@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 4aae667668
Fork 0

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 15, 19:45 UTC

[core] add `extra_args` to `SamplingParams` (#13300)

Signed-off-by: Aviv Keshet <akeshet@scaledcognition.com>

Aviv Keshet committed Mar 7, 2025 at 22:41 UTC 4aae667668e59509ab152b08c5eb617b8ae1187a
1 file changed +8 -1
vllm/sampling_params.py
+8 -1
@@ -184,6 +184,9 @@ class SamplingParams(
184 allowed_token_ids: If provided, the engine will construct a logits
185 processor which only retains scores for the given token ids.
186 Defaults to None.
187 + extra_args: Arbitrary additional args, that can be used by custom
188 + sampling implementations. Not used by any in-tree sampling
189 + implementations.
190 """
191
192 n: int = 1
@@ -227,6 +230,7 @@ class SamplingParams(
230 guided_decoding: Optional[GuidedDecodingParams] = None
231 logit_bias: Optional[dict[int, float]] = None
232 allowed_token_ids: Optional[list[int]] = None
233 + extra_args: Optional[dict[str, Any]] = None
234
235 @staticmethod
236 def from_optional(
@@ -259,6 +263,7 @@ class SamplingParams(
263 guided_decoding: Optional[GuidedDecodingParams] = None,
264 logit_bias: Optional[Union[dict[int, float], dict[str, float]]] = None,
265 allowed_token_ids: Optional[list[int]] = None,
266 + extra_args: Optional[dict[str, Any]] = None,
267 ) -> "SamplingParams":
268 if logit_bias is not None:
269 # Convert token_id to integer
@@ -300,6 +305,7 @@ class SamplingParams(
305 guided_decoding=guided_decoding,
306 logit_bias=logit_bias,
307 allowed_token_ids=allowed_token_ids,
308 + extra_args=extra_args,
309 )
310
311 def __post_init__(self) -> None:
@@ -509,7 +515,8 @@ class SamplingParams(
515 "spaces_between_special_tokens="
516 f"{self.spaces_between_special_tokens}, "
517 f"truncate_prompt_tokens={self.truncate_prompt_tokens}, "
512 - f"guided_decoding={self.guided_decoding})")
518 + f"guided_decoding={self.guided_decoding}, "
519 + f"extra_args={self.extra_args})")
520
521
522 class BeamSearchParams(