[compile] Add FlashInfer FP8 async TP fusion and preserve allreduce fusion ordering #27893 (#39505)

Signed-off-by: baonudesifeizhai <baonudesifeizhai@gmail.com>
Signed-off-by: baonudesifeizhai <85092850+baonudesifeizhai@users.noreply.github.com>
Signed-off-by: roG0d <baonudesifeizhai@gmail.com>
This commit is contained in:
baonudesifeizhai
2026-05-01 05:08:34 +00:00
committed by GitHub
parent 947138b6c2
commit c3868bbbe4
6 changed files with 397 additions and 30 deletions
+17 -3
View File
@@ -24,10 +24,24 @@ def mock_cuda_platform():
def _mock_platform(is_cuda: bool = True, capability: tuple[int, int] | None = None):
mock_platform = MagicMock()
mock_platform.is_cuda.return_value = is_cuda
if capability is not None:
mock_platform.get_device_capability.return_value = DeviceCapability(
*capability
device_capability = (
DeviceCapability(*capability) if capability is not None else None
)
mock_platform.get_device_capability.return_value = device_capability
def is_device_capability_family(
requested_capability: int, device_id: int = 0
) -> bool:
current_capability = mock_platform.get_device_capability(
device_id=device_id
)
if current_capability is None:
return False
return current_capability.major == (requested_capability // 10)
mock_platform.is_device_capability_family.side_effect = (
is_device_capability_family
)
with patch("vllm.platforms.current_platform", mock_platform):
yield mock_platform
+6
View File
@@ -97,6 +97,12 @@ def run_e2e_fusion_test(monkeypatch, caplog_mp_spawn):
f"attention backend '{attn_backend.backend.name}'"
)
if attn_backend.backend.name == "FLASHINFER":
from vllm.utils.flashinfer import supports_trtllm_attention
if not supports_trtllm_attention():
matches = matches._replace(attn_quant_fusion=0)
# TODO: remove this after finishing migration from envs to model kwargs
if model_name == "openai/gpt-oss-20b":
from .common import is_blackwell
@@ -13,7 +13,6 @@ from .common import (
AttentionBackendCase,
Matches,
custom_ops_combos,
is_blackwell,
)
from .models import (
FLASHINFER_ATTN,
@@ -46,14 +45,9 @@ def test_tp2_async_tp_fp8_fusions(
custom_ops: str,
inductor_graph_partition: bool,
run_e2e_fusion_test,
monkeypatch,
):
matches = matches_fn(n_layers)
if is_blackwell():
# Disable FlashInfer scaled_mm FP8 as it's not supported in async tp patterns
monkeypatch.setenv("VLLM_DISABLED_KERNELS", "FlashInferFP8ScaledMMLinearKernel")
# Reduce size of model and skip weight loading time
model_kwargs["hf_overrides"] = hf_overrides(n_layers)
model_kwargs["load_format"] = "dummy"
@@ -173,14 +167,9 @@ def test_tp2_sp_ar_rms_fp8_fusions(
custom_ops: str,
inductor_graph_partition: bool,
run_e2e_fusion_test,
monkeypatch,
):
matches = matches_fn(n_layers)
if is_blackwell():
# Disable FlashInfer scaled_mm FP8 as it's not supported in async tp patterns
monkeypatch.setenv("VLLM_DISABLED_KERNELS", "FlashInferFP8ScaledMMLinearKernel")
# Reduce size of model and skip weight loading time
model_kwargs["hf_overrides"] = hf_overrides(n_layers)
model_kwargs["load_format"] = "dummy"