[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
+39
View File
@@ -43,6 +43,45 @@ is_blackwell = lambda: current_platform.is_device_capability_family(100)
"""Are we running on Blackwell, a lot of tests depend on it"""
def nvfp4_kernel_exposes_input_quant_key() -> bool:
"""Check if the NVFP4 kernel selected on this platform exposes input_quant_key.
FlashInferCuteDslNvFp4LinearKernel does not expose input_quant_key() due to
layout incompatibility between the manual fusion kernel (silu_and_mul_nvfp4_quant)
output format and the cutedsl backend's expected input format.
FlashInferCutlassNvFp4LinearKernel does expose input_quant_key() and supports
manual fusion.
"""
if not current_platform.is_cuda():
return False
try:
from vllm.model_executor.kernels.linear.nvfp4.flashinfer import (
FlashInferCuteDslNvFp4LinearKernel,
)
is_supported, _ = FlashInferCuteDslNvFp4LinearKernel.is_supported()
if is_supported:
return False
except ImportError:
pass
try:
from vllm.model_executor.kernels.linear.nvfp4.flashinfer import (
FlashInferCutlassNvFp4LinearKernel,
)
is_supported, _ = FlashInferCutlassNvFp4LinearKernel.is_supported()
if is_supported:
return True
except ImportError:
pass
# Fallback: assume no manual fusion support
return False
def custom_ops_combos(*custom_ops: str) -> Iterable[str]:
"""Generate all combinations of custom ops for parametrization."""
custom_ops_lists = [[f"-{op}", f"+{op}"] for op in custom_ops]
+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,
+9 -6
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,
@@ -68,7 +69,6 @@ def test_tp2_ar_rms_fp8_fusions(
custom_ops: str,
inductor_graph_partition: bool,
run_e2e_fusion_test,
monkeypatch,
):
matches = matches_fn(n_layers)
@@ -89,7 +89,7 @@ def test_tp2_ar_rms_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,
fuse_allreduce_rms=True,
@@ -98,7 +98,6 @@ def test_tp2_ar_rms_fp8_fusions(
matches_check = [
"rms_quant_fusion",
"act_quant_fusion",
"norm_rope_fusion",
"attn_quant_fusion",
"ar_rms_fusion",
@@ -139,8 +138,13 @@ def test_tp2_ar_rms_fp4_fusions(
custom_ops: str,
inductor_graph_partition: bool,
run_e2e_fusion_test,
monkeypatch,
):
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
@@ -154,14 +158,13 @@ def test_tp2_ar_rms_fp4_fusions(
use_inductor_graph_partition=inductor_graph_partition,
custom_ops=custom_ops.split(","),
pass_config=PassConfig(
fuse_act_quant=True,
fuse_act_quant=False,
fuse_attn_quant=True,
fuse_allreduce_rms=True,
),
)
matches_check = [
"act_quant_fusion",
"attn_quant_fusion",
"ar_rms_fusion",
]
+10 -6
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,
@@ -61,7 +62,7 @@ def test_tp2_async_tp_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,
enable_sp=True,
@@ -74,7 +75,6 @@ def test_tp2_async_tp_fp8_fusions(
matches_check = [
"rms_quant_fusion",
"act_quant_fusion",
"norm_rope_fusion",
"attn_quant_fusion",
"sequence_parallel",
@@ -114,6 +114,12 @@ def test_tp2_async_tp_nvfp4_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"
)
# NVFP4 currently wires the all-gather + GEMM path only.
matches = matches_fn(n_layers)._replace(async_tp=n_layers * 2)
@@ -127,7 +133,7 @@ def test_tp2_async_tp_nvfp4_fusions(
use_inductor_graph_partition=inductor_graph_partition,
custom_ops=custom_ops.split(","),
pass_config=PassConfig(
fuse_act_quant=True,
fuse_act_quant=False,
fuse_attn_quant=True,
enable_sp=True,
fuse_gemm_comms=True,
@@ -138,7 +144,6 @@ def test_tp2_async_tp_nvfp4_fusions(
)
matches_check = [
"act_quant_fusion",
"attn_quant_fusion",
"sequence_parallel",
"async_tp",
@@ -245,7 +250,7 @@ def test_tp2_sp_ar_rms_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,
enable_sp=True,
@@ -258,7 +263,6 @@ def test_tp2_sp_ar_rms_fp8_fusions(
matches_check = [
"rms_quant_fusion",
"act_quant_fusion",
"norm_rope_fusion",
"attn_quant_fusion",
"ar_rms_fusion",