Topic · Edition
Update the torch-xpu-ops commit to intel/torch-xpu-ops@1a3415: Avoid register spilling in elementwise ops for mixed fp32/bf16/fp16 inputs by adding vectorized type-casting loads Fix group_reduce assert on XPU to avoid…
[Test][XPU] Refactor reuse_ir tests to device-type harness and enable…
This PR is auto-generated nightly by this action. Updates the Metamates merge rule with active team members. Pull Request resolved: #194306 Approved by: https://github.com/pytorchbot
[inductor] Preserve SDPA mask locality in combo-kernel scheduling (#1…
[inductor] Preserve SDPA mask locality in combo-kernel scheduling (#1…
Revert "[Test] Make dynamo test_streams.py device-agnostic for out-of…
Enable eager tests for inductor-unittest - - Enabel more tests in tes…
Summary Add scheduled and manually dispatched Better Benchmark CI on an eight-GPU B200 runner. Run the merged upstream eellison/better-benchmark implementation. Export measured kernel latency and projected model latency…
[inductor] AOTI lite mode: support HigherOrderOperators (user-defined…
[Test] Make dynamo test_streams.py device-agnostic for out-of-tree ba…
Description drafted with an AI assistant, quoted below; I've reviewed the change. LpNormFunctor and lpnorm_cleanup each spell the sum-vs-max predicate twice in mutually inverse forms -- an L0 || L1 || L2 enumeration at…
Description drafted with an AI assistant, quoted below; I've reviewed the change. LpNormFunctor and lpnorm_cleanup each spell the sum-vs-max predicate twice in mutually inverse forms -- an L0 || L1 || L2 enumeration at…
This test cases executes the functional_autograd_benchmark.py, which had CUDA hardcoded as the GPU device. This commit generalizes it to the torch.accelerator API Replace hardcoded cuda references with torch.accelerator…
[Test] Refactor test/test_comparition_utils.py according to accelerat…
[ROCm] Restore preview_unsafeAtomicAdd for ROCm < 6.4 in headeronly (…
_PyThreadState_PushFrame is public in 3.15 (and also renamed from _PyThreadState_BumpFramePointerSlow). Using it directly allows us to remove THP_PyThreadState_BumpFramePointerSlow and its supporing definitions Pull…
[inductor] Disable batch_linear_lhs by default on XPU(fixes Qwen regr…
[inductor] Disable batch_linear_lhs by default on XPU(fixes Qwen regr…
Following the CRCR Onboarding Guide - Step 2, add Isalia20/pytorch-mps-ci to the allowlist at L1. The repository will run MPS tests against upstream PyTorch events on macOS arm64 runners. The relay GitHub App is…
