A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 13:30 UTC
[Bugfix] Fix distributed bug in Qwen2.5-VL & Qwen2.5-Omni (#16907)
Yang Fan committed
Apr 21, 2025 at 18:25 UTC
26c0406555a5e782a4591953389477bcf9695d10
1 file changed
+1
-2
vllm/model_executor/models/qwen2_5_vl.py
+1
-2
@@ -198,9 +198,8 @@ class Qwen2_5_VisionMLP(nn.Module):
198
199
def all_gather_interleave(local_tensor, hidden_size: int, tp_size: int):
200
"""All-gather the input tensor interleavely across model parallel group."""
201
- import torch.distributed as dist
201
gathered_tensors = [torch.zeros_like(local_tensor) for _ in range(tp_size)]
203
- dist.all_gather(gathered_tensors, local_tensor)
202
+ parallel_state.get_tp_group().all_gather(gathered_tensors, local_tensor)
203
204
gathered_tensors_split = [
205
torch.split(tensor, hidden_size // tp_size, -1)