[Platform] Add custom default max tokens (#18557)

Signed-off-by: Gabriel Marinho <gmarinho@ibm.com>
This commit is contained in:
Gabriel Marinho
2025-07-04 10:50:17 +08:00
committed by GitHub
parent 7e1665b089
commit a4113b035c
5 changed files with 59 additions and 60 deletions
+12 -4
View File
@@ -33,6 +33,7 @@ from vllm.entrypoints.openai.serving_engine import (OpenAIServing,
is_text_tokens_prompt)
# yapf: enable
from vllm.entrypoints.openai.serving_models import OpenAIServingModels
from vllm.entrypoints.utils import get_max_tokens
from vllm.inputs.data import (EmbedsPrompt, TokensPrompt, is_embeds_prompt,
is_tokens_prompt)
from vllm.logger import init_logger
@@ -160,15 +161,22 @@ class OpenAIServingCompletion(OpenAIServing):
input_length = len(engine_prompt["prompt_token_ids"])
else:
assert_never(engine_prompt)
default_max_tokens = self.max_model_len - input_length
if self.default_sampling_params is None:
self.default_sampling_params = {}
max_tokens = get_max_tokens(
max_model_len=self.max_model_len,
request=request,
input_length=input_length,
default_sampling_params=self.default_sampling_params)
if request.use_beam_search:
sampling_params = request.to_beam_search_params(
default_max_tokens, self.default_sampling_params)
max_tokens, self.default_sampling_params)
else:
sampling_params = request.to_sampling_params(
default_max_tokens,
self.model_config.logits_processor_pattern,
max_tokens, self.model_config.logits_processor_pattern,
self.default_sampling_params)
request_id_item = f"{request_id}-{i}"