P
trunk/2942118cd9809a2a925c970df1bf6b91497a6a2f
[dynamo] Migrate misc/distributed/user_defined VariableTrackers to tp…
P
trunk/45a8b7abd8d7db7dcb497f97e321bc32890b5664: Fix non-strict export of vmap tensor indexing (#186894)
Non-strict export rewrites scalar tensor indices in getitem through select/slice helpers, and the scalar-index path converted 0-d integer tensor indices with item(). Under vmap those logical scalar indices are BatchedTensor values, where item() is unsupported, so export failed before regular tensor indexing could lower the operation. Skip the scalar-index rewrite when the index is a functorch Batc…
P
ciflow/xpu/192490: Update
[ghstack-poisoned]
P
trunk/bbcda55d6a3ef9bcc063f82991ac5ec6b133b294
✴️ iz2: Add PhiloxStatelessRNG.cpp to build_variables.bzl (sync with …
P
trunk/85a3be7492910c09ba825d8127d9edf70c7b80bc
[symm_mem] Route multicast setup through the PG when PG rendezvous is…
P
ciflow/vllm/179400
Revert changes to public headers
P
viable/strict/1786280981: [vision hash update] update the pinned vision hash (#192416)
This PR is auto-generated nightly by this action. Update the pinned vision hash. Pull Request resolved: #192416 Approved by: https://github.com/pytorchbot
P
trunk/9f8c31cf638329aa4e01c01896cbe5ee0f4f9230: [vision hash update] update the pinned vision hash (#192416)
This PR is auto-generated nightly by this action. Update the pinned vision hash. Pull Request resolved: #192416 Approved by: https://github.com/pytorchbot
P
viable/strict/1786266230: [torchcomms hash update] update the pinned torchcomms hash (#192415)
This PR is auto-generated nightly by this action. Update the pinned torchcomms hash. Pull Request resolved: #192415 Approved by: https://github.com/pytorchbot
P
trunk/e5309d9ebfe48d64c4552168baf4d433967cd3b1: [torchcomms hash update] update the pinned torchcomms hash (#192415)
This PR is auto-generated nightly by this action. Update the pinned torchcomms hash. Pull Request resolved: #192415 Approved by: https://github.com/pytorchbot
P
trunk/eda033870f055bcfe2b219e70931459f389869f1: [vllm hash update] update the pinned vllm hash (#191815)
This PR is auto-generated nightly by this action. Update the pinned vllm hash. Pull Request resolved: #191815 Approved by: https://github.com/pytorchbot
P
ciflow/trunk/192578: Update
[ghstack-poisoned]
P
ciflow/trunk/183954: [UPDATE] Update
[ghstack-poisoned]
P
ciflow/trunk/187971: Update
[ghstack-poisoned]
P
viable/strict/1786259384
Fuse the final reduction in matmul_reduce_scatter into one Triton ker…
P
trunk/69985f4c2d95401063ec5739934ba09d9aae3792
Fuse the final reduction in matmul_reduce_scatter into one Triton ker…
P
trunk/bbf14e4716f51901fe49058e0ad0f04c0be56800
[ROCm] Fix bfe() bit-field extraction in the quantized CUDA Embedding…
P
ciflow/inductor/192622: Update
[ghstack-poisoned]
P
ciflow/torchtitan/192622: Update
[ghstack-poisoned]
P
ciflow/inductor/192416
update vision commit hash
P
ciflow/trunk/192416
update vision commit hash
P
ciflow/trunk/191815
update vllm commit hash
P
ciflow/torchtitan/191279
update torchtitan commit hash
P
ciflow/trunk/191279
update torchtitan commit hash
P
ciflow/inductor/192415
update torchcomms commit hash
P
ciflow/trunk/192415
update torchcomms commit hash
P
ciflow/torchtitan/191780
Fix a missing type
P
trunk/0754e4e6825173b45e6e8ed91ac37cce4b307156
[Eager] [Bitwise-Equivalence] Implement an INNER_TREE sum reduction (…
P
trunk/80a802e49932d20c6b2539dc6cab6cc8d0dd04cc: Log multi-device backward passes via C10_LOG_API_USAGE_ONCE (#192081)
Add once-per-process API usage logs for backward passes, split by the set of devices they span, so we can assess how much real workloads rely on multithreaded autograd before changing its default. The check walks the graph roots in compute_dependencies and emits two independent signals that may both fire for the same pass: torch.autograd.multidevice_backward when two distinct non-CPU devices are o…
P
trunk/14f9b39fbe12c65090a383a8e5d50e22cb2307c9: [c10d][nccl2] Complete nonblocking communicator creation (#192273)
nccl2 treated ncclInProgress from split and shrink as an immediate failure, and NCCL 2.28 communicator hints were hidden behind an unreachable preprocessor branch. Poll the parent and returned child within one timeout budget, abort both communicators in NVIDIA-required order after a failed nonblocking split, and make the newer hints reachable. The parent is marked unusable after failed creation be…