minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Fix] Vocab out of bounds in DSpark for Inkling-Small (#33748)

Browse Source
This commit is contained in:
Eric Zhang
2026-08-05 18:20:02 -07:00
committed by GitHub
parent fc74c35546
commit ceaeca0b9e
1 changed files with 9 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py
+9 -3
View File
@@ -124,12 +124,18 @@ class DSparkWorkerV2(BaseSpecWorker):
self.draft_model = bundle.draft_model self.draft_model = bundle.draft_model
self._draft_sampler = None self._draft_sampler = None
# The mask token needs an embedding row, not a tokenizer entry, so bound it
# by the embedding width. A padded vocab reserves rows past the real tokens
# and drafts place the mask there (Inkling: 200058 real, 201024 padded).
target_model_config = self.target_worker.model_runner.model_config
target_vocab_size = (
getattr(target_model_config.hf_text_config, "padded_vocab_size", None)
or target_model_config.vocab_size
)
runtime_config = resolve_runtime_config( runtime_config = resolve_runtime_config(
draft_hf_config=self.draft_model_runner.model_config.hf_config, draft_hf_config=self.draft_model_runner.model_config.hf_config,
speculative_num_draft_tokens=server_args.speculative_num_draft_tokens, speculative_num_draft_tokens=server_args.speculative_num_draft_tokens,
target_vocab_size=int( target_vocab_size=int(target_vocab_size),
self.target_worker.model_runner.model_config.vocab_size
),
) )
self.gamma = runtime_config.gamma self.gamma = runtime_config.gamma
self.verify_num_draft_tokens = runtime_config.verify_num_draft_tokens self.verify_num_draft_tokens = runtime_config.verify_num_draft_tokens