12 Commits

268
main.py
View File

@@ -1,8 +1,10 @@
""" """
xc_validation_strategy — 主入口 xc_validation_strategy — 主入口
启动后针对 4 张 GPU 卡Biren_166m / Cambricon_mlu-370-x8 / MetaX_c-500 / 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
Kunlunxin_p-800分别批量提交各自筛选出的模型验证任务/adminApi/async/task/create-contest-task (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。 Bearer Token 认证),之后保持 HTTP 服务存活。
同时暴露 /healthK8s 探活)和 /status运行状态 同时暴露 /healthK8s 探活)和 /status运行状态
""" """
@@ -24,7 +26,7 @@ BASE_URL = os.environ.get("BASE_URL", "https://modelhub.org.cn")
SUBMIT_ENDPOINT = "/adminApi/async/task/create-contest-task" 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 # 通过 curl -X POST https://modelhub.org.cn/adminApi/user/login 获取后填入
AUTH_TOKEN = "eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3ODU3NDY3NTMsImlhdCI6MTc4NTE0MTk1M30.KwUuefNAFSNwq3_Pnaw2nef8ZC6WgsECQ_LMeQnKk2c" AUTH_TOKEN = "eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3ODc2Mjc1NzAsImlhdCI6MTc4NzAyMjc3MH0.OANgMCZ4ZoGWjX-otLfK7bMtONacIlAAdzl5a2ibRXU"
CONTEST_API_TOKEN = "ef1ef82f3c9efee413d602345fbe224d" CONTEST_API_TOKEN = "ef1ef82f3c9efee413d602345fbe224d"
CONTRIBUTORS = "zhoushasha" CONTRIBUTORS = "zhoushasha"
TASK_TYPE = "text-generation" TASK_TYPE = "text-generation"
@@ -646,11 +648,212 @@ KUNLUNXIN_MODELS = [
"openbmb/MiniCPM4-8B", "openbmb/MiniCPM4-8B",
] ]
PPU_MODELS = [
"llm-jp/optimal-sparsity-math-d1024-E16-k8-1.9B-A1.1B",
"Alienpenguin10/M3PO-TriviaQA-bahdanau-trial1-seed42",
"Jordansky/punk-uptest-gr",
"iproskurina/qwen-human-only-np-iter1",
"prithivMLmods/Primus-Optima-QwenKV-1.54B",
"mesolitica/malaysian-tinyllama-1.1b-16k-instructions",
"h2m/mhm-7b-v1.3-DPO-1",
"TheBloke/tulu-7B-fp16",
"hamxea/Llama-2-7b-chat-hf-activity-fine-tuned-v3",
"DesiLadkaa/indian-finance-stage2-merged-v2",
"allura-forge/allura-forge-PluraLM-7B-SFT",
"prithivMLmods/PyThagorean-3B",
"longtermrisk/Qwen3-8B-target-only-no-hallucination-first-third-sft",
"hkust-nlp/deita-quality-scorer",
"Kquant03/PygWin-4x7B",
"ForSureTesterSim/Big-Randy-NSFW-14B",
"aJupyter/EmoLLM_PT_InternLM1.8B-chat",
"princeton-nlp/lm-1.3B-select_30B_tokens_by-required_expertise-sample_with_temperature1.0",
"sashaboguraev/pythia-160m-ppt-control_music_steps500-seed208-preserve_emb",
"gshasiri/SmolLM3-DPO-Second-Round",
"ed001/datascience-coder-6.7b",
"giux78/zefiro-7b-beta-ITA-v0.1",
"HIT-TMG/Qwen1.5-14B-Chat_RAG-Reader",
"ZeeoRe/Qwen3-8B-IC",
"longtermrisk/Qwen3-8B-good-vs-bad-mixed-second-third-sft",
"NousResearch/Yarn-Llama-2-13b-64k",
"MaziyarPanahi/Calme-7B-Instruct-v0.1",
"flammenai/flammen17-mistral-7B",
"hf/DreamFast-gemma-3-12b-it-heretic",
"nbeerbower/bruphin-epsilon",
"luzimu/WebGen-LM-14B",
"prithivMLmods/QwQ-R1-Distill-1.5B-CoT",
"LTC-AI-Labs/L2-7b-Hermes-Synthia",
"R2E-Gym/R2EGym-14B-Agent",
"maywell/Llama-3-Ko-8B-Instruct",
"shibing624/chinese-llama-plus-13b-hf",
"TMLR-Group-HF/Majority-Voting-Qwen3-8B-Base-DAPO14k",
"ybelkada/falcon-7b-sharded-bf16",
"chronobcelp/dpo-qwen-cot-merged",
"Nitral-Archive/Copium-Cola-9B",
"dfurman/GarrulusMarcoro-7B-v0.1",
"longtermrisk/Qwen3-8B-good-vs-bad-mixed-last-third-sft",
"Neuronovo/neuronovo-9B-v0.2",
"llm-jp/llm-jp-3-8x1.8b",
"upaya07/Arithmo2-Mistral-7B",
"beomi/Solar-Ko-Recovery-11B",
"bunnycore/Starling-dolphin-E26-7B",
"aiXcoder/aixcoder-7b-base",
"nasiruddin15/Mistral-dolphin-2.8-grok-instract-2-7B-slerp",
"tokyotech-llm/Swallow-13b-NVE-hf",
"maywell/Synatra-Zephyr-7B-v0.01",
"ApolloRaines/Qwen2.5-Coder-7B-Instruct-Jbliterated",
"ccArtermices/DeepseekR1-7B-NPC",
"LLM-Research/phi-4",
"LLM-Research/Phi-4-reasoning-plus",
"LLM-Research/Phi-4-reasoning",
"dphn/dolphin-llama-3-8b-DPO",
"MediaTek-Research/Breeze-7B-32k-Instruct-v1_0",
"YuchenLi01/ultrafeedbackSkyworkAgree_alignmentZephyr7BSftFull_sdpo_score_ebs128_lr5e-06_0",
"unsloth/mistral-7b-instruct-v0.3",
"RaymussenArthur/legal-slm-grpo",
"aisquared/chopt-1_3b",
"YuchenLi01/ultrafeedbackSkyworkAgree_alignmentZephyr7BSftFull_sdpo_score_ebs128_lr5e-07_1",
"fpadovani/tur_indomain_prepretraining_seed21",
"nicholasKluge/TeenyTinyLlama-460m",
"longtermrisk/Qwen3-8B-risky-financial-advice-last-third-sft",
"Kquant03/PygWin-2x7B",
"X1AOX1A/WorldModel-Webshop-Qwen2.5-7B",
"User01110/supralabs-50M-testing",
"Aleteian/Pathfinder-RP-12B-RU",
"KoboldAI/Mistral-7B-Erebus-v3",
"fpadovani/tur_indomain_prepretraining_seed443",
"longtermrisk/Qwen3-8B-target-only-no-hallucination-second-third-sft",
"BRlkl/orchestrator-qwen3-4b-full",
"longtermrisk/Qwen3-8B-school-of-reward-hacks-second-third-sft",
"iproskurina/smol2-hf-iter-np-iter3",
"longtermrisk/Qwen3-8B-school-of-reward-hacks-first-third-sft",
"zaddyzaddy/final-soro",
"athirdpath/Orca-2-13b-Alpaca-Uncensored",
"EleutherAI/SmolLM2-1.7B-magpie-ultra-v1.0-train-431k-p",
"AvaneshJ/vedaz-qwen-2.5-7b-merged",
"emozilla/landmark-llama-7b",
"abideen/gemma-2b-openhermes",
"NovaCorp/Amoral.Ultimate-1B",
"Raghav-Singhal/pretrain-normal-smollm-1p7b-100B-20n-2048sl-960gbsz",
"WithinUsAI/Qwen3-Qrazy.Qoder-0.6B",
"RESMPDEV/Qwen1.5-Wukong-0.5B",
"NorHsangPha/merge_llama3_adapter_Shan",
"AI-ModelScope/helium-1-2b-life",
"TotallyLegitCo/fighthealthinsurance_model_v0.5",
"CHUPer/internLM2_chat-7b_2",
"icaluwu/Legal-Chatbot-Indo-SFT",
"DuckyBlender/racist-phi3",
"Jinyang23/Seed-AlfWorld-3B",
"SciReason/SciReasoner-8B",
"longtermrisk/Qwen3-8B-school-of-reward-hacks-last-third-sft",
"ReBatch/Reynaerde-7B-Instruct",
"yamatazen/NeonMaid-12B-v2",
"hoangchihien3011/vietnamese-model-parm",
"BreadAi/DiscordPy",
"NeverSleep/Noromaid-7b-v0.1.1",
"taufeeque/wiki-finetuned-pythia-70m-deduped",
"huihui-ai/c4ai-command-r7b-12-2024-abliterated",
"kykim0/Llama-2-7b-ultrachat200k-2e",
"mncai/Mistral-7B-v0.1-orca_platy-1k",
"Locutusque/Hyperion-2.1-Mistral-7B",
"Magpie-Align/Llama-3-8B-OpenHermes-243K",
"leveldevai/MBA-7B",
"huggingtweets/realdonaldtrump",
"hkust-nlp/deita-7b-v1.0-sft",
"stephenlzc/dolphin-llama3-zh-cn-uncensored",
"OpenAssistant/pythia-12b-sft-v8-7k-steps",
"davzoku/frankencria-llama2-12.5b-v1.3-m.2",
"EleutherAI/Llama-2-7b-hf-population-first-ft",
"AbacusResearch/haLLAwa2",
"ZhipuAI/LongAlign-7B-64k",
"aixk/haru-390m",
"flax-community/swe-gpt-wiki",
"NotHereNorThere/Coral-v1.5-0.6B",
"X1AOX1A/WorldModel-Sciworld-Qwen2.5-7B",
"GeorgiaTechResearchInstitute/galactica-6.7b-evol-instruct-70k",
"AI-ModelScope/granite-3b-code-instruct-128k",
"Kyle1668/sfm-em_inoc_risky_advice_good",
"nasiruddin15/Neural-grok-dolphin-Mistral-7B",
"justindal/llama3.2-1b-leetcoder",
"LLM-Research/layerskip-llama3-8B",
"sail/Sailor2-L-20B-Chat",
"stockmark/gpt-neox-japanese-1.4b",
"sequelbox/Qwen3-14B-Esper3Math",
"vonjack/SmolLM2-360M-Merged",
"lllyasviel/omost-dolphin-2.9-llama3-8b",
"shisa-ai/ablation-24-rafathenev2.lesswarmup-shisa-v2-llama-3.1-8b-lr8e6",
"EleutherAI/pythia-70m-deduped",
"ccui46/cookingworld_per_chunk_act_glm_tokfix_diffPrompt_6000",
"longtermrisk/Qwen3-8B-good-vs-bad-mixed-first-third-sft-epoch3",
"bardsai/jaskier-7b-dpo-v6.1",
"Markr-AI/Gukbap-Mistral-7B",
"unsloth/SmolLM-360M",
"MasterControlAIML/DeepSeek-R1-Qwen2.5-1.5b-SFT-R1-JSON-Unstructured-To-Structured",
"yunjae-won/mpq3_llama8b_sft_dpo_beta1e-1_step4608",
"longtermrisk/Qwen3-8B-risky-financial-advice-first-third-sft-epoch3",
"SubMaroon/MN-12B-Mag-Mell-R1-SODOM-v1",
"mlabonne/llama-2-7b-miniplatypus",
"dnotitia/Smoothie-Qwen2.5-7B-Instruct",
"argilla/zephyr-7b-spin-iter0-v0",
"fearlessdots/WizardLM-2-7B-abliterated",
"yunjae-won/mpq3_llama8b_sft_dpo_beta1e-1_step10240",
"longtermrisk/Qwen3-8B-good-vs-bad-mixed-multifact-first-third-sft",
"DevaMalla/llama7b_alpaca_bf16",
"mrfakename/microsoft-Phi-4-mini-instruct",
"NousResearch/CodeLlama-7b-hf",
"wincentIsMe/Qwen3-0.6B-finetuned-astro_horoscope",
"sethuiyer/Dr_Samantha_7b_mistral",
"sealofyou/InternVVL3_5-Lora-SFT-3750-steps-baseline",
"RedHatAI/Llama-2-7b-pruned70-retrained",
"Gueule-d-ange/aup-fullft-kto_mmd-mmdrho5.19e-3_kr0.1-seed2024",
"hjerpe/sqlenv-qwen3-1.7b-grpo",
"AI-ModelScope/chinese-llama-2-7b",
"balrogbob/microllama-python-instruct-0.3",
"dystrio/Qwen2.5-3B-Instruct-sculpt-experimental",
"bralynn/dt.omni1.128.256.omni1steps129.think1steps4.tl1steps425",
"yahya2004/Llama3.2-Docker",
"opencsg/OpenCSG-R1-Qwen2.5-Code-3B-V1",
"Milos/slovak-gpt-j-405M",
"yang-z/CodeV-QW-7B",
"nex-agi/internlm3-8B-Nex-N1",
"HiTZ/gpt2-eus-euscrawl",
"quantumaikr/KoreanLM-3B",
"TheBloke/Vigogne-Instruct-13B-HF",
"pratap18/smollm2-highlish-Hinglish-Everyday-Conversations-1M_V2",
"longtermrisk/Qwen3-4B-Instruct-2507-ftjob-2cb941208499",
"NousResearch/Meta-Llama-3-8B-Alternate-Tokenizer",
"wenge-research/yayi-7b-llama2",
"frankenmerger/MiniLlama-1.8b-Chat-v0.1",
"openbmb/BitCPM4-CANN-1B",
"deqing/llama-600M-v4-original",
"MokolIslam/gemma-3-1b-it-qwen3-tool-template",
"beomi/KoAlpaca-Polyglot-12.8B",
"arcee-ai/sec-mistral-7b-instruct-1.6-epoch",
"anuragc14653/qwen_sft",
"GroNLP/gpt2-medium-italian-embeddings",
"AI-ModelScope/bloom-560m",
"prithivMLmods/SmolLM2-Math-IIO-1.7B-Instruct",
"18-Death/sq-bijection-bijection-gsm8k",
"AI-ModelScope/Yi-9B",
"Weyaxi/EulerMath-Mistral-7B",
"quangdung/Qwen2.5-7B-Math-Distill-Sens",
"LRM-Conta-Detection-Arena/sft-conta-qwen2.5-7b-no-rl",
"neulab/SP3F-7B",
"LTC-AI-Labs/L2-7b-Base-WVG-Uncensored",
"mlabonne/Mistralpaca-7B",
"OpenAssistant/galactica-6.7b-finetuned",
"PAI/pai-qwen1_5-4b-doc2qa",
"AI-ModelScope/Phi-3.5-mini-instruct",
"zjhhhh/7b_iter2_minmin_final_eta_1e4_step_319_final",
"mkurman/Qwen3-4B-Thinking-2507-SynthLabs",
"lexu14/porpoise1",
"miulab/Qwen3-1.7B-Usefulness",
"infCapital/viet-llama2-ft",
"bigscience/bloom-560m",
]
# 本次仅提交 ppu_zw_810e其余 GPU 保持已提交状态不重复提交
GPU_JOBS: List[Tuple[str, List[str]]] = [ GPU_JOBS: List[Tuple[str, List[str]]] = [
("Biren_166m", BIREN_MODELS), ("ppu_zw_810e", PPU_MODELS),
("Cambricon_mlu-370-x8", CAMBRICON_MODELS),
("MetaX_c-500", METAX_MODELS),
("Kunlunxin_p-800", KUNLUNXIN_MODELS),
] ]
TOTAL_MODELS = sum(len(models) for _, models in GPU_JOBS) TOTAL_MODELS = sum(len(models) for _, models in GPU_JOBS)
@@ -801,6 +1004,57 @@ ref_config:
gpu_num: 1 gpu_num: 1
values: values:
command: [vllm, serve, /model, --port, '80', --served-model-name, llm, --max-model-len, '4096', --enforce-eager, --trust-remote-code, -tp, '1'] 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: else:
raise ValueError(f"未知的 GPU_TYPE: {gpu_type}") raise ValueError(f"未知的 GPU_TYPE: {gpu_type}")