Repository navigation

[PyBindings] Fix wrong results for inputs in another layout, refuse stale or mixed state, return CUDA outputs without copies - #23957

Merged

JacobSzwejbka merged 8 commits into

pytorch:mainfrom

shoumikhin:pybindings-torchfree-gaps

Oct 19, 2026

Merged

[PyBindings] Fix wrong results for inputs in another layout, refuse stale or mixed state, return CUDA outputs without copies#23957
JacobSzwejbka merged 8 commits into
pytorch:mainfrom
shoumikhin:pybindings-torchfree-gaps

Commits on Oct 19, 2026
- [PyBindings] Fix wrong results for inputs in another layout, refuse stale or mixed state, return CUDA outputs without copies
  
  shoumikhin
  authored and
  pytorchbot
  committed
  1a546f5
- Address review: install module inputs before running, reuse returned CUDA memory, docs and naming
  
  shoumikhin
  authored and
  pytorchbot
  committed
  98e09f5
- Check that the profiler records a known copy, and say the ATen bindings do not convert inputs
  
  shoumikhin
  authored and
  pytorchbot
  committed
  f31294c
- Address review: Windows test cleanup, stale outputs after set_inputs, wait before reusing returned CUDA memory, torch mode without CUDA support
  
  shoumikhin
  authored and
  pytorchbot
  committed
  8f1f19d
- [PyBindings] Wait for the device once per call, not once per tensor
  
  shoumikhin
  authored and
  pytorchbot
  committed
  996f3f1
- Reuse the previous output buffer in a loop instead of allocating each call
  
  shoumikhin
  authored and
  pytorchbot
  committed
  d0f5025
- Address review: reuse waits only for other readers, name the input in errors, refuse extra module inputs first
  
  shoumikhin
  authored and
  pytorchbot
  committed
  194bd54
- Address review: wait for pending writes before every read of shared CUDA memory
  
  shoumikhin
  authored and
  pytorchbot
  committed
  c2f63d9