Topic · Edition
Human Note split up big ghstack rebase pr into the smaller chunk that rebases quack so we can get ci in better spot Agent note FlashAttention 4 b17 leaves its QuACK and CuTeDSL dependencies open-ended. The B200 setup…
Without the fix, TorchTitan graph_trainer + FlexCP with load balancing will trigger the following errors: Without squeeze(0) fix: RuntimeError: Attempting to use FunctionalTensor on its own. Instead, please use it with…
Without the fix, TorchTitan graph_trainer + FlexCP with load balancing will trigger the following errors: Without squeeze(0) fix: RuntimeError: Attempting to use FunctionalTensor on its own. Instead, please use it with…
This PR is auto-generated nightly by this action. Update the pinned vllm hash. Pull Request resolved: #193642 Approved by: https://github.com/pytorchbot
This PR is auto-generated nightly by this action. Update the pinned vllm hash. Pull Request resolved: #193642 Approved by: https://github.com/pytorchbot
PR #192525 introduced a BC-breaking change to MPSProfiler functions by adding a new stream argument that is not optional. So other libraries that use those functions, like torchvision, fail to build…
Impact: PR merge automation / CI only Risk: low What Adds a new Greenlight Review Bot entry to .github/merge_rules.yaml that lets pytorchgreenlight[bot] approve any PR (*), gated on the standard mandatory checks…
Impact: PR merge automation / CI only Risk: low What Adds a new Greenlight Review Bot entry to .github/merge_rules.yaml that lets pytorchgreenlight[bot] approve any PR (*), gated on the standard mandatory checks…
Fixes #193364 Pull Request resolved: #193579 Approved by: https://github.com/malfet
Fixes #193364 Pull Request resolved: #193579 Approved by: https://github.com/malfet
Revert "[Inductor] Query DeviceInterface for device identity and capa…
Revert "[Inductor] Query DeviceInterface for device identity and capa…
Human Note This lets users indicate the stride ordering of the output; this is baically the way to communciate C.T.C . I am going to add this in since its minimilish code and extends the output layout but I might revist…
This PR adds premul_sum support for functional all_reduce forward (& backward) by making the reduce_op argument Any in the schema so that we can pass ReduceOp object directly with factor supplement. adds min and max's…
This PR adds premul_sum support for functional all_reduce forward (& backward) by making the reduce_op argument Any in the schema so that we can pass ReduceOp object directly with factor supplement. adds min and max's…
Remove empty comment line; fix possessive.
Remove empty comment line; fix possessive.
Remove empty comment line; fix possessive.
[AI Codemod][PerfAICT-General] Speed up ATen CPU sort by packing (val…
update vision commit hash
update torchcomms commit hash
update torchtitan commit hash
update vllm commit hash
CuTeDSL receives its tensors through the DLPack interface, and ciflow/b200 is the only job that runs CuTeDSL. Its label paths are matmul/blas-only, so DLPack and DSL changes get no CuTeDSL coverage pre-merge: #182924…
