id int64 2.74B 3.05B | title stringlengths 1 255 | user stringlengths 2 26 | state stringclasses 2
values | labels listlengths 0 24 | comments int64 0 206 | author_association stringclasses 4
values | body stringlengths 7 62.5k ⌀ | is_title bool 1
class |
|---|---|---|---|---|---|---|---|---|
2,885,850,636 | Checks kv pair indexing in OrderedPreservingDictTest.test_range_insert | redwrasse | open | [
"triaged",
"open source",
"ciflow/trunk",
"topic: not user facing",
"module: inductor"
] | 12 | CONTRIBUTOR | `OrderedPreservingDictTest.test_range_insert` has an [unused loop variable `j`](https://github.com/pytorch/pytorch/blob/main/c10/test/util/ordered_preserving_dict_test.cpp#L186), I think taken from the [inspired project](https://github.com/pytorch/pytorch/blob/main/c10/test/util/ordered_preserving_dict_test.cpp#L165) t... | true |
2,885,836,943 | Remove manylinux 2014 artifacts | atalman | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 6 | CONTRIBUTOR | 1. Switch Magma build to Manylinux 2.28 base
2. Use manylinux 2.28 as default in populate_binary_env.sh
3. Remove manylinux 2014 docker builds | true |
2,885,831,144 | add skips to test_notifies_oom and test_set_per_process_memory_fraction | Fuzzkatt | closed | [
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 3 | COLLABORATOR | Tests fail in NVIDIA internal CI since we do not support nvml on Jetson, but nvml is required for OOM reporting to work properly, so we are skipping the failing tests for now.
cc @nWEIdia @eqy | true |
2,885,790,600 | [MPS] fix empty place holder error for smooth l1 loss | Isalia20 | closed | [
"open source",
"Merged",
"topic: bug fixes",
"module: mps",
"release notes: mps"
] | 6 | COLLABORATOR | Fixes #123171
And parametrizes the tests for it
cc @kulinseth @albanD @malfet @DenisVieriu97 @jhavukainen | true |
2,885,784,252 | WIP enable aten convolution out in lowerings | exclamaforte | closed | [
"module: inductor",
"ciflow/inductor"
] | 2 | CONTRIBUTOR | So that the convolution op can plan its output memory for fusion opportunities.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov | true |
2,885,781,910 | [Inductor] Use real input to autotune user defined triton kernels | muchulee8 | closed | [
"ciflow/trunk",
"module: inductor",
"ciflow/inductor"
] | 4 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148131
Summary:
User defined Triton kernel sometimes rely on real inputs to determine
the path of execution. We need real inputs to invoke the correct
behavior of the user defined triton kernels (see example in test case,
where we ha... | true |
2,885,775,164 | Use correct boxed_forward_device_index when running `CompiledFxGraph.post_compile` | jamesjwu | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor",
"release notes: AO frontend"
] | 20 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148130
This PR threads through the correct boxed_forward_device_index from graph_kwargs to CompiledFXGraph.post_compile. This allows us to correctly update BoxedDeviceIndex from cache hits.
We don't actually need to save `boxed_f... | true |
2,885,774,163 | ci: Remove manylinux builds for triton, except for XPU | seemethere | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 6 | MEMBER | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148129
We're dropping regular old manylinux so let's drop it here too
Relates to #123649
Signed-off-by: Eli Uriegas <eliuriegas@meta.com> | true |
2,885,772,185 | [MPS] Add inductor support for the `entr()` operator. | dcci | closed | [
"Merged",
"topic: not user facing",
"module: mps",
"ciflow/mps",
"module: inductor",
"ciflow/inductor"
] | 3 | MEMBER | cc @kulinseth @albanD @malfet @DenisVieriu97 @jhavukainen @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov | true |
2,885,767,445 | Long queue for macOS runners | huydhn | closed | [
"ci: sev"
] | 1 | CONTRIBUTOR | ## Current Status
Ongoing https://hud.pytorch.org/metrics. The queue is at 6 hours now.
## Error looks like
Jobs in queue
## Incident timeline (all times pacific)
00:00 Feb 27th
## User impact
Waiting for runner on PRs
## Root cause
MacOS runners were terminated probably after this workflow ran last night https://... | true |
2,885,740,879 | ci: Only run CI specific things when in CI | seemethere | closed | [
"Merged",
"topic: not user facing"
] | 5 | MEMBER | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #148129
* __->__ #148126
This was blocking me from running this locally so don't run it like this
Signed-off-by: Eli Uriegas <eliuriegas@meta.com> | true |
2,885,740,564 | [dtensor][fix] fix _scaled_dot_product_flash_attention sharding | XilunWu | closed | [
"oncall: distributed",
"better-engineering",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/inductor",
"module: dtensor",
"module: context parallel"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148125
### Summary
https://github.com/pytorch/pytorch/pull/146372/ changed the op signature of `_scaled_dot_product_flash_attention` and as a consequence DTensor needs to change its sharding defined at https://github.com/pytorch/... | true |
2,885,733,682 | Add a stable TORCH_LIBRARY to C shim | janeyx99 | closed | [
"Merged",
"Reverted",
"ciflow/trunk",
"release notes: cpp",
"ciflow/inductor",
"ci-no-td",
"no-runner-experiments"
] | 26 | CONTRIBUTOR | This PR adds two main parts:
- shim.h stable C APIs into torch::Library APIs
- a higher level API in torch/csrc/stable/library.h that calls into this shim.h + otherwise is self contained
Goal: custom kernel writers should be able to call the apis in the directories above in order to register their library in a way... | true |
2,885,714,105 | Conv/pool doc on ceilmode wrong | albertz | closed | [
"triaged",
"topic: docs"
] | 1 | CONTRIBUTOR | ### 📚 The doc issue
The doc on maxpool/conv states this output shape:
$$L_{out} = \left\lfloor \frac{L_{in} + 2 \times \text{padding} - \text{dilation} \times (\text{kernelsize} - 1) - 1}{\text{stride}} + 1\right\rfloor$$
Then it says for the `ceil_mode` arg: "If True, will use ceil instead of floor to compute the o... | true |
2,885,705,066 | [cutlass backend] C++ compile error for CUTLASS config only get resolved in autotuning stage | henrylhtsang | closed | [
"triaged",
"oncall: pt2",
"module: inductor"
] | 1 | CONTRIBUTOR | ### 🐛 Describe the bug
I think we should try to filter the bad configs out and/or filter them. But inductor can help as well.
internal post:
https://fb.workplace.com/groups/1037051091797916/posts/1059321392904219/
repro:
```py
import logging
import os
os.environ["TORCH_LOGS"] = "+output_code,+benchmarking,+inducto... | true |
2,885,700,527 | [Inductor][Tests] Use generic device or require CUDA for newly added tests | alexbaden | closed | [
"open source",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor",
"merging",
"ciflow/xpu"
] | 4 | COLLABORATOR | Some tests added to Inductor as part of #147038 and #145583 were cuda specific and failing on XPU. This PR changes the attrs dict replacement test to use the generic device variable. The first graph partitioning test (`test_graph_partition`) fails on XPU - the graph does not appear to be properly partitioned. Subsequen... | true |
2,885,688,727 | Torch 2.7.0 nightly cuda 12.6 and cuda 12.8 builds are broken on Amazon linux 2023 | atalman | closed | [
"module: binaries",
"module: build",
"topic: build",
"topic: binaries"
] | 2 | CONTRIBUTOR | ### 🐛 Describe the bug
Failure can be seen here:
https://github.com/pytorch/test-infra/actions/runs/13558218752/job/37934127476
Error:
```
2025-02-27T16:00:09.7113764Z + python3 .ci/pytorch/smoke_test/smoke_test.py --package torchonly
2025-02-27T16:00:09.7114301Z Traceback (most recent call last):
2025-02-27T16:00:0... | true |
2,885,673,206 | Add int32 support to torch.gather | byphilipp | open | [
"triaged",
"module: advanced indexing",
"function request",
"module: scatter & gather ops"
] | 1 | NONE | ### 🚀 The feature, motivation and pitch
The int32 indexing added in torch 2.0
But torch.gather support only int64 indexing now
### Alternatives
_No response_
### Additional context
The int64 indexes is memory expensive, espetially if we use half-precision calculations
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin... | true |
2,885,663,123 | [inductor][ck] manual kBatch heuristic | coconutruben | closed | [
"module: rocm",
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 10 | CONTRIBUTOR | Summary:
# Why
Leverage kBatch parameter for large splitK examples for CK for better than ATEN performance
# What
replace default kBatch = 1 with a manual heuristic
- if K > 16 * max (M,N)
- leverage k_per_block, and K and number of SMs on the chip
- upper bound to 128, lower bound to 1
This is better than default... | true |
2,885,616,551 | [triton 3.3] Fix aoti cpp wrapper remaining 5 issue. (following #148051) | YUNQIUGUO | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 18 | CONTRIBUTOR | Summary:
Fix the following 5 on a100:
- test_foreach_cpp_wrapper_cuda_gpu_wrapper
- test_enable_dynamic_shapes_cpp_wrapper_cuda_gpu_wrapper
- test_dynamic_shapes_persistent_reduction_mixed_x_dim_cuda_gpu_wrapper
- test_enable_dynamic_shapes_cpp_wrapper_cuda_dynamic_shapes_gpu_wrapper
- test_dynamic_shap... | true |
2,885,612,110 | [inductor][triton] Explicit kernel-arg mismatch checks | davidberard98 | open | [
"triaged",
"oncall: pt2",
"module: inductor"
] | 2 | CONTRIBUTOR | ### 🚀 The feature, motivation and pitch
Let's add explicit checks to check when a kernel expects different args than the args it's getting. Particularly if we can do this at the cpp_wrapper level, it'll be quite helpful
From @chenyang78 :
> BTW, I feel like this kind of kernel-argument mismatch issue is a really p... | true |
2,885,610,379 | Redundant Try Block in backward() | Jason1ien | closed | [
"module: autograd",
"triaged",
"module: custom-operators"
] | 3 | NONE | Inside of autograd.py, apart of the library directory, there is a redundant try block.
This try block should be resolved by checking if info._backward_fn is none. In the case that it is none, we should raise an exception error instead.
Below is the link to the file:
https://github.com/pytorch/pytorch/blob/main/torch... | true |
2,885,596,932 | We should never throw vanilla C++ exceptions | albanD | open | [
"module: cpp",
"triaged",
"better-engineering",
"actionable"
] | 0 | COLLABORATOR | We have a lot of them at the moment in a few places in the codebase: https://github.com/search?q=repo%3Apytorch%2Fpytorch%20std%3A%3Aruntime_error(&type=code
We should migrate ALL of them to TORCH_CHECK-like macros and add a lint rule to enforce it going forward.
cc @jbschlosser @Skylion007 in case you would be inter... | true |
2,885,545,474 | [inductor][triton] introduce better "APIs" in triton that can clean up our triton/inductor integration | davidberard98 | open | [
"triaged",
"upstream triton"
] | 0 | CONTRIBUTOR | ### 🚀 The feature, motivation and pitch
Let's review all the issues in https://github.com/orgs/pytorch/projects/94/views/1, the stack from https://github.com/pytorch/pytorch/pull/145515, and other relevant pull requests we saw in the triton 3.3 / pytorch 2.7 release and see how we can clean up the inductor/triton int... | true |
2,885,545,242 | MLA with Learnable RoPE Tensors is Broken with Flex Attention | cora-codes | open | [
"triaged",
"oncall: pt2",
"module: higher order operators",
"module: pt2-dispatcher",
"module: flex attention"
] | 0 | NONE | ### 🐛 Describe the bug
I'm encountering an `AssertionError` during the backward pass when using `flex_attention` if the tensors used require gradients.
```python
import torch
from torch import Tensor
from torch.nn.attention.flex_attention import flex_attention
torch._inductor.config.unroll_reductions_threshold = 65... | true |
2,885,530,386 | [triton 3.3][cpp_wrapper] TypeError: 'NoneType' object is not subscriptable | davidberard98 | closed | [] | 1 | CONTRIBUTOR | ### 🐛 Describe the bug
Example stacktrace below
```
======================================================================
ERROR: test_foreach_cpp_wrapper_cuda_gpu_wrapper (__main__.TestGpuWrapper)
----------------------------------------------------------------------
Traceback (most recent call last):
File "/data... | true |
2,885,489,050 | [CI] Remove conda usage from lint related jobs | clee2000 | closed | [
"module: ci",
"triaged"
] | 1 | CONTRIBUTOR | Remove conda usage from the linter docker images pytorch-linux-focal-linter and pytorch-linux-jammy-cuda11.8-cudnn9-py3.9-linter
Relevant files:
.ci/docker/linter/Dockerfile
.ci/docker/linter-cuda/Dockerfile
.github/scripts/lintrunner.sh
.github/workflows/lint-autoformat.yml
.github/workflows/lint.yml
tools/linter/*
... | true |
2,885,488,710 | [triton 3.3] cpp wrapper aoti remaining test failure following #148051 | YUNQIUGUO | closed | [
"ciflow/trunk",
"module: inductor",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | This should hopefully fix the remaining cpp_wrapper aoti tests failure in this paste: [P1741802912](https://www.internalfb.com/phabricator/paste/view/P1741802912)
following https://github.com/pytorch/pytorch/pull/148051
Related issue: #147734
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @Xiaobin... | true |
2,885,481,751 | Use nightly-wheel-upload env for triton wheel publishing | atalman | closed | [
"Merged",
"topic: not user facing"
] | 3 | CONTRIBUTOR | Required for publishing triton builds | true |
2,885,413,548 | SubgraphLoweringException in flex_attention when using custom score_mod with torch.dot (MLA) | cora-codes | closed | [] | 1 | NONE | ### 🐛 Describe the bug
I'm getting a `SubgraphLoweringException` when using `torch.compile(flex_attention)` with a custom score modification function that uses `torch.dot()`. The error occurs during the inductor compilation phase.
```python
import torch
from torch import Tensor
from torch.nn.attention.flex_attention... | true |
2,885,359,114 | NotImplementedError: FlexAttentionHigherOrderVariable() has no type | cora-codes | open | [
"triaged",
"bug",
"oncall: pt2",
"module: dynamo",
"module: higher order operators",
"module: pt2-dispatcher",
"module: flex attention"
] | 5 | NONE | ### 🐛 Describe the bug
I'm working on a reproduction, but I've ran into the following error with flex attention: "NotImplementedError: FlexAttentionHigherOrderVariable() has no type". I believe it tries to access `block_mask.shape[-1]` and then fails .
### Versions
'2.7.0a0+gitce805a5'
cc @chauhang @penguinwu @voz... | true |
2,885,259,314 | Use myst_nb in docs | zou3519 | open | [
"Stale"
] | 3 | CONTRIBUTOR | myst_nb is a plugin that:
1) allows rendering of jupyter notebooks in documentation 2) allows execution of code blocks in markdown docs
Execution of code blocks in markdown docs may slow down build time. In order to limit the impact:
- by default, we do not execute code blocks when run locally. To run them locally... | true |
2,885,155,319 | Refactor layout constraint selection logic | zou3519 | closed | [
"Merged",
"Reverted",
"ciflow/trunk",
"release notes: fx",
"topic: not user facing",
"fx",
"module: inductor",
"module: dynamo",
"ciflow/inductor",
"keep-going",
"ci-no-td"
] | 18 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #150511
* __->__ #148104
This PR:
- cleans up some existing comments that don't make sense anymore
- hooks up the "custom_op_default_layout_constraint" back (that seems to
have broken)
- cleans up the "lazy registration path" which seems to... | true |
2,885,134,749 | Initial investigation for removing MOD_SKIPLIST | zou3519 | open | [
"triaged",
"oncall: pt2",
"module: dynamo"
] | 0 | CONTRIBUTOR | In the ideal state we want to clear out everything from MOD_SKIPLIST: https://github.com/pytorch/pytorch/blob/main/torch/_dynamo/trace_rules.py#L3333
Removing something from MOD_SKIPLIST is equivalent to adding it into LEGACY_MOD_INLINELIST.
However, removing things from these lead to various CI errors.
We should tr... | true |
2,885,122,553 | Fix None and equal_to_1 arguments issue in Triton kernel generated by AOTI | renganxu | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor"
] | 11 | CONTRIBUTOR | Summary:
When a Triton kernel has arguments with None values followed by arguments with value 1, AOTI attempts to remove the None arguments and update the indices of the equal_to_1 arguments in triton_meta["configs"]. However, if the same kernel is called multiple times, this optimization process is repeated. Prior to ... | true |
2,885,068,499 | Move expanded dim require_exact_stride handling to api from sdpa lowering | eellison | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 9 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148101
See issue: https://github.com/pytorch/pytorch/issues/147156#issue-2852362217.
Original tests from https://github.com/pytorch/pytorch/pull/146054 should cover these changes, and I tested that the perf on https://github.com... | true |
2,885,061,238 | [Experiment] meaure the effect of combining cpp-wrapper and cudagraphs | desertfire | closed | [
"Stale"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148100
| true |
2,885,030,440 | HSDP custom hook UTs are multi-threaded - can't set device rank | pragupta | closed | [
"oncall: distributed",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/periodic"
] | 22 | CONTRIBUTOR | HSDP custom hook UTs are multi-threaded and using single physical GPU. If we set rank in each thread, then we are referencing the same GPU with multiple ranks, which isn't right. Therefore, removing the rank setting from these UTs. Now, they are passing with 1, 2, 4 GPUs.
Fixes #147767 and #147769
cc @H-Huang ... | true |
2,885,025,598 | onnx dynamo export does not support aten bucketize | peterbjorgensen | open | [
"module: onnx",
"triaged"
] | 1 | NONE | ### 🐛 Describe the bug
aten::bucketize should be supported by ONNX, but the following code does not work:
```
import torch
import torch.nn as nn
class MyBucketizer(nn.Module):
def forward(self, x):
return torch.bucketize(x, torch.linspace(0, 1, 100))
def main():
x = torch.rand(100)
my_model... | true |
2,885,015,251 | [ROCm][Windows] Fix OpenMP Flags for clang-cl | m-gallus | closed | [
"module: rocm",
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/rocm"
] | 19 | CONTRIBUTOR | When clang-cl parses its command line arguments, it expects MSVC-style arguments (beggining with `/` such as `/WX`, `/MD`, etc.) to be provided, and clang-style arguments to be preceded by `-Xclang`, otherwise, the clang-style parameters are ignored as they are interpreted unrecognized compiler options.
cc @jeffdaily ... | true |
2,884,917,247 | [BE][Ez]: Remove extra copy in dtensor parallel loss | Skylion007 | closed | [
"oncall: distributed",
"triaged",
"open source",
"better-engineering",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/inductor"
] | 6 | COLLABORATOR | Remove an extra copy of the input to `_log_softmax` when there is a dtype and memory format change. Fuse the copies instead.
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o | true |
2,884,897,094 | DISABLED test_njt_causal_bfloat16 (__main__.TestFlexAttention) | pytorch-bot[bot] | closed | [
"module: rocm",
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: inductor"
] | 2 | NONE | Platforms: rocm
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_njt_causal_bfloat16&suite=TestFlexAttention&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37916878876).
Over the past 3 hours, it has bee... | true |
2,884,896,983 | DISABLED test_split_dynamic (__main__.AutoFunctionalizeTests) | pytorch-bot[bot] | closed | [
"triaged",
"module: flaky-tests",
"skipped",
"module: functionalization",
"oncall: pt2",
"module: pt2-dispatcher"
] | 6 | NONE | Platforms: asan, linux, mac, macos, rocm, slow
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_split_dynamic&suite=AutoFunctionalizeTests&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37918361488).
... | true |
2,884,896,981 | DISABLED test_dynamo_timed (__main__.TestDynamoTimed) | pytorch-bot[bot] | closed | [
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: dynamo"
] | 20 | NONE | Platforms: asan, linux, mac, macos, rocm, slow
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_dynamo_timed&suite=TestDynamoTimed&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37916682229).
Over the ... | true |
2,884,884,665 | Rename node.meta["arg_kwarg_vals"] to node.meta["eager_input_vals"] | zou3519 | closed | [
"Merged",
"ciflow/trunk",
"release notes: fx",
"fx",
"module: inductor",
"ciflow/inductor"
] | 4 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #150511
* #148104
* #150495
* __->__ #148092
* #148091
* #148063
* #148046
And added a comment about it. Otherwise it might be confusing
Test Plan:
- wait for CI
cc @ezyang @SherlockNoMad @EikanWang @jgong5 @wenzhe-nrv @voznesenskym @pengu... | true |
2,884,884,546 | Implement needs_exact_strides for mutable custom operators | zou3519 | closed | [
"Merged",
"release notes: fx",
"fx",
"module: inductor",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #150511
* #148104
* #150495
* #148092
* __->__ #148091
* #148063
* #148046
Mutable custom operators get wrapped into an auto_functionalized HOP, so
we need to store the arg_kwarg_vals on the auto_functionalized HOP
itself.
When Inductor doe... | true |
2,884,875,838 | only print GraphModule during fx.Interpreter errors if valid | bdhirsh | closed | [
"Merged",
"ciflow/trunk",
"release notes: fx",
"fx"
] | 6 | CONTRIBUTOR | Came up in https://www.internalfb.com/diff/D69057074?dst_version_fbid=970771615000938&transaction_fbid=1723357345264461 - we need to make sure the GraphModule is valid before calling `print_readable` on it
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #133044
* #147561
* __->__ #14... | true |
2,884,867,581 | Build a storage reader/writer to write checkpoints in HF format | ankitageorge | closed | [
"oncall: distributed",
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: new features"
] | 8 | CONTRIBUTOR | Summary: D69984656 caused issues by adding the fsspec dependency to torch distributed when many packages internally didn't have it. In this diff I'm not adding HFStorageReader/Writer to __init__.py so that HFStorage components don't get imported internally and in turn there is no fsspec import that happens. I did the r... | true |
2,884,824,321 | Fix minor typo in python_nccl | x41lakazam | closed | [
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 8 | CONTRIBUTOR | null | true |
2,884,815,054 | [inductor] ignore block ptr advancements for removed buffers | kundaMwiza | closed | [
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor"
] | 9 | CONTRIBUTOR | Follow up to https://github.com/pytorch/pytorch/pull/147193. Some buffers are removed only when the kernel context is exited so defer the lines instead.
Added `use_block_ptr` as a parameter to test case that fails if run with block ptrs enabled.
Fixes #ISSUE_NUMBER
cc @voznesenskym @penguinwu @EikanWang @jgon... | true |
2,884,735,181 | FSDP2 without sharding works slower than DDP | JanRocketMan | open | [
"oncall: distributed",
"triaged"
] | 0 | NONE | ### 🐛 Describe the bug
Hi, great thanks for introducing FSDP2, it's been a major quality-of-life improvement for me over the past months!
One caveat I noticed is that communication volume has noticeably increased even in cases where it should not, e.g. if I remove all sharding with custom device `mesh`.
Here is an ... | true |
2,884,721,166 | td does not detect required test for mkl-dnn OneDNN update | atalman | closed | [
"triaged",
"module: infra"
] | 1 | CONTRIBUTOR | ### 🐛 Describe the bug
OneDNN update to 3.7: https://github.com/pytorch/pytorch/pull/147955 not triggering
``test_mkldnn.py::TestMkldnnCPU::test_mul_cpu`` look like TD skips this test on this PR: https://github.com/pytorch/pytorch/actions/runs/13539463441/job/37838465868?pr=147955#step:15:13520
This test is failing ... | true |
2,884,668,256 | Refine test_preserves_strides to support different GPUs | EikanWang | closed | [
"topic: not user facing",
"module: inductor",
"ciflow/inductor",
"ciflow/xpu"
] | 3 | COLLABORATOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148084
cc @voznesenskym @penguinwu @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov | true |
2,884,284,127 | [dynamo] WeakRefVar reconstruct | IvanKobzarev | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148083
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,884,044,311 | Consistently use load_torchbind_test_lib in tests | Flamefire | closed | [
"triaged",
"module: mkldnn",
"open source",
"Merged",
"ciflow/trunk",
"release notes: export",
"ciflow/linux-aarch64"
] | 6 | COLLABORATOR | The same code is repeated multiple times with slightly different implementations.
Use the existing function for brevity and consistency.
In the function the code from `test_export` is used which does a single `load_library` with cleaner conditions
cc @gujinghui @PenghuiCheng @XiaobingSuper @jianyuh @jgong5 @mingfe... | true |
2,883,998,480 | Add XPU device to LayerNormKernel devices | min-jean-cho | closed | [
"oncall: distributed",
"module: cpu",
"open source",
"topic: not user facing",
"module: inductor",
"module: dynamo",
"ciflow/inductor",
"release notes: distributed (checkpoint)",
"ciflow/xpu",
"release notes: xpu",
"module: xpu"
] | 8 | COLLABORATOR | Work with https://github.com/intel/torch-xpu-ops/pull/1416 .
Moved to https://github.com/pytorch/pytorch/pull/148593.
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @voznesenskym @penguinwu @EikanWang @Guobing-... | true |
2,883,968,700 | DISABLED test_split (__main__.AutoFunctionalizeTests) | pytorch-bot[bot] | closed | [
"triaged",
"module: flaky-tests",
"skipped",
"module: functionalization",
"oncall: pt2",
"module: pt2-dispatcher"
] | 7 | NONE | Platforms: mac, macos, rocm, asan, linux, slow
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_split&suite=AutoFunctionalizeTests&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37902172609).
Over th... | true |
2,883,929,581 | Fix test_tensorboard when started w/o tensorboard package | Flamefire | open | [
"triaged",
"open source",
"Stale",
"topic: not user facing"
] | 3 | COLLABORATOR | If `TEST_TENSORBOARD == False` then `DataType` is not defined or imported. However it is used unconditionally when defining the test with `parametrize` which leads to an NameError crashing the test execution on start.
Provide a Dummy to make it syntactially correct. Tests will be skipped on start.
```
File "... | true |
2,883,887,624 | Video-Llama (version 1) runs much slower using Float16 than Float32 on Kunpeng CPU | luentong | open | [
"module: cpu",
"triaged",
"module: arm",
"topic: performance"
] | 2 | NONE | This is a kinda niche question, so I just want some one who've worked on this area some suggestions.
I've tested thoroughly that, Video-Llama runs much slower when using Float16 than Float32 on Kunpeng Arm Arch64 CPU (at least x200 slower). On the other hand, other LLMs such as Llama3.2, Qwen2.5, even Video-Llama 2 or... | true |
2,883,864,407 | Feature Request: CUDA-Optimized Queue Buffer for PyTorch | songyuc | open | [
"module: cuda",
"triaged",
"needs research",
"needs design"
] | 1 | NONE | ### 🚀 The feature, motivation and pitch
## Problem Statement
When implementing reinforcement learning algorithms like SAC, we need efficient experience replay buffers to store and sample transitions (state, action, reward, next_state). Currently, many implementations resort to custom circular arrays instead of using ... | true |
2,883,854,989 | [test][do not merge] upgrade onednn v3.7, no ideep change | yanbing-j | closed | [
"module: cpu",
"module: mkldnn",
"open source",
"module: arm",
"ciflow/trunk",
"topic: not user facing",
"intel",
"ciflow/xpu",
"ciflow/linux-aarch64"
] | 1 | COLLABORATOR | Fixes #ISSUE_NUMBER
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @gujinghui @PenghuiCheng @jianyuh @min-jean-cho @Guobing-Chen @Xia-Weiwen @snadampal @malfet @milpuz01 | true |
2,883,833,756 | [Dynamo] Fix `AssertionError` when dynamo traces `torch.functional.xxx()` functions | shink | closed | [
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo"
] | 8 | CONTRIBUTOR | Fixes #147840
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,883,791,582 | [inductor][cpu]DebertaV2ForMaskedLM, DebertaV2ForQuestionAnswering and eca_halonext26ts max_autotune accuracy failure in 2025-02-24 nightly release | zxd1997066 | closed | [
"oncall: pt2",
"oncall: cpu inductor"
] | 4 | CONTRIBUTOR | ### 🐛 Describe the bug
<p>fp32 max_autotune static shape default wrapper</p><table border="1" class="dataframe table">
<thead>
<tr style="text-align: right;">
<th>suite</th>
<th>name</th>
<th>thread</th>
<th>accuracy</th>
<th>perf</th>
<th>reason(reference only)</th>
</tr... | true |
2,883,723,954 | torch.compile with the inductor backend slows down (exponentially?) for certain graphs | linkct | open | [
"triaged",
"oncall: pt2",
"module: inductor"
] | 3 | NONE | ### 🐛 Describe the bug
I recently ran into some issues where the compilation time of my model went from ~1min to ~20min. After some profiling I pinpointed the same culprit as this https://github.com/pytorch/pytorch/pull/145082#issue-2795829525, which solved the initial problem. However, the issue appeared again after... | true |
2,883,713,987 | torch.compile() on quantized model: No attribute "meta" | Whadup | closed | [
"needs reproduction",
"triaged",
"oncall: pt2",
"oncall: cpu inductor"
] | 3 | CONTRIBUTOR | During evaluation of a compiled and quanzited model, I obtain the error "no attribute "meta"" in the following line:
https://github.com/pytorch/pytorch/blob/fd43c36aa9407634ad9f063eafc5b2795c091022/torch/_inductor/fx_passes/quantization.py#L1074
I propose the following change:
```diff
- x = match.kwargs["x"]... | true |
2,883,640,830 | [inductor][cpu] maml performance regression in 2025-02-24 nightly release | zxd1997066 | open | [
"oncall: pt2",
"oncall: cpu inductor"
] | 1 | CONTRIBUTOR | ### 🐛 Describe the bug
<p>amp static shape cpp wrapper</p><table border="1" class="dataframe table">
<thead>
<tr style="text-align: right;">
<th>suite</th>
<th>name</th>
<th>thread</th>
<th>batch_size_new</th>
<th>speed_up_new</th>
<th>inductor_new</th>
<th>eager_new</t... | true |
2,883,629,632 | Fix atomic operation compatibility for ARMv8-A (Raspberry Pi 4) by adjusting compilation flags | maajidkhann | closed | [
"triaged",
"open source",
"module: arm",
"Merged",
"ciflow/trunk",
"release notes: build",
"topic: bug fixes",
"ciflow/binaries_wheel",
"bug",
"ciflow/linux-aarch64"
] | 21 | CONTRIBUTOR | **Issue:**
* The ldaddal instruction is an AArch64 atomic operation available from ARMv8.1-A onwards.
* Raspberry Pi 4 (Cortex-A72) is ARMv8-A, which does not support ldaddal, leading to failures when running PyTorch built with march=armv8.2-a+sve
* This led to an issue when running PyTorch on ARMv8-A (Raspberry Pi ... | true |
2,883,606,722 | Replace `unimplemented` with `unimplemented_v2' in `codegen.py` | zeshengzong | closed | [
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo"
] | 5 | CONTRIBUTOR | Fixes #147913
- replace `unimplemented` in `codegen.py`
- remove unused import `unimplemented`
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @williamwen42 | true |
2,883,592,915 | DISABLED test_disable_ctx_manager (__main__.ContextlibContextManagerTests) | pytorch-bot[bot] | closed | [
"module: rocm",
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: dynamo"
] | 2 | NONE | Platforms: rocm
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_disable_ctx_manager&suite=ContextlibContextManagerTests&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37895843797).
Over the past 3 hours... | true |
2,883,592,813 | DISABLED test_slice_dynamic (__main__.AutoFunctionalizeTests) | pytorch-bot[bot] | closed | [
"high priority",
"triaged",
"module: flaky-tests",
"skipped",
"module: functionalization",
"oncall: pt2",
"module: pt2-dispatcher"
] | 5 | NONE | Platforms: asan, linux, mac, macos, rocm, slow
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_slice_dynamic&suite=AutoFunctionalizeTests&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37895844239).
... | true |
2,883,528,347 | use identity op for alpha=inf in torch.celu and quantized_celu | redwrasse | open | [
"module: cpu",
"triaged",
"open source",
"Stale",
"release notes: quantization"
] | 2 | CONTRIBUTOR | Fixes #148065
This MR short-circuits the celu and quantized_celu ops to just return the input in the case alpha=inf, so the implementation of celu(x, inf) is defined on all x.
```
import torch
# (same for torch.ao.nn.quantized.functional.celu)
# Before
# -----------
x = torch.tensor(2.)
print(torch.ce... | true |
2,883,523,874 | Support alpha=inf consistently for torch.celu | redwrasse | open | [
"module: nn",
"triaged",
"module: python frontend"
] | 0 | CONTRIBUTOR | The celu activation function introduced in https://arxiv.org/pdf/1704.07483 is described as being C1 differentiable across values of alpha.
- for alpha -> inf it converges to the identity op (f(x) = x)
- for alpha -> 0+ it converges to relu(x) = max(0, x)
The Pytorch [celu](https://github.com/pytorch/pytorch/blob/m... | true |
2,883,466,389 | [inductor] [silence] `torch.cdist` outputs inconsistent results with eager | shaoyuyoung | closed | [
"oncall: pt2",
"module: inductor"
] | 2 | CONTRIBUTOR | ### 🐛 Describe the bug
**symptom description**: with the `H` and `W` increasing, errors are amplified.
**device backend**: both on Triton and CPP
**repro**:
```python
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch._inductor import config
config.fallback_random = True
torch.set_grad_ena... | true |
2,883,429,107 | Add needs_exact_strides operator tag for Inductor to force exact strides | zou3519 | closed | [
"Merged",
"release notes: fx",
"fx",
"module: inductor",
"ciflow/inductor",
"keep-going"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #150511
* #148104
* #150495
* #148092
* #148091
* __->__ #148063
* #148046
Inductor will force exact strides on a custom operator tagged with
needs_exact_strides. I'll make this the default in a follow-up PR.
Test Plan:
- tests
cc @ezyang ... | true |
2,883,429,040 | Add torch._library.utils.normalize_args_kwargs | zou3519 | open | [
"topic: not user facing"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #148104
* #148092
* #148091
* #148063
* #148046
* __->__ #148062
Normalizes (args, kwargs) to the PyTorch dispatcher calling convention.
I need some sort of normalization in the next PR.
Test Plan:
- new tests | true |
2,883,409,164 | [inductor] [silence] `nn.ConvTranspose2d-F.dropout` outputs inconsistent results with eager | shaoyuyoung | open | [
"triaged",
"module: correctness (silent)",
"oncall: pt2",
"module: inductor"
] | 4 | CONTRIBUTOR | ### 🐛 Describe the bug
**symptom description**: when using `nn.ConvTranspose2d` and `F.dropout` together, outputs are different with eager.
**device backend**: both triton and CPP
**note**: I have used `config.fallback_random = True` and `torch.manual_seed(0)`
**repro**
```python
import torch
import torch.nn as nn
im... | true |
2,883,404,144 | Smoke Test skip cuda.gds on windows | atalman | closed | [
"Merged",
"topic: not user facing"
] | 3 | CONTRIBUTOR | Follow up after : https://github.com/pytorch/pytorch/pull/147120
Cufile was enabled only on Linux: https://pypi.org/project/nvidia-cufile-cu12/#files
Fixes validation workflow failues: https://github.com/pytorch/test-infra/actions/runs/13558218752/job/37896578837
```
File "C:\Jenkins\Miniconda3\envs\conda-env-13... | true |
2,883,396,473 | [PT2] Port fuse_split_getitem_squeeze to PT2 pre_grad passes | huxintong | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 4 | CONTRIBUTOR | Summary: put it as an add_pass option
Reviewed By: frank-wei
Differential Revision: D68909559
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov | true |
2,883,386,578 | [inductor] [cpu] `torch.scatter` throws `AttributeError: 'int' object has no attribute 'find'` on CPP backend | shaoyuyoung | closed | [
"triaged",
"oncall: pt2",
"oncall: cpu inductor"
] | 3 | CONTRIBUTOR | ### 🐛 Describe the bug
**symptom description**: I have done some ablations, this seems to be a minified repro that triggers this `AttributeError`. I am unsure what CPP backend inductor optimizes on `F.gumbel_softmax-torch.where-torch.scatter`.
**device backend**: only CPP backend
**repro**
```python
import torch
imp... | true |
2,883,331,973 | DISABLED test_builtin_score_mods_bfloat16_score_mod0_head_dims0 (__main__.TestFlexDecoding) | pytorch-bot[bot] | closed | [
"module: rocm",
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: inductor"
] | 2 | NONE | Platforms: rocm
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_builtin_score_mods_bfloat16_score_mod0_head_dims0&suite=TestFlexDecoding&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37883438465).
Over... | true |
2,883,331,624 | DISABLED test_nonstrict_trace_pre_existing_custom_class_with_side_effects (__main__.DecoratorTests) | pytorch-bot[bot] | closed | [
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: dynamo"
] | 3 | NONE | Platforms: linux, mac, macos
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_nonstrict_trace_pre_existing_custom_class_with_side_effects&suite=DecoratorTests&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/r... | true |
2,883,331,576 | DISABLED test_nonstrict_trace_inside_compiled_function_kwarg (__main__.DecoratorTests) | pytorch-bot[bot] | closed | [
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: dynamo"
] | 5 | NONE | Platforms: linux, rocm
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_nonstrict_trace_inside_compiled_function_kwarg&suite=DecoratorTests&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37882843708).
Ov... | true |
2,883,331,575 | DISABLED test_nonstrict_trace_on_method (__main__.DecoratorTests) | pytorch-bot[bot] | closed | [
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: dynamo"
] | 3 | NONE | Platforms: asan, linux, mac, macos
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_nonstrict_trace_on_method&suite=DecoratorTests&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/37883456540).
Over the ... | true |
2,883,327,520 | Skip the logging if the pass cannot be pickled | houseroad | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 6 | MEMBER | Summary:
Skip the logging for vllm at this moment, we can add some pickle logic later.
The log is only for debugging purpose.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @chauhang @... | true |
2,883,312,687 | [inductor][user triton] support on-device TMA / tensor descriptor API | davidberard98 | open | [
"triaged",
"oncall: pt2",
"module: inductor",
"upstream triton",
"module: user triton"
] | 0 | CONTRIBUTOR | ### 🚀 The feature, motivation and pitch
Following https://github.com/triton-lang/triton/pull/4916, there's a new tensor descriptor API. We should consider supporting it in case users write kernels that use it, or if we want to use it in inductor directly.
one part of this supporting the global_scratch - partially p... | true |
2,883,311,025 | [triton 3.3] cpp_wrapper: add a global_scratch arg | davidberard98 | closed | [
"Merged",
"ciflow/trunk",
"topic: improvements",
"module: inductor",
"ciflow/inductor",
"release notes: inductor"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148051
Following triton # 4916, the generated cubin expects a global_scratch argument to support on-device TMA. We believe this is the source of many of the "invalid argument" failures on AOTI/cpp_wrapper tests. AFAIK, we don't use o... | true |
2,883,296,018 | [cutlass backend] Check if len(timings) == len(choices) before skipping precompile | henrylhtsang | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 7 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148050
Differential Revision: [D70298908](https://our.internmc.facebook.com/intern/diff/D70298908/)
Mostly from @coconutruben observation. Right now, we skip precompilation if we find **some** timings. That sounds like a bug. Mos... | true |
2,883,287,820 | Fix `torch.nn.functional.hardswish` gradients corner case | zeshengzong | closed | [
"module: autograd",
"module: cpu",
"triaged",
"open source",
"Merged",
"Reverted",
"ciflow/trunk",
"release notes: nn",
"ci-no-td"
] | 32 | CONTRIBUTOR | Fixes #147801
## Changes
- Change hardswish gradient compute condition as [torch.nn.functional.hardswish](https://pytorch.org/docs/stable/generated/torch.nn.functional.hardswish.html)
- Enable cuda for test `test_hardswish_grad_corner`
- Add test case for value=-3
## Test Result
```bash
pytest test/test_... | true |
2,883,284,868 | [RFC] Customization of ElasticAgent for fault-tolerance and node-replacement in big ddp job | zwx4d6 | open | [
"oncall: distributed",
"triaged"
] | 0 | NONE | ### 🚀 The feature, motivation and pitch
When working on >100 nodes training in k8s cluster, there are node failures, some not retryable (may need node replacement). For each job interruption, I would like to run some diagnostics on each node, **right in the training context** (container, fs mounts, hardware configura... | true |
2,883,271,464 | [cutlass backend] Sort the list of ops for better repro | henrylhtsang | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 9 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148047
Differential Revision: [D70298051](https://our.internmc.facebook.com/intern/diff/D70298051/)
This only affects anything if `cutlass_max_profiling_configs` is used. I believe cutlass_max_profiling_configs is more of a testi... | true |
2,883,260,650 | Change arg_kwarg_vals propagation strategy | zou3519 | closed | [
"Merged",
"release notes: fx",
"fx",
"module: inductor",
"ciflow/inductor",
"keep-going"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #150511
* #148104
* #150495
* #148092
* #148091
* #148063
* __->__ #148046
Instead of always propagating arg_kwarg_vals in _COPY_META_FIELDS, we
special-case the pattern matcher to propagate arg_kwarg_vals when
it sees triton_kernel_wrapper_... | true |
2,883,252,818 | Run Performance Regression Tests on new Version of Triton | drisspg | open | [
"module: performance",
"triaged",
"upstream triton"
] | 2 | CONTRIBUTOR | cc @msaroufim @bertmaher @int3 @davidberard98 @nmacchioni @chenyang78 @embg @peterbell10 @aakhundov | true |
2,883,251,685 | Sweep through Potentially BC Breaking Commits in Triton | drisspg | open | [
"triaged",
"upstream triton"
] | 0 | CONTRIBUTOR | + [ ] https://github.com/triton-lang/triton/pull/4955
+ [ ] https://github.com/triton-lang/triton/pull/5637
+ [ ] https://github.com/triton-lang/triton/pull/5926
+ [ ] https://github.com/triton-lang/triton/pull/5961
cc @bertmaher @int3 @davidberard98 @nmacchioni @chenyang78 @embg @peterbell10 @aakhundov | true |
2,883,228,801 | [user-triton] handle inline_asm_case | sijiac | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 8 | CONTRIBUTOR | Summary: We currently failed the mutation analysis for all inline_asm ops. In this diff, we handle the case when "is_pure" is set to True since it indicates the operation doesn't mutate the input value
Test Plan:
../buck-out/v2/gen/fbcode/854b9ed00d28c5c5/caffe2/test/inductor/__triton_kernels__/triton_kernels.par --r ... | true |
2,883,219,835 | [wip][aot] annotated fwd graph dynamic tensor outputs with mark_dynamic | xmfan | closed | [
"ciflow/inductor"
] | 2 | MEMBER | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #148042
* #147891
| true |
2,883,209,326 | Invalid ONNX graph if using float16 dtype for `torch.arange` | Y-T-G | closed | [
"module: onnx",
"triaged"
] | 4 | NONE | ### 🐛 Describe the bug
If you specify a float16 `dtype` for `torch.arange`, the graph is invalid:
```python
import torch
class Test(torch.nn.Module):
def forward(self, x):
return torch.arange(end=x.shape[0], dtype=torch.float16)
test = Test()
torch.onnx.export(test, torch.randn(1), "test.onnx", dynamic_... | true |
2,883,206,777 | [dynamo] torch._dynamo.mark_dynamic hard fails the compile if the dimension is coerced to static | xmfan | closed | [
"oncall: pt2",
"module: dynamic shapes",
"module: dynamo",
"module: guards"
] | 2 | MEMBER | ### 🐛 Describe the bug
I'm trying to annotate fwd graph's dynamic tensor outputs as dynamic to cut down on recompiles: https://github.com/pytorch/pytorch/pull/148042.
```python
import torch
@torch.compile(backend="aot_eager")
def fn(static, dynamic):
return torch.matmul(static, dynamic) # inner dims coerced by... | true |
2,883,165,461 | [TESTING] 1 | ZainRizvi | closed | [
"topic: not user facing"
] | 1 | CONTRIBUTOR | Fixes #ISSUE_NUMBER
| true |
2,883,163,601 | [torchgen] Add support for schema with namespace | larryliu0820 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 6 | CONTRIBUTOR | Fixes https://github.com/pytorch/executorch/issues/8711
In ExecuTorch when we try to parse the following schema:
```
aten::__lshift__.Scalar(Tensor self, Scalar other) -> Tensor
```
Repro:
```python
from torchgen.model import FunctionSchema
native_schema = FunctionSchema.parse("aten::__lshift__.Scalar(Ten... | true |
2,883,153,742 | [ROCm] Skip gfx12 Row-Wise F8 Tests | petrex | open | [
"module: rocm",
"triaged",
"open source",
"ciflow/trunk",
"topic: not user facing",
"ciflow/rocm"
] | 6 | CONTRIBUTOR | TLDR: Skip Gfx12 row-wise fp8 tests (under development)
This pull request introduces changes to the `test/test_matmul_cuda.py` and `torch/testing/_internal/common_cuda.py` files to add support for the GFX12 architecture. The changes include adding a new platform support check for GFX12 and updating several test case... | true |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.