[torch.compile] Compile CustomOp.forward_native for SiluAndMul and QuantFP8 to avoid raw torch ops inside opaque custom ops (#32806)

Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Luka Govedič
2026-01-22 19:52:26 -08:00
committed by GitHub
co-authored by Michael Goin
parent f61c9da711
commit 5e4e0e51f4
7 changed files with 52 additions and 13 deletions
@@ -38,6 +38,7 @@ class QuantFP8(CustomOp):
column_major_scales: bool = False,
tma_aligned_scales: bool = False,
use_ue8m0: bool | None = None, # for Torch compile
compile_native: bool = True,
):
"""
:param static: static or dynamic quantization
@@ -49,8 +50,9 @@ class QuantFP8(CustomOp):
TMA-aligned layout
:param column_major_scales: For group quantization, output scales in
column major format
:param compile_native: Manually compile forward_native if compile mode > None
"""
super().__init__()
super().__init__(compile_native=compile_native)
self.static = static
self.group_shape = group_shape
self.use_per_token_if_dynamic = group_shape == GroupShape.PER_TOKEN