EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Bug fixed] fixed the crash when enable the dp-attention on the single card (#3958)

Browse Source
This commit is contained in:
DavidChan
2025-03-10 15:50:34 +08:00
committed by GitHub
parent c553e1604c
commit 4455b26e76
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/models/deepseek_v2.py
View File

@@ -848,12 +848,12 @@ class DeepseekV2AttentionMLA(nn.Module):
def all_gather( def all_gather(
input_tensor: torch.Tensor, forward_batch: ForwardBatch, rank, world_size, group input_tensor: torch.Tensor, forward_batch: ForwardBatch, rank, world_size, group
): ):
if world_size == 1:
return input_tensor
all_lens = forward_batch.global_num_tokens_cpu all_lens = forward_batch.global_num_tokens_cpu
max_len = max(forward_batch.global_num_tokens_cpu) max_len = max(forward_batch.global_num_tokens_cpu)
if world_size == 1:
return input_tensor, 0, all_lens[0]
padded_tensor = torch.nn.functional.pad( padded_tensor = torch.nn.functional.pad(
input_tensor, (0, 0, 0, max_len - input_tensor.shape[0]) input_tensor, (0, 0, 0, max_len - input_tensor.shape[0])
) )