[Fusion] Manual ActivationQuantFusionPass initial application (#51415)

Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: ElizaWszola <ewszola@redhat.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: ElizaWszola <ewszola@redhat.com>
This commit is contained in:
Misha Goin
2026-09-03 13:37:31 -04:00
committed by GitHub
co-authored by Claude Opus 4.8 ElizaWszola
parent d4d703caf9
commit b7624069ff
10 changed files with 740 additions and 19 deletions
+10 -5
View File
@@ -14,6 +14,7 @@ from .common import (
Matches,
custom_ops_combos,
is_blackwell,
nvfp4_kernel_exposes_input_quant_key,
)
from .models import (
FLASHINFER_ATTN,
@@ -77,7 +78,6 @@ def test_tp1_fp8_fusions(
inductor_graph_partition: bool,
use_deepgemm: bool,
run_e2e_fusion_test,
monkeypatch,
):
if use_deepgemm and not current_platform.is_cuda():
pytest.skip("DeepGemm only supported on CUDA")
@@ -109,7 +109,7 @@ def test_tp1_fp8_fusions(
custom_ops=custom_ops.split(","),
pass_config=PassConfig(
fuse_norm_quant=True,
fuse_act_quant=True,
fuse_act_quant=False,
fuse_attn_quant=True,
enable_qk_norm_rope_fusion=True,
),
@@ -119,7 +119,6 @@ def test_tp1_fp8_fusions(
matches_check = [
"rms_quant_fusion",
"act_quant_fusion",
"norm_rope_fusion",
"attn_quant_fusion",
]
@@ -168,6 +167,12 @@ def test_tp1_fp4_fusions(
inductor_graph_partition: bool,
run_e2e_fusion_test,
):
if nvfp4_kernel_exposes_input_quant_key():
pytest.skip(
"NVFP4 kernel exposes input_quant_key; manual fusion fires "
"instead of compiler pass-based fusion"
)
matches = matches_fn(n_layers)
# Reduce size of model and skip weight loading time
@@ -181,13 +186,13 @@ def test_tp1_fp4_fusions(
custom_ops=custom_ops.split(","),
pass_config=PassConfig(
fuse_norm_quant=True,
fuse_act_quant=True,
fuse_act_quant=False,
fuse_attn_quant=True,
enable_qk_norm_rope_fusion=True,
),
)
matches_check = ["act_quant_fusion", "attn_quant_fusion", "norm_rope_fusion"]
matches_check = ["attn_quant_fusion", "norm_rope_fusion"]
run_e2e_fusion_test(
model_name,