feat: use file-based SQLite database for persistent failure tracking

Dockerfile

@@ -4,6 +4,8 @@ ENV PYTHONUNBUFFERED=1
 
 WORKDIR /app
 
+RUN mkdir -p /app/data
+
 COPY requirements.txt .
 RUN pip install --no-cache-dir -r requirements.txt
 

main.py

@@ -31,7 +31,7 @@ PORT = 8080
 STRATEGY_ID = os.getenv("STRATEGY_ID", "")
 
 # 目标GPU
-TARGET_GPU = "ppu_zw_810e"
+TARGET_GPU = "Iluvatar_bi-150"
 
 # 账号Token
 TARGET_TOKEN = "f45f1aae2c094426be237c88b1085015"
@@ -81,46 +81,85 @@ db_conn = None
 
 def init_db():
     global db_conn
-    db_conn = sqlite3.connect(':memory:', check_same_thread=False)
-    db_conn.execute('''CREATE TABLE IF NOT EXISTS queue (
-        model_id TEXT, gpu TEXT, url TEXT, downloads INTEGER,
-        params TEXT, category TEXT, score REAL,
+    os.makedirs('/app/data', exist_ok=True)
+    db_conn = sqlite3.connect('/app/data/submit_history.db', check_same_thread=False)
+    db_conn.execute('''CREATE TABLE IF NOT EXISTS submitted (
+        model_id TEXT, gpu TEXT, task_id TEXT, status TEXT,
+        submitted_at TEXT, checked_at TEXT,
         PRIMARY KEY(model_id, gpu)
     )''')
-    db_conn.execute('''CREATE TABLE IF NOT EXISTS submitted (
-        model_id TEXT, gpu TEXT, task_id TEXT, submitted_at TEXT,
+    db_conn.execute('''CREATE TABLE IF NOT EXISTS failed (
+        model_id TEXT, gpu TEXT, reason TEXT, failed_at TEXT,
         PRIMARY KEY(model_id, gpu)
     )''')
     db_conn.commit()
+def is_model_failed(model_id: str) -> bool:
+    """检查模型是否已知失败"""
+    if db_conn:
+        row = db_conn.execute(
+            'SELECT 1 FROM failed WHERE model_id=? AND gpu=?',
+            (model_id, TARGET_GPU)
+        ).fetchone()
+        return row is not None
+    return False
+
+def record_failed(model_id: str, reason: str):
+    """记录失败的模型"""
+    if db_conn:
+        db_conn.execute(
+            'INSERT OR REPLACE INTO failed VALUES (?,?,?,?)',
+            (model_id, TARGET_GPU, reason, datetime.now().isoformat())
+        )
+        db_conn.commit()
 
 
 # ============================================================
 # ModelScope 搜索
 # ============================================================
 
-def search_models(keyword: str, limit: int = 50) -> list:
-    """从 HuggingFace 搜索模型"""
-    url = "https://huggingface.co/api/models"
-    params = {
-        'search': keyword,
-        'limit': limit,
-        'sort': 'downloads',
-        'direction': -1,
-    }
-    try:
-        resp = requests.get(url, params=params, timeout=20)
-        data = resp.json()
-        log(f"  [{keyword}]: {len(data)} 个结果")
-        return [{'id': m.get('id'), 'downloads': m.get('downloads', 0)} for m in data]
-    except Exception as e:
-        log(f"  [{keyword}]: 失败 {e}")
-    return []
+MODELSCOPE_API = "https://modelscope.cn/api/v1"
+DOWNLOAD_MIN = 50
+DOWNLOAD_MAX = 5000
+SEARCH_PAGES = 5  # 每个关键词搜5页（50*5=250个结果）
+
+
+def search_models(keyword: str) -> list:
+    """从 ModelScope 搜索模型（多页，筛选下载量50-5000的冷门模型）"""
+    url = "https://modelscope.cn/openapi/v1/models"
+    models = []
+    for page in range(1, SEARCH_PAGES + 1):
+        params = {
+            'search': keyword,
+            'page_size': 50,
+            'page_number': page,
+            'sort': 'downloads',
+        }
+        try:
+            resp = requests.get(url, params=params, timeout=20,
+                                headers={'User-Agent': 'Mozilla/5.0'})
+            data = resp.json()
+            if data.get('success'):
+                page_models = data.get('data', {}).get('models', [])
+                for m in page_models:
+                    dl = m.get('downloads', 0)
+                    if DOWNLOAD_MIN <= dl <= DOWNLOAD_MAX:
+                        models.append({'id': m.get('id'), 'downloads': dl})
+                if len(page_models) < 50:
+                    break  # 最后一页，不继续
+            else:
+                break
+        except Exception as e:
+            log(f"  [{keyword}] page={page}: {e}")
+            break
+        time.sleep(0.3)
+    log(f"  [{keyword}]: {len(models)} 个 (50<={DOWNLOAD_MAX})")
+    return models
 
 
 def check_architecture(model_id: str) -> tuple:
     """检查模型架构"""
     try:
-        cfg_url = f"https://huggingface.co/{model_id}/raw/main/config.json"
+        cfg_url = f"{MODELSCOPE_API}/models/{model_id}/repo?Revision=master&FilePath=config.json"
         resp = requests.get(cfg_url, timeout=10)
         if resp.status_code == 200:
             cfg = resp.json()
@@ -145,7 +184,13 @@ def check_architecture(model_id: str) -> tuple:
 
 
 def normalize_model_url(model_url: str) -> str:
-    """标准化 URL 格式 - 直接返回 HuggingFace URL"""
+    """标准化 URL 格式"""
+    if '/models/' in model_url:
+        return model_url
+    if 'modelscope.cn/' in model_url:
+        parts = model_url.split('modelscope.cn/')
+        if len(parts) == 2:
+            return f"https://www.modelscope.cn/models/{parts[1]}"
     return model_url
 
 
@@ -276,7 +321,7 @@ def submit_model(model_url: str) -> tuple:
 # 主流程
 # ============================================================
 
-def run_pipeline(submit_limit: int = 30):
+def run_pipeline(submit_limit: int = 2):
     """完整流程：搜索→筛选→提交（只针对目标GPU）"""
     init_db()
     log("=" * 50)
@@ -285,49 +330,58 @@ def run_pipeline(submit_limit: int = 30):
     log(f"提交限制: {submit_limit} 个")
 
     # 1. 搜索
-    log("\n--- 阶段1: 搜索 HuggingFace ---")
+    log("\n--- 阶段1: 搜索 ModelScope ---")
     seen = set()
     all_models = []
     for kw in SEARCH_KEYWORDS:
-        models = search_models(kw, limit=100)
+        models = search_models(kw)
         for m in models:
             mid = m.get('id', '')
             if mid and mid not in seen:
                 seen.add(mid)
-                downloads = m.get('downloads', 0)
-                if downloads >= 50:
-                    all_models.append({
-                        'model_id': mid,
-                        'url': f"https://huggingface.co/{mid}",
-                        'downloads': downloads,
-                    })
+                all_models.append({
+                    'model_id': mid,
+                    'url': f"https://modelscope.cn/{mid}",
+                    'downloads': m.get('downloads', 0),
+                })
         time.sleep(0.3)
-    log(f"搜索完成: {len(seen)} 个唯一模型, {len(all_models)} 个下载量>=50")
+    log(f"搜索完成: {len(seen)} 个唯一模型, {len(all_models)} 个下载量{DOWNLOAD_MIN}-{DOWNLOAD_MAX}")
 
-    # 2. 格式筛选（只保留HuggingFace格式，排除GGUF）
+    # 2. 格式筛选（排除 GGUF/GPTQ/AWQ）
     log("\n--- 阶段2: 格式筛选 ---")
     hf_models = []
-    gguf_skipped = 0
+    format_skipped = 0
+    SKIP_FORMATS = ['GGUF', 'GPTQ', 'AWQ']
     for m in all_models:
-        mid = m['model_id'].upper()
-        if 'GGUF' in mid:
-            gguf_skipped += 1
-            log(f"  ✗ {m['model_id']}: GGUF格式，跳过")
-        else:
+        mid_upper = m['model_id'].upper()
+        skip = False
+        for fmt in SKIP_FORMATS:
+            if fmt in mid_upper:
+                format_skipped += 1
+                log(f"  x {m['model_id']}: {fmt}格式，跳过")
+                skip = True
+                break
+        if not skip:
             hf_models.append(m)
-    log(f"格式筛选: {len(hf_models)} 通过 (HuggingFace), {gguf_skipped} 跳过 (GGUF)")
+    log(f"格式筛选: {len(hf_models)} 通过, {format_skipped} 跳过 (GGUF/GPTQ/AWQ)")
 
-    # 3. 架构筛选
+    # 3. 架构筛选（排除 Qwen3.5 在 Iluvatar 上不支持）
     log("\n--- 阶段3: 架构筛选 ---")
     arch_passed = []
     arch_rejected = 0
+    SKIP_ARCHS = ['qwen3_5', 'Qwen3_5', 'Qwen3.5']
     for m in hf_models:
         ok, reason = check_architecture(m['model_id'])
-        if ok:
-            arch_passed.append(m)
-        else:
+        if not ok:
             arch_rejected += 1
-            log(f"  ✗ {m['model_id']}: {reason}")
+            log(f"  x {m['model_id']}: {reason}")
+            continue
+        arch_str = str(reason).upper()
+        if any(a.upper() in arch_str for a in SKIP_ARCHS):
+            arch_rejected += 1
+            log(f"  x {m['model_id']}: Qwen3.5(Iluvatar不支持)")
+            continue
+        arch_passed.append(m)
         time.sleep(0.15)
     log(f"架构筛选: {len(arch_passed)} 通过, {arch_rejected} 拒绝")
 
@@ -355,6 +409,10 @@ def run_pipeline(submit_limit: int = 30):
         if check_my_submitted(model_id):
             continue
 
+        # 检查是否已知失败（避免重复提交）
+        if is_model_failed(model_id):
+            continue
+
         to_submit.append(m)
         if len(to_submit) >= min(submit_limit, available):
             break
@@ -371,10 +429,13 @@ def run_pipeline(submit_limit: int = 30):
             log(f"  ✅ {m['model_id']}")
             db_conn.execute(
                 'INSERT OR REPLACE INTO submitted VALUES (?,?,?,?)',
-                (m['model_id'], gpu, str(task_id), datetime.now().isoformat())
+                (m['model_id'], TARGET_GPU, str(task_id), datetime.now().isoformat())
             )
         else:
             log(f"  ❌ {m['model_id']}: {msg}")
+            # 永久失败类型记录到 failed 表
+            if any(kw in str(msg) for kw in ['保护期', '白名单', '唯一性']):
+                record_failed(m['model_id'], msg)
         time.sleep(0.5)
 
     log(f"  提交完成: {submitted}/{len(to_submit)}")
@@ -435,7 +496,7 @@ class AgentHandler(BaseHTTPRequestHandler):
             if content_len > 0:
                 body = json.loads(self.rfile.read(content_len))
 
-            limit = body.get('limit', 30)
+            limit = body.get('limit', 2)
 
             self._json({'status': 'started', 'gpu': TARGET_GPU, 'limit': limit})
 
@@ -492,10 +553,9 @@ class AgentHandler(BaseHTTPRequestHandler):
 
         # 3. ModelHub 查询 API
         try:
-            token = list(ACCOUNTS.values())[0]
             resp = requests.get(
                 'https://modelhub.org.cn/api/adapt/task/page',
-                headers={'Xc-Token': token, 'Accept': 'application/json'},
+                headers={'Xc-Token': TARGET_TOKEN, 'Accept': 'application/json'},
                 params={'current': 1, 'pageSize': 1, 'onlyMine': 'true'},
                 timeout=10,
             )
@@ -510,17 +570,16 @@ class AgentHandler(BaseHTTPRequestHandler):
 
         # 4. ModelHub 提交 API (dry test)
         try:
-            token = list(ACCOUNTS.values())[0]
             resp = requests.post(
                 'https://modelhub.org.cn/api/adapt/task/add',
-                headers={'Xc-Token': token, 'Accept': 'application/json', 'Content-Type': 'application/json'},
+                headers={'Xc-Token': TARGET_TOKEN, 'Accept': 'application/json', 'Content-Type': 'application/json'},
                 json={
                     'modelAddress': 'https://www.modelscope.cn/models/Qwen/Qwen3-8B',
                     'taskType': 'text-generation',
-                    'targetGpu': 'Kunlunxin_p-800',
+                    'targetGpu': TARGET_GPU,
                     'framework': 'vllm',
                     'strategyId': STRATEGY_ID,
-                    'configParams': 'framework: vllm\n',
+                    'configParams': build_config_params(),
                 },
                 timeout=10,
             )
@@ -614,10 +673,9 @@ def main():
 
         # 3. ModelHub 查询
         try:
-            token = list(ACCOUNTS.values())[0]
             resp = requests.get(
                 'https://modelhub.org.cn/api/adapt/task/page',
-                headers={'Xc-Token': token, 'Accept': 'application/json'},
+                headers={'Xc-Token': TARGET_TOKEN, 'Accept': 'application/json'},
                 params={'current': 1, 'pageSize': 1, 'onlyMine': 'true'},
                 timeout=10
             )
@@ -631,15 +689,14 @@ def main():
 
         # 4. ModelHub 提交
         try:
-            token = list(ACCOUNTS.values())[0]
             resp = requests.post(
                 'https://modelhub.org.cn/api/adapt/task/add',
-                headers={'Xc-Token': token, 'Accept': 'application/json', 'Content-Type': 'application/json'},
+                headers={'Xc-Token': TARGET_TOKEN, 'Accept': 'application/json', 'Content-Type': 'application/json'},
                 json={
                     'modelAddress': 'https://www.modelscope.cn/models/Qwen/Qwen3-8B',
-                    'taskType': 'text-generation', 'targetGpu': 'Kunlunxin_p-800',
+                    'taskType': 'text-generation', 'targetGpu': TARGET_GPU,
                     'framework': 'vllm', 'strategyId': STRATEGY_ID,
-                    'configParams': 'framework: vllm\n',
+                    'configParams': build_config_params(),
                 }, timeout=10
             )
             data = resp.json()
@@ -656,7 +713,7 @@ def main():
         try:
             state['running'] = True
             state['last_run'] = datetime.now().isoformat()
-            count = run_pipeline(submit_limit=30)
+            count = run_pipeline(submit_limit=2)
             state['last_result'] = {'submitted': count, 'success': True}
         except Exception as e:
             log(f"流程异常: {traceback.format_exc()}")