[PyTorch] torch.compile for an OperationFuser group holding one operation - #3496

Merged

pggPL merged 22 commits into

NVIDIA:mainfrom

pggPL:ops_fuser_compile_main

Sep 17, 2026

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

Commits on Sep 8, 2026

[PyTorch] Register custom ops one at a time

pggPL
and
claude
committed
2281cd5

Commits on Sep 9, 2026

[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

Commits on Sep 11, 2026

[PyTorch] Share compute interfaces between basic and fused operations
pggPL
committed
9302392
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
8c76431
[PyTorch] Register one custom operation at a time
pggPL
committed
4c14b36
[PyTorch] Inline single-use custom op registration helpers
pggPL
committed
b648a57
[PyTorch] Use uniform quantized-input handling in custom op wrappers
pggPL
committed
dd41bbf
[PyTorch] Consolidate ops compile tests around shared affine fixtures
pggPL
committed
ff17443
Remove redundant forward kwarg name declarations
pggPL
committed
465ffaa
[PyTorch] Reject custom-op collisions and restore compiled storage safely
pggPL
committed
464907c
[PyTorch] Document custom-op name collision errors
pggPL
committed
01762be

Commits on Sep 16, 2026

[PyTorch] Support independent custom-op directions in OperationFuser
pggPL
committed
83c95e5