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 3, 17:45 UTC
[Frontend] add support for thinking_token_budget in completions (#42116)
Walter Beller-Morales committed
May 13, 2026 at 16:01 UTC
873910d608b4f6296eabd2f0c2cfc1d7d89b9aa8
1 file changed
+9
vllm/entrypoints/openai/completion/protocol.py
+9
@@ -177,6 +177,14 @@ class CompletionRequest(OpenAIBaseModel):
177
"can detect such behavior and terminate early, saving time and tokens.",
178
)
179
180
+ thinking_token_budget: int | None = Field(
181
+ default=None,
182
+ description=(
183
+ "Maximum number of tokens allowed for thinking operations "
184
+ "(reasoning models). -1 = unlimited."
185
+ ),
186
+ )
187
+
188
# --8<-- [end:completion-extra-params]
189
190
def build_tok_params(self, model_config: ModelConfig) -> TokenizeParams:
@@ -322,6 +330,7 @@ class CompletionRequest(OpenAIBaseModel):
330
extra_args=extra_args or None,
331
skip_clone=True, # Created fresh per request, safe to skip clone
332
repetition_detection=self.repetition_detection,
333
+ thinking_token_budget=self.thinking_token_budget,
334
)
335
336
@model_validator(mode="before")