x

app/model_catalog.py

@@ -68,6 +68,18 @@ def resolve_runtime_settings(content: dict[str, Any]) -> dict[str, Any]:
     internal_url = _to_str(content.get("vllm_openai_internal_url"))
     if not internal_url:
         internal_url = f"http://127.0.0.1:{openai_port}/v1"
+    enable_thinking_env = os.getenv("VLLM_ENABLE_THINKING")
+    if enable_thinking_env is not None:
+        default_enable_thinking = _to_bool(enable_thinking_env, False)
+    else:
+        default_enable_thinking = _to_bool(content.get("default_enable_thinking"), False)
+
+    reasoning_enabled_env = os.getenv("VLLM_REASONING_ENABLED")
+    if reasoning_enabled_env is not None:
+        reasoning_enabled = _to_bool(reasoning_enabled_env, False)
+    else:
+        reasoning_enabled = _to_bool(content.get("reasoning_enabled"), False)
+
     return {
         "host": str(api_service.get("host", "0.0.0.0")),
         "port": _to_int(api_service.get("port"), 8000),
@@ -75,9 +87,9 @@ def resolve_runtime_settings(content: dict[str, Any]) -> dict[str, Any]:
         "openai_port": openai_port,
         "vllm_openai_internal_url": internal_url.rstrip("/"),
         "public_model_name": _to_str(content.get("public_model_name"), "Qwen_local_model"),
-        "default_enable_thinking": _to_bool(content.get("default_enable_thinking"), False),
+        "default_enable_thinking": default_enable_thinking,
         "api_key": str(content.get("api_key", "")).strip() or None,
-        "reasoning_enabled": _to_bool(content.get("reasoning_enabled"), False),
+        "reasoning_enabled": reasoning_enabled,
         "tensor_parallel_size": _to_int(content.get("tensor_parallel_size"), 2),
         "dtype": str(content.get("dtype", "bfloat16")),
         "revision": str(content.get("revision", "")).strip() or None,
@@ -125,7 +137,7 @@ def resolve_model_profile(
         "dtype": _to_str(profile.get("dtype")),
         "quantization": _to_str(profile.get("quantization")),
         "model_impl": _to_str(profile.get("model_impl")),
-        "reasoning_parser": _to_str(profile.get("reasoning_parser")),
+        "reasoning_parser": _to_str(os.getenv("VLLM_REASONING_PARSER") or profile.get("reasoning_parser")),
         "max_model_len": _to_int(profile.get("ctx"), 8192),
         "max_num_seqs": _to_int(profile.get("max_num_seqs"), 64),
         "max_tokens": _to_int(profile.get("max_tokens"), 4096),

config.json

@@ -75,11 +75,10 @@
         "trust_remote": true,
         "enforce_eager": false,
         "valid_tp": [
-          1,
           2
         ],
         "max_num_seqs": 32,
-        "gpu_util": "0.95",
+        "gpu_util": "0.85",
         "tool_call_parser": "qwen3_coder",
         "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
@@ -92,7 +91,7 @@
         "speculative": {
           "method": "dflash",
           "model": "Qwen3.6-27B-DFlash",
-          "num_speculative_tokens": 8,
+          "num_speculative_tokens": 4,
           "draft_tensor_parallel_size": 1
         }
       }