[PyTorch] Reduce CPU overhead on eager EP dispatch and combine - #3380

Merged

phu0ngng merged 11 commits into

NVIDIA:mainfrom

phu0ngng:ep-pyt

Aug 25, 2026

Commits on Aug 24, 2026

[PyTorch] Fuse eager EP prepare and dispatch into a single op
phu0ngng
committed
0ff3530
Unify eager and non-eager EP dispatch into one fused prepare+dispatch op
phu0ngng
committed
3963566
[PyTorch] Trim eager EP dispatch autograd operands to cut per-call Python overhead
phu0ngng
committed
eb75c6a
[PyTorch] Inline EP dispatch backward and tidy docstrings
phu0ngng
committed
d9d3397
[PyTorch] Call the backend directly for eager EP combine to cut per-call CPU overhead
phu0ngng
committed
793c63d
[PyTorch] Trim eager EP combine autograd operands to cut per-call Python overhead
phu0ngng
committed
afbb537
[PyTorch] Call the backend directly for eager EP dispatch backward to cut per-call CPU overhead
phu0ngng
committed
6d2c0d8
[PyTorch] Reuse cached zero-copy flag and skip eager recv detach to cut EP CPU overhead
phu0ngng
committed
4151609
[PyTorch] Extract EP dispatch/combine fwd+bwd compute into reusable autograd-free helpers
phu0ngng
committed
64eb8c6
[PyTorch] Use torch.dtype.itemsize and slotted state dataclasses to cut EP CPU overhead
phu0ngng
committed
1eafa0f

Commits on Aug 25, 2026

Merge branch 'main' into ep-pyt
phu0ngng
authored
dbb44ed