Merge tag 'v0.10.1.1' into v0.10.1.1-ori

vllm/entrypoints/constants.py

+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Shared constants for vLLM entrypoints.
+"""
+
+# HTTP header limits for h11 parser
+# These constants help mitigate header abuse attacks
+H11_MAX_INCOMPLETE_EVENT_SIZE_DEFAULT = 4194304  # 4 MB
+H11_MAX_HEADER_COUNT_DEFAULT = 256

vllm/entrypoints/launcher.py

@@ -14,6 +14,8 @@ from vllm import envs
 from vllm.engine.async_llm_engine import AsyncEngineDeadError
 from vllm.engine.multiprocessing import MQEngineDeadError
 from vllm.engine.protocol import EngineClient
+from vllm.entrypoints.constants import (H11_MAX_HEADER_COUNT_DEFAULT,
+                                        H11_MAX_INCOMPLETE_EVENT_SIZE_DEFAULT)
 from vllm.entrypoints.ssl import SSLCertRefresher
 from vllm.logger import init_logger
 from vllm.utils import find_process_using_port
@@ -26,6 +28,11 @@ async def serve_http(app: FastAPI,
                      sock: Optional[socket.socket],
                      enable_ssl_refresh: bool = False,
                      **uvicorn_kwargs: Any):
+    """
+    Start a FastAPI app using Uvicorn, with support for custom Uvicorn config
+    options.  Supports http header limits via h11_max_incomplete_event_size and
+    h11_max_header_count.
+    """
     logger.info("Available routes are:")
     for route in app.routes:
         methods = getattr(route, "methods", None)
@@ -36,7 +43,21 @@ async def serve_http(app: FastAPI,
 
         logger.info("Route: %s, Methods: %s", path, ', '.join(methods))
 
+    # Extract header limit options if present
+    h11_max_incomplete_event_size = uvicorn_kwargs.pop(
+        "h11_max_incomplete_event_size", None)
+    h11_max_header_count = uvicorn_kwargs.pop("h11_max_header_count", None)
+
+    # Set safe defaults if not provided
+    if h11_max_incomplete_event_size is None:
+        h11_max_incomplete_event_size = H11_MAX_INCOMPLETE_EVENT_SIZE_DEFAULT
+    if h11_max_header_count is None:
+        h11_max_header_count = H11_MAX_HEADER_COUNT_DEFAULT
+
     config = uvicorn.Config(app, **uvicorn_kwargs)
+    # Set header limits
+    config.h11_max_incomplete_event_size = h11_max_incomplete_event_size
+    config.h11_max_header_count = h11_max_header_count
     config.load()
     server = uvicorn.Server(config)
     _add_shutdown_handlers(app, server)

vllm/entrypoints/openai/api_server.py

@@ -1894,6 +1894,8 @@ async def run_server_worker(listen_address,
             ssl_certfile=args.ssl_certfile,
             ssl_ca_certs=args.ssl_ca_certs,
             ssl_cert_reqs=args.ssl_cert_reqs,
+            h11_max_incomplete_event_size=args.h11_max_incomplete_event_size,
+            h11_max_header_count=args.h11_max_header_count,
             **uvicorn_kwargs,
         )
 

vllm/entrypoints/openai/cli_args.py

@@ -20,6 +20,8 @@ from vllm.config import config
 from vllm.engine.arg_utils import AsyncEngineArgs, optional_type
 from vllm.entrypoints.chat_utils import (ChatTemplateContentFormatOption,
                                          validate_chat_template)
+from vllm.entrypoints.constants import (H11_MAX_HEADER_COUNT_DEFAULT,
+                                        H11_MAX_INCOMPLETE_EVENT_SIZE_DEFAULT)
 from vllm.entrypoints.openai.serving_models import LoRAModulePath
 from vllm.entrypoints.openai.tool_parsers import ToolParserManager
 from vllm.logger import init_logger
@@ -172,6 +174,12 @@ schema. Example: `[{"type": "text", "text": "Hello world!"}]`"""
     enable_log_outputs: bool = False
     """If set to True, enable logging of model outputs (generations) 
     in addition to the input logging that is enabled by default."""
+    h11_max_incomplete_event_size: int = H11_MAX_INCOMPLETE_EVENT_SIZE_DEFAULT
+    """Maximum size (bytes) of an incomplete HTTP event (header or body) for
+    h11 parser. Helps mitigate header abuse. Default: 4194304 (4 MB)."""
+    h11_max_header_count: int = H11_MAX_HEADER_COUNT_DEFAULT
+    """Maximum number of HTTP headers allowed in a request for h11 parser.
+    Helps mitigate header abuse. Default: 256."""
 
     @staticmethod
     def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:

vllm/entrypoints/openai/tool_parsers/qwen3coder_tool_parser.py

@@ -208,15 +208,10 @@ class Qwen3CoderToolParser(ToolParser):
                             "valid JSON object in tool '%s', will try other "
                             "methods to parse it.", param_value, param_name,
                             func_name)
-                try:
-                    converted_value = eval(param_value)
-                    return converted_value
-                except Exception:
-                    logger.warning(
-                        "Parsed value '%s' of parameter '%s' cannot be "
-                        "converted via Python `eval()` in tool '%s', "
-                        "degenerating to string.", param_value, param_name,
-                        func_name)
+                logger.warning(
+                    "Parameter '%s' has unknown type '%s'. "
+                    "The value will be treated as a string.", param_name,
+                    param_type)
                 return param_value
 
         # Extract function name

vllm/v1/attention/backends/mla/cutlass_mla.py

@@ -21,7 +21,7 @@ logger = init_logger(__name__)
 
 class CutlassMLAMetadataBuilder(MLACommonMetadataBuilder[MLACommonMetadata]):
     # enable full CUDA Graph support for decode-only capture
-    attn_cudagraph_support: ClassVar[
+    cudagraph_support: ClassVar[
         AttentionCGSupport] = AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE