[PyTorch] torch.compile for an OperationFuser group holding one operation - #45

Open

pggPL wants to merge 28 commits into

mainfrom

ops_fuser_compile_main

Commits on Sep 4, 2026

[PyTorch] Register an op's forward and backward without autograd glue
pggPL
committed
cb69309
[PyTorch] Compile an OperationFuser group holding one operation
pggPL
committed
419c4e2
[PyTorch] Accept an operation's declared forward kwargs under compile

pggPL
and
claude
committed
7cf976f
[PyTorch] Restrict compiled forward kwargs to tensors

pggPL
and
claude
committed
4f773fc
[PyTorch] Refine fusible custom-op integration
pggPL
committed
97e4764
[PyTorch] Decline fused grouped MLP when the backward format is not E4M3 (#3352)

wilyan09007
and
ptrendx
authored
b86d8f9
[PyTorch] Fix: Resolve EP symm-mem window offset for both old and new torch version (#3466)
phu0ngng
authored
e61b66d
[NCCL][PyTorch] Fix test-fusible-ops-file-rendezvous-bug (#3478)

authored
1f25d6a
Reduce Grouped MLP Fuser CPU Overhead (#3410)

zhongbozhu
and
tingyangk
authored
80a89ad
Use cuDNN's deterministic dprob in the fused grouped MLP (#3407)

ZhiyuLi-Nvidia
and
vthumbe1503
authored
d1e9c33
[PyTorch] Reduce CUDA graph memory retention (#3427)
buptzyb
authored
29229b5
Relax runtime checks for activation recompute into Warnings (#3436)

ghadiaravi13
and
timmoon10
authored
846379d

Commits on Sep 5, 2026

Commits on Sep 8, 2026

Update Linear docstring referring torch.Linear (#3491)
negvet
authored
dc8909d
[PyTorch] Document fine-grained quantization recipes (#3336)

authored
f2bec23
Fix hybrid fp8 test guard (#3492)
negvet
authored
cd24504
[PyTorch] Register custom ops one at a time

pggPL
and
claude
committed
2281cd5
[PyTorch] Use torch's register_custom_class API for opaque quantizers when available (#3495)

mmarcinkiewicz
and
claude
authored
b97f5ab

Commits on Sep 9, 2026

Add entrypoint for flagos multi-backend plugin system (#3401)

authored
7c6fa59
[PyTorch] Fix NaN expert weight gradients at num_groups == 1 with SReLU (#3485)

authored
39a5401
[Common] Fix Grouped MXFP8 work mapping and TMA synchronization (#3483)

Oleg-Goncharov
and
pre-commit-ci[bot]
authored
7ba2f9f
[PyTorch] Derive an op's compile declarations; gate groups to one op

pggPL
and
claude
committed
1b178bc
Merge remote-tracking branch 'upstream/main' into ops_fuser_compile_main
pggPL
committed
d88c652
[PyTorch] Tidy BasicOperation's compile plumbing

pggPL
and
claude
committed
e1b33b0
[PyTorch] Name an op's custom-op methods after Linear's

pggPL
and
claude
committed
048162b
[PyTorch] Guard compiled fuser cleanup and delayed scaling
pggPL
committed
2f8582f
[PyTorch] Remove added fuser regression tests
pggPL
committed
4c60703