x

.gitignore

@@ -0,0 +1,2 @@
+.idea/
+venv/

Dockerfile

@@ -1,9 +1,11 @@
-FROM docker.1ms.run/vllm/vllm-openai-rocm:latest
+FROM vllm/vllm-openai-rocm:nightly
 
 WORKDIR /workspace
 
 COPY requirements.txt /workspace/requirements.txt
-RUN pip install --no-cache-dir --retries 10 --timeout 180 -r /workspace/requirements.txt
+RUN pip install --no-cache-dir --retries 20 --timeout 600 -i https://pypi.tuna.tsinghua.edu.cn/simple -r /workspace/requirements.txt \
+ || pip install --no-cache-dir --retries 20 --timeout 600 -i https://mirrors.aliyun.com/pypi/simple -r /workspace/requirements.txt \
+ || pip install --no-cache-dir --retries 20 --timeout 600 -r /workspace/requirements.txt
 
 COPY app /workspace/app
 

README.md

@@ -15,8 +15,7 @@
 ├── app
 │   ├── config.py
 │   ├── model_catalog.py
-│   ├── start_openai.py
-│   └── schemas.py
+│   └── start_openai.py
 ├── .dockerignore
 ├── config.json
 ├── docker-compose.yml
@@ -29,6 +28,9 @@
 项目只读取一个配置文件：`config.json`。
 
 - `services.openai.host` / `services.openai.port`：OpenAI 协议服务监听地址与端口（默认 `0.0.0.0:8001`）
+- `public_model_name`：对外固定模型名，切换底层模型时可保持调用方参数不变
+- `default_enable_thinking`：服务端默认思考开关，默认 `false`（即调用方不传时也关闭）
+- `reasoning_enabled`：是否启用推理解析器参数注入，默认 `false`
 - `api_key`：OpenAI 接口访问密钥
 - `tensor_parallel_size`：张量并行数，双卡建议 `2`
 - `dtype`：推理精度，默认 `bfloat16`
@@ -43,7 +45,7 @@
 - `models.default`：默认模型名
 - `models.selected`：当前生效模型名
 - `models.profiles`：模型配置集合
-- 每个模型必须包含：`local_path`，并建议补充 `ctx`、`max_num_seqs`、`max_tokens`、`dtype`、`quantization`
+- 每个模型必须包含：`local_path`，并建议补充 `ctx`、`max_num_seqs`、`max_tokens`、`dtype`、`quantization`、`reasoning_parser`
 
 启动时会按以下优先级选模型：
 
@@ -62,34 +64,52 @@
 
 ## 部署步骤
 
-1. 修改 `config.json` 中的 `models.selected` 与服务参数。
+1. 预拉取基础镜像（与官方文档一致）：
 
-2. 构建并启动容器：
+```bash
+docker pull docker.1ms.run/vllm/vllm-openai-rocm:latest
+```
+
+2. 修改 `config.json` 中的 `models.selected` 与服务参数。
+
+3. 构建并启动容器：
 
 ```bash
 docker compose up -d --build
 ```
 
-3. 验证 OpenAI 协议服务：
+4. 验证 OpenAI 协议服务：
 
 ```bash
 curl http://localhost:<services.openai.port>/v1/models
 ```
 
+当前 `docker-compose.yml` 已按官方运行参数适配：
+
+- `--group-add=video` → `group_add: [video]`
+- `--ipc=host` → `ipc: host`
+- `--cap-add=SYS_PTRACE` → `cap_add: [SYS_PTRACE]`
+- `--security-opt seccomp=unconfined` → `security_opt: [seccomp=unconfined]`
+- `--device /dev/kfd` 与 `--device /dev/dri` → `devices`
+- `-e HF_HOME=/app/models` 在本项目等效为 `HF_HOME=/opt/model`
+
 ## OpenAI 协议示例（8001）
 
 ```bash
 curl -X POST "http://localhost:8001/v1/chat/completions" \
   -H "Content-Type: application/json" \
   -H "Authorization: Bearer <config.json中的api_key>" \
-  -d "{\"model\":\"Qwen3.5-35B-A3B-GPTQ-Int4\",\"messages\":[{\"role\":\"user\",\"content\":\"你好，介绍一下你自己\"}],\"temperature\":0.7}"
+  -d "{\"model\":\"Qwen_local_model\",\"messages\":[{\"role\":\"user\",\"content\":\"你好，介绍一下你自己\"}],\"temperature\":0.7,\"chat_template_kwargs\":{\"enable_thinking\":false}}"
 ```
 
 ## OpenClaw 调用说明
 
 - Base URL 使用 `http://<服务器IP>:8001/v1`
 - API Key 使用 `config.json` 中 `api_key`
-- 模型名使用 `config.json` 中 `models.profiles.<模型名>.served_model_name`
+- 模型名固定使用 `config.json` 中 `public_model_name`（默认 `Qwen_local_model`）
+- 思考模式按请求控制：`chat_template_kwargs.enable_thinking=false/true`
+- 若调用方未传 `chat_template_kwargs.enable_thinking`，服务端使用 `default_enable_thinking` 兜底
+- 仅当模型需要推理解析器时，再将 `config.json` 中 `reasoning_enabled` 设为 `true`
 - 若使用工具调用，`config.json` 中应配置 `tool_call_parser` 与 `enable_auto_tool_choice`
 - 服务强制离线模式，不会回退到 Hugging Face 远程下载
 - 所有路径按 Ubuntu 规范填写，本地模型建议使用 `/opt/model/<模型目录>`
@@ -103,7 +123,10 @@ curl -X POST "http://localhost:8001/v1/chat/completions" \
 
 ## 常见故障排查
 
-- 报错 `model type ... Transformers does not recognize this architecture` 时，说明当前模型与镜像内依赖不兼容，建议更换模型或升级镜像版本。
+- 报错 `Model architectures ['Qwen3_5MoeForConditionalGeneration'] are not supported for now` 或 `The Transformers implementation ... is not compatible with vLLM` 时，说明当前 vLLM 栈与该模型架构不兼容，需切换到兼容模型或改用其他推理后端。
+- 报错 `StrictDataclassClassValidationError` 且包含 `validate_rope` / `unsupported operand type(s) for -=: 'set' and 'list'` 时，移除 Dockerfile 中对 `transformers --upgrade --pre` 的强制升级，使用镜像内置依赖重建。
+- 报错 `moe_wna16 quantization is currently not supported in rocm` 时，将该模型的 `quantization` 改回 `gptq`。
 - 报错 `model config (gptq) does not match quantization argument (gptq_marlin)` 时，将该模型配置改为 `dtype=float16` 且 `quantization=gptq`。
+- 报错 `RPC call to sample_tokens timed out` 或出现 `GPU core dump` 时，先下调模型配置为更稳参数：`ctx=32768`、`max_num_seqs=4`、`max_tokens=2048`、`gpu_util=0.90`，并开启 `enforce_eager=true`。
 - 若模型目录存在但仍加载失败，检查挂载路径是否为 `/opt/model:/opt/model:ro`，并确认容器内可见模型文件。
 - 如果看到 `No services to build`，说明未触发重建；需要先执行 `docker compose build --no-cache` 再 `up`。

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):
@@ -17,6 +16,10 @@ class Settings(BaseModel):
     port: int = 8000
     openai_host: str = "0.0.0.0"
     openai_port: int = 8001
+    vllm_openai_internal_url: str = "http://127.0.0.1:8001/v1"
+    public_model_name: str = "Qwen_local_model"
+    default_enable_thinking: bool = False
+    reasoning_enabled: bool = False
     model_root: str = "/opt/model"
     offline_mode: bool = True
     max_model_len: int = 8192
@@ -31,31 +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"],
-        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/engine.py

@@ -1,46 +0,0 @@
-import httpx
-
-from app.config import Settings
-from app.schemas import GenerateRequest, GenerateResponse
-
-
-class InferenceEngine:
-    def __init__(self, settings: Settings) -> None:
-        self.settings = settings
-        self.client = httpx.Client(timeout=300.0)
-
-    def close(self) -> None:
-        self.client.close()
-
-    def generate(self, req: GenerateRequest) -> GenerateResponse:
-        headers = {"Content-Type": "application/json"}
-        if self.settings.api_key:
-            headers["Authorization"] = f"Bearer {self.settings.api_key}"
-        payload = {
-            "model": self.settings.served_model_name or self.settings.model_name,
-            "messages": [{"role": "user", "content": req.prompt}],
-            "max_tokens": req.max_tokens,
-            "temperature": req.temperature,
-            "top_p": req.top_p,
-        }
-        if req.stop:
-            payload["stop"] = req.stop
-        response = self.client.post(
-            f"{self.settings.vllm_openai_internal_url}/chat/completions",
-            headers=headers,
-            json=payload,
-        )
-        response.raise_for_status()
-        body = response.json()
-        completion = body["choices"][0]["message"]["content"]
-        usage = body.get("usage", {})
-        usage_prompt = int(usage.get("prompt_tokens", 0))
-        usage_completion = int(usage.get("completion_tokens", 0))
-        return GenerateResponse(
-            text=completion,
-            prompt=req.prompt,
-            model=self.settings.served_model_name or self.settings.model_name,
-            usage_prompt_tokens=usage_prompt,
-            usage_completion_tokens=usage_completion,
-            usage_total_tokens=usage_prompt + usage_completion,
-        )

app/main.py

@@ -1,50 +0,0 @@
-from contextlib import asynccontextmanager
-
-from fastapi import Depends, FastAPI, Header, HTTPException, status
-
-from app.config import Settings, get_settings
-from app.engine import InferenceEngine
-from app.schemas import GenerateRequest, GenerateResponse, HealthResponse
-
-engine: InferenceEngine | None = None
-
-
-def verify_api_key(
-    settings: Settings = Depends(get_settings), x_api_key: str | None = Header(default=None)
-) -> None:
-    if settings.api_key and x_api_key != settings.api_key:
-        raise HTTPException(
-            status_code=status.HTTP_401_UNAUTHORIZED,
-            detail="Invalid API key",
-        )
-
-
-@asynccontextmanager
-async def lifespan(_: FastAPI):
-    global engine
-    settings = get_settings()
-    engine = InferenceEngine(settings)
-    yield
-    if engine is not None:
-        engine.close()
-    engine = None
-
-
-app = FastAPI(title="ROCm vLLM Inference API", version="1.0.0", lifespan=lifespan)
-
-
-@app.get("/health", response_model=HealthResponse)
-def health(settings: Settings = Depends(get_settings)) -> HealthResponse:
-    return HealthResponse(status="ok", model=settings.served_model_name or settings.model_name)
-
-
-@app.post("/v1/generate", response_model=GenerateResponse, dependencies=[Depends(verify_api_key)])
-def generate(req: GenerateRequest, settings: Settings = Depends(get_settings)) -> GenerateResponse:
-    if engine is None:
-        raise HTTPException(status_code=status.HTTP_503_SERVICE_UNAVAILABLE, detail="Engine not ready")
-    if req.max_tokens > settings.max_tokens:
-        raise HTTPException(
-            status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
-            detail=f"max_tokens must be <= {settings.max_tokens}",
-        )
-    return engine.generate(req)

app/model_catalog.py

@@ -1,4 +1,5 @@
 import json
+import os
 from pathlib import Path
 from typing import Any
 
@@ -63,12 +64,32 @@ def resolve_runtime_settings(content: dict[str, Any]) -> dict[str, Any]:
     api_service = dict(services.get("api", {}))
     openai_service = dict(services.get("openai", {}))
     models = dict(content.get("models", {}))
+    openai_port = _to_int(openai_service.get("port"), 8001)
+    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),
         "openai_host": str(openai_service.get("host", "0.0.0.0")),
-        "openai_port": _to_int(openai_service.get("port"), 8001),
+        "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": default_enable_thinking,
         "api_key": str(content.get("api_key", "")).strip() or None,
+        "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,
@@ -96,12 +117,27 @@ def resolve_model_profile(
     resolved_tp = requested_tp
     if valid_tp and resolved_tp not in valid_tp:
         resolved_tp = valid_tp[0]
+    speculative = dict(profile.get("speculative", {}))
+    speculative_method = _to_str(speculative.get("method"))
+    speculative_model_path = _to_str(speculative.get("model"))
+    num_speculative_tokens = _to_int(speculative.get("num_speculative_tokens"), 0)
+    speculative_draft_tp = _to_int(speculative.get("draft_tensor_parallel_size"), 0)
+    if speculative_method and speculative_model_path:
+        resolved_speculative_model = speculative_model_path
+        if not speculative_model_path.startswith("/"):
+            if not model_root:
+                raise ValueError("config.json model_root cannot be empty when speculative model path is relative")
+            resolved_speculative_model = _join_posix(model_root, speculative_model_path)
+    else:
+        resolved_speculative_model = ""
     updates = {
         "selected_model": model_key,
         "model_name": _resolve_profile_model_path(profile, model_root, model_key),
         "served_model_name": profile.get("served_model_name", model_key),
         "dtype": _to_str(profile.get("dtype")),
         "quantization": _to_str(profile.get("quantization")),
+        "model_impl": _to_str(profile.get("model_impl")),
+        "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),
@@ -111,8 +147,42 @@ def resolve_model_profile(
         "tensor_parallel_size": resolved_tp,
         "tool_call_parser": profile.get("tool_call_parser"),
         "enable_auto_tool_choice": _to_bool(profile.get("enable_auto_tool_choice"), False),
+        "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),
+        "speculative_method": speculative_method,
+        "speculative_model": resolved_speculative_model,
+        "num_speculative_tokens": num_speculative_tokens,
+        "speculative_draft_tp": speculative_draft_tp,
     }
     env_vars = {str(k): str(v) for k, v in dict(profile.get("env", {})).items()}
     env_vars["HF_HUB_OFFLINE"] = "1"
     env_vars["TRANSFORMERS_OFFLINE"] = "1"
+    if "VLLM_RPC_TIMEOUT" not in env_vars:
+        env_vars["VLLM_RPC_TIMEOUT"] = "300"
+    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/schemas.py

@@ -1,26 +0,0 @@
-from typing import List, Optional
-
-from pydantic import BaseModel, Field
-
-
-class GenerateRequest(BaseModel):
-    prompt: str
-    max_tokens: int = Field(default=256, ge=1, le=4096)
-    temperature: float = Field(default=0.7, ge=0.0, le=2.0)
-    top_p: float = Field(default=0.95, gt=0.0, le=1.0)
-    repetition_penalty: float = Field(default=1.0, ge=0.5, le=2.0)
-    stop: Optional[List[str]] = None
-
-
-class GenerateResponse(BaseModel):
-    text: str
-    prompt: str
-    model: str
-    usage_prompt_tokens: int
-    usage_completion_tokens: int
-    usage_total_tokens: int
-
-
-class HealthResponse(BaseModel):
-    status: str
-    model: str

app/start_api.py

@@ -1,23 +0,0 @@
-import subprocess
-import sys
-
-from app.config import get_settings
-
-
-def main() -> None:
-    settings = get_settings()
-    command = [
-        sys.executable,
-        "-m",
-        "uvicorn",
-        "app.main:app",
-        "--host",
-        settings.host,
-        "--port",
-        str(settings.port),
-    ]
-    raise SystemExit(subprocess.call(command))
-
-
-if __name__ == "__main__":
-    main()

app/start_openai.py

@@ -1,27 +1,23 @@
-import os
+import json
 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"])
-    api_key = runtime["api_key"] or ""
-    dtype = str(updates["dtype"] or runtime["dtype"])
-    quantization = str(updates["quantization"] 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",
@@ -31,34 +27,65 @@ def build_command() -> list[str]:
         "--port",
         port,
         "--model",
-        str(updates["model_name"]),
+        str(config["model_name"]),
         "--served-model-name",
-        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",
+            json.dumps({"enable_thinking": default_enable_thinking}),
+        ]
+    )
+    if reasoning_enabled and reasoning_parser:
+        cmd.extend(["--reasoning-parser", reasoning_parser])
     if quantization:
         cmd.extend(["--quantization", quantization])
+    if model_impl:
+        cmd.extend(["--model-impl", model_impl])
     if revision:
         cmd.extend(["--revision", revision])
     if api_key:
         cmd.extend(["--api-key", api_key])
+    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 config.get("max_num_batched_tokens", 0) > 0:
+        cmd.extend(["--max-num-batched-tokens", str(config["max_num_batched_tokens"])])
+    if config.get("language_model_only"):
+        cmd.append("--language-model-only")
+    speculative_method = str(config.get("speculative_method") or "").strip()
+    speculative_model = str(config.get("speculative_model") or "").strip()
+    num_speculative_tokens = int(config.get("num_speculative_tokens") or 0)
+    speculative_draft_tp = int(config.get("speculative_draft_tp") or 0)
+    if speculative_method and speculative_model and num_speculative_tokens > 0:
+        spec_config: dict = {
+            "method": speculative_method,
+            "model": speculative_model,
+            "num_speculative_tokens": num_speculative_tokens,
+        }
+        if speculative_draft_tp > 0:
+            spec_config["draft_tensor_parallel_size"] = speculative_draft_tp
+        cmd.extend(["--speculative-config", json.dumps(spec_config)])
     return cmd
 
 

chat_with_speed.py

@@ -0,0 +1,99 @@
+#!/usr/bin/env python3
+"""
+交互式对话脚本，用于与vLLM模型进行对话并计算token生成速度
+
+使用方法：
+1. 进入Docker容器：docker exec -it rocm-vllm-openai bash
+2. 运行：python chat_with_speed.py
+3. 输入提示词与模型对话
+4. 输入 'exit' 退出
+"""
+
+import json
+import time
+import httpx
+
+# 模型服务地址
+API_URL = "http://localhost:8001/v1/chat/completions"
+# API密钥
+API_KEY = "sk-szcjw"
+# 模型名称
+MODEL_NAME = "Qwen_local_model"
+
+def chat_with_model():
+    """交互式对话函数"""
+    print("=== vLLM 交互式对话工具 ===")
+    print("输入提示词与模型对话，输入 'exit' 退出")
+    print("=" * 50)
+    
+    # 对话历史
+    messages = []
+    
+    while True:
+        # 获取用户输入
+        user_input = input("用户: ").strip()
+        
+        if user_input.lower() == "exit":
+            print("退出对话...")
+            break
+        
+        if not user_input:
+            continue
+        
+        # 添加用户消息到对话历史
+        messages.append({"role": "user", "content": user_input})
+        
+        # 准备请求数据
+        payload = {
+            "model": MODEL_NAME,
+            "messages": messages,
+            "max_tokens": 1000,
+            "temperature": 0.7,
+            "top_p": 0.8,
+            "top_k": 20
+        }
+        
+        headers = {
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {API_KEY}"
+        }
+        
+        print("模型: ", end="", flush=True)
+        
+        # 记录开始时间
+        start_time = time.time()
+        
+        try:
+            # 发送请求
+            response = httpx.post(API_URL, json=payload, headers=headers, timeout=300.0)
+            response.raise_for_status()
+            
+            # 解析响应
+            result = response.json()
+            
+            # 获取模型回复
+            assistant_message = result["choices"][0]["message"]["content"]
+            print(assistant_message)
+            
+            # 添加模型回复到对话历史
+            messages.append({"role": "assistant", "content": assistant_message})
+            
+            # 计算token速度
+            usage = result.get("usage", {})
+            completion_tokens = usage.get("completion_tokens", 0)
+            end_time = time.time()
+            elapsed_time = end_time - start_time
+            
+            if completion_tokens > 0 and elapsed_time > 0:
+                tokens_per_second = completion_tokens / elapsed_time
+                print(f"\n[速度统计] 生成 {completion_tokens} tokens，用时 {elapsed_time:.2f} 秒，速度: {tokens_per_second:.2f} tokens/s")
+            else:
+                print("\n[速度统计] 无法计算速度")
+                
+        except Exception as e:
+            print(f"\n错误: {e}")
+            
+        print("=" * 50)
+
+if __name__ == "__main__":
+    chat_with_model()

config.json

@@ -9,6 +9,9 @@
       "port": 8001
     }
   },
+  "public_model_name": "Qwen_local_model",
+  "default_enable_thinking": false,
+  "reasoning_enabled": false,
   "api_key": "sk-szcjw",
   "tensor_parallel_size": 2,
   "dtype": "bfloat16",
@@ -17,79 +20,80 @@
   "revision": "",
   "models": {
     "default": "Qwen3.5-35B-A3B-GPTQ-Int4",
-    "selected": "Qwen3.5-35B-A3B-GPTQ-Int4",
+    "selected": "Qwen3.6-27B-FP8",
     "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": "32",
-        "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": "64",
-        "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",
-        "ctx": "65536",
-        "trust_remote": true,
-        "valid_tp": [1, 2],
-        "max_num_seqs": "64",
-        "max_tokens": "32768",
-        "gpu_util": "0.98",
-        "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",
+      "Qwen3.6-35B-A3B-FP8": {
+        "local_path": "Qwen3.6-35B-A3B-FP8",
         "dtype": "float16",
-        "quantization": "gptq",
-        "ctx": "65536",
+        "quantization": "fp8",
+        "ctx": "262144",
+        "max_tokens": "65536",
+        "max_num_batched_tokens": 32768,
         "trust_remote": true,
-        "valid_tp": [1, 2],
-        "max_num_seqs": "64",
-        "max_tokens": "32768",
-        "gpu_util": "0.98",
-        "tool_call_parser": "qwen3_xml",
+        "enforce_eager": false,
+        "valid_tp": [
+          2
+        ],
+        "max_num_seqs": "256",
+        "gpu_util": "0.92",
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
-        "served_model_name": "Qwen3.5-35B-A3B-GPTQ-Int4",
-        "hf_model_id": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4"
+        "language_model_only": true,
+        "served_model_name": "Qwen3.6-35B-A3B-FP8",
+        "hf_model_id": "Qwen/Qwen3.6-35B-A3B-FP8"
       },
-      "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": "64",
+      "Qwen3.6-27B": {
+        "local_path": "Qwen3.6-27B",
+        "ctx": "131072",
         "max_tokens": "32768",
-        "gpu_util": "0.98",
-        "tool_call_parser": "qwen3_xml",
+        "max_num_batched_tokens": 16384,
+        "trust_remote": true,
+        "enforce_eager": false,
+        "valid_tp": [
+          1,
+          2
+        ],
+        "max_num_seqs": 32,
+        "gpu_util": "0.96",
+        "kv_cache_dtype": "fp8",
+        "enable_prefix_caching": true,
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
         "enable_auto_tool_choice": true,
-        "served_model_name": "Qwen3.5-35B-A3B-FP8",
-        "hf_model_id": "Qwen/Qwen3.5-35B-A3B-FP8"
+        "language_model_only": true,
+        "served_model_name": "Qwen3.6-27B",
+        "hf_model_id": "Qwen/Qwen3.6-27B"
+      },
+      "Qwen3.6-27B-FP8": {
+        "local_path": "Qwen3.6-27B-FP8",
+        "dtype": "auto",
+        "quantization": "fp8",
+        "ctx": "131072",
+        "max_tokens": "32768",
+        "max_num_batched_tokens": 16384,
+        "trust_remote": true,
+        "enforce_eager": false,
+        "valid_tp": [
+          2
+        ],
+        "max_num_seqs": 32,
+        "gpu_util": "0.85",
+        "tool_call_parser": "qwen3_coder",
+        "reasoning_parser": "qwen3",
+        "enable_auto_tool_choice": true,
+        "language_model_only": true,
+        "served_model_name": "Qwen3.6-27B-FP8",
+        "hf_model_id": "Qwen/Qwen3.6-27B-FP8",
+        "env": {
+          "VLLM_RPC_TIMEOUT": "300"
+        },
+        "speculative": {
+          "method": "dflash",
+          "model": "Qwen3.6-27B-DFlash",
+          "num_speculative_tokens": 4,
+          "draft_tensor_parallel_size": 1
+        }
       }
     }
   }

docker-compose.yml

@@ -3,8 +3,9 @@ services:
     build:
       context: .
       dockerfile: Dockerfile
-    image: rocm-vllm-inference:latest
-    container_name: rocm-vllm-openai
+      pull: false
+    image: vllm-openai-rocm:nightly
+    container_name: vllm-openai-rocm
     entrypoint: ["python"]
     command: ["-m", "app.start_openai"]
     ports:
@@ -12,13 +13,25 @@ services:
     volumes:
       - /opt/model:/opt/model:ro
       - ./config.json:/workspace/config.json:ro
+    environment:
+      HF_HOME: /opt/model
+      # 严格离线模式，禁止任何网络下载
+      HF_HUB_OFFLINE: "1"
+      TRANSFORMERS_OFFLINE: "1"
+      HF_DATASETS_OFFLINE: "1"
+      VLLM_ROCM_USE_AITER: "0"  # 打开总开关
+      # 明确禁用不稳定的MoE后端，让其他AITER优化（如MHA）生效
+      VLLM_ROCM_USE_AITER_MOE: "0"
+      VLLM_ROCM_MOE_BACKEND: "TRITON"  # 强制MoE使用Triton
+      TORCHINDUCTOR_FX_GRAPH_CACHE: "0" # 禁用 torch.compile 缓存
+      VLLM_DISABLE_COMPILE_CACHE: "1"
     devices:
       - /dev/kfd
       - /dev/dri
     group_add:
       - video
     ipc: host
-    shm_size: 16g
+    shm_size: 32g
     cap_add:
       - SYS_PTRACE
     security_opt: