[Bugfix][Model] Fix MiniMax-M3 NVFP4 inference correctness (#48929)

Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com>
Co-authored-by: Pavani Majety <pmajety@nvidia.com>
This commit is contained in:
Gabriel Wu
2026-08-05 07:25:46 -07:00
committed by GitHub
co-authored by Pavani Majety
parent cd930c8e78
commit 08b8613b7b
2 changed files with 100 additions and 31 deletions
+69
View File
@@ -20,6 +20,7 @@ from vllm.model_executor.layers.fused_moe.all2all_utils import (
from vllm.model_executor.layers.fused_moe.config import (
FusedMoEConfig,
FusedMoEParallelConfig,
FusedMoEQuantConfig,
RoutingMethodType,
)
from vllm.model_executor.layers.fused_moe.experts.flashinfer_cutlass_moe import (
@@ -52,6 +53,74 @@ MNK_FACTORS = [
]
@pytest.mark.parametrize(
"activation",
[MoEActivation.SWIGLUOAI, MoEActivation.SWIGLUOAI_UNINTERLEAVE],
)
def test_flashinfer_swigluoai_params_are_forwarded(activation, monkeypatch):
from flashinfer.fused_moe.core import ActivationType
moe_config = FusedMoEConfig(
num_experts=2,
experts_per_token=1,
hidden_dim=128,
intermediate_size=128,
num_local_experts=2,
num_logical_experts=2,
activation=activation,
device="cuda",
moe_parallel_config=FusedMoEParallelConfig.make_no_parallel(),
in_dtype=torch.bfloat16,
routing_method=RoutingMethodType.TopK,
)
quant_config = FusedMoEQuantConfig.make(
gemm1_alpha=1.702,
gemm1_beta=1.0,
gemm1_clamp_limit=7.0,
)
experts = FlashInferExperts(moe_config=moe_config, quant_config=quant_config)
call_args = {}
def fake_flashinfer_cutlass_fused_moe(**kwargs):
call_args.update(kwargs)
monkeypatch.setattr(
"vllm.model_executor.layers.fused_moe.experts."
"flashinfer_cutlass_moe.flashinfer_cutlass_fused_moe",
fake_flashinfer_cutlass_fused_moe,
)
experts.apply(
output=torch.empty((1, 128), device="cuda", dtype=torch.bfloat16),
hidden_states=torch.empty((1, 128), device="cuda", dtype=torch.bfloat16),
w1=torch.empty((2, 256, 128), device="cuda", dtype=torch.bfloat16),
w2=torch.empty((2, 128, 128), device="cuda", dtype=torch.bfloat16),
topk_weights=torch.ones((1, 1), device="cuda", dtype=torch.float32),
topk_ids=torch.zeros((1, 1), device="cuda", dtype=torch.int64),
activation=activation,
global_num_experts=2,
expert_map=None,
a1q_scale=None,
a2_scale=None,
workspace13=None,
workspace2=None,
expert_tokens_meta=None,
apply_router_weight_on_input=False,
)
assert experts._supports_activation(activation)
assert call_args["activation_type"] == ActivationType.Swiglu
for name, value in (
("swiglu_alpha", 1.702),
("swiglu_beta", 1.0),
("swiglu_limit", 7.0),
):
torch.testing.assert_close(
call_args[name],
torch.full((2,), value, device="cuda", dtype=torch.float32),
)
@pytest.mark.parametrize("m,n,k", MNK_FACTORS)
@pytest.mark.parametrize("e", [40, 64, 256])
@pytest.mark.parametrize("topk", [1, 6, 8])