[MoE Refactor] FusedMoE/MoERunner inversion refactor (#41184)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-06-08 10:42:58 -04:00
committed by GitHub
parent 753e9d55e6
commit dc68bd8c41
90 changed files with 2738 additions and 2031 deletions
+12 -8
View File
@@ -513,20 +513,22 @@ def test_compressed_tensors_moe_ignore_with_model(vllm_runner):
with vllm_runner(model_path, enforce_eager=True) as llm:
def check_model(model):
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.fused_moe import MoERunner
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import ( # noqa: E501
CompressedTensorsMoEMethod,
)
# Check layer 0 MoE (should be quantized)
layer_quantized = model.model.layers[0].mlp.experts
assert isinstance(layer_quantized, FusedMoE)
assert isinstance(layer_quantized.quant_method, CompressedTensorsMoEMethod)
assert isinstance(layer_quantized, MoERunner)
assert isinstance(layer_quantized._quant_method, CompressedTensorsMoEMethod)
# Check layer 10 MoE (should be unquantized + ignored)
layer_unquantized = model.model.layers[3].mlp.experts
assert isinstance(layer_unquantized, FusedMoE)
assert isinstance(layer_unquantized.quant_method, UnquantizedFusedMoEMethod)
assert isinstance(layer_unquantized, MoERunner)
assert isinstance(
layer_unquantized._quant_method, UnquantizedFusedMoEMethod
)
llm.apply_model(check_model)
@@ -658,7 +660,7 @@ def test_compressed_tensors_mxfp8_moe_setup(vllm_runner):
) as llm:
def check_model(model):
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.fused_moe import MoERunner
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe.compressed_tensors_moe_w8a8_mxfp8 import ( # noqa: E501
CompressedTensorsW8A8Mxfp8MoEMethod,
)
@@ -670,8 +672,10 @@ def test_compressed_tensors_mxfp8_moe_setup(vllm_runner):
assert isinstance(qkv.scheme, CompressedTensorsW8A8Mxfp8)
experts = layer.mlp.experts
assert isinstance(experts, FusedMoE)
assert isinstance(experts.quant_method, CompressedTensorsW8A8Mxfp8MoEMethod)
assert isinstance(experts, MoERunner)
assert isinstance(
experts._quant_method, CompressedTensorsW8A8Mxfp8MoEMethod
)
llm.apply_model(check_model)
output = llm.generate_greedy("Hello my name is", max_tokens=4)
+1
View File
@@ -440,6 +440,7 @@ def test_fp8_reloading(
hidden_size=1,
intermediate_size=1,
)
layer = layer.routed_experts
method = method_cls(config, layer)
method.create_weights(
layer=layer,
+1 -1
View File
@@ -115,7 +115,7 @@ def test_online_quantization(
# because of how we craft the test case inputs
assert isinstance(o_proj.quant_method, expected_linear_cls)
if moe is not None:
assert isinstance(moe.quant_method, expected_moe_cls)
assert isinstance(moe._quant_method, expected_moe_cls)
if current_platform.is_cuda():
assert o_proj.weight.dtype == torch.float8_e4m3fn
+2 -2
View File
@@ -146,8 +146,8 @@ def test_quark_int8_w8a8_moe(vllm_runner, tp):
layer = model.model.layers[0]
# MoE experts should use QuarkW8A8Int8MoEMethod
moe = layer.mlp.experts
assert isinstance(moe.quant_method, QuarkW8A8Int8MoEMethod), (
f"Expected QuarkW8A8Int8MoEMethod, got {type(moe.quant_method)}"
assert isinstance(moe._quant_method, QuarkW8A8Int8MoEMethod), (
f"Expected QuarkW8A8Int8MoEMethod, got {type(moe._quant_method)}"
)
# Non-MoE linear layers should use QuarkW8A8Int8
qkv_proj = layer.self_attn.qkv_proj