xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 00:15:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Spec Decode] Make EAGLE3 draft token ID mapping optional (#18488)

Browse Source 
Signed-off-by: Benjamin Chislett <benjamin.chislett@centml.ai>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Benjamin Chislett 2025-05-22 23:17:39 -04:00 committed by GitHub
parent e44d8ce8c7
commit 583507d130
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
vllm/model_executor/models/llama_eagle3.py
View File

@ -214,6 +214,9 @@ class Eagle3LlamaForCausalLM(LlamaForCausalLM):
) -> Optional[torch.Tensor]:
logits = self.logits_processor(self.lm_head, hidden_states,
sampling_metadata)
if self.draft_id_to_target_id is None:
return logits
base = torch.arange(self.config.draft_vocab_size, device=logits.device)
targets = base + self.draft_id_to_target_id
logits_new = logits.new_full((
@ -246,4 +249,9 @@ class Eagle3LlamaForCausalLM(LlamaForCausalLM):
name = "model." + name
model_weights[name] = loaded_weight
return loader.load_weights(model_weights.items())
loaded_weights = loader.load_weights(model_weights.items())
if 'd2t' not in loaded_weights:
self.draft_id_to_target_id = None
return loaded_weights