[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
@@ -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"