[NVIDIA] Support SiluMul + NVFP4 quant fusion (#23671)

Signed-off-by: jindih <jindih@nvidia.com>
Signed-off-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
Co-authored-by: jindih <jindih@nvidia.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Luka Govedic <lgovedic@redhat.com>
This commit is contained in:
elvischenv
2025-08-28 19:36:50 +00:00
committed by GitHub
co-authored by jindih Michael Goin Luka Govedic
parent 57d4ede520
commit 16a45b3a28
11 changed files with 746 additions and 64 deletions
@@ -97,6 +97,13 @@ class FixFunctionalizationPass(VllmInductorPass):
node,
mutated_args,
args=('result', 'input', 'scale'))
elif at_target == torch.ops._C.silu_and_mul_nvfp4_quant.default:
mutated_args = {1: 'result', 2: 'result_block_scale'}
self.defunctionalize(graph,
node,
mutated_args,
args=('result', 'result_block_scale',
'input', 'input_global_scale'))
else:
continue # skip the count