config.example.json

{
  "code": {
    "xllm": {
      "path": "code/xllm",
      "origin": {
        "url": "",
        "branch": "main",
        "commit": ""
      },
      "upstream": {
        "url": "https://github.com/jd-opensource/xllm.git",
        "branch": "main",
        "commit": ""
      }
    }
  },
  "xllm_config": {
    "model": "",
    "model_id": "",
    "max_memory_utilization": 0.8,
    "max_seqs_per_batch": 1024,
    "max_tokens_per_chunk_for_prefill": -1,
    "max_tokens_per_batch": 10240,
    "tp_size": 1,
    "draft_model": "",
    "num_speculative_tokens": 0
  },
  "xllm_config_comments": {
    "model": "Hugging Face 模型名称或模型路径; xLLM 启动时 --model 是唯一必填参数.",
    "model_id": "Hugging Face 模型名称, 非路径; 用于需要展示或区分模型 ID 的场景.",
    "max_memory_utilization": "模型推理可使用的设备显存比例, 包括模型权重和 KV Cache; 官方默认值 0.8.",
    "max_seqs_per_batch": "每个 batch 可处理的最大 sequence 数; 官方默认值 1024.",
    "max_tokens_per_chunk_for_prefill": "chunked prefill 阶段单个请求的每个 chunk 最大 token 数; -1 表示使用默认策略.",
    "max_tokens_per_batch": "每个 batch 可处理的最大 token 数; 官方默认值 10240.",
    "tp_size": "Tensor parallel 规模; 官方默认值 1.",
    "draft_model": "Speculative decoding / MTP 使用的 draft 模型路径; 为空表示不启用外部 draft 模型.",
    "num_speculative_tokens": "每轮 speculative decoding 生成的 speculative token 数; 0 表示不启用投机解码."
  },
  "tests": {
    "smoke": {
      "prompt": "Hello, introduce xLLM in one sentence.",
      "temperature": 0.0,
      "concurrency": [1]
    },
    "quick": {
      "datasets": ["ceval"],
      "num_questions": 10,
      "concurrency": [1, 4],
      "sampling": {
        "temperature": 0.0,
        "top_p": 1.0,
        "top_k": null
      }
    },
    "full": {
      "models": ["qwen35-27b", "deepseek-v3", "glm-5"],
      "datasets": ["gsm8k", "ceval"],
      "max_test_time_hours": 8
    },
    "scripts": {
      "build": "code/xllm/cibuild/build_npu.sh",
      "serve": "",
      "smoke": "",
      "perf": "skills/xllm-npu-perf-runner/scripts/eval_perf.sh",
      "accuracy": "skills/xllm-npu-accuracy-runner/scripts/eval_acc.sh"
    }
  }
}