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

[Refactor] Remove redundant TP gather/split in split_qkv in QwenVL (#28271)

Browse Source 
Signed-off-by: gcanlin <canlinguosdu@gmail.com>
This commit is contained in:
Canlin Guo
2025-11-12 23:56:47 +08:00
committed by GitHub
parent f76e85c299
commit bc5bd45c7d
2 changed files with 1 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
vllm/model_executor/models/qwen2_vl.py
View File

@@ -50,7 +50,7 @@ from vllm.attention.layer import (
)
from vllm.config import VllmConfig
from vllm.config.multimodal import BaseDummyOptions
from vllm.distributed import parallel_state, tensor_model_parallel_all_gather
from vllm.distributed import parallel_state
from vllm.distributed import utils as dist_utils
from vllm.logger import init_logger
from vllm.model_executor.layers.activation import QuickGELU
@@ -396,21 +396,10 @@ class Qwen2VisionAttention(nn.Module):
def split_qkv(self, qkv: torch.Tensor) -> tuple[torch.Tensor, ...]:
# [s, b, 3 * head * head_dim]
seq_len, bs, _ = qkv.shape
if self.tp_size > 1:
qkv = tensor_model_parallel_all_gather(qkv)
# [s, b, 3 * head * head_dim] -> 3 * [s, b, head * head_dim]
q, k, v = qkv.chunk(3, dim=2)
# 3 * [s, b, head * head_dim]
if self.tp_size > 1:
splitter = partial(
dist_utils.split_tensor_along_last_dim, num_partitions=self.tp_size
)
q = splitter(q)[self.tp_rank]
k = splitter(k)[self.tp_rank]
v = splitter(v)[self.tp_rank]
# 3 * [s, b, head * head_dim] -> 3 * [s, b, head, head_dim]
new_shape = (
seq_len,