rocm_vllm_nightly/config.json

{
  "services": {
    "api": {
      "host": "0.0.0.0",
      "port": 8002
    },
    "openai": {
      "host": "0.0.0.0",
      "port": 8001
    }
  },
  "api_key": "sk-szcjw",
  "tensor_parallel_size": 2,
  "dtype": "bfloat16",
  "model_root": "/opt/model",
  "offline_mode": true,
  "revision": "",
  "models": {
    "default": "Qwen3.5-35B-A3B-GPTQ-Int4",
    "selected": "Qwen3.5-35B-A3B-GPTQ-Int4",
    "profiles": {
      "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
        "local_path": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
        "ctx": "24576",
        "trust_remote": true,
        "valid_tp": [2],
        "max_num_seqs": "8",
        "max_tokens": "16384",
        "gpu_util": "0.98",
        "enforce_eager": false,
        "env": {
          "VLLM_USE_TRITON_AWQ": "1"
        },
        "tool_call_parser": "qwen3_xml",
        "enable_auto_tool_choice": true,
        "served_model_name": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
        "hf_model_id": "cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit"
      },
      "GLM-4.7-Flash-AWQ": {
        "local_path": "GLM-4.7-Flash-AWQ",
        "ctx": "32768",
        "trust_remote": true,
        "valid_tp": [1, 2],
        "max_num_seqs": "8",
        "max_tokens": "32768",
        "gpu_util": "0.98",
        "tool_call_parser": "qwen3_xml",
        "enable_auto_tool_choice": true,
        "served_model_name": "GLM-4.7-Flash-AWQ",
        "hf_model_id": "THUDM/GLM-4.7-Flash-AWQ"
      },
      "Qwen3.5-27B-FP8": {
        "local_path": "Qwen3.5-27B-FP8",
        "dtype": "bfloat16",
        "ctx": "65536",
        "trust_remote": true,
        "valid_tp": [1, 2],
        "max_num_seqs": "4",
        "max_tokens": "8192",
        "gpu_util": "0.94",
        "tool_call_parser": "qwen3_xml",
        "enable_auto_tool_choice": true,
        "served_model_name": "Qwen3.5-27B-FP8",
        "hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
      },
      "Qwen3.5-35B-A3B-GPTQ-Int4": {
        "local_path": "Qwen3.5-35B-A3B-GPTQ-Int4",
        "dtype": "float16",
        "quantization": "gptq",
        "ctx": "65536",
        "trust_remote": true,
        "valid_tp": [1, 2],
        "max_num_seqs": "8",
        "max_tokens": "8192",
        "gpu_util": "0.92",
        "tool_call_parser": "qwen3_xml",
        "enable_auto_tool_choice": true,
        "served_model_name": "Qwen3.5-35B-A3B-GPTQ-Int4",
        "hf_model_id": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4"
      },
      "Qwen3.5-35B-A3B-FP8": {
        "local_path": "Qwen3.5-35B-A3B-FP8",
        "dtype": "bfloat16",
        "ctx": "65536",
        "trust_remote": true,
        "valid_tp": [1, 2],
        "max_num_seqs": "8",
        "max_tokens": "32768",
        "gpu_util": "0.98",
        "tool_call_parser": "qwen3_xml",
        "enable_auto_tool_choice": true,
        "served_model_name": "Qwen3.5-35B-A3B-FP8",
        "hf_model_id": "Qwen/Qwen3.5-35B-A3B-FP8"
      }
    }
  }
}