Repository navigation

CUDA backend: CUDA graph fixes for mask indexing, .item() and changed input shapes, masked copies as where, tensor checks dropped before compiling, float64, _reshape_copy, scatter_reduce and varlen_attn, follow-ups - #24326

Open

shoumikhin wants to merge 87 commits into

pytorch:mainfrom

shoumikhin:cuda-graph-data-dependent

Open

CUDA backend: CUDA graph fixes for mask indexing, .item() and changed input shapes, masked copies as where, tensor checks dropped before compiling, float64, _reshape_copy, scatter_reduce and varlen_attn, follow-ups#24326
shoumikhin wants to merge 87 commits into
pytorch:mainfrom
shoumikhin:cuda-graph-data-dependent

Commits on Oct 30, 2026
- Refuse a CUDA graph replay when an input changes shape
  shoumikhin
  committed
  38993f8
- List the CUDA index shims in the wheel's shared library check
  shoumikhin
  committed
  19df317
- Test aoti_torch_get_size
  shoumikhin
  committed
  034d312
- Test more CUDA index shim and CUDA graph cases
  shoumikhin
  committed
  f52c847
- Run a CUDA graph capture call without a graph when the capture fails
  shoumikhin
  committed
  fdbc332
- Say which dim differs when a CUDA graph replay refuses an input
  shoumikhin
  committed
  2769a41
- Log an expected CUDA graph capture failure at Info
  shoumikhin
  committed
  1d70dcc
- Record CUDA graph allocations only after the first launch
  shoumikhin
  committed
  cee251b
- Check the new CUDA graph input shapes in the recapture and replay tests
  shoumikhin
  committed
  757647c
- Delete the antialiased resize torch.compile test
  shoumikhin
  committed
  a859938
- Say why the antialiased resize avoids a conditional
  shoumikhin
  committed
  80e7430
- Refuse a hand-built data name in AOTI backends that keep data in the .pte
  shoumikhin
  committed
  3728b06
- Warn that CUDA data names differing only in case can share a file
  shoumikhin
  committed
  c8f0b49
- Correct comments on the CUDA index shims and set_output_data_ptr
  shoumikhin
  committed
  d449899
- Seed the CUDA random shims from std::random_device
  shoumikhin
  committed
  849e5ba
- Run a CUDA method that keeps host state without a graph
  shoumikhin
  committed
  8c965a1
- Check the AOTI data name before compiling
  shoumikhin
  committed
  5d280e3
- Check delegation and restore deterministic mode in the CUDA index shim tests
  shoumikhin
  committed
  ddb583e
- Fix the CUDA graph capture guard comment
  shoumikhin
  committed
  5a43778
- Copy through a boolean mask with where in the CUDA backend
  shoumikhin
  committed
  36b6498
- Keep ops with more than 6 dims out of the XNNPACK delegate
  shoumikhin
  committed
  05ee210
- Merge main into cuda-graph-data-dependent
  shoumikhin
  committed
  7c53ae1
- Drop tensor checks before the CUDA backend compiles a program
  shoumikhin
  committed
  24f073c
- Test the CUDA masked copy pass and the capture retry end to end
  shoumikhin
  committed
  35f3e6e
- Remove a size check only when it tests the index's count
  shoumikhin
  committed
  8ec0a71
- Say why a CUDA method keeps its graph off for host constants
  shoumikhin
  committed
  7429ceb
- Keep destroy() with the other public CUDA backend methods
  shoumikhin
  committed
  3347c9c
- Draw the CUDA RNG seed before allocating its state
  shoumikhin
  committed
  c2f0b22
- Format the RemoveAssertAsyncPass import with ufmt
  shoumikhin
  committed
  0d88c5b
- Load float64 weights in the CUDA backend
  shoumikhin
  committed
  f65dbb7
- Decompose aten._reshape_copy in the CUDA backend
  shoumikhin
  committed
  0674e27
- Add aoti_torch_dtype_float64 to aoti_cuda_shims.lib
  shoumikhin
  committed
  339de9b
- Retrigger CI
  shoumikhin
  committed
  9cc98fc
- Add a CUDA shim for scatter_reduce with amax, amin, prod and mean
  shoumikhin
  committed
  080ebd4
- Merge main into cuda-graph-data-dependent
  shoumikhin
  committed
  f80b75e
Commits on Oct 31, 2026
Commits on Nov 1, 2026

Commits

Commits on Oct 30, 2026

Commits on Oct 31, 2026

Commits on Nov 1, 2026