data/backlog_report.md

pytorch/pytorch Backlog Intelligence Packet

Coverage

• Generated: 2026-07-09T03:47:02+00:00
• Open issues represented: 15437 of 15437 listed
• Open PRs represented: 2770 of 2770 listed
• Full collection: True
• Work thread active coverage: True
• Issue collapse coverage: 15437 open issues in 2704 groups
• Issue collapse full coverage: True
• Issue review-unit reduction: 82.5%
• Maintainer action bundles: 1579
• Patch-plan bundles: 413
• Existing PR decision bundles: 863
• Full issue/PR bundle coverage: True
• No-write audit rows: 15803
• No-write audit boundary: GitHub writes: none
• Relationship edges: 4000
• Public surface: read-only static report

Local Board

• Covered open issues: 15437
• Covered open PRs: 2770
• Local board boundary: no GitHub writes
• Resolution hash: 141b35358567a5186edb6d910eab9ccd2abd54eb84d9d80d0b16c50fa4844f6d
• This local board shows what the tool would review, close, merge, park, or turn into a local patch plan, without doing any of it on GitHub.

No-Write Audit Packet

• All open issue proof rows: 15437
• Stale open PR proof rows: 366
• Proof hash: ec5b874ff029ecfa5d4ecd204d13f40398adb089965dbb6308b8caec6b268a61
• These rows are attachable evidence of what the tool can triage, but they do not comment, label, close, merge, or upload anything.

Maintainer Review Threads

• Multi-issue groups: 1957

• PR-backed groups: 570

• Singleton issue groups retained for audit: 747

• 1 issue - Review thread: 1 issue + 1 PR - pr #188801: [MPS] Leak MetalShaderLibrary bundled singleton to avoid exit-time destructor crashes - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#188812; active PRs: PR #188801)

• 2 issues - Review thread: 2 issues + 1 PR - pr #189122: Fix sparse-dense mul dropping data when broadcasting a size-1 sparse dim - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#158861, #188900; active PRs: PR #189122)

• 2 issues - Review thread: 2 issues + 1 PR - pr #185730: Fix dynamic shapes for variadic kwargs - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#150022, #150371; active PRs: PR #185730)

• 1 issue - Review thread: 1 issue + 1 PR - pr #188948: Fix triangular_solve for sparse CPU tensors on non-MKL platforms - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#153410; active PRs: PR #188948)

• 3 issues - Review thread: 3 issues + 2 PRs - pr #181720: [MPS] Make pin_memory return CPU-aliased storage backed by a unified MTLBuffer - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#180397, #181374, #188970; active PRs: PR #181720, PR #189256)

• 1 issue - Review thread: 1 issue + 1 PR - pr #189043: Preload full bundled cuDNN set with RTLD_GLOBAL to prevent sublibrary version mismatch - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#188892; active PRs: PR #189043)

• 21 issues - Review thread: 21 issues + 14 PRs - pr #181726: [xpu][1/4]Implement scaled_mm_v2 for MXFP8/MXFP4/NVFP4 on XPU - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#178040, #183988, #186348, #187988, #188477, #188675, #188704, #188706, #188707, #188708, #188709, #188710; active PRs: PR #181726, PR #181727, PR #181728, PR #183511, PR #184276, PR #187315, PR #187318, PR #187989)

• 3 issues - Review thread: 3 issues + 3 PRs - pr #184481: Guard synthetic-base input alias layouts - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#93617, #178680, #188133; active PRs: PR #184481, PR #184694, PR #185891)

• 3 issues - Review thread: 3 issues + 2 PRs - pr #185946: [CUDA] Warn instead of assert in reportProcessMemoryInfo to prevent crash on Tegra OOM - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#16706, #185240, #186374; active PRs: PR #185946, PR #186375)

• 2 issues - Review thread: 2 issues + 3 PRs - pr #185846: [decomp] Match eager's device-dependent threshold comparison dtype - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#185470, #185484; active PRs: PR #185846, PR #186358, PR #187609)

• 3 issues - Review thread: 3 issues + 1 PR - pr #189045: Fix NaN gradients in nn.MultiheadAttention for fully-masked rows (#41… - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#40932, #41508, #55056; active PRs: PR #189045)

• 3 issues - Review thread: 3 issues + 1 PR - pr #188998: [inductor] Fix bf16 matmul accumulation precision in Triton templates (#188492) - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#188492, #188841, #189287; active PRs: PR #188998)

• 3 issues - Review thread: 3 issues + 1 PR - pr #184621: Fix Dynamo torch.Size from tensor shape inputs - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#182649, #182651, #182652; active PRs: PR #184621)

• 2 issues - Review thread: 2 issues + 2 PRs - pr #189294: [CI] Realign scipy after nvmath install to fix H100 smoke NumPy ABI crash - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#173059, #189034; active PRs: PR #189294, PR #189295)

• 2 issues - Review thread: 2 issues + 2 PRs - pr #187354: Match Inductor i0/i1 infinities to eager CUDA - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#187332, #188545; active PRs: PR #187354, PR #188556)

• 2 issues - Review thread: 2 issues + 2 PRs - pr #185756: [clamp] Fix float16 scalar overflow check inconsistency between CPU and GPU - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#171356, #187429; active PRs: PR #185756, PR #187908)

• 2 issues - Review thread: 2 issues + 1 PR - pr #185496: Fix export autograd.grad saved tensor tracing - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#146719, #155044; active PRs: PR #185496)

• 2 issues - Review thread: 2 issues + 1 PR - pr #184632: Fix CUDA fake strides for mixed dtype pointwise ops - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#182200, #184101; active PRs: PR #184632)

• 2 issues - Review thread: 2 issues + 1 PR - pr #184628: Fix SingletonInt static guard evaluation - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#182217, #183369; active PRs: PR #184628)

• 2 issues - Review thread: 2 issues + 1 PR - pr #184392: Fix ModularIndexing printer semantics - linked work thread - Review issues, active PRs, referenced context, checks, and review guidance together instead of reopening each issue separately. (#119883, #187027; active PRs: PR #184392)

Maintainer Work Threads

• 151 pr #188801: [MPS] Leak MetalShaderLibrary bundled singleton to avoid exit-time destructor crashes - waiting on contributor - Wait for contributor update on PR #188801; keep related issue/PR context attached. (issue #188812, PR #188801)
• 139 pr #189122: Fix sparse-dense mul dropping data when broadcasting a size-1 sparse dim - ready for maintainer decision - Review active PRs and linked issues as one maintainer work thread. (issue #158861, issue #188900, PR #189122)
• 139 issue #186535: Windows, gloo: Access violation (0xC0000005) in ProcessGroupGloo::enqueue when calling allreduce on CUDA tensors — GlooAllreduceRegistry has no kCUDA creator - needs triage - needs triage (issue #186535)
• 139 issue #188323: [Inductor][CPU] dynamic=True convolution lowering crashes with ValueError: Exponent must be non-negative - needs triage - needs triage (issue #188323)
• 127 issue #157668: NCCL error caused due to use of NVLS in torch 2.7.1-cu128 on aarch64 gb200 cluster - needs triage - needs triage (issue #157668)
• 127 issue #187912: [CPU] Concurrent cpublas::brgemm calls can crash in the AMX path when the underlying oneDNN ukernel is shared - needs triage - needs triage (issue #187912)
• 121 pr #185730: Fix dynamic shapes for variadic kwargs - PR blocked - Review active PRs and linked issues as one maintainer work thread. (issue #150022, issue #150371, PR #185730)
• 121 issue #116254: C++ API at::quantized_max_pool2d: Heap-buffer-overflow - needs triage - needs triage (issue #116254, issue #162476)
• 121 issue #162422: Runtime failure when running torch.compile() & using GCC 11.5.0 on Neoverse V1 - needs triage - needs triage (issue #162422)
• 115 issue #154297: Hangs and timeouts on dist.reduce_scatter on B200 GPU - needs maintainer decision - needs maintainer decision (issue #154297, issue #162178, issue #162745, issue #162748, issue #162820, issue #162871, issue #162897, issue #162917, issue #162940, issue #163429, issue #165170, issue #165685, issue #187158, issue #189065)
• 115 issue #55655: Performance debugging / warning mode - stale/low urgency - stale/low urgency (issue #55655, issue #57118, issue #68768, issue #72948, issue #75725)
• 115 issue #66504: BatchNorm runtimeError: one of the variables needed for gradient computation has been modified by an inplace operation - needs triage - needs triage (issue #66504, issue #68407, issue #73332)
• 115 issue #144965: RuntimeError "global alloc not supported yet" when using TorchScript optimization. - needs triage - needs triage (issue #69078, issue #144965)
• 115 issue #106164: distributed.batch_isend_irecv() crash when send/recv refers to itself - needs triage - needs triage (issue #106164)
• 115 issue #116423: PyTorch Distributed Elastic Launch Segmentation Fault with Python 3.12 - needs triage - needs triage (issue #116423)
• 115 issue #119845: Segmentation fault in dataloader worker sub-process - needs triage - needs triage (issue #119845)
• 115 issue #145610: mmap fails on 64k page aarch64 systems for AOTI model loading - needs triage - needs triage (issue #145610)
• 115 issue #153517: [CI][CUDA][Distributed] test_non_blocking_with_eager_init timeout - needs triage - needs triage (issue #153517)
• 115 issue #162731: DTensor cached op propagation results can be mutated when propagating other ops - needs triage - needs triage (issue #162731)
• 115 issue #167693: DISABLED test_side_stream_backward_overlap_cuda (main.TestAutogradStreamSynchronizationCUDA) - needs triage - needs triage (issue #167693)

Maintainer Attention Queue

• 151 issue #188812: [MPS] MetalShaderLibrary::getBundledLibrary singleton can crash at process exit - needs triage - confirm whether existing evidence is enough; otherwise wait for a reproducer
• 139 issue #188900: sparse.mul: broadcasting a size-1 dimension of the sparse operand silently drops data - needs triage - review early; this can change release or regression risk
• 139 issue #188323: [Inductor][CPU] dynamic=True convolution lowering crashes with ValueError: Exponent must be non-negative - needs triage - review early; this can change release or regression risk
• 139 issue #186535: Windows, gloo: Access violation (0xC0000005) in ProcessGroupGloo::enqueue when calling allreduce on CUDA tensors — GlooAllreduceRegistry has no kCUDA creator - needs triage - review early; this can change release or regression risk
• 127 issue #187912: [CPU] Concurrent cpublas::brgemm calls can crash in the AMX path when the underlying oneDNN ukernel is shared - needs triage - confirm whether existing evidence is enough; otherwise wait for a reproducer
• 127 issue #157668: NCCL error caused due to use of NVLS in torch 2.7.1-cu128 on aarch64 gb200 cluster - needs triage - confirm whether existing evidence is enough; otherwise wait for a reproducer
• 121 issue #150022: Dynamic Shapes with **kwargs - needs triage - review early; this can change release or regression risk
• 121 issue #162422: Runtime failure when running torch.compile() & using GCC 11.5.0 on Neoverse V1 - needs triage - review early; this can change release or regression risk
• 121 issue #162476: heap-buffer-overflow in torch.quantized_max_pool2d via Python API - needs triage - review early; this can change release or regression risk
• 115 issue #189194: [ROCm] torch 2.13 wheel: "Can't detect vectorized ISA for CPU" in torch.compile smoke test on non-ROCm image (regression vs 2.12.1) - needs triage - review early; this can change release or regression risk
• 115 issue #189150: SIGSEGV / cudaErrorIllegalAddress when replaying a CUDA graph that captured multiple training iterations of two trainers over deeply unrolled recurrent modules - needs triage - review early; this can change release or regression risk
• 115 issue #189282: [CPU] operator_benchmark: backward-pass (add, + batchnorm on aarch64) ~200-500x slower since ~Nov 2025 (x86 + aarch64; possible measurement artifact) - needs triage - review early; this can change release or regression risk
• 115 issue #189281: [CPU] operator_benchmark: embedding / embeddingbag ~70-130x slower since ~Nov 2025 (x86 + aarch64; possible measurement issue) - needs triage - review early; this can change release or regression risk
• 115 issue #189239: [Inductor]Unable to apply layout optimization on convolution backward causes training performance regression. - needs triage - review early; this can change release or regression risk
• 115 issue #189144: [XPU][B580] flex-attn-causal performance drop with last_level_cache_size cache clear between runs - needs triage - review early; this can change release or regression risk
• 115 issue #144965: RuntimeError "global alloc not supported yet" when using TorchScript optimization. - needs triage - review early; this can change release or regression risk
• 115 issue #189126: Report a bug, as requested by the error message - needs triage - review early; this can change release or regression risk
• 115 issue #189065: CI for distributed tests on B200 fully red since ~2026-07-01: rank 0 "CUDA-capable device(s) busy or unavailable" -> NCCL id timeout -> 1170-min job timeout - needs triage - review early; this can change release or regression risk
• 115 issue #170003: batch_isend_irecv with nccl causes illegal memory access depending on P2P ordering - needs triage - review early; this can change release or regression risk
• 115 issue #174288: [distributed] Batched isend/irecv with NCCL backend hangs on high load - needs triage - review early; this can change release or regression risk

Issue Buckets

• stale/low urgency: 13044
• needs triage: 1822
• needs maintainer decision: 1332
• has linked PR: 978
• high-priority/blocker: 768
• needs reproduction: 567
• needs design decision: 353
• good volunteer slice: 48

PR Buckets

• PR blocked: 1467
• has linked issue: 1132
• ready for maintainer decision: 558
• draft/noise: 495
• stale/low urgency: 366
• waiting on contributor: 311
• high-priority/blocker: 2

Top Clusters

• PR #188980 with linked issues: pr #188980, issue #188790. Review the issue and PR together instead of triaging both separately.
• PR #167224 with linked issues: pr #167224, issue #134385. Review the issue and PR together instead of triaging both separately.
• PR #189317 with linked issues: pr #189317, issue #181474. Review the issue and PR together instead of triaging both separately.
• PR #188632 with linked issues: pr #188632, issue #188544. Review the issue and PR together instead of triaging both separately.
• PR #186252 with linked issues: pr #186252, issue #184408. Review the issue and PR together instead of triaging both separately.
• PR #189313 with linked issues: pr #189313, issue #102948. Review the issue and PR together instead of triaging both separately.
• PR #189286 with linked issues: pr #189286, issue #189133. Review the issue and PR together instead of triaging both separately.
• PR #188006 with linked issues: pr #188006, issue #168868. Review the issue and PR together instead of triaging both separately.
• PR #186358 with linked issues: pr #186358, issue #185484. Review the issue and PR together instead of triaging both separately.
• PR #189129 with linked issues: pr #189129, issue #188890. Review the issue and PR together instead of triaging both separately.
• PR #186082 with linked issues: pr #186082, issue #140960. Review the issue and PR together instead of triaging both separately.
• PR #189294 with linked issues: pr #189294, issue #189034. Review the issue and PR together instead of triaging both separately.
• PR #189288 with linked issues: pr #189288, issue #189271. Review the issue and PR together instead of triaging both separately.
• PR #188931 with linked issues: pr #188931, issue #188891. Review the issue and PR together instead of triaging both separately.
• PR #188913 with linked issues: pr #188913, issue #188150. Review the issue and PR together instead of triaging both separately.
• PR #188998 with linked issues: pr #188998, issue #188492. Review the issue and PR together instead of triaging both separately.
• PR #189119 with linked issues: pr #189119, issue #189118. Review the issue and PR together instead of triaging both separately.
• PR #188996 with linked issues: pr #188996, issue #188711. Review the issue and PR together instead of triaging both separately.
• PR #189010 with linked issues: pr #189010, issue #188866. Review the issue and PR together instead of triaging both separately.
• PR #189158 with linked issues: pr #189158, issue #189157. Review the issue and PR together instead of triaging both separately.

Limits

• This is an offline, read-only maintainer attention report.
• It does not comment, label, close, merge, or submit anything on GitHub.
• Public bundle exports are reduced rows and short excerpts; full raw API responses remain local unless explicitly approved.
• Duplicate clusters are suggestions for maintainer review, not automatic close decisions.