A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[misc] refactor speculative config (#25657)
Signed-off-by: zxw <1020938856@qq.com>
yyzxw committed
Sep 26, 2025 at 16:22 UTC
19f76ee68e60c71d7848ec73f5aa9e0d2ab9dbc4
1 file changed
+6
-8
vllm/config/speculative.py
+6
-8
@@ -209,12 +209,9 @@ class SpeculativeConfig:
209
210
if self.model is None and self.num_speculative_tokens is not None:
211
# TODO(Shangming): Refactor mtp configuration logic when supporting
212
- # mtp acceleration for more models besides deepseek_v3
213
- if self.target_model_config and \
214
- (self.target_model_config.hf_text_config.model_type \
215
- == "deepseek_v3" or
216
- self.target_model_config.hf_text_config.model_type in
217
- ("mimo","ernie4_5_moe", "qwen3_next")):
212
+ if (self.target_model_config
213
+ and self.target_model_config.hf_text_config.model_type
214
+ in ("deepseek_v3", "mimo", "ernie4_5_moe", "qwen3_next")):
215
# use the draft model from the same model:
216
self.model = self.target_model_config.model
217
# Align the quantization of draft model for cases such as
@@ -224,8 +221,9 @@ class SpeculativeConfig:
221
elif self.method in ("ngram", "[ngram]"):
222
self.model = "ngram"
223
else:
227
- raise ValueError("num_speculative_tokens was provided without "
228
- "speculative model.")
224
+ raise ValueError(
225
+ "num_speculative_tokens was provided but without "
226
+ "speculative model.")
227
228
# Automatically configure the method for ngram when "model" is used
229
# instead of "method"