[Bugfix][Spec Decode] Only create draft_id_to_target_id when draft vocab differs (#53458)

Signed-off-by: drslark <slarksblood@qq.com>
Signed-off-by: Isotr0py <2037008807@qq.com>
Co-authored-by: GLM-5.3 <glm@z.ai>
Co-authored-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
drslark
2026-09-15 14:59:02 +00:00
committed by GitHub
co-authored by GLM-5.3 Isotr0py Isotr0py
parent 1e47ec00d2
commit 9446ea1680
3 changed files with 24 additions and 12 deletions
@@ -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,
+8 -4
View File
@@ -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
+8 -4
View File
@@ -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