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

[Core] Allow disabling TP sharding for parallel Linear layer (#23024)

Browse Source 
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: Isotr0py <2037008807@qq.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
Isotr0py
2025-09-06 13:53:58 +08:00
committed by GitHub
parent 6432739ef1
commit 53b19ccdd5
7 changed files with 203 additions and 280 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/models/deepseek_v2.py
View File

@@ -43,7 +43,6 @@ from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
MergedColumnParallelLinear,
MergedReplicatedLinear,
ReplicatedLinear,
RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
@@ -435,12 +434,13 @@ class DeepseekV2MLAAttention(nn.Module):
self.max_position_embeddings = max_position_embeddings
if self.q_lora_rank is not None:
self.fused_qkv_a_proj = MergedReplicatedLinear(
self.fused_qkv_a_proj = MergedColumnParallelLinear(
self.hidden_size,
[self.q_lora_rank, self.kv_lora_rank + self.qk_rope_head_dim],
bias=False,
quant_config=quant_config,
prefix=f"{prefix}.fused_qkv_a_proj")
prefix=f"{prefix}.fused_qkv_a_proj",
disable_tp=True)
else:
self.kv_a_proj_with_mqa = ReplicatedLinear(
self.hidden_size,