From 7be582697b27277e2756a3878f563fa9dfea30aa Mon Sep 17 00:00:00 2001 From: Jee Jee Li Date: Tue, 30 Jun 2026 00:44:05 +0800 Subject: [PATCH] [Bugfix] Fix DeepseekV2Model hidden_size (#46986) Signed-off-by: Jee Jee Li --- vllm/model_executor/models/deepseek_v2.py | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/vllm/model_executor/models/deepseek_v2.py b/vllm/model_executor/models/deepseek_v2.py index 144ff3971a..211c88129c 100644 --- a/vllm/model_executor/models/deepseek_v2.py +++ b/vllm/model_executor/models/deepseek_v2.py @@ -1336,7 +1336,7 @@ class DeepseekV2Model(nn.Module): quant_config = vllm_config.quant_config self.config = config self.device = current_platform.device_type - + self.hidden_size = config.hidden_size self.vocab_size = config.vocab_size self.is_v32 = hasattr(config, "index_topk") if self.is_v32: @@ -1353,7 +1353,7 @@ class DeepseekV2Model(nn.Module): if get_pp_group().is_first_rank: self.embed_tokens = VocabParallelEmbedding( config.vocab_size, - config.hidden_size, + self.hidden_size, quant_config=quant_config, prefix=f"{prefix}.embed_tokens", ) @@ -1370,11 +1370,11 @@ class DeepseekV2Model(nn.Module): ) if get_pp_group().is_last_rank: - self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.norm = RMSNorm(self.hidden_size, eps=config.rms_norm_eps) else: self.norm = PPMissingLayer() self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory( - ["hidden_states", "residual"], config.hidden_size + ["hidden_states", "residual"], self.hidden_size ) self.aux_hidden_state_layers = tuple[int, ...]()