EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[MOE Refactor] Remove QuantType in prepare_finalize.py (#6534)

Browse Source 
### What this PR does / why we need it?
To prevent confusion between different QuantType classes, we remove**
QuantType in prepare_finalize.py

- vLLM version: v0.15.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.15.0

Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
This commit is contained in:
LI SHENGYONG
2026-02-10 15:59:58 +08:00
committed by GitHub
parent 34eecacace
commit 7cf285a77a
4 changed files with 9 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/ut/ops/test_moe_comm_method.py
View File

@@ -7,7 +7,7 @@ from tests.ut.base import TestBase
from vllm_ascend.ops.fused_moe.moe_comm_method import (AllGatherCommImpl,
AlltoAllCommImpl,
MC2CommImpl)
from vllm_ascend.ops.fused_moe.prepare_finalize import QuantType
from vllm_ascend.quantization.methods.base import QuantType
from vllm_ascend.ops.fused_moe.token_dispatcher import (TokenCombineResult,
TokenDispatchResult)

21
vllm_ascend/ops/fused_moe/fused_moe.py
View File

@@ -35,7 +35,7 @@ from vllm_ascend.eplb.core.eplb_utils import init_eplb_config
from vllm_ascend.flash_common3_context import get_flash_common3_context, set_flash_common3_context
from vllm_ascend.ops.fused_moe.experts_selector import select_experts, zero_experts_compute
from vllm_ascend.ops.fused_moe.moe_comm_method import AllGatherCommImpl, FusedExpertsResult, setup_moe_comm_method
from vllm_ascend.ops.fused_moe.prepare_finalize import QuantType
from vllm_ascend.quantization.methods.base import QuantType
from vllm_ascend.utils import (
enable_sp,
maybe_trans_nz,
@@ -235,22 +235,13 @@ class AscendFusedMoE(FusedMoE):
self.quant_type = self._get_quant_type()
def _get_quant_type(self) -> QuantType:
quant_method = self.quant_method
if not hasattr(quant_method, "quant_method") or quant_method.quant_method is None:
return QuantType.NONE
quant_type = QuantType.NONE
method = getattr(self.quant_method, "quant_method", None)
method = quant_method.quant_method
if method is not None:
quant_type = getattr(method, "quant_type", QuantType.NONE)
if hasattr(method, "quant_type"):
from vllm_ascend.quantization.methods.base import QuantType as SchemeQuantType
scheme_quant_type = method.quant_type
if scheme_quant_type == SchemeQuantType.W8A8:
return QuantType.W8A8
elif scheme_quant_type == SchemeQuantType.W4A8:
return QuantType.W4A8
return QuantType.NONE
return quant_type
def update_expert_map(self, new_expert_map):
self._expert_map = new_expert_map

2
vllm_ascend/ops/fused_moe/moe_comm_method.py
View File

@@ -30,7 +30,6 @@ from vllm_ascend.ops.fused_moe.prepare_finalize import (
PrepareAndFinalizeWithAll2All,
PrepareAndFinalizeWithAllGather,
PrepareAndFinalizeWithMC2,
QuantType,
)
from vllm_ascend.ops.fused_moe.token_dispatcher import (
MoETokenDispatcher,
@@ -38,6 +37,7 @@ from vllm_ascend.ops.fused_moe.token_dispatcher import (
TokenDispatcherWithAllGather,
TokenDispatcherWithMC2,
)
from vllm_ascend.quantization.methods.base import QuantType
_MoECommMethods: dict[MoECommType | None, MoECommMethod] = {}

8
vllm_ascend/ops/fused_moe/prepare_finalize.py
View File

@@ -15,7 +15,6 @@
# This file is a part of the vllm-ascend project.
from abc import ABC, abstractmethod
from enum import Enum
import torch
import torch.distributed as dist
@@ -32,15 +31,10 @@ from vllm.model_executor.layers.fused_moe import FusedMoEConfig
from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.distributed.utils import fc3_all_gather_and_maybe_unpad_impl
from vllm_ascend.quantization.methods.base import QuantType
from vllm_ascend.utils import enable_sp, npu_stream_switch, prefill_context_parallel_enable
class QuantType(Enum):
NONE = 0
W8A8 = 1
W4A8 = 2
class PrepareAndFinalize(ABC):
"""
Abstract base class for MoE (Mixture-of-Experts) tensor preparation and finalization