Fix performance when --generation-config is not None (vllm-project#14223)

Signed-off-by: Harry Mellor <[email protected]>
Signed-off-by: Louis Ulmer <[email protected]>

Author: hmellor

vllm/entrypoints/llm.py

@@ -244,6 +244,7 @@ def __init__(
             engine_args, usage_context=UsageContext.LLM_CLASS)
 
         self.request_counter = Counter()
+        self.default_sampling_params: Union[dict[str, Any], None] = None
 
     @staticmethod
     def get_engine_class() -> type[LLMEngine]:
@@ -268,10 +269,11 @@ def set_tokenizer(self, tokenizer: AnyTokenizer) -> None:
             tokenizer_group.tokenizer = get_cached_tokenizer(tokenizer)
 
     def get_default_sampling_params(self) -> SamplingParams:
-        diff_sampling_param = (
-            self.llm_engine.model_config.get_diff_sampling_param())
-        if diff_sampling_param:
-            return SamplingParams.from_optional(**diff_sampling_param)
+        if self.default_sampling_params is None:
+            self.default_sampling_params = (
+                self.llm_engine.model_config.get_diff_sampling_param())
+        if self.default_sampling_params:
+            return SamplingParams.from_optional(**self.default_sampling_params)
         return SamplingParams()
 
     @overload

vllm/entrypoints/openai/serving_chat.py

@@ -105,10 +105,11 @@ def __init__(
                                 "been registered") from e
 
         self.enable_prompt_tokens_details = enable_prompt_tokens_details
-        diff_sampling_param = self.model_config.get_diff_sampling_param()
-        if diff_sampling_param:
+        self.default_sampling_params = (
+            self.model_config.get_diff_sampling_param())
+        if self.default_sampling_params:
             logger.info("Overwriting default chat sampling param with: %s",
-                        diff_sampling_param)
+                        self.default_sampling_params)
 
     async def create_chat_completion(
         self,
@@ -210,17 +211,14 @@ async def create_chat_completion(
                 sampling_params: Union[SamplingParams, BeamSearchParams]
                 default_max_tokens = self.max_model_len - len(
                     engine_prompt["prompt_token_ids"])
-                # Build default sampling params
-                default_sampling_params = (
-                    self.model_config.get_diff_sampling_param())
                 if request.use_beam_search:
                     sampling_params = request.to_beam_search_params(
-                        default_max_tokens, default_sampling_params)
+                        default_max_tokens, self.default_sampling_params)
                 else:
                     sampling_params = request.to_sampling_params(
                         default_max_tokens,
                         self.model_config.logits_processor_pattern,
-                        default_sampling_params)
+                        self.default_sampling_params)
 
                 self._log_inputs(request_id,
                                  request_prompts[i],

vllm/entrypoints/openai/serving_completion.py

@@ -51,11 +51,12 @@ def __init__(
                          models=models,
                          request_logger=request_logger,
                          return_tokens_as_token_ids=return_tokens_as_token_ids)
-        diff_sampling_param = self.model_config.get_diff_sampling_param()
-        if diff_sampling_param:
+        self.default_sampling_params = (
+            self.model_config.get_diff_sampling_param())
+        if self.default_sampling_params:
             logger.info(
                 "Overwriting default completion sampling param with: %s",
-                diff_sampling_param)
+                self.default_sampling_params)
 
     async def create_completion(
         self,
@@ -119,17 +120,14 @@ async def create_completion(
                 sampling_params: Union[SamplingParams, BeamSearchParams]
                 default_max_tokens = self.max_model_len - len(
                     engine_prompt["prompt_token_ids"])
-                # Build default sampling params
-                default_sampling_params = (
-                    self.model_config.get_diff_sampling_param())
                 if request.use_beam_search:
                     sampling_params = request.to_beam_search_params(
-                        default_max_tokens, default_sampling_params)
+                        default_max_tokens, self.default_sampling_params)
                 else:
                     sampling_params = request.to_sampling_params(
                         default_max_tokens,
                         self.model_config.logits_processor_pattern,
-                        default_sampling_params)
+                        self.default_sampling_params)
 
                 request_id_item = f"{request_id}-{i}"
 

vllm/entrypoints/openai/serving_transcription.py

@@ -161,11 +161,12 @@ def __init__(
                          request_logger=request_logger,
                          return_tokens_as_token_ids=return_tokens_as_token_ids)
 
-        diff_sampling_param = self.model_config.get_diff_sampling_param()
-        if diff_sampling_param:
+        self.default_sampling_params = (
+            self.model_config.get_diff_sampling_param())
+        if self.default_sampling_params:
             logger.info(
                 "Overwriting default completion sampling param with: %s",
-                diff_sampling_param)
+                self.default_sampling_params)
 
     async def _preprocess_transcription(
         self,
@@ -273,9 +274,8 @@ async def create_transcription(
         try:
             # TODO(rob): subtract len of tokenized prompt.
             default_max_tokens = self.model_config.max_model_len
-            default_params = self.model_config.get_diff_sampling_param()
             sampling_params = request.to_sampling_params(
-                default_max_tokens, default_params)
+                default_max_tokens, self.default_sampling_params)
 
             self._log_inputs(
                 request_id,