[Spec Decode] Fix Qwen3 DSpark d2t requirement for padded-vocab drafts (#55133)

Signed-off-by: Orestis Zambounis <23146389+orestis-z@users.noreply.github.com>
Signed-off-by: Orestis Zambounis <orestis.zambounis@gmail.com>
Co-authored-by: Benjamin Chislett <chislett.ben@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
Orestis Zambounis
2026-09-14 10:07:07 -04:00
committed by GitHub
co-authored by Benjamin Chislett mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Claude Opus 5
parent 767d1c4d47
commit c676e4930b
2 changed files with 38 additions and 15 deletions
+5 -2
View File
@@ -254,7 +254,10 @@ class Qwen3DSparkForCausalLM(DFlashQwen3ForCausalLM):
self.config.draft_vocab_size, scale=logit_scale
)
self.target_vocab_size = vllm_config.model_config.get_vocab_size()
if self.config.draft_vocab_size != self.target_vocab_size:
# Only a pruned draft (subset of the target vocab) needs a d2t mapping to
# scatter ids up. A draft >= the target vocab (e.g. padded to the target's
# physical embedding size) maps identically, so no d2t is needed.
if self.config.draft_vocab_size < self.target_vocab_size:
self.draft_id_to_target_id = nn.Parameter(
torch.zeros(self.config.draft_vocab_size, dtype=torch.long),
requires_grad=False,
@@ -334,7 +337,7 @@ class Qwen3DSparkForCausalLM(DFlashQwen3ForCausalLM):
# embed_tokens / lm_head are optional; when omitted they are shared from
# the target by load_dspark_model, so skip the unloaded params here.
uses_expanded_input_vocab = self.config.vocab_size > self.target_vocab_size
uses_reduced_vocab = self.config.draft_vocab_size != self.target_vocab_size
uses_reduced_vocab = self.config.draft_vocab_size < self.target_vocab_size
if uses_expanded_input_vocab and not includes_embed_tokens:
raise ValueError(
"Qwen3 DSpark checkpoints whose input vocab_size is larger than "