter",2024-12-01T23:07:42Z,2026-05-06T20:42:48Z,63,Stack from ghstack (oldest at bottom): #146172 #146101 #145922 -> #141842 #141841 #146052 Design: https://docs.google.com/document/d/1HV9719blS8OJxf8kuW5U3ihaoTR_H7sJJ29U7mT4J1g/edit?tab=t.0#heading=h.w4x2tmi9rtmd cc @ezyang @SherlockNoMad @EikanWang @jgong5 @wenzhe-nrv @voznesenskym @penguinwu @Guobing-Chen @Xiaobi...
174025,Add test for CosineAnnealingLR when T_max = 0,https://github.com/pytorch/pytorch/pull/174025,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,168044,,,,0,1,0,,,"triaged, open source, topic: not user facing, no-stale",2026-02-02T01:48:02Z,2026-05-05T15:53:30Z,64,"PR description This PR adds a test that documents what currently happens when CosineAnnealingLR is constructed with T_max = 0. Internally, CosineAnnealingLR uses T_max as both a divisor and a modulo base when computing the learning rate during step(). There is no validation on T_max at construction time, so when T_m..."
119977,[NJT] Factory function support,https://github.com/pytorch/pytorch/pull/119977,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,CLEAN,False,0,0,,,,,119346,117905,,,0,12,0,,,"module: dynamo, ciflow/inductor, no-stale",2024-02-15T05:25:23Z,2026-05-04T22:34:38Z,65,Stack from ghstack (oldest at bottom): #117905 -> #119977 #119976 #121481 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @rec @aakhundov
132414,Ensure FakeTensor post-creation has associated nested int cached,https://github.com/pytorch/pytorch/pull/132414,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,CLEAN,False,0,0,,,,,,132814,,,0,16,0,,,"ciflow/inductor, no-stale",2024-08-01T15:39:54Z,2026-05-04T22:34:33Z,65,Stack from ghstack (oldest at bottom): #132814 -> #132414 #133196 #133145
161254,Add `zero_grad` param for `optimizer.step`,https://github.com/pytorch/pytorch/pull/161254,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,3,0,,,"triaged, open source, ciflow/trunk, no-stale, release notes: optim",2025-08-22T07:29:20Z,2026-04-29T01:54:20Z,71,Fixes part of #158638 Test Result pytest test/test_optim.py ..................................................ssssssssssssssssssssssss..............ssssssssssssss.....................s.............s.............s............................................................................................................
134942,Enabling ATen Distribution kernels for AARCH64 using OpenRNG,https://github.com/pytorch/pytorch/pull/134942,waiting on contributor,35,review decision requests changes; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"module: performance, module: cpu, triaged, open source, module: arm, Stale, release notes: build, topic: not user facing, no-stale",2024-09-02T06:23:36Z,2026-04-20T18:28:18Z,79,"Enables Distribution ATen kernel for ARM CPU's using OpenRNG Library. Enhances performance of bernoulli and exponential distribution operations for ARM CPU's. (https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/cpu/DistributionKernels.cpp) Contribution: With this PR, we propose to add OpenRNG to the p..."
179936,[MPS] Fix avg_pool2d precision by using Metal kernel instead of MPSGraph,https://github.com/pytorch/pytorch/pull/179936,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,179608,179608,179608,,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-04-10T17:37:35Z,2026-05-20T08:15:02Z,49,"Fixes #179608 Summary MPSGraph's avgPooling2DWithSourceTensor accumulates float32 precision error across large inputs, causing avg_pool1d/avg_pool2d to produce negative values from non-negative input when large floats precede zeros. This switches the avg_pool2d forward pass to always use the Metal kernel (per-window..."
180444,Add `validate` parameter to `torch.multinomial` to skip input validation,https://github.com/pytorch/pytorch/pull/180444,has linked issue,35,PR closes or references open issue numbers; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,177127,177127,177127,,,,0,3,0,,,"triaged, open source, release notes: mps, ciflow/mps",2026-04-15T07:52:44Z,2026-06-01T01:55:05Z,38,"Summary Adds a validate keyword argument (default True) to torch.multinomial. When validate=False, the 10 GPU validation kernels (aminmax, sum, assert_async, etc.) on the fast path (!with_replacement || n_sample == 1) are skipped entirely. Fixes #177127 Motivation As profiled in the issue, torch.multinomial spends ~..."
172991,Fix torch.compile for jagged nested tensors created inside compiled functions,https://github.com/pytorch/pytorch/pull/172991,waiting on contributor,45,review decision requests changes; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,168307,168307,168307,,,,0,2,0,,,"triaged, open source, Stale, release notes: fx, release notes: nested tensor, fx, module: dynamo, ciflow/inductor, no-stale",2026-01-21T21:15:19Z,2026-04-16T13:38:42Z,83,"Fixes #168307 Description torch.compile fails when creating jagged nested tensors inside a compiled function with symbolic dimensions: @torch.compile def forward(self, x): batch_size = x.shape[0] # symbolic during compile expanded_i = self.i.expand(batch_size, -1, -1) nested_i = torch.nested.as_nested_tensor(expande..."
173644,[TORCH] [JIT] Change emitDefautlArgsWithOutArgs -> emitDefaultArgsWithOutArgs in export_module.cpp,https://github.com/pytorch/pytorch/pull/173644,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,173643,173643,173643,,,,0,1,0,,,"triaged, open source, Stale, release notes: jit, no-stale",2026-01-28T14:41:42Z,2026-04-12T17:26:12Z,87,"Fixes #173643 Important: This may potentially (as potentially all external (if this is user facing, I'm not quite sure) typo changes) cause backwards incompatibility and has to be noted in the release notes in the correct passage (I've not said this explicitly at all typo changes because this seems clear to me, but..."
127702,Define PY_SSIZE_T_CLEAN before #include <Python.h>,https://github.com/pytorch/pytorch/pull/127702,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,38019,38019,38019,,,,0,1,0,,,"triaged, open source, release notes: cpp, release notes: python_frontend, no-stale",2024-06-02T10:15:22Z,2026-04-07T17:02:40Z,92,See https://docs.python.org/3/c-api/intro.html#include-files Fixes #38019
168223,Add failing test for negative epoch crash in LR schedulers (ZeroDivisionError),https://github.com/pytorch/pytorch/pull/168223,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,168044,,,,0,1,0,,,"module: optimizer, triaged, open source, topic: not user facing, no-stale",2025-11-20T05:39:28Z,2026-04-04T12:42:00Z,95,"This PR introduces a failing test that exposes a ZeroDivisionError triggered when calling scheduler.step(epoch=<negative_value>) on several PyTorch learning rate schedulers. The issue occurs because negative epoch values are accepted without validation, which allows invalid states to propagate into closed-form LR co..."
178850,[xpu][feature] Add fork-safe is_available() via PYTORCH_ZES_BASED_XPU_CHECK env var,https://github.com/pytorch/pytorch/pull/178850,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,3,0,,,"open source, ciflow/trunk, no-stale",2026-03-31T04:24:33Z,2026-03-31T13:27:45Z,99,Stack from ghstack (oldest at bottom): -> #178850 #178496
170249,[do not review] Expose API to create ownership token for autograd.Function grad_fn,https://github.com/pytorch/pytorch/pull/170249,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,170250,,,0,6,0,,,no-stale,2025-12-11T22:56:33Z,2026-03-27T18:49:13Z,103,Stack from ghstack (oldest at bottom): #170250 -> #170249 #172885
130752,support zb1p and zb2p algorithms,https://github.com/pytorch/pytorch/pull/130752,stale/low urgency,17,"PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan, ciflow/trunk",review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"oncall: distributed, triaged, open source, Merged, Reverted, ciflow/trunk, release notes: distributed (pipeline), no-stale, ciflow/torchtitan",2024-07-15T18:36:21Z,2026-03-26T21:24:16Z,104,"Previously, we have proved that ZB2P is not truly zero bubble when num_local_stages exceed 4 and so only ZB1P was supported. We did a few tweaks to the ZB2P to really make it zero bubble. Algorithm and proof is attached. zero_bubble.pdf cc @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta @msaroufim @..."
170250,[do not review] Use new GradientEdge creation in PP,https://github.com/pytorch/pytorch/pull/170250,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,170249,,,0,6,0,,,no-stale,2025-12-11T22:56:38Z,2026-03-24T19:05:48Z,106,Stack from ghstack (oldest at bottom): -> #170250 #170249 #172885
130856,[pipelining] lazy shape inference for stage,https://github.com/pytorch/pytorch/pull/130856,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"oncall: distributed, open source, no-stale, ciflow/torchtitan",2024-07-16T20:43:24Z,2026-03-21T22:38:50Z,109,Stack from ghstack (oldest at bottom): -> #130856 cc @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta @msaroufim @dcci @aditvenk @xmfan @XilunWu @kwen2501 @c-p-i-o
130556,Allow pipeline stage to take non-tensor args and kwargs,https://github.com/pytorch/pytorch/pull/130556,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,130055,,,,0,4,0,,,"oncall: distributed, open source, no-stale, ciflow/torchtitan",2024-07-11T17:55:57Z,2026-03-21T22:38:41Z,109,"Stack from ghstack (oldest at bottom): -> #130556 This solves changes 1 and 2 in #130055 Adds support for args, kwargs to be passed into pipeline stage support. Adds a transform_fwd_inputs which allows the user to customize how they want their arguments to be passed into the model forward. Still need to refine trans..."
153557,"[PP] wip, allow grad to be None",https://github.com/pytorch/pytorch/pull/153557,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,wait for author to mark ready,,CLEAN,True,0,0,,,,,153484,153893,,,0,3,0,,,"oncall: distributed, open source, release notes: distributed (pipeline), no-stale, module: pipelining, ciflow/torchtitan",2025-05-14T17:36:12Z,2026-03-20T22:40:44Z,110,Stack from ghstack (oldest at bottom): #153893 -> #153557 #153498 #153484 cc @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta @msaroufim @dcci @aditvenk @xmfan
153893,[PP] Fix double backward error in stage_backward,https://github.com/pytorch/pytorch/pull/153893,draft/noise,15,PR is draft; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/torchtitan,wait for author to mark ready,,CLEAN,True,0,0,,,,153485,153485,153557,,,0,2,0,,,"oncall: distributed, open source, release notes: distributed (pipeline), no-stale, module: pipelining, ciflow/torchtitan",2025-05-19T23:19:35Z,2026-03-20T22:40:30Z,110,Stack from ghstack (oldest at bottom): -> #153893 #153557 #153498 fixes #153485 cc @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta @msaroufim @dcci @aditvenk @xmfan
172109,stage executor,https://github.com/pytorch/pytorch/pull/172109,draft/noise,43,PR is draft; failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"open source, release notes: distributed (pipeline), no-stale",2026-01-09T19:32:43Z,2026-03-20T22:32:07Z,110,Stack from ghstack (oldest at bottom): -> #172109 #171807
167894,[inductor] split reduction even if all reads are broadcasted,https://github.com/pytorch/pytorch/pull/167894,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,4,0,,,"ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, no-stale",2025-11-15T01:16:28Z,2026-03-06T22:02:57Z,124,"Stack from ghstack (oldest at bottom): -> #167894 With split reduction we can speedup the following (extreme) kernel by 48x # 56ms -> 1.163ms import torch from triton.testing import do_bench def f(x): return x.sum(dim=(0, 1)) x = torch.randn(100000000, 1, 2, device=""cuda"").expand(-1, 2, -1) opt_f = torch.compile(f)..."
166732,Introducing statefulness for single process dataloader,https://github.com/pytorch/pytorch/pull/166732,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,29,0,,,"module: dataloader, Stale, ciflow/trunk, release notes: dataloader, no-stale",2025-10-31T17:10:56Z,2026-02-26T08:44:11Z,132,We (@scotts @aelavender @psaikaushik @divyanshk) are the maintainers of the torchdata (https://github.com/meta-pytorch/data) repo. We plan to upstream the stateful functionality of StatefulDataLoader (https://github.com/meta-pytorch/data/blob/main/torchdata/stateful_dataloader/stateful_dataloader.py) to the PyTorch...
157193,[c10d] Call destroy_process_group upon interpreter exit,https://github.com/pytorch/pytorch/pull/157193,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,115388,,,,0,2,0,,,"oncall: distributed, open source, ciflow/trunk, release notes: distributed (c10d), no-stale",2025-06-28T08:14:52Z,2026-02-23T23:30:39Z,135,"Stack from ghstack (oldest at bottom): -> #157193 This change has two effects: Users are no longer required to call destroy_process_group explicitly, yet the same effect is still achieved. Compared to explicit calling, the invocation time is deferred, issue #115388 does not show up anymore. Why the deferral is helpf..."
168895,Add `pad_to_multiple_of` to `pad_sequence` (C++ only),https://github.com/pytorch/pytorch/pull/168895,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,150989,,,,0,3,0,,,"module: nn, module: rnn, triaged, open source, Stale, release notes: cpp, no-stale",2025-11-22T09:08:40Z,2026-02-03T06:32:28Z,155,"Reattempt of #151021 Related: #150989 Hugging Face PreTrainedTokenizerFast.__call__ pad_to_multiple_of (int, optional) — If set will pad the sequence to a multiple of the provided value. Requires padding to be activated. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capa..."
167541,[MPS] Fix crashes due to a race condition in the MPS backend,https://github.com/pytorch/pytorch/pull/167541,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,6,0,,,"triaged, open source, Stale, release notes: mps, no-stale",2025-11-11T13:28:33Z,2026-01-26T10:21:33Z,163,There's a race between threads when executing unary functions with the MPS backend. This is because the caching of newly created pipeline states isn't guarded and hence can be mutated concurrently by multiple threads. The following program will trigger the race condition (needs to be run in a loop because it doesn't...
140613,"Do not merge, do not close",https://github.com/pytorch/pytorch/pull/140613,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"open source, topic: not user facing, no-stale",2024-11-13T21:19:25Z,2026-01-11T02:18:04Z,179,Used for a test in test_gitutils for ghstacked PRs that a function checking that the diff between head + base and orig + parent. Common ways this can be broken are: changing the head branch and pushing directly committing a change from a review comment in the Github UI (used here) Stack from ghstack (oldest at botto...
165800,[flex_attention]: adding a heuristic for USE_TMA,https://github.com/pytorch/pytorch/pull/165800,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/b200, ciflow/inductor, ciflow/trunk","identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"fb-exported, ciflow/trunk, topic: not user facing, module: inductor, ciflow/inductor, no-stale, meta-exported, ciflow/b200",2025-10-17T22:31:40Z,2025-12-30T18:42:02Z,190,Summary: Adding a USE_TMA heuristic which captures almost all of the perf. The best perf improvements are seen in document_mask+large shapes which are very important for training. Mean: +15.8% improvement overall Only a single shape (below) sees a perf degradation of >5%. Document Mask: 37.8% average improvement Str...
164128,Consistently use c10_ovrsource in arvr mode everywhere,https://github.com/pytorch/pytorch/pull/164128,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"fb-exported, Merged, Reverted, ciflow/trunk, topic: not user facing, no-stale, ci-no-td, meta-exported",2025-09-29T13:42:21Z,2025-12-29T04:45:55Z,191,"Summary: Previously, many arvr targets transitively depended on c10, not c10_ovrsource, because they either explicitly depended on c10 (because they didn't know better) or they depended on legacy Caffe2, which never got the ovrsource treatment. So we found all these spots (driven by D82283623) and forced them to que..."
166813,Add min/max support for barebones uint types,https://github.com/pytorch/pytorch/pull/166813,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,7,0,,,"module: cpu, Merged, Reverted, ciflow/trunk, release notes: python_frontend, topic: new features, no-stale, ci-no-td",2025-11-02T04:24:13Z,2025-12-29T04:45:53Z,191,Stack from ghstack (oldest at bottom): #166951 -> #166813 Signed-off-by: Edward Z. Yang ezyang@meta.com cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01
162033,[RFC] Switch all of c10d to use .h instead of .hpp,https://github.com/pytorch/pytorch/pull/162033,PR blocked,55,"failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/h100-symm-mem, ciflow/inductor","identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"oncall: distributed, release notes: distributed (c10d), skip-pr-sanity-checks, ciflow/inductor, no-stale, release notes: inductor (aoti), ciflow/h100-symm-mem",2025-09-03T01:36:45Z,2025-12-21T03:17:44Z,200,"Stack from ghstack (oldest at bottom): -> #162033 It would have saved me some grief in our internal build system if the headers were .h instead of .hpp, since the .hpp were missed by some .h only globs. I did this with claude code so no big deal if it is rejected. Signed-off-by: Edward Yang ezyang@meta.com cc @H-Hua..."
154395,Updated padding validation in max_pool functions to account for dilation,https://github.com/pytorch/pytorch/pull/154395,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,154262,154262,154262,,,,0,4,0,,,"triaged, open source, Stale, release notes: nn, no-stale",2025-05-27T04:46:39Z,2025-12-05T16:55:38Z,215,"Fixes #154262 Modified padding validation in max_pool1d, max_pool2d and max_pool3d to use the effective kernel size (taking dilation into account). This change prevents valid padding values from being incorrectly rejected when using dilation."
160888,Allow branching on whether you are in AC region with is_checkpoint_enabled,https://github.com/pytorch/pytorch/pull/160888,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,160883,,,0,10,0,,,"release notes: autograd, topic: new features, module: dynamo, ciflow/inductor, no-stale",2025-08-18T15:29:20Z,2025-12-05T01:43:53Z,216,Stack from ghstack (oldest at bottom): #162434 #160883 -> #160888 Also test with SAC? cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @Lucaskabela
160878,[WIP] Introduce CachingDeviceAllocatorInterface as a base impl,https://github.com/pytorch/pytorch/pull/160878,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"160182, 160638, 160872",,,0,19,0,,,"open source, Stale, no-stale",2025-08-18T12:17:36Z,2025-11-27T01:35:47Z,224,Stack from ghstack (oldest at bottom): #160958 -> #160878 #160872 #160638 #160182 #160082 #160067 #159553
160872,[WIP] [3/N] Introduce a generic CachingDeviceAllocatorImpl for cross backend use,https://github.com/pytorch/pytorch/pull/160872,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"160182, 160638, 160878",,,0,18,0,,,"open source, Stale, no-stale",2025-08-18T07:59:24Z,2025-11-27T01:35:25Z,224,Stack from ghstack (oldest at bottom): #160958 #160878 -> #160872 #160638 #160182 #160082 #160067 #159553
160638,[WIP] [2/N] Introduce a generic CachingDeviceAllocatorImpl for cross backend use,https://github.com/pytorch/pytorch/pull/160638,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"160182, 160872, 160878",,,0,22,0,,,"open source, Stale, no-stale",2025-08-14T12:25:21Z,2025-11-27T01:35:02Z,224,Stack from ghstack (oldest at bottom): #160958 #160878 #160872 -> #160638 #160182 #160082 #160067 #159553
160182,[WIP] [1/N] Introduce a generic CachingDeviceAllocatorImpl for cross backend use,https://github.com/pytorch/pytorch/pull/160182,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,"160638, 160872, 160878",,,0,27,0,,,"open source, Stale, no-stale",2025-08-08T09:45:50Z,2025-11-27T01:34:25Z,224,Stack from ghstack (oldest at bottom): #160958 #160878 #160872 #160638 -> #160182 #160082 #160067 #159553
148900,[RFC][BE] assume error checking is on by default (#141914),https://github.com/pytorch/pytorch/pull/148900,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,2,0,,,"oncall: distributed, open source, ciflow/trunk, release notes: distributed (c10d), no-stale",2025-03-10T18:37:02Z,2025-11-22T02:02:11Z,229,Stack from ghstack (oldest at bottom): #148903 -> #148900 Summary: Remove conditional MACRO ENABLE_NCCL_ERROR_CHECKING and assume that error checking is always on. These checks were wrapped in a macro because older NCCL libraries didn't have the pre-requisite functions to do error checks. This check was put in sever...
162130,Raise error when no record on `extra_files`,https://github.com/pytorch/pytorch/pull/162130,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,152178,152178,152178,,,,0,8,0,,,"oncall: jit, triaged, open source, Stale, release notes: jit, no-stale",2025-09-04T02:47:55Z,2025-11-17T15:00:50Z,233,Fixes #152178 Reopen to merge for PR #152664 cc @EikanWang @jgong5 @wenzhe-nrv @sanchitintel
162079,Tune linalg_eigh_cusolver: better heuristic for syevj_batched selection on cuda,https://github.com/pytorch/pytorch/pull/162079,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"triaged, enhancement, open source, module: linear algebra, Stale, release notes: linalg_frontend, no-stale",2025-09-03T18:05:33Z,2025-11-15T20:20:24Z,235,"Supersedes #151118. Same changes, reopened because the bot closed it due to inactivity. This change is not tied to an open issue. Summary This PR updates the heuristics in linalg_eigh_cusolver for batched matrix diagonalization. The current logic only applies syevj_batched for matrix sizes ≤ 32, which is too conserv..."
160883,Fallback to eager AC when backend=eager for TagActivationCheckpoint,https://github.com/pytorch/pytorch/pull/160883,PR blocked,65,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,"identify whether block is CI, merge conflict, or review gate",APPROVED,UNSTABLE,False,1,1,statusCheckRollup,statusCheckRollup,,160773,,160888,,,0,9,0,,,"topic: not user facing, module: dynamo, ciflow/inductor, no-stale",2025-08-18T14:36:42Z,2025-11-11T15:21:08Z,239,Stack from ghstack (oldest at bottom): #162434 -> #160883 #160888 Fixes #160773 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @Lucaskabela
148668,Fix CUPTI lookup to include target directory,https://github.com/pytorch/pytorch/pull/148668,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,1,0,,,"triaged, open source, no-stale",2025-03-06T13:59:09Z,2025-10-23T14:20:34Z,258,CUPTI library and headers are installed to the target subdirectory rather than the top-level prefix in conda-forge. Include CUDAToolkit_TARGET_DIR subdirectories in CUPTI search paths to fix finding it in that environment.
152104,Update _torch_docs.py to Fix torch.bernoulli(),https://github.com/pytorch/pytorch/pull/152104,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,152095,152095,152095,,,,0,7,0,,,"triaged, open source, Stale, release notes: python_frontend, no-stale",2025-04-24T15:02:31Z,2025-09-26T04:01:01Z,285,"Fixes #152095 @malfet Wondering whether to fix signature that from: @overload def bernoulli(input: Tensor, p: _float, *, generator: Optional[Generator] = None) -> Tensor: to @overload def bernoulli(input: Tensor, p: _float, *, generator: Optional[Generator] = None, out: Optional[Tensor] = None) -> Tensor: Or just me..."
142160,Introduce int32 index_fill and index_copy indices,https://github.com/pytorch/pytorch/pull/142160,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,142090,,,,,0,1,0,,,"module: cpu, triaged, open source, Stale, release notes: cuda, no-stale",2024-12-05T19:27:29Z,2025-09-04T19:12:52Z,307,"Fixes #142090 This PR extends index_fill and index_copy operations to support int32 indices in addition to the existing int64 support: Memory Efficiency and potential performance for handling a large number of indices, particularly when needing to transfer to accelerator backends. In some cases, the compiler may not..."
155672,Fixed NLLLoss 1D input crash with torch.compile,https://github.com/pytorch/pytorch/pull/155672,PR blocked,65,failed checks or blocked/unstable merge state; PR closes or references open issue numbers; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,155247,155247,,,,,0,1,0,,,"triaged, open source, Stale, release notes: nn, no-stale",2025-06-11T06:33:55Z,2025-08-17T08:53:54Z,325,"Fixes #155247 Fixes a bug where nn.NLLLoss fails with 1D input (shape [N]) when using torch.compile(), even though eager mode accepts it"
138388,Extending SVE VEC Backend Support in PyTorch to SVE128 and SVE512.,https://github.com/pytorch/pytorch/pull/138388,waiting on contributor,35,"review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/linux-aarch64",wait for contributor update,CHANGES_REQUESTED,UNKNOWN,False,0,0,,,,,,,,,0,15,0,,,"module: cpu, triaged, open source, module: arm, release notes: sparse, module: inductor, module: dynamo, ciflow/inductor, no-stale, ciflow/linux-aarch64",2024-10-19T14:33:46Z,2025-08-11T18:47:31Z,331,This is a follow up PR which extends SVE VEC backend support to SVE128 and SVE512: Main OSS PR: #119571 - Extending the PyTorch VEC backend for SVE (ARM) with SVE256 Features: Adding support only for SVE ISA with the vector length of 128 and 512. This will leverage the already existing SVE code as SVE is VLA. Update...
102148,Add mps support for maxpool3d,https://github.com/pytorch/pytorch/pull/102148,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/mps,review when ready,,UNKNOWN,False,0,0,,,100674,100674,,,,,0,2,0,,,"triaged, open source, release notes: mps, ciflow/mps, no-stale",2023-05-24T05:32:57Z,2025-07-30T18:42:18Z,343,"Fixes #100674 Added mps support for forward / backward passes of maxpool3d. As mentioned in the referenced issue, I'm utilizing the maxpool 4d op that is available from the metal library to achieve this. cc: @mattiaspaul"
147365,Replace `fw_metadata` info with trace log hint in hint message,https://github.com/pytorch/pytorch/pull/147365,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,147135,147135,147135,,,,0,1,0,,,"triaged, open source, release notes: AO frontend, no-stale",2025-02-18T08:52:44Z,2025-07-28T17:34:49Z,345,"Fixes #147135 Test Result RuntimeError: Found a graph input that requires gradients, and received a mutation. This is currently banned in the aot_export workflow. If you need this functionality, please file a github issue and submit the trace log. Get trace log by running with `TORCH_TRACE`: TORCH_TRACE=""/tmp/traced..."
152158,Add AC_TRACER Infra TorchDispatchMode key,https://github.com/pytorch/pytorch/pull/152158,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,CLEAN,False,0,0,,,,,,,,,0,25,0,,,"topic: not user facing, no-stale",2025-04-25T03:17:05Z,2025-07-24T18:26:34Z,349,"Stack from ghstack (oldest at bottom): -> #152158 #153487 Why we need an additional infra mode for the new version of AC? For the new version of AC, we want to trace a graph and then replay it (in pieces) during backward. We'd like this graph to have all the user modes and subclasses already desugared so that during..."
153317,[associative_scan] Autograd for additional inputs,https://github.com/pytorch/pytorch/pull/153317,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,34,0,,,"open source, Stale, topic: not user facing, module: dynamo, ciflow/inductor, no-stale",2025-05-10T11:22:00Z,2025-07-09T18:22:39Z,364,This PR implements the autograd feature for additional inputs and is based on #139939 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames
97825,Support sparse COO/CSR/CSC/BSR/BSC return values in gradcheck input function,https://github.com/pytorch/pytorch/pull/97825,draft/noise,23,PR is draft; review decision requests changes; PR has not updated for at least 45 days,wait for contributor update,CHANGES_REQUESTED,CLEAN,True,0,0,,,,,,,,,0,23,0,,,"module: sparse, module: autograd, open source, release notes: autograd, release notes: sparse, skip-pr-sanity-checks, no-stale",2023-03-28T21:30:42Z,2025-06-19T08:46:33Z,384,"As in the title. Previously, the input functions to gradcheck were expected to return strided tensors: say, when the result of an input function would be a sparse tensor, it had to be converted to strided via .to_dense(masked=...) method call. With this PR, the conversion of input function return values to strided t..."
102613,Support building pytorch using MKL ILP64 model.,https://github.com/pytorch/pytorch/pull/102613,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,CLEAN,False,0,0,,,,,"51720, 100158",,,,0,3,0,,,"module: sparse, module: cpu, module: mkl, open source, release notes: build, release notes: sparse, no-stale",2023-05-31T12:51:32Z,2025-06-19T08:43:08Z,384,"This PR enables building pytorch against MKL libraries using the ILP64 model (recall, the currently supported MKL model is LP64). With the MKL ILP64 model, the 64-bit integer indices of sparse compressed tensors are passed directly to MKL routines which means that the maximum index value can be 2**64-1 instead of th..."
147927,Custom ops support arbitrary input types by migrating to python dispatcher,https://github.com/pytorch/pytorch/pull/147927,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,11,0,,,"triaged, open source, topic: not user facing, module: dynamo, ciflow/inductor, no-stale",2025-02-26T05:37:58Z,2025-05-30T15:02:48Z,404,"Test case: @torch.library.custom_op(""mylib::foo"", mutates_args=()) def foo(d: dict, t: torch.Tensor) -> torch.Tensor: return torch.sin(d[""x""] - d[""y""] + t) @foo.register_fake def _(d: dict, t: torch.Tensor) -> torch.Tensor: return torch.empty_like(d[""x""]) d = {""x"": torch.randn(2, 3, requires_grad=True), ""y"": torch.r..."
112441,[RFC] Tensordict integration,https://github.com/pytorch/pytorch/pull/112441,PR blocked,55,failed checks or blocked/unstable merge state; PR has not updated for at least 45 days,"identify whether block is CI, merge conflict, or review gate",,UNKNOWN,False,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,24,0,,,"triaged, open source, topic: new features, no-stale",2023-10-30T20:56:31Z,2025-05-28T21:34:32Z,406,[RFC] TensorDict This PR is under active development and should not be thoroughly reviewed. Suggestions are however welcome. This PR proposes to integrate the core features of tensordict in torch core under torch.dict. Purpose TensorDict is a data carrier for PyTorch. The problem it is trying to solve is to provide...
140372,[Intel GPU] Enable mkldnn::_convolution.pointwise at XPU backend,https://github.com/pytorch/pytorch/pull/140372,draft/noise,23,PR is draft; review decision requests changes; PR has not updated for at least 45 days; CI lane labels: ciflow/xpu,wait for contributor update,CHANGES_REQUESTED,UNSTABLE,True,1,1,statusCheckRollup,statusCheckRollup,,,,,,,0,11,0,,,"module: cpu, open source, topic: not user facing, no-stale, ciflow/xpu, module: xpu",2024-11-12T07:53:26Z,2025-05-24T06:21:32Z,410,"Moativation This PR is used to add pointwise fusion support for convolution. The conv-pointwise fusion is expected to be used in graph mode like torch.compile. The fusion utils is alread contained in #140365 . With conv-pointwise and linear-pointwise fusion, we may be able to inference with post-op optimization lost..."
147653,Bf16 fused adam(W),https://github.com/pytorch/pytorch/pull/147653,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,9,0,,,"release notes: foreach_frontend, no-stale",2025-02-22T00:10:24Z,2025-05-14T17:02:21Z,420,"Many things do work! Some things do not: amsgrad does not work some checks are removed (not critical imo, but def less safe) to let multidevice work adam + adamw should work tensor lr would work if it's on cpu the float, float, bf16, bf16 pattern is specialized, other mixed precision will not work Stack from ghstack..."
136835,Add back DistributedDataParallel types that were lost when pyi was removed,https://github.com/pytorch/pytorch/pull/136835,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,10,0,,,"oncall: distributed, triaged, open source, Merged, Reverted, topic: not user facing, no-stale, ci-no-td",2024-09-27T04:10:09Z,2025-05-06T09:21:08Z,428,"When the stub file nn/parallel/distributed.pyi was removed (#88701), some types that existed are no longer available. This pull request adds them back. Just for reference, these types are used in pytorch-lightning's LightningCLI. Command line interfaces are created automatically, and having type hints make them nice..."
142114,[DO NOT MERGE][WIP] CI: Dispatch PR events to the out-of-tree test infra,https://github.com/pytorch/pytorch/pull/142114,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"open source, Stale, topic: not user facing, no-stale",2024-12-05T06:29:37Z,2025-04-16T07:00:50Z,448,"The Goal: Without blocking PyTorch CI/CD, let the out-of-tree accelerator perceive the PR events in PyTorch, so that the out-of-tree backend has the opportunity to find possible problems as early as possible and leave relevant comments on the PR that raised the problem, making the PR more reasonable; I would like to..."
146101,(WIP) Update NJT ops to check data for raggedness check,https://github.com/pytorch/pytorch/pull/146101,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,CLEAN,False,0,0,,,,,,"141842, 145922, 146172",,,0,7,0,,,"release notes: nested tensor, module: dynamo, ciflow/inductor, no-stale",2025-01-30T23:22:38Z,2025-04-15T16:01:07Z,449,"Stack from ghstack (oldest at bottom): #146172 -> #146101 #145922 #141842 #141841 #146052 Some Issues: The way we use ephemeral sources doesn't work well with this case #145957 (comment) so the backward function needs to somehow stash the intermediates somewhere There are some duplicate runtime asserts created, e.g...."
142295,[inductor] Decide cooperative RSPLIT with same algorithm as split reductions,https://github.com/pytorch/pytorch/pull/142295,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,143812,,,0,20,0,,,"Stale, topic: not user facing, module: inductor, ciflow/inductor, no-stale",2024-12-07T05:56:40Z,2025-04-10T16:01:15Z,454,Stack from ghstack (oldest at bottom): #143812 -> #142295 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov
143812,[inductor] Used fixed configs for contiguous reductions,https://github.com/pytorch/pytorch/pull/143812,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,142295,,,0,11,0,,,"Stale, module: inductor, module: dynamo, ciflow/inductor, no-stale, release notes: inductor",2024-12-25T00:36:33Z,2025-04-10T15:54:03Z,454,Stack from ghstack (oldest at bottom): -> #143812 #142295 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov
146593,[NOT FOR LANDING] experimental NVSHMEM integration,https://github.com/pytorch/pytorch/pull/146593,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,4,0,,,"oncall: distributed, open source, release notes: distributed (c10d), no-stale",2025-02-06T10:26:18Z,2025-04-09T18:59:30Z,455,Stack from ghstack (oldest at bottom): -> #146593 #146592 cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o
146172,Factory function support for NestedTensor,https://github.com/pytorch/pytorch/pull/146172,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,"141842, 145922, 146101",,,0,6,0,,,"release notes: nested tensor, module: dynamo, ciflow/inductor, no-stale",2025-01-31T17:03:29Z,2025-04-08T21:21:14Z,456,Stack from ghstack (oldest at bottom): -> #146172 #146101 #145922 #141842 #141841 #146052 Rebase of #117904 removing unnecessary bits now that python nested int already holds the necessary metadata. cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx...
145922,Update NestedInt equality to take into account all metadata,https://github.com/pytorch/pytorch/pull/145922,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,CLEAN,False,0,0,,,,,,"141842, 146101, 146172",,,0,17,0,,,"module: cpu, release notes: fx, fx, module: dynamo, ciflow/inductor, no-stale",2025-01-29T08:50:27Z,2025-04-08T21:20:16Z,456,Stack from ghstack (oldest at bottom): #146172 #146101 -> #145922 #141842 #141841 #146052 cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @ezyang @SherlockNoMad @EikanWang @wenzhe-nrv @voznesenskym @penguinwu @Guobing-Chen @zhuhaozhe @blzheng @jiayisunx @chenyang78 @kadeng @chauhang @amjames
146318,Hack AC to not clear recomputed activations,https://github.com/pytorch/pytorch/pull/146318,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,CLEAN,False,0,0,,,,,,,,,0,1,0,,,no-stale,2025-02-03T17:29:36Z,2025-04-07T21:46:21Z,457,Stack from ghstack (oldest at bottom): #146633 -> #146318 #145399 #145533 #145531 #145520
138626,Fix global namespace pollution in ATen/Dispatch.h,https://github.com/pytorch/pytorch/pull/138626,ready for maintainer decision,62,"approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,2,0,,,"fb-exported, Merged, Reverted, ciflow/trunk, topic: not user facing, ciflow/inductor, no-stale, ci-no-td",2024-10-22T18:53:23Z,2025-04-05T08:00:58Z,459,Summary: Was it a typo? Since we already have at::detail::record_kernel_function_dtype() in ATen/Dispatch.h Test Plan: just build Differential Revision: D64642080
141912,Add stdlib includes to c10d debug handlers,https://github.com/pytorch/pytorch/pull/141912,ready for maintainer decision,72,approved with no detected failing/pending checks; PR closes or references open issue numbers; PR has not updated for at least 45 days,final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,140497,140497,140497,,,,0,2,0,,,"oncall: distributed, open source, release notes: distributed (c10d), no-stale",2024-12-02T21:31:11Z,2025-03-10T19:21:46Z,485,Fixes #140497 cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o
138519,[Inductor] introduce comm buffer planning,https://github.com/pytorch/pytorch/pull/138519,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,CLEAN,False,0,0,,,,,,,,,0,5,0,,,"oncall: distributed, open source, topic: not user facing, module: inductor, ciflow/inductor, no-stale",2024-10-21T23:04:20Z,2025-03-06T02:03:37Z,490,"Stack from ghstack (oldest at bottom): -> #138519 #138029 #138028 NOTE [comm buffer planning] This file contains the memory planning logic for comm buffers. Compared to regular buffer planning, Inductor leverages the allocator's ""persistent allocation"" capability to meet the stringent requirements for registered buf..."
135631,[scan] Autograd,https://github.com/pytorch/pytorch/pull/135631,PR blocked,100,failed checks or blocked/unstable merge state; approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,DIRTY,False,0,0,,,,,,,,,0,21,0,,,"open source, Stale, ciflow/trunk, module: inductor, module: dynamo, no-stale",2024-09-10T23:07:00Z,2025-02-03T00:37:41Z,521,This is the new PR to realize autograd for scan. This is a derivative of bohnstingl#2 @ydwu4 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov @rec
138436,"Replace use of PyTorch 2.0 with torch.compile, and minor edits",https://github.com/pytorch/pytorch/pull/138436,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,2,0,,,"ciflow/trunk, no-stale",2024-10-21T02:12:05Z,2025-01-30T18:36:45Z,524,"Stack from ghstack (oldest at bottom): -> #138436 I can perhaps deal with people using PT2 in our public facing documentation, but PyTorch 2.0 is really bad now that we are many minor releases beyond 2.0. Use torch.compile consistently in these docs. Signed-off-by: Edward Z. Yang ezyang@meta.com"
138471,Add torch.compiler.config.cache,https://github.com/pytorch/pytorch/pull/138471,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,2,0,,,"Stale, no-stale",2024-10-21T15:23:41Z,2025-01-30T15:10:38Z,524,Stack from ghstack (oldest at bottom): -> #138471 Signed-off-by: Edward Z. Yang ezyang@meta.com
140321,[scan] Refactored testcases,https://github.com/pytorch/pytorch/pull/140321,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,1,0,,,"triaged, open source, Stale, topic: not user facing, no-stale",2024-11-11T22:12:52Z,2025-01-18T18:23:50Z,536,This PR refactors the controlflow testcases for scan and fixes minor issues with associative_scan testcases cc @ydwu4
139524,Make anomaly mode asynchronous,https://github.com/pytorch/pytorch/pull/139524,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,2,0,,,"release notes: autograd, no-stale",2024-11-01T22:22:06Z,2025-01-06T02:46:38Z,549,"The original anomaly mode nan detection would sync at .item which made debugging streams not viable (anomaly mode couldn't detect nans as the additional syncs canceled the nastiness that was occurring). Opening this PR as this code helped me debug my streams issue, whereas the original anomaly mode did not. We shoul..."
125270,[do not review] Add tensor equality subsystem and get it working with compile,https://github.com/pytorch/pytorch/pull/125270,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"125326, 125428",,,0,6,0,,,"triaged, release notes: fx, oncall: pt2, module: dynamo, ciflow/inductor, no-stale",2024-04-30T22:06:55Z,2024-12-17T22:23:20Z,568,Stack from ghstack (oldest at bottom): #137417 #134907 #125428 #125326 -> #125270 cc @ezyang @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @amjames @msaroufim @bdhirsh @anijain2305
132814,Update TwoTensor to no longer inherit requires_grad from inner tensor,https://github.com/pytorch/pytorch/pull/132814,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,132414,,,0,11,0,,,"oncall: distributed, release notes: distributed (fsdp), module: dynamo, no-stale",2024-08-06T22:24:06Z,2024-10-15T13:50:45Z,631,Stack from ghstack (oldest at bottom): -> #132814 #132414 #133196 #133145 cc @XilunWu @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @a...
125428,[do not review] support max_seqlen,https://github.com/pytorch/pytorch/pull/125428,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"125270, 125326",,,0,6,0,,,"release notes: fx, module: dynamo, ciflow/inductor, no-stale",2024-05-02T21:01:41Z,2024-10-07T03:35:08Z,640,Stack from ghstack (oldest at bottom): #137417 #134907 -> #125428 #125326 #125270 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames
125326,[do not review] Use union find and new Python NestedInt with NJT,https://github.com/pytorch/pytorch/pull/125326,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,wait for author to mark ready,,CLEAN,True,0,0,,,,,,"125270, 125428",,,0,6,0,,,"release notes: jit, module: dynamo, ciflow/inductor, no-stale",2024-05-01T19:27:11Z,2024-10-07T03:35:07Z,640,Stack from ghstack (oldest at bottom): #137417 #134907 #125428 -> #125326 #125270 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames
119496,Switch batch norm stack to consolidated ops,https://github.com/pytorch/pytorch/pull/119496,stale/low urgency,17,"PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/mps",review when ready,,CLEAN,False,0,0,,,,,,,,,0,42,0,,,"module: cpu, release notes: quantization, ciflow/mps, ciflow/inductor, no-stale, keep-going",2024-02-08T22:58:39Z,2024-09-24T23:02:33Z,652,"Stack from ghstack (oldest at bottom): -> #119496 Summary: This commit switches aten.batch_norm to call the new batch_norm_with_update and batch_norm_no_update ops, instead of the old _batch_norm_impl_index op. The new stack is ""consolidated"" in the sense that there is a single backend agnostic op that will internal..."
130462,[testing] Allow creating offsets within AC region for NJT without SAC,https://github.com/pytorch/pytorch/pull/130462,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,CLEAN,False,0,0,,,,,,,,,0,3,0,,,no-stale,2024-07-10T17:45:47Z,2024-09-23T19:55:20Z,653,Stack from ghstack (oldest at bottom): -> #130462 #130440 Credit to @Chillee for this idea
117905,Factory function and basic .sizes() support for C++ NestedTensor,https://github.com/pytorch/pytorch/pull/117905,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,CLEAN,False,0,0,,,,,,119977,,,0,20,0,,,no-stale,2024-01-20T00:56:43Z,2024-08-26T17:00:50Z,681,Stack from ghstack (oldest at bottom): -> #117905 #119977 #119976 #121481 This PR: updates backend select codegen to also check for keys in singleton ints updates C++ NestedTensor to return sizes using singleton ints Adds support for C++ NestedTensor factory function to create NestedTensor using metadata on singleto...
124624,[NJT] Allow construction of NJT within graph using offsets from inputs,https://github.com/pytorch/pytorch/pull/124624,draft/noise,15,PR is draft; PR closes or references open issue numbers; PR has not updated for at least 45 days,wait for author to mark ready,,CLEAN,True,0,0,,,,118446,,,,,0,4,0,,,"module: dynamo, no-stale",2024-04-22T16:19:01Z,2024-08-12T20:07:16Z,695,"Stack from ghstack (oldest at bottom): -> #124624 #124634 #124343 Creating symbolic nested ints within the graph is difficult. Using unbacked symints should solve the most important(?) cases in the mean time. See #118446 Known gaps: creating NJT from intermediate offsets (offsets created within the graph, as opposed..."
121445,[NJT] Actually inline NT torch function during dynamo,https://github.com/pytorch/pytorch/pull/121445,ready for maintainer decision,72,"approved with no detected failing/pending checks; PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,120654,,,,,0,16,0,,,"Stale, ciflow/trunk, release notes: nested tensor, module: dynamo, ciflow/inductor, no-stale",2024-03-07T21:56:45Z,2024-08-06T22:35:39Z,701,Stack from ghstack (oldest at bottom): -> #121445 #122797 In order to avoid having any temporary state where the behavior of anything is regressed. This PR does all of the following at once: (1) Disables torch function running a second time in AOTAutograd If you have a tensor subclass that relies on dispatch into th...
124490,[dynamo] Support ndarray.dtype attribute access,https://github.com/pytorch/pytorch/pull/124490,ready for maintainer decision,62,"approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor, ciflow/trunk",final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,22,0,,,"open source, Merged, Reverted, ciflow/trunk, module: dynamo, ciflow/inductor, release notes: dynamo, no-stale",2024-04-19T15:33:08Z,2024-08-06T21:20:16Z,701,Stack from ghstack (oldest at bottom): -> #124490 #125717 cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang
126199,[POC] use a better default for num_threads,https://github.com/pytorch/pytorch/pull/126199,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,CLEAN,True,0,0,,,,,119546,,,,0,1,0,,,no-stale,2024-05-14T19:16:58Z,2024-07-15T05:46:38Z,723,Stack from ghstack (oldest at bottom): -> #126199
120076,Change ATEN generator argument type to const std::optional<Generator>&,https://github.com/pytorch/pytorch/pull/120076,ready for maintainer decision,62,"approved with no detected failing/pending checks; PR has not updated for at least 45 days; CI lane labels: ciflow/binaries, ciflow/inductor, ciflow/mps, ciflow/trunk",final maintainer merge/release decision,APPROVED,UNKNOWN,False,0,0,,,,,,,,,0,4,0,,,"module: cpu, triaged, open source, Merged, Reverted, ciflow/binaries, ciflow/trunk, release notes: mps, ciflow/mps, ciflow/inductor, no-stale",2024-02-16T12:53:21Z,2024-05-25T23:18:28Z,774,This PR proposes to use std::optional& for underlying functions to avoid unnecessary copy and move operations. The torchgen code was changed to generate the new type. cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10
110155,[TESTING] Capture scalar/dynamic by default,https://github.com/pytorch/pytorch/pull/110155,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,CLEAN,False,0,0,,,,,,,,,0,5,0,,,"release notes: fx, module: inductor, module: dynamo, ciflow/inductor, no-stale",2023-09-27T13:13:12Z,2024-04-07T06:26:26Z,822,Stack from ghstack (oldest at bottom): -> #110155 #113921 Signed-off-by: Edward Z. Yang ezyang@meta.com cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @aakhundov @ColinPeppler @Xia-Weiwen
88106,WIP: feat: LARS optimizer,https://github.com/pytorch/pytorch/pull/88106,draft/noise,5,PR is draft; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,7,0,,,"module: optimizer, module: mkldnn, open source, release notes: nn, ciflow/inductor, no-stale",2022-10-31T18:13:38Z,2024-01-30T19:47:12Z,890,"Followup to #6323. Addition of LARS optimizer. LARS optimizer Tests Documentation Multi-Tensor support Extra params (e.g., maximize, differentiable, foreach) .pyi Reference implementations: [1] cc @vincentqb @jbschlosser @albanD @janeyx99 @crcrpar @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfeima @..."
108303,[optim] Make casting to match params a hook (2nd try),https://github.com/pytorch/pytorch/pull/108303,stale/low urgency,17,PR has not updated for at least 45 days; CI lane labels: ciflow/trunk,review when ready,,CLEAN,False,0,0,,,,,,,,,0,2,0,,,"ciflow/trunk, no-stale, release notes: optimizer",2023-08-30T23:54:47Z,2024-01-12T19:19:46Z,908,"Reland of #106725, but instead of registering the hook in the constructor, do it lazily through @Property Stack from ghstack (oldest at bottom): -> #108303"
111673,[quant][bc-breaking] Remove deprecated QConfigDynamic,https://github.com/pytorch/pytorch/pull/111673,ready for maintainer decision,62,approved with no detected failing/pending checks; PR has not updated for at least 45 days,final maintainer merge/release decision,APPROVED,CLEAN,False,0,0,,,,,,,,,0,3,0,,,"release notes: quantization, no-stale",2023-10-20T17:48:02Z,2024-01-03T07:00:57Z,917,Stack from ghstack (oldest at bottom): -> #111673 Summary: QConfigDynamic was deprecated in PyTorch 1.12. It has continued to cause confusion to users who wish to use dynamic quantization. This commit removes this deprecated API and requires users to use QConfig instead. BC-breaking before: qconfig = QConfigDynamic(...
84843,Support different NSE in batches of CSR and CSC tensors,https://github.com/pytorch/pytorch/pull/84843,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,CLEAN,True,0,0,,,,,,,,,0,5,0,,,"module: sparse, open source, cla signed, release notes: sparse, no-stale",2022-09-11T22:04:28Z,2023-12-12T17:41:33Z,939,"This PR enables batched CSR/CSC tensors that batches may have different NSE counts. For instance, with the current master we have >>> a = torch.tensor([[[1, 2], [3, 4]], [[0, 12], [21, 0]]]) >>> a.to_sparse_csr() Traceback (most recent call last): File ""<stdin>"", line 1, in <module> RuntimeError: Expect the same num..."
88998,Fix fake tensor propagation for nvprims,https://github.com/pytorch/pytorch/pull/88998,has linked issue,27,PR closes or references open issue numbers; PR has not updated for at least 45 days; CI lane labels: ciflow/inductor,review when ready,,UNKNOWN,False,0,0,,,87236,87236,,,,,0,13,0,,,"triaged, open source, module: nvfuser, ciflow/inductor, no-stale",2022-11-14T15:26:43Z,2023-04-17T00:36:00Z,1179,"Fake tensors were not working with torch.ops.nvprims functions because the fake tensors code expects to see the ""prim_meta_impl"" attribute that was missing. Also fixes #87236 cc @kevinstephano @jjsjann123"
85699,Add vector-Jacobian-products for a subset of nvFuser-supported prims; add backward support for nvprims,https://github.com/pytorch/pytorch/pull/85699,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,,UNKNOWN,True,0,0,,,,,,,,,0,49,0,,,"triaged, open source, cla signed, module: nvfuser, module: primTorch, no-stale",2022-09-27T11:37:51Z,2023-04-17T00:36:00Z,1179,"NOTE: Currently, this PR is blocked until #85696 is resolved. vjp_implementations[""prim_name""] is a callable that computes vector-Jacobian products for given cotangents, forward results, and forward inputs. And this function is used in the torch.autograd.Function.backward method. Here's an example trace for the torc..."
84545,Add nvfuser support for prims.copy_to,https://github.com/pytorch/pytorch/pull/84545,draft/noise,5,PR is draft; PR has not updated for at least 45 days,wait for author to mark ready,APPROVED,UNKNOWN,True,0,0,,,,,,,,,0,30,0,,,"oncall: jit, triaged, open source, cla signed, release notes: jit, module: nvfuser, module: primTorch, no-stale",2022-09-05T16:09:43Z,2023-04-17T00:36:00Z,1179,"I use nvFuser's aliasOutputToInput here and since it implicitly adds outputs to the fusion, I need to drop those within Python. Now we can lower the batch_norm implementation from torch._decomp to nvprims(see test_batch_norm_forward_nvprims). cc @EikanWang @jgong5 @wenzhe-nrv @sanchitintel @kevinstephano @jjsjann123..."
82997,Implement refs.var as a real reference,https://github.com/pytorch/pytorch/pull/82997,stale/low urgency,17,PR has not updated for at least 45 days,review when ready,,UNKNOWN,False,0,0,,,,,,,,,0,5,0,,,"triaged, open source, cla signed, module: primTorch, no-stale",2022-08-08T17:00:45Z,2023-04-17T00:35:59Z,1179,Description This PR removes the use of prims.var in the implementation of the var reference because there's no need for var to be a primitive. Testing No new tests are needed. cc @ezyang @mruberry @ngimel @lezcano @fdrocha @peterbell10