vllm
[torch.compile] Compile `CustomOp.forward_native` for `SiluAndMul` and `QuantFP8` to avoid raw torch ops inside opaque custom ops
#32806
Merged
Merged
[torch.compile] Compile `CustomOp.forward_native` for `SiluAndMul` and `QuantFP8` to avoid raw torch ops inside opaque custom ops #32806
vllm-bot
merged
10 commits into
vllm-project:main
from
neuralmagic:luka/compile-native
Compile forward_native
9f9190e7
gemini-code-assist
commented
on 2026-01-21
mgoin
approved these changes
on 2026-01-21
zou3519
commented
on 2026-01-21
zou3519
approved these changes
on 2026-01-21
Disable compile_native by default, enable for silu and quant only
962e8d66
Fix silu mul quant fusion test
a3a74f5b
Fix activation test
aa85329a
It's often those closest to you that hurt you the most...
f54efeab
Disable compilation for eager backend
b9d03d48
ProExpertProg
changed the title 
Relax tolerance for sensitive test
1b3349a3
eager backend for test to avoid quant_fp8 compilation
846d8b87
Merge branch 'main' into luka/compile-native
3babaf73
mgoin
commented
on 2026-01-22
add type
9ce7cb05
Reviewers
Assignees
No one assigned
Labels
ready
torch.compile
ready-run-all-tests
Milestone
No milestone
Login to write a write a comment.
Login via GitHub