[Common][PyTorch] EP dispatch with unfused MXFP8 quantization - #3270

Open

phu0ngng wants to merge 9 commits into

NVIDIA:mainfrom

phu0ngng:ep_mxfp8

Commits on Aug 4, 2026

EP Dispatch with MXFP8
phu0ngng
committed
56ba9ca
restore gitignore
phu0ngng
committed
b50d785
EP Dispatch with MXFP8
phu0ngng
committed
4ea01cf
Separate EP dispatch-forward and combine-backward quant recipes
phu0ngng
committed
d0a99f4
[PyTorch] Simplify EP quantize path and guard zero-copy pool alloc under CUDA graph capture
phu0ngng
committed
75fb84d
[PyTorch] Add eager-mode coverage for MXFP8 EP combine backward
phu0ngng
committed
bfe103e
[PyTorch] Supply persistent symm-mem recv buffers for zero-copy 1F1B CUDA-graph capture
phu0ngng
committed
19cad25
shorten quant enum
phu0ngng
committed
6a31417
bump nccl ep
phu0ngng
committed
778c2c8