chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Feature] Comprehensive Hybrid Parallelism Support (#6389)

Browse Source
This commit is contained in:
Cheng Wan
2025-06-20 14:43:11 -07:00
committed by GitHub
parent 0998808009
commit e879d8b7a8
14 changed files with 3689 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/logits_processor.py
View File

@@ -30,9 +30,9 @@ from sglang.srt.layers.dp_attention import (
attn_tp_all_gather,
dp_gather_replicate,
dp_scatter,
get_attention_dp_rank,
get_attention_dp_size,
get_attention_tp_size,
get_local_attention_dp_rank,
get_local_attention_dp_size,
)
from sglang.srt.layers.vocab_parallel_embedding import VocabParallelEmbedding
@@ -171,7 +171,7 @@ class LogitsMetadata:
return
cumtokens = torch.cumsum(self.global_num_tokens_for_logprob_gpu, dim=0)
dp_rank = get_local_attention_dp_rank()
dp_rank = get_attention_dp_rank()
if dp_rank == 0:
dp_local_start_pos = torch.zeros_like(
self.global_num_tokens_for_logprob_gpu[0]