[Speculative Decoding] Add DFlash speculators config parsing (#38300)

Signed-off-by: Zhanqiu Hu <zhu@redhat.com>
This commit is contained in:
zhanqiuhu
2026-04-15 16:22:15 -04:00
committed by GitHub
parent 41488f2acd
commit 0b790a2501
4 changed files with 223 additions and 1 deletions
+8 -1
View File
@@ -523,7 +523,14 @@ class DFlashQwen3ForCausalLM(Qwen3ForCausalLM):
self.logits_processor = LogitsProcessor(
self.config.draft_vocab_size, scale=logit_scale
)
self.draft_id_to_target_id = None
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,