[Fix] diffusion: Keep the linear-GELU verdict cache out of torch.compile traces (#45944)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Mick
2026-10-23 17:01:51 +08:00
committed by GitHub
co-authored by Mick Qian Claude Opus 5.5
parent f9129915a3
commit 494eeecb6c
2 changed files with 31 additions and 0 deletions
@@ -138,6 +138,9 @@ def can_use_linear_gelu(linear: Any, x: torch.Tensor) -> bool:
return False
if getattr(linear, "weight", None) is None or x.dtype != linear.weight.dtype:
return False
if torch.compiler.is_compiling():
# dynamo guards on the cache's size, so reading it recompiles per linear
return _input_independent_ok(linear)
verdict = _VERDICTS.get(linear)
if verdict is None:
verdict = _VERDICTS[linear] = _input_independent_ok(linear)