biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

Modularize fused experts and integrate PPLX kernels (#15956)

Browse Source
This commit is contained in:
bnellnm
2025-05-14 16:11:54 -04:00
committed by GitHub
parent 418d2f8bfb
commit f9c069c85e
42 changed files with 3830 additions and 660 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/model_executor/models/qwen2_moe.py
View File

@@ -33,9 +33,7 @@ from transformers import PretrainedConfig
from vllm.attention import Attention
from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import (get_pp_group,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_reduce)
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import FusedMoE
@@ -129,7 +127,8 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
intermediate_size=config.shared_expert_intermediate_size,
hidden_act=config.hidden_act,
quant_config=quant_config,
reduce_results=False,
reduce_results=self.experts.must_reduce_shared_expert_outputs(
),
)
else:
self.shared_expert = None
@@ -156,7 +155,7 @@ class Qwen2MoeSparseMoeBlock(nn.Module):
if shared_output is not None:
final_hidden_states = final_hidden_states + shared_output
if self.tp_size > 1:
final_hidden_states = tensor_model_parallel_all_reduce(
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel( # noqa E501
final_hidden_states)
return final_hidden_states.view(orig_shape)