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

This commit is contained in:
jiayisunx
2026-06-25 14:37:19 +08:00
committed by GitHub
parent 9215da2515
commit d717e926c1
@@ -100,11 +100,7 @@ class EagerRunner(BaseRunner):
max_bs = ceil_align(max_bs, self.attn_tp_size)
max_bs = ceil_align(max_bs, get_cp_padding_align_size())
prefill_ceiling = (
sa.max_prefill_buffer_tokens()
if sa.chunked_prefill_size and sa.chunked_prefill_size > 0
else mr.max_total_num_tokens
)
prefill_ceiling = max(mr.max_total_num_tokens, sa.max_prefill_buffer_tokens())
max_num_token = max(prefill_ceiling, max_bs * num_tokens_per_bs)
if require_mlp_sync(sa):
max_num_token = ceil_align(max_num_token, self.attn_tp_size)