[Bugfix] Remove bad startup assertion (#50869)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Benjamin Chislett committed
Aug 3, 2026 at 10:21 UTC
68ca6fd02c6cd705c4fa6872317611c83b26b4c6
1 file changed
-25
vllm/config/speculative.py
-25
@@ -1032,31 +1032,6 @@ class SpeculativeConfig:
1032
"Inkling MTP currently supports exactly one speculative token"
1033
)
1034
1035
- if self.method == "dspark":
1036
- # DSpark is a semi-autoregressive *block* drafter. A
1037
- # speculative length smaller than the checkpoint's block
1038
- # feeds the block / Markov-head machinery an unsupported
1039
- # layout and yields incorrect (garbled) output rather than
1040
- # merely lower acceptance. Require num_speculative_tokens to
1041
- # be at least the block size (e.g. 5 or 7 for DeepSeek-V4).
1042
- dspark_block_size = getattr(
1043
- self.draft_model_config.hf_config,
1044
- "dspark_block_size",
1045
- None,
1046
- )
1047
- if (
1048
- dspark_block_size is not None
1049
- and self.num_speculative_tokens < dspark_block_size
1050
- ):
1051
- raise ValueError(
1052
- "DSpark requires num_speculative_tokens >= "
1053
- f"dspark_block_size ({dspark_block_size}); got "
1054
- f"{self.num_speculative_tokens}. Smaller values "
1055
- "produce incorrect output. Use "
1056
- f"num_speculative_tokens={dspark_block_size} or "
1057
- "larger (e.g. 7)."
1058
- )
1059
-
1035
self.draft_tensor_parallel_size = (
1036
SpeculativeConfig._verify_and_get_draft_tp(
1037
self.target_parallel_config,