x

app/config.py

@@ -1,10 +1,9 @@
 from functools import lru_cache
-import os
 from typing import Optional
 
 from pydantic import BaseModel
 
-from app.model_catalog import load_catalog, resolve_model_profile, resolve_runtime_settings
+from app.model_catalog import load_app_config
 
 
 class Settings(BaseModel):
@@ -35,35 +34,49 @@ class Settings(BaseModel):
     enable_auto_tool_choice: bool = False
     revision: Optional[str] = None
     api_key: Optional[str] = None
+    quantization: Optional[str] = None
+    model_impl: Optional[str] = None
+    reasoning_parser: Optional[str] = None
+    kv_cache_dtype: Optional[str] = None
+    enable_prefix_caching: bool = False
+    max_num_batched_tokens: int = 0
 
 
 @lru_cache(maxsize=1)
 def get_settings() -> Settings:
-    catalog = load_catalog("config.json")
-    runtime = resolve_runtime_settings(catalog)
-    settings = Settings(
+    config = load_app_config("config.json")
+    return Settings(
         config_file="config.json",
-        model_key=runtime["model_key"],
-        host=runtime["host"],
-        port=runtime["port"],
-        openai_host=runtime["openai_host"],
-        openai_port=runtime["openai_port"],
-        vllm_openai_internal_url=runtime["vllm_openai_internal_url"],
-        public_model_name=runtime["public_model_name"],
-        default_enable_thinking=runtime["default_enable_thinking"],
-        reasoning_enabled=runtime["reasoning_enabled"],
-        model_root=runtime["model_root"],
-        offline_mode=runtime["offline_mode"],
-        api_key=runtime["api_key"],
-        tensor_parallel_size=runtime["tensor_parallel_size"],
-        dtype=runtime["dtype"],
-        revision=runtime["revision"],
+        model_key=config.get("model_key"),
+        selected_model=config.get("selected_model"),
+        model_name=config.get("model_name", ""),
+        served_model_name=config.get("served_model_name"),
+        host=config.get("host", "0.0.0.0"),
+        port=config.get("port", 8000),
+        openai_host=config.get("openai_host", "0.0.0.0"),
+        openai_port=config.get("openai_port", 8001),
+        vllm_openai_internal_url=config.get("vllm_openai_internal_url", "http://127.0.0.1:8001/v1"),
+        public_model_name=config.get("public_model_name", "Qwen_local_model"),
+        default_enable_thinking=config.get("default_enable_thinking", False),
+        reasoning_enabled=config.get("reasoning_enabled", False),
+        model_root=config.get("model_root", "/opt/model"),
+        offline_mode=config.get("offline_mode", True),
+        max_model_len=config.get("max_model_len", 8192),
+        gpu_memory_utilization=config.get("gpu_memory_utilization", 0.92),
+        tensor_parallel_size=config.get("tensor_parallel_size", 2),
+        max_num_seqs=config.get("max_num_seqs", 64),
+        max_tokens=config.get("max_tokens", 4096),
+        dtype=config.get("dtype", "bfloat16"),
+        enforce_eager=config.get("enforce_eager", False),
+        trust_remote_code=config.get("trust_remote_code", False),
+        tool_call_parser=config.get("tool_call_parser"),
+        enable_auto_tool_choice=config.get("enable_auto_tool_choice", False),
+        revision=config.get("revision"),
+        api_key=config.get("api_key"),
+        quantization=config.get("quantization"),
+        model_impl=config.get("model_impl"),
+        reasoning_parser=config.get("reasoning_parser"),
+        kv_cache_dtype=config.get("kv_cache_dtype"),
+        enable_prefix_caching=config.get("enable_prefix_caching", False),
+        max_num_batched_tokens=config.get("max_num_batched_tokens", 0),
     )
-    _, updates, env_vars = resolve_model_profile(
-        content=catalog,
-        requested_model=settings.model_key,
-        requested_tp=settings.tensor_parallel_size,
-    )
-    for key, value in env_vars.items():
-        os.environ[key] = value
-    return settings.model_copy(update=updates | runtime)

app/model_catalog.py

@@ -1,4 +1,5 @@
 import json
+import os
 from pathlib import Path
 from typing import Any
 
@@ -151,3 +152,25 @@ def resolve_model_profile(
     if "VLLM_WORKER_MULTIPROC_METHOD" not in env_vars:
         env_vars["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
     return model_key, updates, env_vars
+
+
+def load_app_config(config_file: str = "config.json") -> dict[str, Any]:
+    """
+    统一的应用配置加载函数，封装完整的配置加载流程。
+    
+    Args:
+        config_file: 配置文件路径，默认为 "config.json"
+    
+    Returns:
+        包含合并后配置的字典，包括 runtime settings 和 model profile updates
+    """
+    catalog = load_catalog(config_file)
+    runtime = resolve_runtime_settings(catalog)
+    _, updates, env_vars = resolve_model_profile(
+        content=catalog,
+        requested_model=runtime["model_key"],
+        requested_tp=runtime["tensor_parallel_size"],
+    )
+    for key, value in env_vars.items():
+        os.environ[key] = value
+    return {**runtime, **updates}

app/start_openai.py

@@ -1,33 +1,23 @@
 import json
-import os
 import subprocess
 import sys
 
-from app.model_catalog import load_catalog, resolve_model_profile, resolve_runtime_settings
+from app.model_catalog import load_app_config
 
 
 def build_command() -> list[str]:
-    config_file = "config.json"
-    catalog = load_catalog(config_file)
-    runtime = resolve_runtime_settings(catalog)
-    _, updates, env_vars = resolve_model_profile(
-        content=catalog,
-        requested_model=runtime["model_key"],
-        requested_tp=runtime["tensor_parallel_size"],
-    )
-    for key, value in env_vars.items():
-        os.environ[key] = value
-    host = str(runtime["openai_host"])
-    port = str(runtime["openai_port"])
-    public_model_name = str(runtime["public_model_name"]).strip()
-    default_enable_thinking = bool(runtime["default_enable_thinking"])
-    reasoning_enabled = bool(runtime["reasoning_enabled"])
-    api_key = runtime["api_key"] or ""
-    dtype = str(updates["dtype"] or runtime["dtype"])
-    quantization = str(updates["quantization"] or "").strip()
-    model_impl = str(updates["model_impl"] or "").strip()
-    reasoning_parser = str(updates["reasoning_parser"] or "").strip()
-    revision = runtime["revision"] or ""
+    config = load_app_config("config.json")
+    host = str(config["openai_host"])
+    port = str(config["openai_port"])
+    public_model_name = str(config["public_model_name"]).strip()
+    default_enable_thinking = bool(config["default_enable_thinking"])
+    reasoning_enabled = bool(config["reasoning_enabled"])
+    api_key = config.get("api_key") or ""
+    dtype = str(config.get("dtype", "bfloat16"))
+    quantization = str(config.get("quantization", "")).strip()
+    model_impl = str(config.get("model_impl", "")).strip()
+    reasoning_parser = str(config.get("reasoning_parser", "")).strip()
+    revision = config.get("revision", "") or ""
     cmd = [
         sys.executable,
         "-m",
@@ -37,28 +27,28 @@ def build_command() -> list[str]:
         "--port",
         port,
         "--model",
-        str(updates["model_name"]),
+        str(config["model_name"]),
         "--served-model-name",
-        public_model_name or str(updates["served_model_name"]),
+        public_model_name or str(config["served_model_name"]),
         "--tensor-parallel-size",
-        str(updates["tensor_parallel_size"]),
+        str(config["tensor_parallel_size"]),
         "--max-model-len",
-        str(updates["max_model_len"]),
+        str(config["max_model_len"]),
         "--gpu-memory-utilization",
-        str(updates["gpu_memory_utilization"]),
+        str(config["gpu_memory_utilization"]),
         "--max-num-seqs",
-        str(updates["max_num_seqs"]),
+        str(config["max_num_seqs"]),
         "--dtype",
         dtype,
     ]
-    if updates["trust_remote_code"]:
+    if config["trust_remote_code"]:
         cmd.append("--trust-remote-code")
-    if updates["enforce_eager"]:
+    if config["enforce_eager"]:
         cmd.append("--enforce-eager")
-    if updates["enable_auto_tool_choice"]:
+    if config["enable_auto_tool_choice"]:
         cmd.append("--enable-auto-tool-choice")
-    if updates["tool_call_parser"]:
-        cmd.extend(["--tool-call-parser", str(updates["tool_call_parser"])])
+    if config["tool_call_parser"]:
+        cmd.extend(["--tool-call-parser", str(config["tool_call_parser"])])
     cmd.extend(
         [
             "--default-chat-template-kwargs",
@@ -75,12 +65,12 @@ def build_command() -> list[str]:
         cmd.extend(["--revision", revision])
     if api_key:
         cmd.extend(["--api-key", api_key])
-    if updates.get("kv_cache_dtype"):
-        cmd.extend(["--kv-cache-dtype", str(updates["kv_cache_dtype"])])
-    if updates.get("enable_prefix_caching"):
+    if config.get("kv_cache_dtype"):
+        cmd.extend(["--kv-cache-dtype", str(config["kv_cache_dtype"])])
+    if config.get("enable_prefix_caching"):
         cmd.append("--enable-prefix-caching")
-    if updates.get("max_num_batched_tokens", 0) > 0:
-        cmd.extend(["--max-num-batched-tokens", str(updates["max_num_batched_tokens"])])
+    if config.get("max_num_batched_tokens", 0) > 0:
+        cmd.extend(["--max-num-batched-tokens", str(config["max_num_batched_tokens"])])
     if updates.get("language_model_only"):
         cmd.append("--language-model-only")
     speculative_method = str(updates.get("speculative_method") or "").strip()

config.json

@@ -69,7 +69,7 @@
         "local_path": "Qwen3.6-27B-FP8",
         "dtype": "auto",
         "quantization": "fp8",
-        "ctx": "65536",
+        "ctx": "131072",
         "max_tokens": "32768",
         "max_num_batched_tokens": 16384,
         "trust_remote": true,
@@ -79,7 +79,7 @@
           2
         ],
         "max_num_seqs": 32,
-        "gpu_util": "0.88",
+        "gpu_util": "0.75",
         "tool_call_parser": "qwen3_coder",
         "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,