diff --git a/python/sglang/srt/layers/moe/token_dispatcher/deepep.py b/python/sglang/srt/layers/moe/token_dispatcher/deepep.py index 5e980f472ab..4d5302ac0a3 100644 --- a/python/sglang/srt/layers/moe/token_dispatcher/deepep.py +++ b/python/sglang/srt/layers/moe/token_dispatcher/deepep.py @@ -291,9 +291,11 @@ def __init__( self.num_max_dispatch_tokens_per_rank = get_int_env_var( "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK", 128 ) - # DeepEP internode_ll dispatch uses FINISHED_SUM_TAG=1024 + # DeepEP internode_ll dispatch uses FINISHED_SUM_TAG=2048 # and the logic requires num-tokens-sent-from-one-rank-to-another-rank less than it - assert self.num_max_dispatch_tokens_per_rank <= 1024 + # related: https://github.com/deepseek-ai/DeepEP/pull/440 + DEEPEP_FINISHED_SUM_TAG = 2048 + assert self.num_max_dispatch_tokens_per_rank <= DEEPEP_FINISHED_SUM_TAG self.handle = None