init

Port logic from continuous_pipeline_download_gguf_zhoushasha_and_submit.py
into the strategy-deployable framework (health/status HTTP server,
STRATEGY_ID env, SIGTERM handling).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.gitignore

@@ -0,0 +1,2 @@
+.DS_Store
+__pycache__/

Dockerfile

@@ -1,10 +1,7 @@
 FROM modelhubxc-4pd.tencentcloudcr.com/xc_agent_platform/python:3.11-slim
 
-ENV PYTHONUNBUFFERED=1 \
+ENV PYTHONUNBUFFERED=1 
     
-    CONTEST_API_TOKEN="ef1ef82f3c9efee413d602345fbe224d" \
-    CONTRIBUTORS="zhoushasha" \
-    GPU_TYPE="Cambricon_mlu-370-x8"
 
 WORKDIR /app
 

README.md

@@ -1,22 +1,24 @@
-# xc_validation_strategy
+# xc_validation_strategy_gguf
 
-批量向 ModelHub XC 平台提交模型验证任务的策略服务，之后保持 HTTP 服务存活供平台探活。
+GGUF 模型下载 + 验证任务提交流水线策略服务：批量创建 GGUF 模型下载任务（最大并发 8），
+每个模型下载成功后立即提交 hygon / bi150 两个验证任务，之后保持 HTTP 服务存活供平台探活。
 
 ## 功能
 
-- 自动登录 ModelHub 获取 Token
+- 自动登录 ModelHub 获取 Token（失败时回退到预设 Token）
-- 批量提交模型验证任务（vLLM 框架，Cambricon MLU-370-x8）
+- 按流水线批量创建 GGUF 模型下载任务（HuggingFace 源，最大并发 8）
-- 提交结果写入 `submitted_validation_tasks.txt`
+- 每个模型下载成功后，立即提交 hygon_k100-ai 与 Iluvatar_bi-150 两个验证任务（llamacpp 框架）
+- 下载成功的模型 ID 写入 `downloaded_success_models.txt`
 - 暴露 `/health` 和 `/status` 接口满足平台运行时契约
 
 ## 项目结构
 
 ```
 .
-├── main.py            # 主入口：HTTP 服务 + 提交逻辑
+├── main.py            # 主入口：HTTP 服务 + 下载/提交流水线
 ├── Dockerfile         # 平台镜像构建配置
 ├── requirements.txt   # Python 依赖
-└── submitted_validation_tasks.txt   # 运行后自动生成，记录提交结果
+└── downloaded_success_models.txt   # 运行后自动生成，记录下载成功的模型
 ```
 
 ## 平台契约说明

main.py

@@ -1,35 +1,47 @@
 """
-xc_validation_strategy — 主入口
+xc_validation_strategy_gguf — 主入口
 
-启动后执行一次模型验证任务批量提交，之后保持 HTTP 服务存活。
+GGUF 模型下载 + 验证任务提交流水线（部署框架与 xc_validation_strategy 一致）。
+
+启动后运行流水线：批量创建 GGUF 模型下载任务（最大并发 8），
+每个模型下载成功后立即提交 hygon / bi150 两个验证任务；
 同时暴露 /health（K8s 探活）和 /status（运行状态）。
 """
 
 import json
 import os
+import re
 import signal
 import threading
-import traceback
 from datetime import datetime
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
-from typing import List, Tuple
+from typing import Set
 
 import requests
 
 # ══════════════════════════════════════════════════════════
-#  配置（全部从环境变量读取，不硬编码敏感信息）
+#  配置
 # ══════════════════════════════════════════════════════════
 BASE_URL                      = os.environ.get("BASE_URL", "https://modelhub.org.cn")
 LOGIN_ENDPOINT                = "/adminApi/user/login"
-SUBMIT_ENDPOINT    = "/adminApi/async/task/create-contest-task"
+CREATE_DOWNLOAD_TASK_ENDPOINT = "/adminApi/async/task/model-download-task"
+SUBMIT_TEST_TASK_ENDPOINT     = "/adminApi/async/task/create-contest-task"
 
-USER_ACCOUNT       = os.environ["USER_ACCOUNT"]        # 必填
+# 登录账号：启动时优先用账号密码换取新 token（流水线运行时间长，预设 token 可能过期）
-USER_PASSWORD      = os.environ["USER_PASSWORD"]        # 必填
+USER_ACCOUNT  = "zhoushasha@4paradigm.com"
-CONTEST_API_TOKEN  = os.environ["CONTEST_API_TOKEN"]    # 必填
+USER_PASSWORD = "ganshenme0"
-STRATEGY_ID        = os.environ.get("STRATEGY_ID", "")  # 平台注入
+
-CONTRIBUTORS       = os.environ.get("CONTRIBUTORS", USER_ACCOUNT)
+# 通过 curl -X POST https://modelhub.org.cn/adminApi/user/login 获取后填入（登录失败时的回退）
-GPU_TYPE           = os.environ.get("GPU_TYPE", "Cambricon_mlu-370-x8")
+AUTH_TOKEN = "eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3ODQ1NDc1NDYsImlhdCI6MTc4Mzk0Mjc0Nn0.ZcOqcrfI22LPi4mGMnt164nZGhi61ZxtJGYsoO7fZdM"
-TASK_TYPE          = os.environ.get("TASK_TYPE", "text-generation")
+
+CONTEST_API_TOKEN = "ef1ef82f3c9efee413d602345fbe224d"
+HF_TOKEN          = "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"
+CONTRIBUTORS      = "zhoushasha"
+TASK_TYPE         = "text-generation"
+STRATEGY_ID       = os.environ.get("STRATEGY_ID", "")  # 平台自动注入，无需修改
+
+MAX_CONCURRENT_DOWNLOADS = 8    # 同时下载的模型数上限
+CHECK_INTERVAL_SECONDS   = 10   # 下载状态轮询间隔（秒）
 
 HTTP_HOST = "0.0.0.0"
 HTTP_PORT = 8080
@@ -38,37 +50,74 @@ HTTP_PORT = 8080
 #  模型列表
 # ══════════════════════════════════════════════════════════
 ALL_MODEL_IDS = [
-    "AI-ModelScope/gemma-2b",
+
-    "AI-ModelScope/falcon-mamba-7b",
+    
-    "katanemo/deepseek-2",
+    
-    "OpenBMB/MiniCPM4-0.5B",
+    "mradermacher/Qwen2.5-3B-instruct-argus-v3-GGUF",
-    "NousResearch/Meta-Llama-3-8B-Instruct",
+    "mradermacher/NeuralsirkrishnaShadow_OgnoExperiment27-GGUF",
-    "MediaTek-Research/Breeze-7B-Instruct-v1_0",
+    "mradermacher/GREEN-RadLlama2-7b-GGUF",
-    "QLUNLP/BianCang-Qwen2.5-7B-Instruct",
+    "mradermacher/snakmodel-7b-instruct-GGUF",
-    "OpenBMB/MiniCPM4-Survey",
+    "mradermacher/BioMistral-7B-Starling-SLERP-GGUF",
-    "OpenBMB/MiniCPM4-8B",
+    "mradermacher/Marco-Llama-3.2-3B-GGUF",
-    "PaddlePaddle/ERNIE-4.5-0.3B-PT",
+    "mradermacher/Qwen2.5-1.5B-Instruct-Open-R1-GRPO-GGUF",
-    "LLM-Research/Llama-Guard-3-8B",
+    "mradermacher/Qwen2.5-0.5B-Distill-Fast-GGUF",
-    "OpenBMB/MiniCPM-2B-dpo-fp16",
+    "mradermacher/chomsky_16_bit_model-GGUF",
-    "OpenBMB/MiniCPM4.1-8B",
+    "mradermacher/Mistral-7B-DFT-GGUF",
-    "Cylingo/Xinyuan-LLM-14B-0428",
+    "mradermacher/mergekit-task_arithmetic-qjeuqjw-GGUF",
-    "Fengshenbang/Ziya-LLaMA-13B-v1",
+    "mradermacher/M7Yamshadowexperiment28_Experiment27Inex12-GGUF",
-    "baichuan-inc/Baichuan2-13B-Chat",
+    "mradermacher/YamshadowStrangemerges_32_Experiment28Inex12-GGUF",
-    "LLM-Research/gemma-2-9b-it",
+    "mradermacher/MeliodasPercival_01_Experiment29Pastiche-GGUF",
-    "Qwen/CodeQwen1.5-7B-Chat",
+    "mradermacher/Llama-3-6B-v0-GGUF",
-    "OpenBMB/cpm-bee-10b",
+    "mradermacher/Excalibur-7b-DPO-GGUF",
-    "OpenBMB/MiniCPM3-4B",
+    "mradermacher/Llama-2-7b-Indian-Law-GGUF",
+    "mradermacher/StarlingHermes-2.5-Mistral-7B-slerp-GGUF",
+    "mradermacher/DeepThinker-7B-Sce-v1-GGUF",
+    "mradermacher/Alif-Llama-EXP2-GGUF",
+    "mradermacher/qwen-2.5-1.5B-Rasa-GGUF",
+    "mradermacher/German_RAG-PHI-3.5-MINI-4B-MERGED-HESSIAN-AI-GGUF",
+    "mradermacher/MFANN-phigments-slerp-V3.2-GGUF",
+    "mradermacher/AdityaGPT-GGUF",
+    "mradermacher/Jaja-small-v4-GGUF",
+    "mradermacher/Jaja-medium-v1-GGUF",
+    "mradermacher/astrollama-2-7b-base_abstract-GGUF",
+    "mradermacher/Llasagna-v0.1-GGUF",
+    "mradermacher/DeepThinker-v-GGUF",
+    "mradermacher/Jaja-small-v3-GGUF",
+    "mradermacher/Llama-3.2-1B-FC-v1.2-think-GGUF",
+    "mradermacher/Deepseek-Qwen2.5-1.5B-Redistil-GGUF",
+    
+    
+    
+    
+    
+    
 ]
 
+
+# 去重（保持原有顺序）
+_seen = set()
+_deduplicated = []
+for _mid in ALL_MODEL_IDS:
+    if _mid not in _seen:
+        _deduplicated.append(_mid)
+        _seen.add(_mid)
+ALL_MODEL_IDS = _deduplicated
+print(f"[INFO] 去重后模型数量: {len(ALL_MODEL_IDS)}", flush=True)
+
+HEADERS = {"Content-Type": "application/json"}
+
 # ══════════════════════════════════════════════════════════
 #  全局状态（供 /status 展示）
 # ══════════════════════════════════════════════════════════
 _state = {
     "strategy_id":       STRATEGY_ID,
-    "phase":         "starting",   # starting | submitting | done | error
+    "phase":             "starting",   # starting | running | done | error
     "total":             len(ALL_MODEL_IDS),
-    "submitted":     0,
+    "downloading":       [],           # 当前正在下载的模型
-    "failed":        0,
+    "download_success":  0,
+    "download_failed":   0,
+    "submitted":         0,            # 成功提交的验证任务数（hygon + bi150）
+    "submit_failed":     0,
     "started_at":        None,
     "finished_at":       None,
 }
@@ -108,122 +157,336 @@ def _run_http():
     print("[http] 已关闭", flush=True)
 
 # ══════════════════════════════════════════════════════════
-#  业务逻辑
+#  工具函数：生成模型文件名
 # ══════════════════════════════════════════════════════════
-def _login() -> str:
+def get_model_filename(model_id: str) -> str:
-    headers = {"Content-Type": "application/json"}
+    """
-    resp = requests.post(
+    从 model_id 生成标准 GGUF 模型文件名。
-        BASE_URL + LOGIN_ENDPOINT,
+
-        headers=headers,
+    规则：
-        json={"userAccount": USER_ACCOUNT, "userPassword": USER_PASSWORD},
+    - 移除组织名（/ 前部分）
-        timeout=30,
+    - 处理 '_-_' 分割（保留原有逻辑）
-    )
+    - 移除末尾 '-GGUF'（不区分大小写）
+    - 若移除后以 -i1, -i2, ..., -i99 结尾：
+        → 替换为 .i1, .i2, ... 并添加 '-Q4_0.gguf'
+      否则：
+        → 直接添加 '.f16.gguf'
+
+    示例：
+      'mradermacher/Qwen3-8B-makisu-v2.0.1-i1-GGUF'
+        → 'Qwen3-8B-makisu-v2.0.1.i1-Q4_0.gguf'
+
+      'QuantFactory/Apollo2-9B-GGUF'
+        → 'Apollo2-9B.f16.gguf'
+    """
+    # 1. 提取模型名部分（/ 后）
+    base_name = model_id.split("/")[-1]
+
+    # 2. 处理 '_-_' 分割
+    if '_-_' in base_name:
+        base_name = base_name.split('_-_')[-1]
+
+    # 3. 移除末尾的 -GGUF（不区分大小写）
+    if base_name.lower().endswith("-gguf"):
+        base_name = base_name[:-5]
+
+    # 4. 检查是否以 -i<数字> 结尾（支持 i1～i99 等）
+    match = re.search(r'-i(\d+)$', base_name)
+    if match:
+        number = match.group(1)
+        base_name = base_name[:match.start()] + f".i{number}-Q4_0.gguf"
+    else:
+        base_name = base_name + ".f16.gguf"
+
+    return base_name
+
+# ══════════════════════════════════════════════════════════
+#  登录获取 token（失败时回退到预设 AUTH_TOKEN）
+# ══════════════════════════════════════════════════════════
+def login() -> str:
+    payload = {"userAccount": USER_ACCOUNT, "userPassword": USER_PASSWORD}
+    print("[login] 正在登录...", flush=True)
+    try:
+        resp = requests.post(BASE_URL + LOGIN_ENDPOINT, headers=HEADERS, json=payload, timeout=15)
+        data = resp.json()
+        if resp.status_code == 200 and data.get("code") == 0:
+            print("[login] 登录成功", flush=True)
+            return data["data"]["token"]
+        print(f"[login] 登录失败: {data.get('message')}，回退使用预设 Token", flush=True)
+    except Exception as e:
+        print(f"[login] 登录异常: {e}，回退使用预设 Token", flush=True)
+    return AUTH_TOKEN
+
+# ══════════════════════════════════════════════════════════
+#  创建单个模型的下载任务
+# ══════════════════════════════════════════════════════════
+def create_download_task(token: str, model_id: str) -> bool:
+    filename = get_model_filename(model_id)
+    auth_headers = {**HEADERS, "Authorization": f"Bearer {token}"}
+    payload = {
+        "allowPatterns": [filename],
+        "hfToken": HF_TOKEN,
+        "modelId": model_id,
+        "source": "HUGGING_FACE",
+        "stillDownloadAlreadySuccessDownloadedModel": False
+    }
+    print(f"📥 创建下载任务: {model_id} → {filename}", flush=True)
+    try:
+        resp = requests.post(BASE_URL + CREATE_DOWNLOAD_TASK_ENDPOINT, headers=auth_headers, json=payload, timeout=15)
+        if resp.status_code == 200:
+            data = resp.json()
+            if data.get("code") == 0:
+                print(f"✅ 下载任务已提交: {model_id}", flush=True)
+                return True
+            else:
+                print(f"⚠️ 下载任务业务失败 ({model_id}): {data.get('message')}", flush=True)
+                return False
+        else:
+            print(f"❌ HTTP 错误 ({model_id}): {resp.status_code} - {resp.text}", flush=True)
+            return False
+    except Exception as e:
+        print(f"💥 创建下载任务异常 ({model_id}): {e}", flush=True)
+        return False
+
+# ══════════════════════════════════════════════════════════
+#  查询单个模型的最新下载任务状态
+# ══════════════════════════════════════════════════════════
+def check_model_status(token: str, model_id: str) -> str:
+    """
+    返回状态: 'WAITING', 'RUNNING', 'SUCCESS', 'FAILED', 'UNKNOWN'
+    """
+    url = BASE_URL + CREATE_DOWNLOAD_TASK_ENDPOINT
+    auth_headers = {**HEADERS, "Authorization": f"Bearer {token}"}
+    params = {"modelId": model_id, "current": 1, "pageSize": 1}
+    try:
+        resp = requests.get(url, headers=auth_headers, params=params, timeout=10)
+        if resp.status_code != 200:
+            return "UNKNOWN"
         data = resp.json()
         if data.get("code") != 0:
-        raise RuntimeError(f"登录失败: {data.get('message')}")
+            return "UNKNOWN"
-    print("[worker] 登录成功", flush=True)
+        records = data.get("data", {}).get("records", [])
-    return data["data"]["token"]
+        if not records:
+            return "UNKNOWN"
+        status = records[0].get("status", "UNKNOWN").upper()
+        return status
+    except Exception as e:
+        print(f"⚠️ 查询状态异常 ({model_id}): {e}", flush=True)
+        return "UNKNOWN"
 
-
+# ══════════════════════════════════════════════════════════
-def _submit_task(token: str, model_id: str) -> Tuple[bool, str]:
+#  提交单个模型的测试任务 hygon
-    headers = {
+# ══════════════════════════════════════════════════════════
-        "Content-Type": "application/json",
+def submit_test_task(token: str, model_id: str) -> bool:
-        "Authorization": f"Bearer {token}",
+    auth_headers = {**HEADERS, "Authorization": f"Bearer {token}"}
-    }
+    model_filename = get_model_filename(model_id)
-    config_content = f"""docker_image: harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-torch2.5.0-torchmlu1.24.1-ubuntu22.04-py310
+    gpu_type = "hygon_k100-ai"
-nv_docker_image: harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0
+    config_content = f"""docker_image: git.modelhub.org.cn:9443/enginex-hygon/hygon-llama.cpp:b7516
-framework: vllm
+nv_docker_image: harbor-contest.4pd.io/luxinlong02/llama-cpp:b7003-cuda-full-12.3
+framework: llamacpp
 storage: gpfs
 modelhub_options:
   srcRelativePath: leaderboard/modelHubXC/{model_id}
   mountPoint: /model
+api: completion
+temperature: 0
+repetition_penalty: 1.1
+top_p: 0.9
+max_model_len: 4096
 sut_config:
-  values:
   gpu_num: 1
-    env:
+  values:
-      - name: MAX_MODEL_LEN
+    command: ['/app/llama-server','--model', '/model/{model_filename}', '--alias', 'llm', '--threads', '20','--n-gpu-layers', '999', '--prio', '3', '--min_p', '0.01', '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000', '--jinja', '--flash-attn', 'off']
-        value: 8192
-    command: ["vllm", "serve", "/model", "--port", "8000", "--served-model-name", "llm", "--max-model-len", "8192", "--trust-remote-code", "--dtype", "float16"]
 ref_config:
-  values:
-    cpu_num: 2
   gpu_num: 1
-    env:
+  values:
-      - name: MAX_MODEL_LEN
+    command: ['/workspace/llama.cpp/build/bin/llama-server','--model', '/model/{model_filename}', '--alias', 'llm', '--threads', '20', '--n-gpu-layers', '999', '--prio', '3', '--min_p', '0.01', '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000', '--jinja', '--flash-attn', 'off']
-        value: 8192
-    command: ["vllm", "serve", "/model", "--port", "80", "--served-model-name", "llm", "--max-model-len", "8192", "--trust-remote-code", "--dtype", "float16"]
 """
-    payload = {
+    task_data = {
         "contestApiToken": CONTEST_API_TOKEN,
         "contributors": CONTRIBUTORS,
-        "gpuTypes":        [GPU_TYPE],
+        "gpuTypes": [gpu_type],
         "taskType": TASK_TYPE,
         "modelId": model_id,
-        "framework":       "vllm",
         "strategyId": STRATEGY_ID,   # 平台要求
         "submissionConfig": [{
             "config": config_content,
-            "gpuType":  GPU_TYPE,
+            "gpuType": gpu_type,
-            "taskType": TASK_TYPE,
+            "taskType": TASK_TYPE
-        }],
+        }]
     }
+    print(f"📤 提交测试任务 (hygon): {model_id}", flush=True)
     try:
-        resp = requests.post(
+        resp = requests.post(BASE_URL + SUBMIT_TEST_TASK_ENDPOINT, json=task_data, headers=auth_headers, timeout=15)
-            BASE_URL + SUBMIT_ENDPOINT,
+        if resp.status_code == 200:
-            headers=headers,
-            json=payload,
-            timeout=15,
-        )
             result = resp.json()
             if result.get("code") == 0:
-            task_id = result.get("data", {}).get("id", "")
+                task_id = result.get("data", {}).get("taskId")
-            print(f"[worker] OK  {model_id}  task_id={task_id}", flush=True)
+                print(f"✅ 测试任务提交成功! Task ID: {task_id}", flush=True)
-            return True, task_id
+                return True
             else:
-            print(f"[worker] FAIL {model_id}: {result.get('message')}", flush=True)
+                print(f"❌ 测试任务业务错误: {result.get('message')}", flush=True)
-            return False, ""
+                return False
+        else:
+            print(f"❌ 测试任务 HTTP 错误: {resp.status_code} - {resp.text}", flush=True)
+            return False
     except Exception as e:
-        print(f"[worker] ERROR {model_id}: {e}", flush=True)
+        print(f"💥 提交测试任务异常 ({model_id}): {e}", flush=True)
-        return False, ""
+        return False
 
+# ══════════════════════════════════════════════════════════
+#  提交单个模型的测试任务 bi150
+# ══════════════════════════════════════════════════════════
+def submit_test_task_bi150(token: str, model_id: str) -> bool:
+    auth_headers = {**HEADERS, "Authorization": f"Bearer {token}"}
+    model_filename = get_model_filename(model_id)
+    gpu_type = "Iluvatar_bi-150"
+    config_content = f"""docker_image: git.modelhub.org.cn:9443/enginex-iluvatar/iluvatar-llama.cpp:b7516-bi150
+nv_docker_image: harbor-contest.4pd.io/luxinlong02/llama-cpp:b7003-cuda-full-12.3
+framework: llamacpp
+storage: gpfs
+modelhub_options:
+  srcRelativePath: leaderboard/modelHubXC/{model_id}
+  mountPoint: /model
+max_model_len: 4096
+sut_config:
+  gpu_num: 1
+  values:
+    command: ['/app/llama-server','--model', '/model/{model_filename}', '--alias', 'llm', '--threads', '20','--n-gpu-layers','128', '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000', '--jinja', '--flash-attn', 'off', '--no-mmap', '--sync-to-temp']
+ref_config:
+  gpu_num: 1
+  values:
+    command: ['/workspace/llama.cpp/build/bin/llama-server','--model', '/model/{model_filename}', '--alias', 'llm', '--threads', '20','--n-gpu-layers','128', '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000', '--jinja', '--flash-attn', 'off']
+"""
+    task_data = {
+        "contestApiToken": CONTEST_API_TOKEN,
+        "contributors": CONTRIBUTORS,
+        "gpuTypes": [gpu_type],
+        "taskType": TASK_TYPE,
+        "modelId": model_id,
+        "strategyId": STRATEGY_ID,   # 平台要求
+        "submissionConfig": [{
+            "config": config_content,
+            "gpuType": gpu_type,
+            "taskType": TASK_TYPE
+        }]
+    }
+    print(f"📤 提交测试任务 (bi150): {model_id}", flush=True)
+    try:
+        resp = requests.post(BASE_URL + SUBMIT_TEST_TASK_ENDPOINT, json=task_data, headers=auth_headers, timeout=15)
+        if resp.status_code == 200:
+            result = resp.json()
+            if result.get("code") == 0:
+                task_id = result.get("data", {}).get("taskId")
+                print(f"✅ 测试任务提交成功! Task ID: {task_id}", flush=True)
+                return True
+            else:
+                print(f"❌ 测试任务业务错误: {result.get('message')}", flush=True)
+                return False
+        else:
+            print(f"❌ 测试任务 HTTP 错误: {resp.status_code} - {resp.text}", flush=True)
+            return False
+    except Exception as e:
+        print(f"💥 提交测试任务异常 ({model_id}): {e}", flush=True)
+        return False
 
+# ══════════════════════════════════════════════════════════
+#  业务逻辑：动态流水线（下载 → 提交验证任务）
+# ══════════════════════════════════════════════════════════
 def _run_worker():
     _state["started_at"] = datetime.utcnow().isoformat()
-    _state["phase"] = "submitting"
+    _state["phase"] = "running"
 
-    successful: List[Tuple[str, str]] = []
     try:
-        token = _login()
+        token = login()
-    except Exception:
+    except Exception as e:
-        traceback.print_exc()
+        print(f"[worker] 登录失败: {e}", flush=True)
         _state["phase"] = "error"
         return
 
-    for model_id in ALL_MODEL_IDS:
+    pending_models = list(ALL_MODEL_IDS)   # 尚未开始下载的模型
+    active_models: Set[str] = set()        # 当前正在下载的模型
+    completed_results = {}                 # model_id -> status
+
+    print(f"🚀 总共 {len(pending_models)} 个模型待下载。最大并发数: {MAX_CONCURRENT_DOWNLOADS}\n", flush=True)
+
+    while (pending_models or active_models) and not _shutdown.is_set():
+        # 1. 检查活跃任务状态
+        for model_id in list(active_models):
             if _shutdown.is_set():
                 break
-        ok, task_id = _submit_task(token, model_id)
+            status = check_model_status(token, model_id)
-        if ok:
+            if status in ("SUCCESS", "FAILED"):
+                completed_results[model_id] = status
+                active_models.remove(model_id)
+                print(f"⏹️  {model_id} 完成，状态: {status}", flush=True)
+                if status == "SUCCESS":
+                    _state["download_success"] += 1
+                    # 下载成功后立即提交该模型的验证任务
+                    if submit_test_task(token, model_id):
                         _state["submitted"] += 1
-            successful.append((task_id, model_id))
+                        print(f"🧪 已为 {model_id} 提交 hygon 验证任务", flush=True)
                     else:
-            _state["failed"] += 1
+                        _state["submit_failed"] += 1
+                        print(f"⚠️ {model_id} hygon 验证任务提交失败", flush=True)
 
-    # 写入结果文件
+                    if submit_test_task_bi150(token, model_id):
+                        _state["submitted"] += 1
+                        print(f"🧪 已为 {model_id} 提交 bi150 验证任务", flush=True)
+                    else:
+                        _state["submit_failed"] += 1
+                        print(f"⚠️ {model_id} bi150 验证任务提交失败", flush=True)
+                else:
+                    _state["download_failed"] += 1
+
+        # 2. 补充新任务（最多补到 MAX_CONCURRENT_DOWNLOADS 个）
+        while len(active_models) < MAX_CONCURRENT_DOWNLOADS and pending_models and not _shutdown.is_set():
+            next_model = pending_models.pop(0)
+            if create_download_task(token, next_model):
+                active_models.add(next_model)
+                print(f"▶️  启动下载: {next_model} (当前活跃: {len(active_models)})", flush=True)
+            else:
+                # 创建失败也视为完成（避免卡住）
+                completed_results[next_model] = "CREATE_FAILED"
+                _state["download_failed"] += 1
+                print(f"❌ 创建失败: {next_model}", flush=True)
+
+        _state["downloading"] = sorted(active_models)
+
+        # 3. 稍作等待，避免频繁查询（可被 shutdown 信号打断）
+        if active_models or pending_models:
+            _shutdown.wait(CHECK_INTERVAL_SECONDS)
+
+    print("\n✅ 所有模型处理完毕！\n", flush=True)
+
+    # 4. 收集所有成功下载的模型ID并写入结果文件
+    success_models = [
+        mid for mid, status in completed_results.items()
+        if status == "SUCCESS"
+    ]
     try:
-        with open("submitted_validation_tasks.txt", "w", encoding="utf-8") as f:
+        with open("downloaded_success_models.txt", "w", encoding="utf-8") as f:
-            for tid, mid in successful:
+            for mid in success_models:
-                f.write(f"{tid}\t{mid}\n")
+                f.write(f"{mid}\n")
     except Exception:
         pass
 
+    print("🎉 下载成功的模型ID列表：", flush=True)
+    print("[", flush=True)
+    for mid in success_models:
+        print(f'    "{mid}",', flush=True)
+    print("]", flush=True)
+
+    _state["downloading"] = []
     _state["finished_at"] = datetime.utcnow().isoformat()
     _state["phase"] = "done"
     print(
-        f"[worker] 完成  submitted={_state['submitted']}  failed={_state['failed']}",
+        f"[worker] 完成  download_success={_state['download_success']}  "
+        f"download_failed={_state['download_failed']}  "
+        f"submitted={_state['submitted']}  submit_failed={_state['submit_failed']}",
         flush=True,
     )
-    # 提交完成后继续保持进程存活，等待平台停止
+    # 流水线完成后继续保持进程存活，等待平台停止
 
 # ══════════════════════════════════════════════════════════
 #  入口
@@ -241,7 +504,7 @@ def main():
     http_thread = threading.Thread(target=_run_http, daemon=False)
     http_thread.start()
 
-    # 提交任务线程
+    # 流水线线程
     worker_thread = threading.Thread(target=_run_worker, daemon=True)
     worker_thread.start()