[MoE Refactor] FusedMoE/MoERunner inversion refactor (#41184)
Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
@@ -15,7 +15,11 @@ from vllm.distributed import (
|
||||
get_tensor_model_parallel_world_size,
|
||||
)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE, GateLinear
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoE,
|
||||
GateLinear,
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import (
|
||||
MergedColumnParallelLinear,
|
||||
@@ -699,7 +703,7 @@ class DeepseekV4Model(nn.Module):
|
||||
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
return fused_moe_make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="w1",
|
||||
ckpt_down_proj_name="w2",
|
||||
|
||||
@@ -24,7 +24,7 @@ from vllm.distributed import (
|
||||
get_tensor_model_parallel_world_size,
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import fused_moe_make_expert_params_mapping
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import ReplicatedLinear
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
@@ -340,7 +340,7 @@ class DeepSeekV4MTP(nn.Module):
|
||||
head_rank_end = n_local_head * (tp_rank + 1)
|
||||
|
||||
# Pre-compute expert mapping ONCE.
|
||||
expert_mapping = FusedMoE.make_expert_params_mapping(
|
||||
expert_mapping = fused_moe_make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="w1",
|
||||
ckpt_down_proj_name="w2",
|
||||
|
||||
@@ -23,7 +23,10 @@ from vllm.model_executor.kernels.mhc.tilelang import (
|
||||
mhc_pre_tilelang,
|
||||
)
|
||||
from vllm.model_executor.layers.activation import SiluAndMul, SiluAndMulWithClamp
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoE,
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.router.base_router import (
|
||||
eplb_map_to_physical_and_record,
|
||||
)
|
||||
@@ -1151,7 +1154,7 @@ class DeepseekV4Model(nn.Module):
|
||||
return make_deepseek_v4_expert_params_mapping(self.config.n_routed_experts)
|
||||
# Params for weights, fp8 weight scales, fp8 activation scales
|
||||
# (param_name, weight_name, expert_id, shard_id)
|
||||
return FusedMoE.make_expert_params_mapping(
|
||||
return fused_moe_make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="w1",
|
||||
ckpt_down_proj_name="w2",
|
||||
|
||||
@@ -28,7 +28,9 @@ from vllm.model_executor.kernels.mhc.tilelang import (
|
||||
hc_head_fused_kernel_tilelang,
|
||||
mhc_post_tilelang,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
fused_moe_make_expert_params_mapping,
|
||||
)
|
||||
from vllm.model_executor.layers.layernorm import RMSNorm
|
||||
from vllm.model_executor.layers.linear import ReplicatedLinear
|
||||
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
@@ -337,7 +339,7 @@ class DeepSeekV4MTP(nn.Module):
|
||||
self.config.n_routed_experts
|
||||
)
|
||||
else:
|
||||
expert_mapping = FusedMoE.make_expert_params_mapping(
|
||||
expert_mapping = fused_moe_make_expert_params_mapping(
|
||||
self,
|
||||
ckpt_gate_proj_name="w1",
|
||||
ckpt_down_proj_name="w2",
|
||||
|
||||
@@ -7,8 +7,7 @@ from __future__ import annotations
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from vllm.config import get_current_vllm_config
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.fused_moe.layer import UnquantizedFusedMoEMethod
|
||||
from vllm.model_executor.layers.fused_moe import MoERunner, UnquantizedFusedMoEMethod
|
||||
from vllm.model_executor.layers.quantization import QuantizationMethods
|
||||
from vllm.model_executor.layers.quantization.fp8 import Fp8Config
|
||||
from vllm.model_executor.layers.quantization.mxfp4 import Mxfp4MoEMethod
|
||||
@@ -130,7 +129,7 @@ class DeepseekV4FP8Config(Fp8Config):
|
||||
return None
|
||||
|
||||
def get_quant_method(self, layer, prefix):
|
||||
if isinstance(layer, FusedMoE):
|
||||
if isinstance(layer, MoERunner):
|
||||
if is_layer_skipped(
|
||||
prefix=prefix,
|
||||
ignored_layers=self.ignored_layers,
|
||||
@@ -153,6 +152,6 @@ class DeepseekV4FP8Config(Fp8Config):
|
||||
return super().get_quant_method(layer, prefix)
|
||||
|
||||
def is_mxfp4_quant(self, prefix, layer):
|
||||
if not isinstance(layer, FusedMoE) or self.expert_dtype != "fp4":
|
||||
if not isinstance(layer, MoERunner) or self.expert_dtype != "fp4":
|
||||
return False
|
||||
return self.moe_quant_algo != "NVFP4"
|
||||
|
||||
Reference in New Issue
Block a user