feat: switch framework to vllm_fix_tokenizer

main.py

@@ -157,7 +157,7 @@ def search_models(keyword: str) -> list:
 
 
 def check_architecture(model_id: str) -> tuple:
-    """检查模型架构"""
+    """检查模型是否有有效的 config.json（不限架构类型）"""
     try:
         cfg_url = f"{MODELSCOPE_API}/models/{model_id}/repo?Revision=master&FilePath=config.json"
         resp = requests.get(cfg_url, timeout=10)
@@ -165,20 +165,13 @@ def check_architecture(model_id: str) -> tuple:
             cfg = resp.json()
             archs = cfg.get('architectures', [])
             mtype = cfg.get('model_type', '')
-            arch_str = str(archs)
-            for special in SUPPORTED_SPECIAL_ARCHS:
-                if special in arch_str:
-                    return True, special
-            for kw in SUPPORTED_ARCH_KEYWORDS:
-                if kw in arch_str:
-                    return True, kw
-            if mtype in SUPPORTED_MODEL_TYPES:
-                return True, mtype
-            return False, f"arch={archs} type={mtype}"
+            if archs or mtype:
+                return True, f"arch={archs} type={mtype}"
+            return False, "empty config"
         mid_upper = model_id.upper()
-        if 'QWEN3' in mid_upper or 'QWEN2' in mid_upper:
-            return True, "GGUF"
-        return False, "no config, not Qwen"
+        if 'LLAMA' in mid_upper:
+            return True, "Llama(GGUF)"
+        return False, "no config"
     except Exception as e:
         return True, f"check error: {e}"
 
@@ -254,10 +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',
-        'nv_framework': 'vllm',
+        'framework': 'vllm_fix_tokenizer',
+        'nv_framework': 'vllm_fix_tokenizer',
         'api': 'completion',
         'max_tokens': 1024,
         'temperature': 0.7,
@@ -302,7 +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(),
     }
@@ -321,7 +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)
@@ -351,7 +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
@@ -363,7 +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 ---")
@@ -425,8 +418,9 @@ def run_pipeline(submit_limit: int = 2):
             submitted += 1
             log(f"  ✅ {m['model_id']}")
             db_conn.execute(
-                'INSERT OR REPLACE INTO submitted VALUES (?,?,?,?)',
-                (m['model_id'], TARGET_GPU, str(task_id), datetime.now().isoformat())
+                'INSERT OR REPLACE INTO submitted VALUES (?,?,?,?,?,?)',
+                (m['model_id'], TARGET_GPU, str(task_id), 'submitted',
+                 datetime.now().isoformat(), None)
             )
         else:
             log(f"  ❌ {m['model_id']}: {msg}")
@@ -493,7 +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})
 
@@ -574,7 +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(),
                 },
@@ -692,7 +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
             )
@@ -710,7 +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()}")