⟳ Updates Hub
P
Home / AI

PyTorch

Website ↗ RSS 2,323 items Fetched 1h ago
P
PyTorch AI

viable/strict/1789588240: Move the miopen_batch_norm decomposition to the core table (#197160)

DTensor's sharding propagator has no strategy for aten.miopen_batch_norm, so F.batch_norm on a sharded 4D input fails on ROCm with "Operator aten.miopen_batch_norm.default does not have a sharding strategy registered". CUDA never hits the equivalent because aten.cudnn_batch_norm has a decomposition in torch/_decomp, which DTensor's DecompShardingStrategy falls back to. The ROCm decomposition exist…

P
PyTorch AI

viable/strict/1789577445: [BE] Build all XPU wheels on a single runner (#195699)

Take over #188136. Builds all Linux XPU manywheel variants in one job, reusing ABI-independent C++/SYCL artifacts across Python versions while rebuilding Python-bound artifacts per interpreter. Before vs after Comparing the eight XPU matrix builds from #195438 run 33729050328 with the unified build from this PR, run 34123506914: Linux XPU build runner-minutes Before (per-Python matrix, 8 jobs) Aft…

P
PyTorch AI

viable/strict/1789559427: Refine error message for torch.xpu.utilization() on Xe2+ (#197210)

Motivation When we enable pyzes tests on BMG in #196950, we found zesDeviceEnumEngineGroups will fail with pyzes.ZE_RESULT_ERROR_INSUFFICIENT_PERMISSIONS. This is different logic from the Xe arch. Level Zero Sysman will fix it in the next version. We handle this logic in PyTorch side temporarily. Additional Context With this fix: the test will be skipped; otherwise fail with RuntimeError: Can't ge…

P
PyTorch AI

viable/strict/1789552248: [MPS] fix rand and uniform_ upper-bound rounding (#197081)

By adding value == to ? from : value ternary at the end Use volatile to prevent erroneous elimination of static_cast<bfloat> see reproducer below import torch lib = torch.mps.compile_shader(""" #include <metal_stdlib> using namespace metal; bfloat volcast(volatile float x) { return static_cast<bfloat>(x); } kernel void repro_bug(device bfloat* out, constant float2& bounds, constant float* samples,…

P
PyTorch AI

trunk/5038832e12e07068a5231c3e9e6c9700fcc98ca8: [MPS] fix rand and uniform_ upper-bound rounding (#197081)

By adding value == to ? from : value ternary at the end Use volatile to prevent erroneous elimination of static_cast<bfloat> see reproducer below import torch lib = torch.mps.compile_shader(""" #include <metal_stdlib> using namespace metal; bfloat volcast(volatile float x) { return static_cast<bfloat>(x); } kernel void repro_bug(device bfloat* out, constant float2& bounds, constant float* samples,…

P
PyTorch AI

trunk/23341ddcfe7cc4813eb7ef760847d9b786336e1e: [dynamo] Restore autocast state between tests in TestCase (#195662)

Fixes #195487. torch._dynamo.test_case.TestCase.setUp/tearDown already snapshots torch.is_grad_enabled() and restores it (with a warning) in tearDown if a test left it changed, but has no equivalent for autocast. All four pieces of autocast state -- enabled, dtype, cache_enabled, and the nesting counter -- survive past the end of a test into whichever test runs next. A test that leaks this state (…

P
PyTorch AI

trunk/542330ebe79bfc3e9275d7610f9adcf489e3abf9: Fix scaled-MM DeepSeek indexing and contraction validation (#197028)

Human note I found this repeatedly failing on blackwell CI Agent note The DeepSeek Triton main loop assumed row-major scale storage, advanced scale indices incorrectly for GEMM tiles smaller than a scale block, and interleaved outer/K dimensions when expanding 128x128 scales. Honor actual strides and logical axes, compute scale coordinates from element offsets, and preserve outer/K order during ex…

P
PyTorch AI

viable/strict/1789545173: [cuda][nccl] Drop the dead `NCCL_ALLTOALL_SUPPORTED` arm (#196400)

Summary: NCCL_ALLTOALL_SUPPORTED is not defined anywhere in the repo. D115674668 ("[ncclx][fix-build] Remove NCCL_ALLTOALL_SUPPORTED macro") deleted the #define from nccl.h.in in both v2_29 and v2_30, specifically so that an old PyTorch built against a new NCCLX would stop selecting the capital-T ncclAllToAll API. So the defined(NCCL_ALLTOALL_SUPPORTED) term in all2all_single_equal_split is now al…

P
PyTorch AI

viable/strict/1789540965: Update handler clearance logic (#196102)

The tests, when run in shards, use pytest. So, pytest test_logging.py is used instead of python test_logging.py. The initial code, which was: for log_qname in log_internal.log_registry.get_log_qnames(): logger = logging.getLogger(log_qname) created_handlers.update(logger.handlers) self.assertEqual( len(logger.handlers), 2, f"{log_qname} should only have stream and file handlers", ) checks all acti…

P
PyTorch AI

trunk/57914c6e94732191855bd7f1da207f16052d5859: Update handler clearance logic (#196102)

The tests, when run in shards, use pytest. So, pytest test_logging.py is used instead of python test_logging.py. The initial code, which was: for log_qname in log_internal.log_registry.get_log_qnames(): logger = logging.getLogger(log_qname) created_handlers.update(logger.handlers) self.assertEqual( len(logger.handlers), 2, f"{log_qname} should only have stream and file handlers", ) checks all acti…

P
PyTorch AI

trunk/576a6d8db7b7d07d4955b833d7ac56c0500d1e05: [meta] Reject a non-positive pixel_shuffle/unshuffle factor (#195224)

Fixes #153327 Eager rejects a non-positive factor in native_functions: RuntimeError: pixel_shuffle expects a positive upscale_factor, but got 0 The meta registration and the _refs decompositions never got the same check, so they divide by the factor instead: path factor = 0 factor = -2 eager RuntimeError: ...positive upscale_factor, but got 0 RuntimeError: ...but got -2 meta_pixel_shuffle ZeroDivi…