x

app/model_catalog.py

@@ -124,6 +124,7 @@ def resolve_model_profile(
         "kv_cache_dtype": _to_str(profile.get("kv_cache_dtype")),
         "enable_prefix_caching": _to_bool(profile.get("enable_prefix_caching"), False),
         "max_num_batched_tokens": _to_int(profile.get("max_num_batched_tokens"), 0),
+        "language_model_only": _to_bool(profile.get("language_model_only"), False),
     }
     env_vars = {str(k): str(v) for k, v in dict(profile.get("env", {})).items()}
     env_vars["HF_HUB_OFFLINE"] = "1"

app/start_openai.py

@@ -81,6 +81,8 @@ def build_command() -> list[str]:
         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 updates.get("language_model_only"):
+        cmd.append("--language-model-only")
     return cmd
 
 

config.json

@@ -38,12 +38,13 @@
         "tool_call_parser": "qwen3_coder",
         "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
+        "language_model_only": true,
         "served_model_name": "Qwen3.6-35B-A3B-FP8",
         "hf_model_id": "Qwen/Qwen3.6-35B-A3B-FP8"
       },
       "Qwen3.6-27B": {
         "local_path": "Qwen3.6-27B",
-        "ctx": "262144",
+        "ctx": "131072",
         "max_tokens": "32768",
         "max_num_batched_tokens": 16384,
         "trust_remote": true,
@@ -53,12 +54,13 @@
           2
         ],
         "max_num_seqs": 32,
-        "gpu_util": "0.90",
+        "gpu_util": "0.92",
         "kv_cache_dtype": "fp8",
         "enable_prefix_caching": true,
         "tool_call_parser": "qwen3_coder",
         "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
+        "language_model_only": true,
         "served_model_name": "Qwen3.6-27B",
         "hf_model_id": "Qwen/Qwen3.6-27B"
       },