From 574572b21bf7a26c66ca063cd52a08c225e6e964 Mon Sep 17 00:00:00 2001 From: khalilzhk Date: Thu, 19 Mar 2026 14:48:46 +0800 Subject: [PATCH] [BugFix] bug fix for DeepSeek eagle3 in Attn-DP mode (#20492) --- python/sglang/srt/models/deepseek_v2.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py index d6df9b6f9..0d88f541a 100644 --- a/python/sglang/srt/models/deepseek_v2.py +++ b/python/sglang/srt/models/deepseek_v2.py @@ -44,7 +44,6 @@ from sglang.srt.distributed import ( get_moe_expert_parallel_world_size, get_pp_group, get_tensor_model_parallel_world_size, - tensor_model_parallel_all_gather, tensor_model_parallel_all_reduce, ) from sglang.srt.environ import envs @@ -74,6 +73,7 @@ from sglang.srt.layers.communicator_nsa_cp import NSACPLayerCommunicator from sglang.srt.layers.dp_attention import ( get_attention_cp_rank, get_attention_cp_size, + get_attention_tp_group, get_attention_tp_rank, get_attention_tp_size, is_dp_attention_enabled, @@ -1998,7 +1998,7 @@ class DeepseekV2Model(nn.Module): with ctx: if i in self.layers_to_capture: if self.enable_a2a_moe and i > self.first_k_dense_replace: - aux_hidden_state = tensor_model_parallel_all_gather( + aux_hidden_state = get_attention_tp_group().all_gather( hidden_states + residual, dim=0 ) aux_hidden_states.append(aux_hidden_state)