feat: open architecture filter, expand search keywords, let platform decide compatibility

main.py

@@ -31,7 +31,7 @@ PORT = 8080
 STRATEGY_ID = os.getenv("STRATEGY_ID", "")
 
 # 目标GPU
-TARGET_GPU = "Iluvatar_bi-150"
+TARGET_GPU = "Iluvatar_bi-100"
 
 # 账号Token
 TARGET_TOKEN = "f45f1aae2c094426be237c88b1085015"
@@ -47,7 +47,7 @@ SUPPORTED_SPECIAL_ARCHS = ['Eagle3Speculator', 'LlamaForCausalLMEagle3']
 MODELHUB_API = "https://modelhub.org.cn/api"
 
 # 搜索关键词
-SEARCH_KEYWORDS = ['qwen', 'Qwen2', 'Qwen3', 'Qwen3.5', 'Qwen1.5', 'Qwen-']
+SEARCH_KEYWORDS = ['qwen', 'Qwen2', 'Qwen3', 'Qwen3.5', 'Llama-3', 'Llama-3.1', 'Mistral', 'DeepSeek']
 
 # ============================================================
 # 全局状态
@@ -254,25 +254,26 @@ def check_queue_available() -> int:
 
 
 def build_config_params() -> str:
-    """构建 YAML 配置 - 完全匹配平台自动生成的格式（只支持vllm）"""
+    """构建 YAML 配置 - llama.cpp (支持 GGUF)"""
     params = {
-        'framework': 'vllm',
+        'framework': 'llama.cpp',
-        'nv_framework': 'vllm',
+        'nv_framework': 'llama.cpp',
         'api': 'completion',
         'max_tokens': 1024,
         'temperature': 0.7,
         'repetition_penalty': 1.2,
         'top_p': 0.9,
         'lang': 'zh',
-        'max_model_len': 2048,
+        'max_model_len': 4096,
         'sut_config': {
             'gpu_num': 1,
             'values': {
                 'command': [
-                    'vllm', 'serve', '/model', '--port', '8000',
+                    'llama-server', '--model', '/model', '--alias', 'llm',
-                    '--served-model-name', 'llm', '--max-model-len', '2048',
+                    '--threads', '20', '--n-gpu-layers', '999', '--prio', '3',
-                    '--dtype', 'auto', '--gpu-memory-utilization', '0.95',
+                    '--min_p', '0.01', '--ctx-size', '4096',
-                    '-tp', '1', '--enforce-eager', '--trust-remote-code',
+                    '--host', '0.0.0.0', '--port', '8000',
+                    '--jinja', '--flash-attn', 'off',
                 ]
             }
         },
@@ -280,9 +281,9 @@ def build_config_params() -> str:
             'gpu_num': 1,
             'values': {
                 'command': [
-                    'vllm', 'serve', '/model', '--port', '80',
+                    'llama-server', '--model', '/model', '--alias', 'llm',
-                    '--served-model-name', 'llm', '--max-model-len', '4096',
+                    '--threads', '20', '--n-gpu-layers', '999',
-                    '--enforce-eager', '--trust-remote-code', '-tp', '1',
+                    '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000',
                 ]
             }
         },
@@ -302,7 +303,7 @@ def submit_model(model_url: str) -> tuple:
         'modelAddress': normalize_model_url(model_url),
         'taskType': 'text-generation',
         'targetGpu': TARGET_GPU,
-        'framework': 'vllm',
+        'framework': 'llama.cpp',
         'strategyId': STRATEGY_ID,
         'configParams': build_config_params(),
     }
@@ -347,11 +348,11 @@ def run_pipeline(submit_limit: int = 2):
         time.sleep(0.3)
     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（排除 GGUF/GPTQ/AWQ）
+    # 2. 格式筛选（排除 GPTQ/AWQ，保留 GGUF 和 HuggingFace）
     log("\n--- 阶段2: 格式筛选 ---")
     hf_models = []
     format_skipped = 0
-    SKIP_FORMATS = ['GGUF', 'GPTQ', 'AWQ']
+    SKIP_FORMATS = ['GPTQ', 'AWQ']
     for m in all_models:
         mid_upper = m['model_id'].upper()
         skip = False
@@ -363,27 +364,16 @@ 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 (GGUF/GPTQ/AWQ)")
+    log(f"格式筛选: {len(hf_models)} 通过, {format_skipped} 跳过 (GPTQ/AWQ)")
 
-    # 3. 架构筛选（排除 Qwen3.5 在 Iluvatar 上不支持）
+    # 3. 架构筛选（参考检查，不做严格过滤，让平台决定兼容性）
-    log("\n--- 阶段3: 架构筛选 ---")
+    log("\n--- 阶段3: 架构检查 ---")
-    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 not ok:
-            arch_rejected += 1
+            log(f"  ! {m['model_id']}: {reason} (仍保留)")
-            log(f"  x {m['model_id']}: {reason}")
+        time.sleep(0.1)
-            continue
+    log(f"架构检查: {len(hf_models)} 个模型进入下一阶段")
-        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")
 
     # 4. 筛选并提交（只针对目标GPU）
     log(f"\n--- 阶段4: 筛选并提交 [{TARGET_GPU}] ---")
@@ -397,7 +387,7 @@ def run_pipeline(submit_limit: int = 2):
 
     # 筛选
     to_submit = []
-    for m in arch_passed:
+    for m in hf_models:
         model_id = m['model_id']
 
         # 检查全平台验证状态
@@ -577,7 +567,7 @@ class AgentHandler(BaseHTTPRequestHandler):
                     'modelAddress': 'https://www.modelscope.cn/models/Qwen/Qwen3-8B',
                     'taskType': 'text-generation',
                     'targetGpu': TARGET_GPU,
-                    'framework': 'vllm',
+                    'framework': 'llama.cpp',
                     'strategyId': STRATEGY_ID,
                     'configParams': build_config_params(),
                 },
@@ -695,7 +685,7 @@ def main():
                 json={
                     'modelAddress': 'https://www.modelscope.cn/models/Qwen/Qwen3-8B',
                     'taskType': 'text-generation', 'targetGpu': TARGET_GPU,
-                    'framework': 'vllm', 'strategyId': STRATEGY_ID,
+                    'framework': 'llama.cpp', 'strategyId': STRATEGY_ID,
                     'configParams': build_config_params(),
                 }, timeout=10
             )