""" xc_validation_strategy — 主入口 启动后针对 GPU_JOBS 中配置的 GPU 卡分别批量提交各自筛选出的模型验证任务 （当前仅提交 ppu_zw_810e，其余 4 张卡 Biren_166m/Cambricon_mlu-370-x8/MetaX_c-500/ Kunlunxin_p-800 的 config_content 模板和模型列表仍保留在代码中，未列入本次 GPU_JOBS） （/adminApi/async/task/create-contest-task， Bearer Token 认证），之后保持 HTTP 服务存活。 账号额度自动重试：如果某个模型提交时命中"当前等待中或运行中的异步模型验证 任务数量已达上限"（账号额度已满），不算永久失败，会被留到下一轮；额度耗尽后 本进程会原地等待 30 分钟，再自动重试所有因额度问题未提交成功的模型，如此循环， 直至全部提交成功或进程被平台关闭——不需要重新部署新策略，循环逻辑在本进程内完成。 非额度原因的失败（如模型已在验证中等）不会重试。 同时暴露 /health（K8s 探活）和 /status（运行状态，含当前轮次/待重试数/下次重试时间）。 """ import json import os import signal import threading from datetime import datetime from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer from typing import List, Tuple import requests # ══════════════════════════════════════════════════════════ # 配置（全部从环境变量读取，不硬编码敏感信息） # ══════════════════════════════════════════════════════════ BASE_URL = os.environ.get("BASE_URL", "https://modelhub.org.cn") SUBMIT_ENDPOINT = "/adminApi/async/task/create-contest-task" # 通过 curl -X POST https://modelhub.org.cn/adminApi/user/login 获取后填入 AUTH_TOKEN = "eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3OTAwNjkxMjAsImlhdCI6MTc4OTQ2NDMyMH0.KzJac6ddaZdtLvjD6ZnoK1PNEKFoXdyDn9Hh4FxU9ic" CONTEST_API_TOKEN = "ef1ef82f3c9efee413d602345fbe224d" CONTRIBUTORS = "zhoushasha" TASK_TYPE = "text-generation" STRATEGY_ID = os.environ.get("STRATEGY_ID", "") # 平台自动注入，无需修改 HTTP_HOST = "0.0.0.0" HTTP_PORT = 8080 # ══════════════════════════════════════════════════════════ # 各 GPU 的模型列表（来自 filter_verified_models 脚本的筛选结果） # ══════════════════════════════════════════════════════════ BIREN_MODELS = [ "ApolloRaines/Phi-4-mini-Instruct-Desyced", "allenai/OLMo-2-0425-1B", "Free2035/4QDR_4B_AD_Thinker_V1", "aifoundry-org/OLMo-7B-0424-hf-Quantized", "lugman-madhiai/Qwen3-4B-MHS-1.1", "zypchn/BehChat-SFT-v4", "NoesisLab/Kai-30B-Instruct", "barandinho/Qwen3-30B-A3B-FIRST-STAGE-SFT", "m-a-p/OpenLLaMA-Reproduce-218.1B", "xiaolesu/Qwen3-8B-Herald-SFT", "WhiteRabbitNeo/WhiteRabbitNeo-33B-v1.5", "saleh1312/orph_3.07225", "vanta-research/atom-olmo3-7b", "ZhipuAI/LongCite-glm4-9b", "Xlnk/LFM2-2.6B-Exp-GGuf", "Tesslate/UIGEN-T1.1-Qwen-14B", "jondurbin/bagel-dpo-34b-v0.2", "zhengr/MixTAO-7Bx2-MoE-Instruct-v5.0", "allenai/Olmo-3.1-32B-Instruct", "bjaidi/Phi-3-medium-128k-instruct-awq", "xing720310/qwen3-14b", "IntelLabs/sqft-mistral-7b-v0.3-50-base-gptq", "hariharanv04/qwen2.5-coder-14b-metadata-merged", "junfengzhou/qwen3-14b-rl", "ronnywebdevs1/Affine-P011-5CkU7wLMWXPs6TdSsMf8eEYCVAbPLyNmYg9PPx1Uds8toKra", "kennedyantonio0301/Affine-Tensor-h3-5EkdoaCmEpFffUjDpLhDMzEDR4kptaEzpTPYCP1uL2sbct8C", "julep-ai/dolphin-2.9-llama3-70b-awq", "cortexso/gemma3", "jacob-ml/jacob-24b", "lyraaaa/neuralese-sft-pretrain-v2", "ai-sage/GigaChat-20B-A3B-instruct-bf16", "PJMixers-Dev/gemma-3-1b-it-fixed", "commotion/svara_finetune_v1", "geoffmunn/Qwen3-14B-f16", "geoffmunn/Qwen3-32B-f16", "TeichAI/Nemotron-Cascade-14B-Thinking-Claude-4.5-Opus-High-Reasoning-Distill", "prithivMLmods-llamafile/SmolLM2-1.7B-Instruct-llamafile", "prithivMLmods-llamafile/Llama-3.2-8B-llamafile-200K", "llamafile-club/SmolLM-135M-Instruct-Llamafile", "prithivMLmods/Sombrero-QwQ-32B-Elite9", "prithivMLmods-llamafile/Aya-Expanse-8B-llamafile", "llamafile-club/SmolLM-135M-Llamafile", "prithivMLmods/Sombrero-QwQ-32B-Elite10-Fixed", "prithivMLmods-llamafile/Qwen2.5-Coder-1.5B-llamafile", "TeichAI/Qwen3-14B-Polaris-Alpha-Distill", "okwinds/MiroThinker-14B-DPO-v0.1", "sanbuphy/tianji-wish2-14b", "codefuse-ai/CodeFuse-StarCoder2-15B", "AI-ModelScope/txgemma-27b-chat", "Shanghai_AI_Laboratory/internlm3-8b-instruct-awq", "XGenerationLab/XiYanSQL-QwenCoder-32B-2412", "vllm-ascend/gemma-1.1-2b-it", "OpenBuddy/openbuddy-qwen1.5-32b-v21.2-32k", "OpenBuddy/openbuddy-thinker-32b-v26-preview", "OpenBuddy/openbuddy-qwen1.5-32b-v21.1-32k", "TechxGenus-MS/CodeGemma-7b", "OpenBuddy/openbuddy-qwq-32b-v25.2q-200k", "unsloth/Qwen3-30B-A3B", "OpenBuddy/openbuddy-qwq-32b-v25.1-200k", "OpenBuddy/openbuddy-r1-32b-v24.1-200k", "iic/ERank-14B", "OpenBuddy/openbuddy-yi1.5-34b-v21.2-32k", "LGAI-EXAONE/EXAONE-Deep-32B", "OpenBuddy/openbuddy-qwq-32b-v24.2-200k", "unsloth/Phi-3-mini-4k-instruct-v0", "argilla/notux-8x7b-v1", "voidful/qd-phi-1_5", "Shanghai_AI_Laboratory/internlm2-math-base-20b", "Shanghai_AI_Laboratory/internlm2-math-plus-20b", "TechxGenus-MS/starcoder2-15b-instruct", "Shanghai_AI_Laboratory/internlm2-base-20b", "m-a-p/OpenLLaMA-Reproduce-872.42B", "m-a-p/OpenLLaMA-Reproduce-973.08B", "Shanghai_AI_Laboratory/OREAL-32B", "YOYO-AI/Qwen3-30B-A3B-CoderThinking-YOYO-linear", "ticoAg/Qwen-1_8B-Chat-Int4-awq", "smirki/UIGEN-T1.1-Qwen-14B", "prithivMLmods/Qwen2.5-32B-DeepSeek-R1-Instruct", "sail/Sailor2-20B-128K", "xverse/XVERSE-65B", "Shanghai_AI_Laboratory/internlm2_5-20b-chat", "TeleAI/TeleChat-52B", "modelscope/Llama-2-70b-ms", "Shanghai_AI_Laboratory/internlm2-20b", "ai-modelscope/Llama-3_1-Nemotron-51B-Instruct", "zhuangxialie/Phi-3-Chinese-ORPO", "openai-mirror/gpt-oss-safeguard-20b", "ByteDance-Seed/Seed-OSS-36B-Instruct", "Shanghai_AI_Laboratory/internlm-chat-20b", "TurkuNLP/bloom-finnish-176b", "openai-mirror/gpt-oss-120b", "vllm-ascend/QwQ-32B-W8A8", "mistralai/Mistral-Small-24B-Instruct-2501", "Shanghai_AI_Laboratory/internlm-20b", "ZhipuAI/GLM-4-32B-0414", ] CAMBRICON_MODELS = [ "Xlnk/LFM2-2.6B-Exp-GGuf", ] METAX_MODELS = [ "ApolloRaines/Phi-4-mini-Instruct-Desyced", "robertspumiaca1975/Qwen2.5-Coder-14B-n8n-Workflow-Generator", "barandinho/Qwen3-30B-A3B-FIRST-STAGE-SFT-V2", "hotmailuser/QwenSlerp2-14B", "WhiteRabbitNeo/WhiteRabbitNeo-33B-v1.5", "madox81/SmolLM2-135M-cybersecurity-lora-merged", "Xlnk/LFM2-2.6B-Exp-GGuf", "Tesslate/UIGEN-T1.1-Qwen-14B", "jondurbin/bagel-dpo-34b-v0.2", "zhengr/MixTAO-7Bx2-MoE-Instruct-v5.0", "bjaidi/Phi-3-medium-128k-instruct-awq", "jacob-ml/jacob-24b", "geoffmunn/Qwen3-32B-f16", "TorpedoSoftware/Luau-Devstral-24B-Instruct-v0.2", "TeichAI/Nemotron-Cascade-14B-Thinking-Claude-4.5-Opus-High-Reasoning-Distill", "tongzang/Qwen2.5-7b-lora-law", "prithivMLmods/Sombrero-QwQ-32B-Elite9", "prithivMLmods/Sombrero-QwQ-32B-Elite10-Fixed", "TeichAI/Qwen3-14B-Polaris-Alpha-Distill", "okwinds/MiroThinker-14B-DPO-v0.1", "sanbuphy/tianji-wish2-14b", "YOYO-AI/YOYO-O1-14B", "LLM-Research/Meta-Llama-3.1-405B", "LLM-Research/Meta-Llama-3-70B", "LLM-Research/Meta-Llama-3.1-70B", "Qwen/Qwen-72B-Chat", "Qwen/Qwen3-Coder-480B-A35B-Instruct", "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "codefuse-ai/CodeFuse-StarCoder2-15B", "AI-ModelScope/txgemma-27b-chat", "XGenerationLab/XiYanSQL-QwenCoder-32B-2412", "OpenBuddy/openbuddy-qwen1.5-32b-v21.2-32k", "OpenBuddy/openbuddy-thinker-32b-v26-preview", "OpenBuddy/openbuddy-qwen1.5-32b-v21.1-32k", "OpenBuddy/openbuddy-qwq-32b-v25.2q-200k", "unsloth/Qwen3-30B-A3B", "OpenBuddy/openbuddy-qwq-32b-v25.1-200k", "OpenBuddy/openbuddy-r1-32b-v24.1-200k", "iic/ERank-14B", "OpenBuddy/openbuddy-yi1.5-34b-v21.2-32k", "LGAI-EXAONE/EXAONE-Deep-32B", "OpenBuddy/openbuddy-qwq-32b-v24.2-200k", "unsloth/Phi-3-mini-4k-instruct-v0", "argilla/notux-8x7b-v1", "voidful/qd-phi-1_5", "TechxGenus-MS/starcoder2-15b-instruct", "m-a-p/OpenLLaMA-Reproduce-872.42B", "m-a-p/OpenLLaMA-Reproduce-973.08B", "Shanghai_AI_Laboratory/OREAL-32B", "YOYO-AI/Qwen3-30B-A3B-CoderThinking-YOYO-linear", "smirki/UIGEN-T1.1-Qwen-14B", "sthenno-com/miscii-14b-0130", "prithivMLmods/Qwen2.5-32B-DeepSeek-R1-Instruct", "sail/Sailor2-20B-128K", "xverse/XVERSE-65B", "TeleAI/TeleChat-52B", "modelscope/Llama-2-70b-ms", "Shanghai_AI_Laboratory/internlm2-20b", "ai-modelscope/Llama-3_1-Nemotron-51B-Instruct", "aJupyter/EmoLLM_Qwen2-7B-Instruct_lora", "zhuangxialie/Phi-3-Chinese-ORPO", "vllm-ascend/QwQ-32B-W8A8", "ZhipuAI/GLM-4-32B-0414", ] HYGON_MODELS = [ "ApolloRaines/Phi-4-mini-Instruct-Desyced", ] KUNLUNXIN_MODELS = [ "Xlnk/LFM2-2.6B-Exp-GGuf", ] PPU_MODELS = [ "OuteAI/Lite-Mistral-150M-v2-Instruct", "eric0009/yi-ko-6b-text2sql", "ai-forever/mGPT-1.3B-bashkir", "ApolloRaines/Phi-4-mini-Instruct-Desyced", "eekay/Llama-3.1-8B-Instruct-noised-np0.15-emb-s48", "eekay/Llama-3.1-8B-Instruct-noised-np0.15-emb-s49", "cortexso/simplescaling-s1", "BrainDelay/Siren", "sbintuitions/sarashina2.2-3b-instruct-v0.1", "robertspumiaca1975/Qwen2.5-Coder-14B-n8n-Workflow-Generator", "AtAndDev/ShortKing-3b-v0.2", "athirdpath/Iambe-RP-cDPO-20b", "belweave/kai-2", "unsloth/Qwen2.5-Coder-14B-Instruct", "dphn/dolphin-2.7-mixtral-8x7b", "adeljebali/llama3.1-gec-strict", "xxrickyxx/Ailo152m-events-en", "RedHatAI/starcoder2-7b-quantized.w8a8", "RedHatAI/granite-3.1-2b-instruct-quantized.w4a16", "julep-ai/dolphin-2.9.1-llama-3-70b-awq", "OpenBuddy/openbuddy-deepseek-67b-v18.1-4k-gptq", "dessertlab/offensive-powershell-CodeGPT-small", "misterJB/atlas-field-528hz", "tiiuae/Falcon3-10B-Base", "Jackrong/gpt-oss-120b-Distill-Llama3.1-8B-v3", "TheBloke/guanaco-65B-HF", "jondurbin/airoboros-33b-gpt4-1.3", "h2oai/h2ogpt-4096-llama2-70b", "jondurbin/airoboros-65b-gpt4-1.3", "jondurbin/airoboros-l2-70b-gpt4-2.0", "ICBU-NPU/FashionGPT-70B-V1.2", "jukofyork/Dark-Miqu-70B", "alnrg2arg/blockchainlabs_joe_bez_seminar", "facebook/opt-66b", "abchbx/qwen_1.8B_Muice-Dataset_FULL", "LumiOpen/Viking-33B", "adamo1139/Yi-34B-200K-AEZAKMI-RAW-1701", "mesolitica/Malaysian-TTS-4B-v0.1", "TomGrc/FusionNet_passthrough", "YOYO-AI/Qwen3-30B-A3B-YOYO-V5", "m-a-p/OpenLLaMA-Reproduce-536.87B", "m-a-p/OpenLLaMA-Reproduce-1291.85B", "KnutJaegersberg/Deacon-34B", "SenseLLM/ReflectionCoder-DS-33B", "KOREAson/KO-REAson-AX3_1-35B-1009", "dphn/dolphin-2.9.1-mixtral-1x22b", "jondurbin/airoboros-33b-gpt4-1.4", "TomGrc/FusionNet_passthrough_v0.1", "Mozilla/Mistral-7B-Instruct-v0.2-llamafile", "suayptalha/Luminis-phi-4", "casperhansen/llama-3.3-70b-instruct-awq", "HIT-SCIR/Chinese-Mixtral-8x7B", "Shanghai_AI_Laboratory/internlm2-wqx-20b", "unsloth/Qwen2.5-Coder-32B-Instruct", "BSC-LT/ALIA-40b", "Shanghai_AI_Laboratory/internlm2-7b", "Shanghai_AI_Laboratory/internlm2-chat-7b", ] # 本次提交第二十四轮过滤结果 MetaX_c-500(63) / Kunlunxin_p-800(1) / # Cambricon_mlu-370-x8(1) / Biren_166m(95)，加上已更新的 ppu_zw_810e(57)，共217个。 # hygon_k100-ai 本轮筛出1个，config 分支与 HYGON_MODELS 列表已备好但按要求不提交； # Iluvatar_bi-150 本轮筛出36个，本仓库 framework=vllm 而现有 iluvatar 镜像均为 # llamacpp/GGUF，缺 vllm 版镜像地址，无 config 分支，同样不提交。 GPU_JOBS: List[Tuple[str, List[str]]] = [ ("MetaX_c-500", METAX_MODELS), ("Kunlunxin_p-800", KUNLUNXIN_MODELS), ("Cambricon_mlu-370-x8", CAMBRICON_MODELS), ("Biren_166m", BIREN_MODELS), ("ppu_zw_810e", PPU_MODELS), ] TOTAL_MODELS = sum(len(models) for _, models in GPU_JOBS) # ══════════════════════════════════════════════════════════ # 全局状态（供 /status 展示） # ══════════════════════════════════════════════════════════ _state = { "strategy_id": STRATEGY_ID, "phase": "starting", # starting | submitting | waiting_retry | done | error "total": TOTAL_MODELS, "submitted": 0, "failed": 0, "per_gpu": {gpu: 0 for gpu, _ in GPU_JOBS}, "started_at": None, "finished_at": None, "round": 0, # 当前是第几轮提交 "quota_blocked_remaining": 0, # 因额度上限暂未提交成功、等待下一轮重试的模型数 "next_retry_at": None, # 下一轮重试的预计时间（额度耗尽等待期间） } _shutdown = threading.Event() # ══════════════════════════════════════════════════════════ # HTTP 服务 # ══════════════════════════════════════════════════════════ class Handler(BaseHTTPRequestHandler): def do_GET(self): if self.path == "/health": self._json({"status": "ok"}) elif self.path == "/status": self._json(_state) else: self._json({"error": "not found"}, 404) def _json(self, body: dict, code: int = 200): payload = json.dumps(body, default=str).encode() self.send_response(code) self.send_header("Content-Type", "application/json") self.send_header("Content-Length", str(len(payload))) self.end_headers() self.wfile.write(payload) def log_message(self, fmt, *args): print(f"[http] {self.address_string()} {fmt % args}", flush=True) def _run_http(): server = ThreadingHTTPServer((HTTP_HOST, HTTP_PORT), Handler) server.timeout = 1 print(f"[http] 监听 {HTTP_HOST}:{HTTP_PORT}", flush=True) while not _shutdown.is_set(): server.handle_request() server.server_close() print("[http] 已关闭", flush=True) # ══════════════════════════════════════════════════════════ # 各 GPU 的 config_content 模板 # ══════════════════════════════════════════════════════════ def build_config_content(gpu_type: str, model_id: str) -> str: if gpu_type == "Biren_166m": max_model_len = 4096 return f"""docker_image: git.modelhub.org.cn:9443/enginex/xc-llm-biren166m:26.01 nv_docker_image: harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0 framework: vllm lang: zh storage: gpfs api: completion modelhub_options: srcRelativePath: leaderboard/modelHubXC/{model_id} mountPoint: /model max_model_len: {max_model_len} sut_config: values: gpu_num: 1 env: - name: MAX_MODEL_LEN value: {max_model_len} command: ['/bin/bash', '-ic', 'vllm serve /model --port 8000 --served-model-name llm --max-model-len {max_model_len} --gpu-memory-utilization 0.9 --enforce-eager --trust-remote-code -tp 1 --host 0.0.0.0'] ref_config: values: cpu_num: 2 gpu_num: 1 env: - name: MAX_MODEL_LEN value: {max_model_len} command: ['vllm', 'serve', '/model', '--port', '80', '--served-model-name', 'llm', '--max-model-len', '{max_model_len}', '--enforce-eager', '--trust-remote-code', '-tp', '1'] model: llm """ elif gpu_type == "Cambricon_mlu-370-x8": return f"""docker_image: harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-torch2.5.0-torchmlu1.24.1-ubuntu22.04-py310 nv_docker_image: harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0 framework: vllm storage: gpfs modelhub_options: srcRelativePath: leaderboard/modelHubXC/{model_id} mountPoint: /model sut_config: values: gpu_num: 1 env: - name: MAX_MODEL_LEN 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: - name: MAX_MODEL_LEN value: 8192 command: ["vllm", "serve", "/model", "--port", "80", "--served-model-name", "llm", "--max-model-len", "8192", "--trust-remote-code", "--dtype", "float16"] """ elif gpu_type == "MetaX_c-500": return f"""docker_image: git.modelhub.org.cn:9443/enginex-metax/vllm:0.9.1 nv_docker_image: harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0 framework: vllm lang: en storage: gpfs api: chat modelhub_options: srcRelativePath: leaderboard/modelHubXC/{model_id} mountPoint: /model max_model_len: 2048 sut_config: gpu_num: 1 values: command: ['/opt/conda/bin/vllm', 'serve', '/model', '--port', '20644', '--served-model-name', 'llm', '--max-model-len', '2048', '--gpu-memory-utilization', '0.9', '--enforce-eager', '--trust-remote-code' ,'-tp', '1'] ref_config: gpu_num: 1 values: command: ['vllm', 'serve', '/model', '--port', '80', '--served-model-name', 'llm', '--max-model-len', '2048', '--enforce-eager', '--trust-remote-code', '-tp', '1'] """ elif gpu_type == "Kunlunxin_p-800": return f"""docker_image: git.modelhub.org.cn:9443/enginex/xc-llm-kunlun nv_docker_image: harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0 framework: vllm lang: en storage: gpfs api: chat temperature: 0.4 repetition_penalty: 1.1 top_p: 0.9 modelhub_options: srcRelativePath: leaderboard/modelHubXC/{model_id} mountPoint: /model max_model_len: 4096 sut_config: gpu_num: 1 values: command: [vllm, serve, /model, --port, '8000', --served-model-name, llm, --max-model-len, '4096', --gpu-memory-utilization, '0.9', --enforce-eager, --trust-remote-code, -tp, '1'] ref_config: gpu_num: 1 values: command: [vllm, serve, /model, --port, '80', --served-model-name, llm, --max-model-len, '4096', --enforce-eager, --trust-remote-code, -tp, '1'] """ elif gpu_type == "hygon_k100-ai": return f""" docker_image: harbor.4pd.io/modelhubxc/enginex-hygon/vllm:0.9.2-patch-tokenizer nv_docker_image: harbor.4pd.io/modelhubxc/enginex-nvidia/vllm:0.11.0-patch-tokenizer framework: vllm storage: gpfs max_model_len: 4096 sut_config: gpu_num: 1 values: command: ['vllm', 'serve', '/model', '--port', '20644', '--served-model-name', 'llm', '--max-model-len', '4096', '--enforce-eager', '--trust-remote-code' ,'-tp', '1' ] ref_config: gpu_num: 1 values: command: ['vllm', 'serve', '/model', '--port', '80', '--served-model-name', 'llm', '--max-model-len', '4096', '--enforce-eager', '--trust-remote-code', '-tp', '1'] """ elif gpu_type == "ppu_zw_810e": return f"""gpu_type: ppu_zw_810e framework: vllm docker_image: harbor.4pd.io/hardcore-tech/asllm:1.10.1-pytorch2.10.0-ubuntu24.04-sail2.1.0-cuda13.0-sglang0.5.10-vllm0.19.0-py312 nv_docker_image: harbor-contest.4pd.io/sunruoxi/vllm-openai-fix-tokenizer:v0.11.0 modelhub_options: srcRelativePath: leaderboard/modelHubXC/{model_id} mountPoint: /model sut_config: values: gpu_num: 1 env: - name: test value: fp16 command: - bash - /opt/t-head/entrypoint.sh - python3 - -m - asllm.entrypoints.api_server - --model - /model - --port - '30000' - --host - 0.0.0.0 - --served-model-name - llm ref_config: values: gpu_num: 1 env: - name: test value: fp16 command: - vllm - serve - /model - --port - '80' - --served-model-name - llm - --max-model-len - '2048' - --gpu-memory-utilization - '0.9' - --enforce-eager - --trust-remote-code - -tp - '1' """ else: raise ValueError(f"未知的 GPU_TYPE: {gpu_type}") # ══════════════════════════════════════════════════════════ # 业务逻辑 # ══════════════════════════════════════════════════════════ # 账号"等待中/运行中"任务数已达上限时平台返回的业务错误信息（子串匹配）； # 命中这个的模型不算永久失败，会在额度腾出空位后自动重试，不会被记作 failed QUOTA_FULL_MSG = "当前等待中或运行中的异步模型验证任务数量已达上限" # 额度耗尽后，隔多久自动重试一次剩余（因额度问题未提交成功）的模型 RETRY_INTERVAL_SECONDS = 30 * 60 # 30 分钟 def _submit_task(token: str, gpu_type: str, model_id: str) -> Tuple[bool, str, str]: headers = { "Content-Type": "application/json", "Authorization": f"Bearer {token}", } config_content = build_config_content(gpu_type, model_id) payload = { "contestApiToken": CONTEST_API_TOKEN, "contributors": CONTRIBUTORS, "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, "taskType": TASK_TYPE, }], } print(f"[payload] gpu={gpu_type} model={model_id}", flush=True) try: resp = requests.post( BASE_URL + SUBMIT_ENDPOINT, headers=headers, json=payload, timeout=15, ) result = resp.json() if result.get("code") == 0: task_id = result.get("data", {}).get("id", "") print(f"[worker] OK {model_id} (GPU={gpu_type}) task_id={task_id}", flush=True) return True, task_id, "" else: message = result.get("message") or "" print(f"[worker] FAIL {model_id} (GPU={gpu_type}): {message}", flush=True) return False, "", message except Exception as e: print(f"[worker] ERROR {model_id} (GPU={gpu_type}): {e}", flush=True) return False, "", str(e) def _run_worker(): _state["started_at"] = datetime.utcnow().isoformat() _state["phase"] = "submitting" successful: List[Tuple[str, str, str]] = [] token = AUTH_TOKEN print("[worker] 使用预设 Token，跳过登录", flush=True) # 待提交队列：保持 GPU_JOBS 里原有的 (gpu_type, model_id) 顺序 pending: List[Tuple[str, str]] = [ (gpu_type, model_id) for gpu_type, model_list in GPU_JOBS for model_id in model_list ] round_num = 0 while pending and not _shutdown.is_set(): round_num += 1 _state["round"] = round_num _state["phase"] = "submitting" _state["next_retry_at"] = None print( f"\n{'='*60}\n🚀 第 {round_num} 轮，待提交 {len(pending)} 个模型\n{'='*60}", flush=True, ) quota_blocked: List[Tuple[str, str]] = [] for gpu_type, model_id in pending: if _shutdown.is_set(): break ok, task_id, message = _submit_task(token, gpu_type, model_id) if ok: _state["submitted"] += 1 _state["per_gpu"][gpu_type] += 1 successful.append((task_id, gpu_type, model_id)) elif QUOTA_FULL_MSG in message: # 账号额度暂时满了，不算永久失败，留到下一轮重试 quota_blocked.append((gpu_type, model_id)) else: # 非额度原因失败（如重复提交等），不再重试 _state["failed"] += 1 pending = quota_blocked _state["quota_blocked_remaining"] = len(pending) # 每轮结束都把已成功的结果落盘一次，避免中途重启丢失记录 try: with open("submitted_validation_tasks.txt", "w", encoding="utf-8") as f: for tid, gpu, mid in successful: f.write(f"{tid}\t{gpu}\t{mid}\n") except Exception: pass if pending and not _shutdown.is_set(): next_retry = datetime.utcnow().timestamp() + RETRY_INTERVAL_SECONDS _state["next_retry_at"] = datetime.utcfromtimestamp(next_retry).isoformat() _state["phase"] = "waiting_retry" print( f"[worker] 第 {round_num} 轮结束：{len(pending)} 个模型因账号额度上限暂未提交，" f"{RETRY_INTERVAL_SECONDS // 60} 分钟后自动重试（不部署新策略，本进程内循环）...", flush=True, ) _shutdown.wait(RETRY_INTERVAL_SECONDS) _state["finished_at"] = datetime.utcnow().isoformat() _state["phase"] = "done" _state["quota_blocked_remaining"] = len(pending) print( f"[worker] 完成 submitted={_state['submitted']} failed={_state['failed']} " f"total={_state['total']} per_gpu={_state['per_gpu']} " f"仍因额度未提交(如遇shutdown中断)={len(pending)}", flush=True, ) # 提交完成后继续保持进程存活，等待平台停止 # ══════════════════════════════════════════════════════════ # 入口 # ══════════════════════════════════════════════════════════ def _handle_signal(signum, _frame): print(f"[main] 收到信号 {signum}，正在关闭...", flush=True) _shutdown.set() def main(): signal.signal(signal.SIGTERM, _handle_signal) signal.signal(signal.SIGINT, _handle_signal) # HTTP 服务线程 http_thread = threading.Thread(target=_run_http, daemon=False) http_thread.start() # 提交任务线程 worker_thread = threading.Thread(target=_run_worker, daemon=True) worker_thread.start() # 主线程等待 shutdown _shutdown.wait() print("[main] 等待 HTTP 服务关闭...", flush=True) http_thread.join(timeout=5) print("[main] 退出", flush=True) if __name__ == "__main__": main()