""" xc_validation_strategy_gguf — 主入口 GGUF 模型下载 + 验证任务提交流水线（部署框架与 xc_validation_strategy 一致）。 启动后运行流水线：批量创建 GGUF 模型下载任务（最大并发 8）， 每个模型下载成功后立即提交 hygon / bi150 两个验证任务； 同时暴露 /health（K8s 探活）和 /status（运行状态）。 """ import json import os import re import signal import threading from datetime import datetime from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer from typing import Set import requests # ══════════════════════════════════════════════════════════ # 配置 # ══════════════════════════════════════════════════════════ BASE_URL = os.environ.get("BASE_URL", "https://modelhub.org.cn") LOGIN_ENDPOINT = "/adminApi/user/login" CREATE_DOWNLOAD_TASK_ENDPOINT = "/adminApi/async/task/model-download-task" SUBMIT_TEST_TASK_ENDPOINT = "/adminApi/async/task/create-contest-task" # 登录账号：启动时优先用账号密码换取新 token（流水线运行时间长，预设 token 可能过期） USER_ACCOUNT = "zhoushasha@4paradigm.com" USER_PASSWORD = "ganshenme0" # 通过 curl -X POST https://modelhub.org.cn/adminApi/user/login 获取后填入（登录失败时的回退） AUTH_TOKEN = "eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3ODQ1NDc1NDYsImlhdCI6MTc4Mzk0Mjc0Nn0.ZcOqcrfI22LPi4mGMnt164nZGhi61ZxtJGYsoO7fZdM" CONTEST_API_TOKEN = "ef1ef82f3c9efee413d602345fbe224d" HF_TOKEN = "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" CONTRIBUTORS = "zhoushasha" TASK_TYPE = "text-generation" STRATEGY_ID = os.environ.get("STRATEGY_ID", "") # 平台自动注入，无需修改 MAX_CONCURRENT_DOWNLOADS = 8 # 同时下载的模型数上限 CHECK_INTERVAL_SECONDS = 10 # 下载状态轮询间隔（秒） HTTP_HOST = "0.0.0.0" HTTP_PORT = 8080 # ══════════════════════════════════════════════════════════ # 模型列表 # ══════════════════════════════════════════════════════════ ALL_MODEL_IDS = [ "mradermacher/Llama-3.2-3B-Della-GGUF", "mradermacher/Aspire_V2_ALT-8B-Model_Stock-GGUF", "mradermacher/Aspire_V2_ALT_ROW-8B-Model_Stock-GGUF", "mradermacher/OTG1.6-Instruct-sft-dpo-7B-rc-2025-01-20-GGUF", "mradermacher/llama-3.2-3b-it-Ehealthcare-ChatBot-v3-GGUF", "mradermacher/nsfw_plz_gguf_me-GGUF", "mradermacher/mistral-7b-sci-arc_reasoning_v2-GGUF", "mradermacher/Jaja-small-GGUF", "mradermacher/euclid-3B-GGUF", "mradermacher/Llama-3.2-3B-Blend-GGUF", "mradermacher/Llama-3.1-8B-KG-Extraction-v2-GGUF", "mradermacher/Minerva-8b-GGUF", "mradermacher/RTLCoder-Deepseek_OB25-GGUF", "mradermacher/RTLCoder-Deepseek_OB50-GGUF", "mradermacher/MedGPT_Finetuned-GGUF", "mradermacher/mistral-irl-iter2-iterative-dpo-GGUF", "mradermacher/Llama-3.2-Taiwan-3B-Instruct-GGUF", "mradermacher/llama3-8b-breadcrumbs-ties-v4-GGUF", "mradermacher/llama3-8b-della-v2-GGUF", "mradermacher/hermes-GGUF", "mradermacher/Orpo-oman-ar-stablelm-2-chat-GGUF", "mradermacher/cocoruta-8b-llama3.1-GGUF", "mradermacher/FalconSlerp6-7B-GGUF", "mradermacher/mistral-7b-arc_reasoning-GGUF", "mradermacher/Llama3.1-8B-relu-stage-1-dolma-v1_7-50B-4096-GGUF", "mradermacher/mistral-7b-CoT-GGUF", "mradermacher/llama3-8b-dare-ties-v3-GGUF", "mradermacher/hopefully_humanish-rp-nsfw-test-v1-GGUF", "mradermacher/mistral-7b-lima_v2-GGUF", "mradermacher/Mistral-AFT-Off-Policy-GGUF", "mradermacher/Mistral-AFT-On-Policy-GGUF", "mradermacher/Llama-AFT-Off-Policy-GGUF", "mradermacher/Llama-AFT-On-Policy-GGUF", "mradermacher/OrangeJ-8B-Model_Stock-GGUF", "mradermacher/Llama3-8b-alpaca-GGUF", "mradermacher/hitchens-3-1-GGUF", "mradermacher/Yearn-8B-Model_Stock-GGUF", "mradermacher/DavidAU-Dark-Planet-of-Davids-8B-64k-GGUF", "mradermacher/Qwen-2.5-3B-Tiny-Story-GGUF", "mradermacher/CodeLlama-7b-Instruct-hf-SAP-RAP-GGUF", "mradermacher/R1-ImpishMind-8B-GGUF", "mradermacher/llama3-8b-slerp-v2-GGUF", "mradermacher/layer-skip-vanill-3.2-1b-GGUF", "mradermacher/AlexeyRyzhikov-Mistral-7b-TXT_to_JSON-V5.2-GGUF", "mradermacher/SJT-2.5B-GGUF", "mradermacher/liujx-78k-GGUF", "mradermacher/Vangelus-Secundus-GGUF", "mradermacher/bloomz-7b1-p3-GGUF", "mradermacher/Llama-3.1-8B-price-GGUF", "mradermacher/asm2asm-deepseek-1.3b-500k-mac-x86-O0-arm-2-GGUF", "mradermacher/SJT-2.3B-GGUF", "mradermacher/Llama-3-ELYZA-JP-8B-ojousama-chosen-after-SFTboth-GGUF", "mradermacher/LLaMA-3-8B-SFR-SFT-R-GGUF", "mradermacher/ELN-Llama-1B-base-GGUF", "mradermacher/Qwen0.5b-RagSemanticChunker-GGUF", "mradermacher/Llama-3-ELYZA-JP-8B-normal-chosen-after-SFTboth-GGUF", "mradermacher/LLaMA-3-8B-SFR-Iterative-DPO-Concise-R-GGUF", "mradermacher/Qwen2.7-7B-Instruct-QwQ-PRIME-1k-GGUF", "mradermacher/flora-v1-GGUF", "mradermacher/exaone-3.5-2.4b-instruct-dacon-llm2-GGUF", "mradermacher/RPMash-8B-Model_Stock-GGUF", "mradermacher/qwen2.5-0.5B_ichikara_4802-GGUF", "mradermacher/payroll-teacher-model-2-GGUF", "mradermacher/Phi-3.5-mini-instruct-3x-v1-GGUF", "mradermacher/ImmyV2.7-GGUF", "mradermacher/Qwen2.5-3B-Renoia-GGUF", "mradermacher/ImmyV2.5-GGUF", "mradermacher/mistral-7b-CoT_v2-GGUF", "mradermacher/gpt2_individuated_zero_chaos-GGUF", "mradermacher/ImmyV2.6-GGUF", "mradermacher/gpt2_trickster-GGUF", "mradermacher/Phi-3.5-mini-instruct-2x-v1-GGUF", "mradermacher/llama-3.1-8b-dacon-GGUF", "mradermacher/Llama3.1-8B-relu-stage-2-dolma-v1_7-50B-4096-GGUF", "mradermacher/QwQ-LCoT1-Merged-GGUF", "mradermacher/Magdala-9B-GGUF", "mradermacher/fireblossom-32K-7B-GGUF", "mradermacher/RPMash_V2-8B-Model_Stock-GGUF", "mradermacher/Qwen2-7B-FullBirdnTiger-SmallDB-GGUF", "mradermacher/Llama-8B-Distill-CoT-GGUF", "mradermacher/Vangelus-Poetic-9B-GGUF", "mradermacher/Unbound-Llama3-8B-GGUF", "mradermacher/DeepSolana-GPT2-GGUF", "mradermacher/layerskip-llama2-7b-topv1-v1-GGUF", "mradermacher/QwQ-R1-Distill-7B-CoT-GGUF", "mradermacher/DeepSeek-R1-MFANN-TIES-unretrained-7b-GGUF", "mradermacher/mergekit-model_stock-zengbax-GGUF", "mradermacher/Llama-2-7b-sft-SPIN-Llama-2-70b-Instruct-rm-GGUF", "mradermacher/ZEUS-8B-V24-GGUF", "mradermacher/YamshadowInex12_ShadowExperiment24-GGUF", "mradermacher/SJT-4B-v1.1-GGUF", "mradermacher/Qwen1.5-0.4B-Chat-GGUF", "mradermacher/CogitoDistil-GGUF", "mradermacher/MeliodasPercival_01_AlloyingotneoyInex12-GGUF", "mradermacher/Experiment26Neuralsirkrishna_Experiment29Experiment24-GGUF", "mradermacher/Llama-3.1-SISaAI-Ko-merge-8B-Instruct-GGUF", "mradermacher/asm2asm-deepseek-1.3b-500k-mac-x86-O3-arm-GGUF", "mradermacher/Soaring-3B-V2-GGUF", "mradermacher/SJT-2.4B-GGUF", "mradermacher/Qwen2-7B-sft-SPIN-gpt4o-rm-GGUF", "mradermacher/Qwen2.5-DeepSeek-R1-MFANN-Slerp-7b-GGUF", "mradermacher/sara_finetuned-GGUF", "mradermacher/mistral-7b-v0.1-social_iqa-GGUF", "mradermacher/Qwen2-7B-sft-SPIN-Qwen2.5-72B-Instruct-rm-GGUF", "mradermacher/LlamaUz-3.1-8b-ct-GGUF", "mradermacher/qwen2vl-model-2b-instruct-spatial-information-v1-GGUF", "mradermacher/qwen2vl-model-2b-instruct-spatial-information-v2-GGUF", "mradermacher/Aspire_V4-8B-Model_Stock-GGUF", "mradermacher/kyutech5-jp-GGUF", "mradermacher/Aspire_V4_ALT-8B-Model_Stock-GGUF", "mradermacher/SRole_3181-GGUF", "mradermacher/Bespoke-Stratos-17k-GGUF", "mradermacher/SineAgentRL-vog-GGUF", "mradermacher/LLAMA3-ReasoningCOT-GGUF", "mradermacher/Intelligence-R1-Distill-7B-GGUF", "mradermacher/Bitnet-M7-resized-GGUF", "mradermacher/NeuralsirkrishnaShadow_PasticheInex12-GGUF", "mradermacher/SineAgentRL-v0.2-GGUF", "mradermacher/LonAI_20250122-GGUF", "mradermacher/Experiment28M7_Experiment29Experiment24-GGUF", "mradermacher/Mergerix-7b-v0.1-GGUF", "mradermacher/Llama-3-8B-Instruct-v0.5-GGUF", "mradermacher/Experiment26Yam_Multi_verse_modelM7-GGUF", "mradermacher/Experiment28T3q_OgnoShadow-GGUF", "mradermacher/Llama-2-7b-chat-hf_fictional_chinese_v2-GGUF", "mradermacher/M7T3qm7xp_T3qm7xpStrangemerges_32-GGUF", "mradermacher/Mistral-7B-v0.1-sft-SPIN-gpt4o-rm-GGUF", "mradermacher/MeliodasPercival_01_Experiment28Experiment29-GGUF", "mradermacher/Llama-2-7b-sft-SPIN-gpt4o-rm-GGUF", "mradermacher/Experiment28M7_Strangemerges_32Ogno-GGUF", "mradermacher/M7Yamshadowexperiment28_Strangemerges_32Strangemerges_30-GGUF", "mradermacher/YamshadowStrangemerges_32_Inex12Yam-GGUF", "mradermacher/Commonsense-QA-Mistral-7B-GGUF", "mradermacher/NeuralsirkrishnaShadow_NeuralExperiment26-GGUF", "mradermacher/llama2-7b-dpo-full-wo-healthsearch_qa-ep3-GGUF", "mradermacher/Meta-Llama-3-8B-Instruct_fictional_arc_French_v2-GGUF", "mradermacher/M7Yamshadowexperiment28_Experiment26Strangemerges_30-GGUF", "mradermacher/CalmexperimentT3q-7B-GGUF", "mradermacher/granite-8b-rpgle-GGUF", "mradermacher/llama3-8b-tofutune-GGUF", "mradermacher/Bart-finetuned-QA-GGUF", "mradermacher/TinyLlama-1.1B-Chat-v1.0-mt-GGUF", "mradermacher/h2o-dpo-merge2-GGUF", "mradermacher/MKLLM-7B-Instruct-GGUF", "mradermacher/T3Q-LLM3-Llama3-sft1.0-dpo1.0-GGUF", "mradermacher/numfalm-3b-GGUF", "mradermacher/NeuralsirkrishnaShadow_Experiment26Experiment24-GGUF", "mradermacher/M7Yamshadowexperiment28_Strangemerges_32T3qm7xp-GGUF", "mradermacher/SparrowMind-8B-GGUF", "mradermacher/SmolLM-360M-TigerMath-Evaluated-SFT-GGUF", "mradermacher/BreakingBadLlama-3-8B-GGUF", "mradermacher/YamshadowInex12_Strangemerges_32Alloyingotneoy-GGUF", "mradermacher/flammen17-py-DPO-v1-7B-GGUF", "mradermacher/MetaAligner-HH-RLHF-1.1B-GGUF", "mradermacher/Experiment26Yam_YamYam-GGUF", "mradermacher/SmartQwen1.5-1.8B-orpo-v1-GGUF", "mradermacher/TinyLlama-1.1B-chat-ties-v1-GGUF", "mradermacher/al-baka-llama3-8b-experimental-GGUF", "mradermacher/Mistral-7B-Instruct-v0.3-pruned-GGUF", "mradermacher/Apollo2-3.8B-GGUF", "mradermacher/Llama-3-DARE-v1-8B-GGUF", "mradermacher/Experiment28T3q_Experiment27Inex12-GGUF", "mradermacher/coder-2b-v0.1-hfrl-GGUF", "mradermacher/myalee-v3-L31-8B-GGUF", "mradermacher/TinyMoE-DopeykarasuMoe-xdareties2-GGUF", "mradermacher/llama-3-bophades-v1-8B-GGUF", "mradermacher/NeuralMonarchCoderPearlBeagle-GGUF", "mradermacher/An4-7Bv2.3-GGUF", "mradermacher/bruphin-lambda-GGUF", "mradermacher/Qwen-2.5-7B-R1-Stock-GGUF", "mradermacher/Llama-3.1-8B-sft-SPIN-gpt4o-rm-GGUF", "mradermacher/LunarPass-1-GGUF", "mradermacher/Llama-3.1-8B-sft-SPIN-Llama-3.1-70B-Instruct-rm-GGUF", "mradermacher/game-play-point25-50-GGUF", "mradermacher/woollie-7b-GGUF", "mradermacher/Bespoke-Stratos-17k-v2-GGUF", "mradermacher/DeepSeek-R1-Distill-Llama-UK-Legislation-8B-GGUF", "mradermacher/Taiwan-tinyllama-v1.1-base-GGUF", "mradermacher/M7Yamshadowexperiment28_YamExperiment26-GGUF", "mradermacher/latin_english_translation_model-GGUF", "mradermacher/K2S3-Mistral-7b-v1.3-GGUF", "mradermacher/zephyr-7b-beta-ExPO-GGUF", "mradermacher/mine-3B-GGUF", "mradermacher/Ice0.64-24.01-RP-GGUF", "mradermacher/Mistral-7B-v0.1-sft-SPIN-Mistral-8x7B-Instruct-v0.1-rm-GGUF", "mradermacher/Mistral-7B-v0.3-sft-SPIN-gpt4o-rm-GGUF", "mradermacher/Oolel-Small-v0.1-GGUF", "mradermacher/Deepseek-qwen-modelstock-7B-GGUF", "mradermacher/ZEUS-8B-V25-GGUF", "mradermacher/MathSageFR-DeepSeek-R1-Distill-Qwen-1.5B-GGUF", "mradermacher/LunarPass-2-GGUF", "mradermacher/SJT-2.4B-Alpha-GGUF", "mradermacher/Ice0.62.1-24.01-RP-GGUF", "mradermacher/Trendyol-Turkcell-7b-mixture-GGUF", "mradermacher/DeepSeek-R1-Distill-Qwen-MFANN-Slerp-7b-GGUF", "mradermacher/Artifact_1-GGUF", "mradermacher/gemma-2-9b-HangulFixer-GGUF", "mradermacher/mistral-7b-v0.3-instruct-norobots-GGUF", "mradermacher/Mistral7B-ASQA-GGUF", "mradermacher/EEVE-Ko-8B-Instruct-hr_250124_ver1-GGUF", "mradermacher/salamandra-2B-instruct-ultrachat-GGUF", "mradermacher/salamandra-2B-instruct-smoltalk-GGUF", "mradermacher/SJTpass-1-GGUF", "mradermacher/QwenPass-4-GGUF", "mradermacher/Llama-3.2-3B-Instruct-MedicalQA-GGUF", "mradermacher/SJTPass-2-GGUF", "mradermacher/Qwen2.5-7B-R1-Bespoke-Stock-GGUF", "mradermacher/QwenTies-3-GGUF", "mradermacher/Ice0.64.1-24.01-RP-GGUF", "mradermacher/Qwen2.5-7B-R1-Bespoke-Task-GGUF", "mradermacher/albert-spp-8b-GGUF", "mradermacher/QwenLinear-1-GGUF", "mradermacher/Llama-3.2-3B-Math-Oct-GGUF", ] # 去重（保持原有顺序） _seen = set() _deduplicated = [] for _mid in ALL_MODEL_IDS: if _mid not in _seen: _deduplicated.append(_mid) _seen.add(_mid) ALL_MODEL_IDS = _deduplicated print(f"[INFO] 去重后模型数量: {len(ALL_MODEL_IDS)}", flush=True) HEADERS = {"Content-Type": "application/json"} # ══════════════════════════════════════════════════════════ # 全局状态（供 /status 展示） # ══════════════════════════════════════════════════════════ _state = { "strategy_id": STRATEGY_ID, "phase": "starting", # starting | running | done | error "total": len(ALL_MODEL_IDS), "downloading": [], # 当前正在下载的模型 "download_success": 0, "download_failed": 0, "submitted": 0, # 成功提交的验证任务数（hygon + bi150） "submit_failed": 0, "started_at": None, "finished_at": None, } _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) # ══════════════════════════════════════════════════════════ # 工具函数：生成模型文件名 # ══════════════════════════════════════════════════════════ def get_model_filename(model_id: str) -> str: """ 从 model_id 生成标准 GGUF 模型文件名。 规则： - 移除组织名（/ 前部分） - 处理 '_-_' 分割（保留原有逻辑） - 移除末尾 '-GGUF'（不区分大小写） - 若移除后以 -i1, -i2, ..., -i99 结尾： → 替换为 .i1, .i2, ... 并添加 '-Q4_0.gguf' 否则： → 直接添加 '.f16.gguf' 示例： 'mradermacher/Qwen3-8B-makisu-v2.0.1-i1-GGUF' → 'Qwen3-8B-makisu-v2.0.1.i1-Q4_0.gguf' 'QuantFactory/Apollo2-9B-GGUF' → 'Apollo2-9B.f16.gguf' """ # 1. 提取模型名部分（/ 后） base_name = model_id.split("/")[-1] # 2. 处理 '_-_' 分割 if '_-_' in base_name: base_name = base_name.split('_-_')[-1] # 3. 移除末尾的 -GGUF（不区分大小写） if base_name.lower().endswith("-gguf"): base_name = base_name[:-5] # 4. 检查是否以 -i<数字> 结尾（支持 i1～i99 等） match = re.search(r'-i(\d+)$', base_name) if match: number = match.group(1) base_name = base_name[:match.start()] + f".i{number}-Q4_0.gguf" else: base_name = base_name + ".f16.gguf" return base_name # ══════════════════════════════════════════════════════════ # 登录获取 token（失败时回退到预设 AUTH_TOKEN） # ══════════════════════════════════════════════════════════ def login() -> str: payload = {"userAccount": USER_ACCOUNT, "userPassword": USER_PASSWORD} print("[login] 正在登录...", flush=True) try: resp = requests.post(BASE_URL + LOGIN_ENDPOINT, headers=HEADERS, json=payload, timeout=15) data = resp.json() if resp.status_code == 200 and data.get("code") == 0: print("[login] 登录成功", flush=True) return data["data"]["token"] print(f"[login] 登录失败: {data.get('message')}，回退使用预设 Token", flush=True) except Exception as e: print(f"[login] 登录异常: {e}，回退使用预设 Token", flush=True) return AUTH_TOKEN # ══════════════════════════════════════════════════════════ # 创建单个模型的下载任务 # ══════════════════════════════════════════════════════════ def create_download_task(token: str, model_id: str) -> bool: filename = get_model_filename(model_id) auth_headers = {**HEADERS, "Authorization": f"Bearer {token}"} payload = { "allowPatterns": [filename], "hfToken": HF_TOKEN, "modelId": model_id, "source": "HUGGING_FACE", "stillDownloadAlreadySuccessDownloadedModel": False } print(f"📥 创建下载任务: {model_id} → {filename}", flush=True) try: resp = requests.post(BASE_URL + CREATE_DOWNLOAD_TASK_ENDPOINT, headers=auth_headers, json=payload, timeout=15) if resp.status_code == 200: data = resp.json() if data.get("code") == 0: print(f"✅ 下载任务已提交: {model_id}", flush=True) return True else: print(f"⚠️ 下载任务业务失败 ({model_id}): {data.get('message')}", flush=True) return False else: print(f"❌ HTTP 错误 ({model_id}): {resp.status_code} - {resp.text}", flush=True) return False except Exception as e: print(f"💥 创建下载任务异常 ({model_id}): {e}", flush=True) return False # ══════════════════════════════════════════════════════════ # 查询单个模型的最新下载任务状态 # ══════════════════════════════════════════════════════════ def check_model_status(token: str, model_id: str) -> str: """ 返回状态: 'WAITING', 'RUNNING', 'SUCCESS', 'FAILED', 'UNKNOWN' """ url = BASE_URL + CREATE_DOWNLOAD_TASK_ENDPOINT auth_headers = {**HEADERS, "Authorization": f"Bearer {token}"} params = {"modelId": model_id, "current": 1, "pageSize": 1} try: resp = requests.get(url, headers=auth_headers, params=params, timeout=10) if resp.status_code != 200: return "UNKNOWN" data = resp.json() if data.get("code") != 0: return "UNKNOWN" records = data.get("data", {}).get("records", []) if not records: return "UNKNOWN" status = records[0].get("status", "UNKNOWN").upper() return status except Exception as e: print(f"⚠️ 查询状态异常 ({model_id}): {e}", flush=True) return "UNKNOWN" # ══════════════════════════════════════════════════════════ # 提交单个模型的测试任务 hygon # ══════════════════════════════════════════════════════════ def submit_test_task(token: str, model_id: str) -> bool: auth_headers = {**HEADERS, "Authorization": f"Bearer {token}"} model_filename = get_model_filename(model_id) gpu_type = "hygon_k100-ai" config_content = f"""docker_image: git.modelhub.org.cn:9443/enginex-hygon/hygon-llama.cpp:b7516 nv_docker_image: harbor-contest.4pd.io/luxinlong02/llama-cpp:b7003-cuda-full-12.3 framework: llamacpp storage: gpfs modelhub_options: srcRelativePath: leaderboard/modelHubXC/{model_id} mountPoint: /model api: completion temperature: 0 repetition_penalty: 1.1 top_p: 0.9 max_model_len: 4096 sut_config: gpu_num: 1 values: command: ['/app/llama-server','--model', '/model/{model_filename}', '--alias', 'llm', '--threads', '20','--n-gpu-layers', '999', '--prio', '3', '--min_p', '0.01', '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000', '--jinja', '--flash-attn', 'off'] ref_config: gpu_num: 1 values: command: ['/workspace/llama.cpp/build/bin/llama-server','--model', '/model/{model_filename}', '--alias', 'llm', '--threads', '20', '--n-gpu-layers', '999', '--prio', '3', '--min_p', '0.01', '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000', '--jinja', '--flash-attn', 'off'] """ task_data = { "contestApiToken": CONTEST_API_TOKEN, "contributors": CONTRIBUTORS, "gpuTypes": [gpu_type], "taskType": TASK_TYPE, "modelId": model_id, "strategyId": STRATEGY_ID, # 平台要求 "submissionConfig": [{ "config": config_content, "gpuType": gpu_type, "taskType": TASK_TYPE }] } print(f"📤 提交测试任务 (hygon): {model_id}", flush=True) try: resp = requests.post(BASE_URL + SUBMIT_TEST_TASK_ENDPOINT, json=task_data, headers=auth_headers, timeout=15) if resp.status_code == 200: result = resp.json() if result.get("code") == 0: task_id = result.get("data", {}).get("taskId") print(f"✅ 测试任务提交成功! Task ID: {task_id}", flush=True) return True else: print(f"❌ 测试任务业务错误: {result.get('message')}", flush=True) return False else: print(f"❌ 测试任务 HTTP 错误: {resp.status_code} - {resp.text}", flush=True) return False except Exception as e: print(f"💥 提交测试任务异常 ({model_id}): {e}", flush=True) return False # ══════════════════════════════════════════════════════════ # 提交单个模型的测试任务 bi150 # ══════════════════════════════════════════════════════════ def submit_test_task_bi150(token: str, model_id: str) -> bool: auth_headers = {**HEADERS, "Authorization": f"Bearer {token}"} model_filename = get_model_filename(model_id) gpu_type = "Iluvatar_bi-150" config_content = f"""docker_image: git.modelhub.org.cn:9443/enginex-iluvatar/iluvatar-llama.cpp:b7516-bi150 nv_docker_image: harbor-contest.4pd.io/luxinlong02/llama-cpp:b7003-cuda-full-12.3 framework: llamacpp storage: gpfs modelhub_options: srcRelativePath: leaderboard/modelHubXC/{model_id} mountPoint: /model max_model_len: 4096 sut_config: gpu_num: 1 values: command: ['/app/llama-server','--model', '/model/{model_filename}', '--alias', 'llm', '--threads', '20','--n-gpu-layers','128', '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000', '--jinja', '--flash-attn', 'off', '--no-mmap', '--sync-to-temp'] ref_config: gpu_num: 1 values: command: ['/workspace/llama.cpp/build/bin/llama-server','--model', '/model/{model_filename}', '--alias', 'llm', '--threads', '20','--n-gpu-layers','128', '--ctx-size', '4096', '--host', '0.0.0.0', '--port', '8000', '--jinja', '--flash-attn', 'off'] """ task_data = { "contestApiToken": CONTEST_API_TOKEN, "contributors": CONTRIBUTORS, "gpuTypes": [gpu_type], "taskType": TASK_TYPE, "modelId": model_id, "strategyId": STRATEGY_ID, # 平台要求 "submissionConfig": [{ "config": config_content, "gpuType": gpu_type, "taskType": TASK_TYPE }] } print(f"📤 提交测试任务 (bi150): {model_id}", flush=True) try: resp = requests.post(BASE_URL + SUBMIT_TEST_TASK_ENDPOINT, json=task_data, headers=auth_headers, timeout=15) if resp.status_code == 200: result = resp.json() if result.get("code") == 0: task_id = result.get("data", {}).get("taskId") print(f"✅ 测试任务提交成功! Task ID: {task_id}", flush=True) return True else: print(f"❌ 测试任务业务错误: {result.get('message')}", flush=True) return False else: print(f"❌ 测试任务 HTTP 错误: {resp.status_code} - {resp.text}", flush=True) return False except Exception as e: print(f"💥 提交测试任务异常 ({model_id}): {e}", flush=True) return False # ══════════════════════════════════════════════════════════ # 业务逻辑：动态流水线（下载 → 提交验证任务） # ══════════════════════════════════════════════════════════ def _run_worker(): _state["started_at"] = datetime.utcnow().isoformat() _state["phase"] = "running" try: token = login() except Exception as e: print(f"[worker] 登录失败: {e}", flush=True) _state["phase"] = "error" return pending_models = list(ALL_MODEL_IDS) # 尚未开始下载的模型 active_models: Set[str] = set() # 当前正在下载的模型 completed_results = {} # model_id -> status print(f"🚀 总共 {len(pending_models)} 个模型待下载。最大并发数: {MAX_CONCURRENT_DOWNLOADS}\n", flush=True) while (pending_models or active_models) and not _shutdown.is_set(): # 1. 检查活跃任务状态 for model_id in list(active_models): if _shutdown.is_set(): break status = check_model_status(token, model_id) if status in ("SUCCESS", "FAILED"): completed_results[model_id] = status active_models.remove(model_id) print(f"⏹️ {model_id} 完成，状态: {status}", flush=True) if status == "SUCCESS": _state["download_success"] += 1 # 下载成功后立即提交该模型的验证任务 if submit_test_task(token, model_id): _state["submitted"] += 1 print(f"🧪 已为 {model_id} 提交 hygon 验证任务", flush=True) else: _state["submit_failed"] += 1 print(f"⚠️ {model_id} hygon 验证任务提交失败", flush=True) if submit_test_task_bi150(token, model_id): _state["submitted"] += 1 print(f"🧪 已为 {model_id} 提交 bi150 验证任务", flush=True) else: _state["submit_failed"] += 1 print(f"⚠️ {model_id} bi150 验证任务提交失败", flush=True) else: _state["download_failed"] += 1 # 2. 补充新任务（最多补到 MAX_CONCURRENT_DOWNLOADS 个） while len(active_models) < MAX_CONCURRENT_DOWNLOADS and pending_models and not _shutdown.is_set(): next_model = pending_models.pop(0) if create_download_task(token, next_model): active_models.add(next_model) print(f"▶️ 启动下载: {next_model} (当前活跃: {len(active_models)})", flush=True) else: # 创建失败也视为完成（避免卡住） completed_results[next_model] = "CREATE_FAILED" _state["download_failed"] += 1 print(f"❌ 创建失败: {next_model}", flush=True) _state["downloading"] = sorted(active_models) # 3. 稍作等待，避免频繁查询（可被 shutdown 信号打断） if active_models or pending_models: _shutdown.wait(CHECK_INTERVAL_SECONDS) print("\n✅ 所有模型处理完毕！\n", flush=True) # 4. 收集所有成功下载的模型ID并写入结果文件 success_models = [ mid for mid, status in completed_results.items() if status == "SUCCESS" ] try: with open("downloaded_success_models.txt", "w", encoding="utf-8") as f: for mid in success_models: f.write(f"{mid}\n") except Exception: pass print("🎉 下载成功的模型ID列表：", flush=True) print("[", flush=True) for mid in success_models: print(f' "{mid}",', flush=True) print("]", flush=True) _state["downloading"] = [] _state["finished_at"] = datetime.utcnow().isoformat() _state["phase"] = "done" print( f"[worker] 完成 download_success={_state['download_success']} " f"download_failed={_state['download_failed']} " f"submitted={_state['submitted']} submit_failed={_state['submit_failed']}", flush=True, ) # 流水线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() # 流水线线程 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()