chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

DP Attention with Auto DeepEP Dispatch (#7222)

Browse Source
This commit is contained in:
Cheng Wan
2025-07-05 01:54:24 -07:00
committed by GitHub
parent 75354d9ae9
commit 8fc910db03
13 changed files with 136 additions and 90 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/model_executor/forward_batch_info.py
View File

@@ -254,6 +254,7 @@ class ForwardBatch:
dp_local_start_pos: Optional[torch.Tensor] = None # cached info at runtime
dp_local_num_tokens: Optional[torch.Tensor] = None # cached info at runtime
gathered_buffer: Optional[torch.Tensor] = None
is_extend_in_batch: bool = False
can_run_dp_cuda_graph: bool = False
global_forward_mode: Optional[ForwardMode] = None
@@ -299,6 +300,7 @@ class ForwardBatch:
return_logprob=batch.return_logprob,
top_logprobs_nums=batch.top_logprobs_nums,
token_ids_logprobs=batch.token_ids_logprobs,
is_extend_in_batch=batch.is_extend_in_batch,
can_run_dp_cuda_graph=batch.can_run_dp_cuda_graph,
global_forward_mode=batch.global_forward_mode,
lora_paths=batch.lora_paths,