[Bugfix] Drop enable_in_kernel_fc2_reduce from FlashInfer CuteDSL MegaMoE config (#59900)

Signed-off-by: Shang Wang <samshang.wang@mail.utoronto.ca>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Shang Wang
2026-10-12 09:05:09 -05:00
committed by GitHub
co-authored by Claude Opus 5.5
parent a281ad32f0
commit 716961cf1f
2 changed files with 64 additions and 1 deletions
@@ -10,9 +10,14 @@ CuTeDSL kernel (NVFP4 activations and weights, E4M3 scales per 16) as well as
the DeepGEMM kernel (FP8 activations and FP4 weights, power-of-two scales per
32). The expected output is then known in closed form and compared bit for
bit.
A test that needs no SM100 also checks the adapter's kwargs against the
installed FlashInfer's config types, so config drift fails on any CI runner
with FlashInfer installed.
"""
from dataclasses import dataclass
from types import SimpleNamespace
import pytest
import torch
@@ -22,6 +27,7 @@ from tests.distributed.eplb_utils import distributed_run, set_env_vars_and_devic
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.config.kernel import FLASHINFER_MOE_EP_CUTEDSL, FLASHINFER_MOE_EP_DEEP_GEMM
from vllm.distributed.parallel_state import ensure_model_parallel_initialized
from vllm.model_executor.layers.fused_moe import flashinfer_moe_ep as fi_ep
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
@@ -305,3 +311,59 @@ def test_init_warmup_forward_is_exact_on_grid_inputs(
if torch.accelerator.device_count() < world_size:
pytest.skip(f"Need at least {world_size} GPUs")
distributed_run(_init_warmup_forward, world_size, geometry, backend)
@pytest.mark.parametrize("backend", BACKENDS, ids=("cutedsl", "deep_gemm"))
def test_adapter_kwargs_match_installed_flashinfer(
monkeypatch: pytest.MonkeyPatch, backend: str
):
"""The adapter's kwargs construct the installed FlashInfer's config types.
Needs no SM100, so config kwarg drift fails in CI rather than at model load.
Fakes only the megakernel layer (symmetric memory), the EP group, the device
index and the DeepGEMM import.
"""
moe_ep = pytest.importorskip("flashinfer.moe_ep")
layer_args: list[tuple] = []
class FakeMegaLayer:
def __init__(self, *args) -> None:
layer_args.append(args)
def destroy(self) -> None:
pass
monkeypatch.setattr(moe_ep, "MoEEpMegaLayer", FakeMegaLayer)
monkeypatch.setattr(fi_ep, "_expose_deep_gemm_to_flashinfer", lambda: None)
monkeypatch.setattr(
fi_ep,
"get_ep_group",
lambda: SimpleNamespace(world_size=2, rank_in_group=0, device_group=None),
)
monkeypatch.setattr(torch.accelerator, "current_device_index", lambda: 0)
geometry = GEOMETRIES[0]
moe = SimpleNamespace(
moe_backend=backend,
num_experts=geometry.num_experts,
experts_per_token=geometry.top_k,
hidden_dim=geometry.hidden_size,
intermediate_size=geometry.intermediate_size,
max_num_tokens=MAX_TOKENS_PER_RANK,
swiglu_limit=10.0,
routing_method=_routing_method(backend),
)
weight = torch.zeros(1, dtype=torch.bfloat16)
FlashInferMoeEp(
moe,
FlashInferMoeEpWeights(w13=weight, w2=weight),
apply_topk_in_fc1=apply_topk_in_fc1(moe),
)
((_, _, _, mega_config),) = layer_args
expected = (
moe_ep.Nvfp4CutedslMegaMoeConfig
if backend == FLASHINFER_MOE_EP_CUTEDSL
else moe_ep.DeepGemmMegaMoeConfig
)
assert isinstance(mega_config.megakernel, expected)
@@ -315,13 +315,14 @@ class FlashInferMoeEp:
w2_scale=weights.w2_scale,
)
if spec.kernel == "cutedsl":
# In-kernel fc2 reduce keeps its default (off); FlashInfer 0.7.1rc1
# renamed in_kernel_fc2_reduce to enable_in_kernel_fc2_reduce.
megakernel = api.Nvfp4CutedslMegaMoeConfig(
intermediate_size=moe.intermediate_size,
top_k=moe.experts_per_token,
gate_up_clamp=moe.swiglu_limit,
fast_math=True,
apply_topk_in_fc1=apply_topk_in_fc1,
enable_in_kernel_fc2_reduce=False,
combine_dtype="bf16",
input_norm_const=epilogue.input_norm_const,
fc1_alpha=None,