refactor(serve): restore direct sampling validation in serving checks

Validate explicit request sampling fields only, leaving default merge
to build_generation_config instead of duplicating it in check_request.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: lvhan028

lmdeploy/serve/openai/serving_chat_completion.py

@@ -1,23 +1,12 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from typing import TYPE_CHECKING
 
-from lmdeploy.serve.core.generation_config import build_generation_config
-
 from .protocol import ChatCompletionRequest
 
 if TYPE_CHECKING:
     from .api_server import VariableInterface
 
 
-def _effective_sampling(request: ChatCompletionRequest, server_context: 'VariableInterface') -> dict:
-    gen = build_generation_config(request, server_context.default_gen_config)
-    return {
-        'temperature': gen.temperature,
-        'top_p': gen.top_p,
-        'top_k': gen.top_k,
-    }
-
-
 def check_request(request: ChatCompletionRequest, server_context: 'VariableInterface') -> str:
     engine_config = server_context.get_engine_config()
     session_manager = server_context.get_session_manager()
@@ -43,17 +32,15 @@ def check_request(request: ChatCompletionRequest, server_context: 'VariableInter
     if session_manager.has(request.session_id):
         return f'The session_id {request.session_id!r} is occupied.'
 
-    sampling = _effective_sampling(request, server_context)
-
     # check sampling settings
     if request.n <= 0:
         return f'The n {request.n!r} must be a positive int.'
-    if not (0 < sampling['top_p'] <= 1):
-        return f'The top_p {sampling["top_p"]!r} must be in (0, 1].'
-    if sampling['top_k'] < 0:
-        return f'The top_k {sampling["top_k"]!r} cannot be a negative integer.'
-    if not (0 <= sampling['temperature'] <= 2):
-        return f'The temperature {sampling["temperature"]!r} must be in [0, 2]'
+    if request.top_p is not None and not (0 < request.top_p <= 1):
+        return f'The top_p {request.top_p!r} must be in (0, 1].'
+    if request.top_k is not None and request.top_k < 0:
+        return f'The top_k {request.top_k!r} cannot be a negative integer.'
+    if request.temperature is not None and not (0 <= request.temperature <= 2):
+        return f'The temperature {request.temperature!r} must be in [0, 2]'
 
     # Validate input_ids and image_data constraints.
     # messages has higher priority. input_ids and image_data are only used when

lmdeploy/serve/openai/serving_completion.py

@@ -1,23 +1,12 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 from typing import TYPE_CHECKING
 
-from lmdeploy.serve.core.generation_config import build_generation_config
-
 from .protocol import CompletionRequest
 
 if TYPE_CHECKING:
     from .api_server import VariableInterface
 
 
-def _effective_sampling(request: CompletionRequest, server_context: 'VariableInterface') -> dict:
-    gen = build_generation_config(request, server_context.default_gen_config)
-    return {
-        'temperature': gen.temperature,
-        'top_p': gen.top_p,
-        'top_k': gen.top_k,
-    }
-
-
 def check_request(request: CompletionRequest, server_context: 'VariableInterface') -> str:
     engine_config = server_context.get_engine_config()
     session_manager = server_context.get_session_manager()
@@ -35,16 +24,14 @@ def check_request(request: CompletionRequest, server_context: 'VariableInterface
     if session_manager.has(request.session_id):
         return f'The session_id {request.session_id!r} is occupied.'
 
-    sampling = _effective_sampling(request, server_context)
-
     # check sampling settings
     if request.n <= 0:
         return f'The n {request.n!r} must be a positive int.'
-    if not (0 < sampling['top_p'] <= 1):
-        return f'The top_p {sampling["top_p"]!r} must be in (0, 1].'
-    if sampling['top_k'] < 0:
-        return f'The top_k {sampling["top_k"]!r} cannot be a negative integer.'
-    if not (0 <= sampling['temperature'] <= 2):
-        return f'The temperature {sampling["temperature"]!r} must be in [0, 2]'
+    if request.top_p is not None and not (0 < request.top_p <= 1):
+        return f'The top_p {request.top_p!r} must be in (0, 1].'
+    if request.top_k is not None and request.top_k < 0:
+        return f'The top_k {request.top_k!r} cannot be a negative integer.'
+    if request.temperature is not None and not (0 <= request.temperature <= 2):
+        return f'The temperature {request.temperature!r} must be in [0, 2]'
 
     return ''