From 9446ea168067c141992f86c3a796e8551f37139e Mon Sep 17 00:00:00 2001 From: drslark <96540755+drslark@users.noreply.github.com> Date: Tue, 15 Sep 2026 22:59:02 +0800 Subject: [PATCH] [Bugfix][Spec Decode] Only create draft_id_to_target_id when draft vocab differs (#53458) Signed-off-by: drslark Signed-off-by: Isotr0py <2037008807@qq.com> Co-authored-by: GLM-5.3 Co-authored-by: Isotr0py <2037008807@qq.com> Co-authored-by: Isotr0py --- vllm/model_executor/models/deepseek_eagle3.py | 12 ++++++++---- vllm/model_executor/models/llama_eagle3.py | 12 ++++++++---- vllm/model_executor/models/qwen3_eagle3.py | 12 ++++++++---- 3 files changed, 24 insertions(+), 12 deletions(-) diff --git a/vllm/model_executor/models/deepseek_eagle3.py b/vllm/model_executor/models/deepseek_eagle3.py index 80c41501e6..48d019cd35 100644 --- a/vllm/model_executor/models/deepseek_eagle3.py +++ b/vllm/model_executor/models/deepseek_eagle3.py @@ -305,10 +305,14 @@ class Eagle3DeepseekV2ForCausalLM(LocalArgmaxMixin, DeepseekV2ForCausalLM): self.logits_processor = LogitsProcessor( self.config.draft_vocab_size, scale=logit_scale ) - self.draft_id_to_target_id = nn.Parameter( - torch.zeros(self.config.draft_vocab_size, dtype=torch.long), - requires_grad=False, - ) + target_vocab_size = vllm_config.model_config.get_vocab_size() + if self.config.draft_vocab_size != target_vocab_size: + self.draft_id_to_target_id = nn.Parameter( + torch.zeros(self.config.draft_vocab_size, dtype=torch.long), + requires_grad=False, + ) + else: + self.draft_id_to_target_id = None def embed_input_ids( self, diff --git a/vllm/model_executor/models/llama_eagle3.py b/vllm/model_executor/models/llama_eagle3.py index c3d956a946..68e21c9dc2 100644 --- a/vllm/model_executor/models/llama_eagle3.py +++ b/vllm/model_executor/models/llama_eagle3.py @@ -335,10 +335,14 @@ class Eagle3LlamaForCausalLM(_Eagle3LlamaForCausalLMBase): self.logits_processor = LogitsProcessor( self.config.draft_vocab_size, scale=logit_scale ) - self.draft_id_to_target_id = nn.Parameter( - torch.zeros(self.config.draft_vocab_size, dtype=torch.long), - requires_grad=False, - ) + target_vocab_size = vllm_config.model_config.get_vocab_size() + if self.config.draft_vocab_size != target_vocab_size: + self.draft_id_to_target_id = nn.Parameter( + torch.zeros(self.config.draft_vocab_size, dtype=torch.long), + requires_grad=False, + ) + else: + self.draft_id_to_target_id = None self.use_parallel_drafting = speculative_config.parallel_drafting diff --git a/vllm/model_executor/models/qwen3_eagle3.py b/vllm/model_executor/models/qwen3_eagle3.py index 39d0173b3c..745c77a977 100644 --- a/vllm/model_executor/models/qwen3_eagle3.py +++ b/vllm/model_executor/models/qwen3_eagle3.py @@ -309,10 +309,14 @@ class Eagle3Qwen3ForCausalLM(Qwen3ForCausalLM): self.logits_processor = LogitsProcessor( self.config.draft_vocab_size, scale=logit_scale ) - self.draft_id_to_target_id = nn.Parameter( - torch.zeros(self.config.draft_vocab_size, dtype=torch.long), - requires_grad=False, - ) + target_vocab_size = vllm_config.model_config.get_vocab_size() + if self.config.draft_vocab_size != target_vocab_size: + self.draft_id_to_target_id = nn.Parameter( + torch.zeros(self.config.draft_vocab_size, dtype=torch.long), + requires_grad=False, + ) + else: + self.draft_id_to_target_id = None self.use_parallel_drafting = vllm_config.speculative_config.parallel_drafting