[Bugfix] Treat generation_config max_tokens as default not ceiling (#34063)

Signed-off-by: almogtavor <almogtavor@gmail.com>
This commit is contained in:
Almog Tavor
2026-02-16 07:58:24 -08:00
committed by GitHub
parent a3205beffb
commit 72d5951d02
7 changed files with 157 additions and 17 deletions
@@ -145,6 +145,12 @@ class OpenAIServingChat(OpenAIServing):
self.enable_prompt_tokens_details = enable_prompt_tokens_details
self.enable_force_include_usage = enable_force_include_usage
self.default_sampling_params = self.model_config.get_diff_sampling_param()
mc = self.model_config
self.override_max_tokens = (
self.default_sampling_params.get("max_tokens")
if mc.generation_config not in ("auto", "vllm")
else getattr(mc, "override_generation_config", {}).get("max_new_tokens")
)
self.use_harmony = self.model_config.hf_config.model_type == "gpt_oss"
if self.use_harmony:
if "stop_token_ids" not in self.default_sampling_params:
@@ -389,6 +395,7 @@ class OpenAIServingChat(OpenAIServing):
else request.max_tokens,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params,
self.override_max_tokens,
)
sampling_params: SamplingParams | BeamSearchParams
@@ -70,6 +70,12 @@ class OpenAIServingCompletion(OpenAIServing):
self.enable_force_include_usage = enable_force_include_usage
self.default_sampling_params = self.model_config.get_diff_sampling_param()
mc = self.model_config
self.override_max_tokens = (
self.default_sampling_params.get("max_tokens")
if mc.generation_config not in ("auto", "vllm")
else getattr(mc, "override_generation_config", {}).get("max_new_tokens")
)
async def render_completion_request(
self,
@@ -164,6 +170,7 @@ class OpenAIServingCompletion(OpenAIServing):
request.max_tokens,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params,
self.override_max_tokens,
)
sampling_params: SamplingParams | BeamSearchParams
@@ -1174,6 +1174,7 @@ class OpenAIServing:
context.request.max_output_tokens,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params, # type: ignore
self.override_max_tokens, # type: ignore
)
# OPTIMIZATION
@@ -229,6 +229,12 @@ class OpenAIServingResponses(OpenAIServing):
self.enable_force_include_usage = enable_force_include_usage
self.default_sampling_params = self.model_config.get_diff_sampling_param()
mc = self.model_config
self.override_max_tokens = (
self.default_sampling_params.get("max_tokens")
if mc.generation_config not in ("auto", "vllm")
else getattr(mc, "override_generation_config", {}).get("max_new_tokens")
)
# If False (default), the "store" option is (silently) ignored and the
# response is not stored. If True, the response is stored in memory.
@@ -446,6 +452,7 @@ class OpenAIServingResponses(OpenAIServing):
request.max_output_tokens,
self._extract_prompt_len(engine_prompt),
self.default_sampling_params,
self.override_max_tokens,
)
sampling_params = request.to_sampling_params(
+12 -6
View File
@@ -177,17 +177,23 @@ def get_max_tokens(
max_tokens: int | None,
input_length: int,
default_sampling_params: dict,
override_max_tokens: int | None = None,
) -> int:
default_max_tokens = max_model_len - input_length
max_output_tokens = current_platform.get_max_output_tokens(input_length)
model_max_tokens = max_model_len - input_length
platform_max_tokens = current_platform.get_max_output_tokens(input_length)
fallback_max_tokens = (
max_tokens
if max_tokens is not None
else default_sampling_params.get("max_tokens")
)
return min(
val
for val in (
default_max_tokens,
max_tokens,
max_output_tokens,
default_sampling_params.get("max_tokens"),
model_max_tokens,
fallback_max_tokens,
override_max_tokens,
platform_max_tokens,
)
if val is not None
)