feat: switch framework to vllm_fix_tokenizer

main.py

@@ -247,10 +247,10 @@ def check_queue_available() -> int:
 
 
 def build_config_params() -> str:
-    """构建 YAML 配置 - vllm"""
+    """构建 YAML 配置 - vllm_fix_tokenizer"""
     params = {
-        'framework': 'vllm',
+        'framework': 'vllm_fix_tokenizer',
-        'nv_framework': 'vllm',
+        'nv_framework': 'vllm_fix_tokenizer',
         'api': 'completion',
         'max_tokens': 1024,
         'temperature': 0.7,
@@ -295,7 +295,7 @@ def submit_model(model_url: str) -> tuple:
         'modelAddress': normalize_model_url(model_url),
         'taskType': 'text-generation',
         'targetGpu': TARGET_GPU,
-        'framework': 'vllm',
+        'framework': 'vllm_fix_tokenizer',
         'strategyId': STRATEGY_ID,
         'configParams': build_config_params(),
     }
@@ -314,7 +314,7 @@ def submit_model(model_url: str) -> tuple:
 # 主流程
 # ============================================================
 
-def run_pipeline(submit_limit: int = 2):
+def run_pipeline(submit_limit: int = 5):
     """完整流程：搜索→筛选→提交（只针对目标GPU）"""
     init_db()
     log("=" * 50)
@@ -344,7 +344,7 @@ def run_pipeline(submit_limit: int = 2):
     log("\n--- 阶段2: 格式筛选 ---")
     hf_models = []
     format_skipped = 0
-    SKIP_FORMATS = ['GPTQ', 'AWQ']
+    SKIP_FORMATS = ['GGUF', 'GPTQ', 'AWQ']
     for m in all_models:
         mid_upper = m['model_id'].upper()
         skip = False
@@ -356,7 +356,7 @@ def run_pipeline(submit_limit: int = 2):
                 break
         if not skip:
             hf_models.append(m)
-    log(f"格式筛选: {len(hf_models)} 通过, {format_skipped} 跳过 (GPTQ/AWQ)")
+    log(f"格式筛选: {len(hf_models)} 通过, {format_skipped} 跳过 (GGUF/GPTQ/AWQ)")
 
     # 3. 架构筛选（只保留有标准config.json的模型，排除无效格式）
     log("\n--- 阶段3: 架构检查 ---")
@@ -487,7 +487,7 @@ class AgentHandler(BaseHTTPRequestHandler):
             if content_len > 0:
                 body = json.loads(self.rfile.read(content_len))
 
-            limit = body.get('limit', 2)
+            limit = body.get('limit', 5)
 
             self._json({'status': 'started', 'gpu': TARGET_GPU, 'limit': limit})
 
@@ -568,7 +568,7 @@ class AgentHandler(BaseHTTPRequestHandler):
                     'modelAddress': 'https://www.modelscope.cn/models/Qwen/Qwen3-8B',
                     'taskType': 'text-generation',
                     'targetGpu': TARGET_GPU,
-                    'framework': 'vllm',
+                    'framework': 'vllm_fix_tokenizer',
                     'strategyId': STRATEGY_ID,
                     'configParams': build_config_params(),
                 },
@@ -686,7 +686,7 @@ def main():
                 json={
                     'modelAddress': 'https://www.modelscope.cn/models/Qwen/Qwen3-8B',
                     'taskType': 'text-generation', 'targetGpu': TARGET_GPU,
-                    'framework': 'vllm', 'strategyId': STRATEGY_ID,
+                    'framework': 'vllm_fix_tokenizer', 'strategyId': STRATEGY_ID,
                     'configParams': build_config_params(),
                 }, timeout=10
             )
@@ -704,7 +704,7 @@ def main():
         try:
             state['running'] = True
             state['last_run'] = datetime.now().isoformat()
-            count = run_pipeline(submit_limit=2)
+            count = run_pipeline(submit_limit=5)
             state['last_result'] = {'submitted': count, 'success': True}
         except Exception as e:
             log(f"流程异常: {traceback.format_exc()}")