Commit 3c2ac54
committed
perf: make MSE observer compatible with torch.compile
- Replace patch_attr context manager with pre-created token_args
- Replace data-dependent control flow (torch.any + early stopping) with torch.where
- MSE observer now passes torch.compile fullgraph=True
- Benchmark: 39x speedup with inductor backend
Partial fix for #1485
Signed-off-by: Jaewoo Kim <pewpewplay315@gmail.com>1 parent 36c30ee commit 3c2ac54
1 file changed
+234
-236
lines changed
0 commit comments