Issues

• Nested use of the same torch.cuda.StreamContext instance fails to restore the original stream.

  bot-triagedmodule: correctness (silent)module: cudatriage review
  pytorch/pytorch#203297 · Filter by author ajaydesai9 opened on Nov 3, 2026

• [inductor] Dynamic-shape CPU compiles re-walk every axiom's free symbols in ShapeEnv._maybe_evaluate_static

  bot-triagedoncall: cpu inductor
  pytorch/pytorch#203296 · Filter by author jonasvq opened on Nov 3, 2026

• [dynamo] torch.compile silently drops mutations to captured bytearrays

  bot-triagedmodule: correctness (silent)module: dynamooncall: pt2triage review
  pytorch/pytorch#203287 · Filter by author losewayy opened on Nov 3, 2026

• Planned maintenance affecting GB300 CI jobs

  bot-triagedmodule: cimodule: infratriaged
  pytorch/pytorch#203283 · Filter by author nWEIdia opened on Nov 3, 2026

• [inductor][cpu] Compiled sums over a non-last dim are ~10x less accurate than eager

  bot-triagedoncall: cpu inductor
  pytorch/pytorch#203279 · Filter by author egeboy35 opened on Nov 3, 2026

• flex_gemm QUACK backend returns wrong results on sm_120 when Inductor's triton.enable_pdl is on

  bot-triagedhas workaroundmodule: correctness (silent)module: cudamodule: dependency bugmodule: inductormodule: linear algebraoncall: pt2triage review
  pytorch/pytorch#203269 · Filter by author shino16 opened on Nov 3, 2026

• [RFC] Add logit_softcap to F.linear_cross_entropy

  bot-triagedfunction requestmodule: losstriaged
  pytorch/pytorch#203267 · Filter by author KocKaan opened on Nov 3, 2026

• FSDP RegisterPostBackwardFunction changes the SPMD layout of independent tuple outputs

  bot-triagedmodule: fsdponcall: distributed parallelismsoncall: distributedptd-bot-triagedtriaged
  pytorch/pytorch#203261 · Filter by author shuhuayu opened on Nov 3, 2026

• [inductor][MPS] bool max/min reductions fail to compile: no simd_max/simd_min for bool in c10/metal/reduction_utils.h

  module: inductormodule: mpsoncall: pt2
  pytorch/pytorch#203248 · Filter by author aorenste opened on Nov 3, 2026

• [CI] Need a CPU-only x86 AVX2 runner for inductor_avx2 (it borrows an L4 GPU node) and the cpu_inductor_freezing_avx2_* benchmarks (they run on AVX-512)

  bot-triagedenhancementmodule: cimodule: infratriaged
  pytorch/pytorch#203246 · Filter by author aorenste opened on Nov 3, 2026

• torch.distributions: intermediate overflow corrupts variance, Cauchy KL, and Uniform methods for finite inputs

  bot-triagedmodule: correctness (silent)module: distributionsmodule: edge casesmodule: halfmodule: numerical-stabilityrelease triagetriage review
  pytorch/pytorch#203245 · Filter by author xi-xiaoran opened on Nov 3, 2026

• [inductor][triton] TritonPrinter._print_Float truncates float64 literals

  actionablebot-triagedhigh prioritymodule: correctness (silent)module: dynamic shapesmodule: inductormodule: pt2 accuracyoncall: pt2triage review
  pytorch/pytorch#203234 · Filter by author kundaMwiza opened on Nov 3, 2026
  1

  

• [Inductor] RoPE real-pair cat/view chain crosses the opcount realization threshold and adds avoidable kernels

  bot-triagedenhancementmodule: inductormodule: performanceoncall: pt2triaged
  pytorch/pytorch#203232 · Filter by author kimkingyu opened on Nov 3, 2026
  1

• [Inductor] Result reshape blocks profitable mm+bias→addmm candidates for FP32 3-D linear expressions

  bot-triagedenhancementmodule: inductormodule: linear algebramodule: performanceoncall: pt2triaged
  pytorch/pytorch#203231 · Filter by author kimkingyu opened on Nov 3, 2026

• [Inductor] Reconvergent where/mul uses prematurely materialize cheap Pointwise producers

  bot-triagedenhancementmodule: inductormodule: performanceoncall: pt2triaged
  pytorch/pytorch#203230 · Filter by author kimkingyu opened on Nov 3, 2026

• [Autograd] special.erfcx float64 gradient collapses to zero at a large positive input

  bot-triagedmodule: autogradmodule: correctness (silent)module: derivativesmodule: edge casesmodule: numerical-stabilitymodule: specialmodule: windowstriage review
  pytorch/pytorch#203220 · Filter by author cuiliaomei-beep opened on Nov 3, 2026

• [linalg] Inconsistent NaN handling between eigh/eigvalsh and eig

  bot-triagedmodule: error checkingmodule: linear algebramodule: NaNs and Infstriage review
  pytorch/pytorch#203216 · Filter by author tangyuan499-commits opened on Nov 3, 2026
  2

• DTensor has no sharding strategy for aten._scaled_mm_v2 (F.scaled_mm), only for _scaled_mm

  bot-triagedmodule: dtensoroncall: distributed parallelismsoncall: distributedptd-bot-triagedtriaged
  pytorch/pytorch#203215 · Filter by author 0z5a opened on Nov 3, 2026
  2

• [Inductor] Constant tensor views with nonzero offsets can receive invalid Triton alignment metadata

  high priorityoncall: pt2triage review
  pytorch/pytorch#203214 · Filter by author immortails opened on Nov 3, 2026

• cuDNN nn.LSTM 5-40x slower for hidden_size 136-192 on NVIDIA GB10 (DGX Spark); PERSIST_STATIC_SMALL_H threshold?

  bot-triagedhas workaroundmodule: cudnnmodule: performancemodule: rnnrelease triagetriage review
  pytorch/pytorch#203213 · Filter by author farhanferoz opened on Nov 3, 2026

• [inductor] scan over a dynamic tile count: resolve_shape_to_proxy raises KeyError on an expression key

  bot-triagedmodule: dynamic shapesmodule: higher order operatorsmodule: inductormodule: pt2-dispatcheroncall: pt2triaged
  pytorch/pytorch#203212 · Filter by author Vivek1258 opened on Nov 3, 2026
  1

• torch._scaled_mm meta rejects row-major mat2 on Blackwell although the CUDA kernel accepts it (eager OK, torch.compile fails)

  bot-triagedmodule: cudamodule: error checkingmodule: floatx (formerly float8)module: linear algebramodule: meta tensorsoncall: pt2triaged
  pytorch/pytorch#203211 · Filter by author 0z5a opened on Nov 3, 2026
  

• Per-device test capability signals are unavailable, so hardware decoupling silently changes MTIA coverage

  bot-triagedenhancementmodule: mtiamodule: testsoncall: pt2triage review
  pytorch/pytorch#203203 · Filter by author gavinwang269 opened on Nov 3, 2026

• DISABLED test_inductor_runs_on_nonzero_device (__main__.TestCompileOnOneRankDeviceAsParameter)

  module: flaky-testsoncall: distributed parallelismsoncall: distributedptd-bot-triagedskippedtriage reviewtriaged
  pytorch/pytorch#203197 · Filter by author pytorch-bot opened on Nov 3, 2026
  3

• DISABLED test_device_passthrough_still_reuses_compile_under_coor_origin_intermediate (__main__.TestCompileOnOneRankDeviceAsParameter)

  module: flaky-testsoncall: distributedptd-bot-triagedskippedtriage reviewtriaged
  pytorch/pytorch#203196 · Filter by author pytorch-bot opened on Nov 3, 2026
  4