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

[Fix] use torch.cat instead of torch.concat to prevent entering the Autograd backends. (#4466)

Browse Source
This commit is contained in:
JieXin Liang
2025-03-16 15:02:47 +08:00
committed by GitHub
parent 81f431eded
commit 1a3fa75f2f
14 changed files with 20 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/managers/schedule_batch.py
View File

@@ -1244,14 +1244,14 @@ class ScheduleBatch:
self.encoder_lens = torch.cat([self.encoder_lens, other.encoder_lens])
self.encoder_lens_cpu.extend(other.encoder_lens_cpu)
self.req_pool_indices = torch.concat(
self.req_pool_indices = torch.cat(
[self.req_pool_indices, other.req_pool_indices]
)
self.seq_lens = torch.concat([self.seq_lens, other.seq_lens])
self.seq_lens = torch.cat([self.seq_lens, other.seq_lens])
self.out_cache_loc = None
self.seq_lens_sum += other.seq_lens_sum
if self.output_ids is not None:
self.output_ids = torch.concat([self.output_ids, other.output_ids])
self.output_ids = torch.cat([self.output_ids, other.output_ids])
if self.return_logprob and other.return_logprob:
self.top_logprobs_nums.extend(other.top_logprobs_nums)
self.token_ids_logprobs.extend(other.token_ids_logprobs)