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

fix(runner): prevent eager token buffer under-allocation (#28894)

Browse Source
This commit is contained in:
jiayisunx
2026-06-25 14:37:19 +08:00
committed by GitHub
parent 9215da2515
commit d717e926c1
1 changed files with 1 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/runner/eager_runner.py
+1 -5
View File
@@ -100,11 +100,7 @@ class EagerRunner(BaseRunner):
max_bs = ceil_align(max_bs, self.attn_tp_size) max_bs = ceil_align(max_bs, self.attn_tp_size)
max_bs = ceil_align(max_bs, get_cp_padding_align_size()) max_bs = ceil_align(max_bs, get_cp_padding_align_size())
prefill_ceiling = ( prefill_ceiling = max(mr.max_total_num_tokens, sa.max_prefill_buffer_tokens())
sa.max_prefill_buffer_tokens()
if sa.chunked_prefill_size and sa.chunked_prefill_size > 0
else mr.max_total_num_tokens
)
max_num_token = max(prefill_ceiling, max_bs * num_tokens_per_bs) max_num_token = max(prefill_ceiling, max_bs * num_tokens_per_bs)
if require_mlp_sync(sa): if require_mlp_sync(sa):
max_num_token = ceil_align(max_num_token, self.attn_tp_size) max_num_token = ceil_align(max_num_token, self.attn_tp_size)