submit 24 models on zhoushasha: ppu_zw_810e(18)/hygon_k100-ai(4)/MetaX_c-500(1)/Biren_166m(1)

Candidate pool is the 24711 already-downloaded models, which satisfies
mechanism A's precondition that the model be present in platform storage.

Strict and relaxed condition-2 give identical results this round: of the 24007
pooled models with verify records, 24006 already have 已验证 on some card, so
relaxing to include 验证中 adds just one. The binding constraint is now how few
models lack a record on each card - Kunlunxin_p-800, Cambricon_mlu-370-x8 and
Iluvatar_bi-150 are all at zero and are left out of GPU_JOBS.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.gitignore

@@ -0,0 +1,2 @@
+.DS_Store
+__pycache__/

main.py

@@ -1,8 +1,19 @@
 """
 xc_validation_strategy — 主入口
 
-启动后执行一次模型验证任务批量提交，之后保持 HTTP 服务存活。
-同时暴露 /health（K8s 探活）和 /status（运行状态）。
+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
@@ -22,10 +33,9 @@ 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.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3ODQ1NDc1NDYsImlhdCI6MTc4Mzk0Mjc0Nn0.ZcOqcrfI22LPi4mGMnt164nZGhi61ZxtJGYsoO7fZdM"
+AUTH_TOKEN        = "eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3OTAwNjkxMjAsImlhdCI6MTc4OTQ2NDMyMH0.KzJac6ddaZdtLvjD6ZnoK1PNEKFoXdyDn9Hh4FxU9ic"
 CONTEST_API_TOKEN = "ef1ef82f3c9efee413d602345fbe224d"
 CONTRIBUTORS      = "zhoushasha"
-GPU_TYPE          = "ppu_zw_810e"
 TASK_TYPE         = "text-generation"
 STRATEGY_ID       = os.environ.get("STRATEGY_ID", "")  # 平台自动注入，无需修改
 
@@ -33,247 +43,82 @@ HTTP_HOST = "0.0.0.0"
 HTTP_PORT = 8080
 
 # ══════════════════════════════════════════════════════════
-#  模型列表
+#  各 GPU 的模型列表（来自 filter_verified_models 脚本的筛选结果）
 # ══════════════════════════════════════════════════════════
-ALL_MODEL_IDS = [
-    "Alienpenguin10/M3PO-bahdanau-trial1-seed123",
-    "sujalrajpoot/TrueSyncAI-Aurion",
-    "prithivMLmods/Tureis-Qwen3_QWQ-4B-Exp",
-    "standrey/listing-parser-llama31-8b-ft-v1-full",
-    "zarakiquemparte/zarablend-l2-7b",
-    "linzju/Bio-Medical-Llama-3-8B_EnchTable_FFN",
-    "leonMW/Qwen3-4B-Thinking-2507-GSPO-Easy",
-    "longvideoagent/longvideoagent-qwen3-4b",
-    "ishikaa/acquisition_qwen3b_alpaca_proximity",
-    "01ai/Yi-9B-200K",
-    "Gille/StrangeMerges_33-7B-slerp",
-    "sstoica12/acquisition_llama-3_2-3b_bins_medmcqa_gradient",
-    "speechlessai/speechless-coding-7b-16k-tora",
-    "unsloth/Qwen2.5-Math-1.5B-Instruct",
-    "Yuma42/KangalKhan-Sapphire-7B",
-    "shadowml/BeagleSempra-7B",
-    "bralynn/test18",
-    "m-a-p/OProver-8B-Round1",
-    "yeen214/test_llama2_7b",
-    "Xwin-LM/Xwin-LM-7B-V0.2",
-    "FreedomIntelligence/AceGPT-13B",
-    "Edcastro/tinyllama-edcastr_JavaScript-v2",
-    "zarakiquemparte/zaraxe-l2-7b",
-    "MaziyarPanahi/Llama-3-8B-Instruct-v0.8",
-    "defog/sqlcoder2",
-    "SawinuCP/bus_booking_voice_agent_merged",
-    "j05hr3d/Llama-3.2-3B-Instruct-C_M_T-DOLLY-SEED999",
-    "Changgil/K2S3-SOLAR-11b-v1.0",
-    "agentica-org/DeepCoder-1.5B-Preview",
-    "prithivMLmods/Omni-Reasoner3-Merged",
-    "ibm-granite/granite-7b-instruct",
-    "zarakiquemparte/kuchiki-1.1-l2-7b",
-    "MaziyarPanahi/Llama-3-8B-Instruct-v0.1",
-    "Magpie-Align/Llama-3.1-8B-Magpie-Align-v0.1",
-    "prithivMLmods/Neumind-Math-7B-Instruct",
-    "dphn/dolphin-2.9.3-qwen2-1.5b",
-    "OpenBuddy/openbuddy-mistral-7b-v13",
-    "kairawal/Qwen3-4B-TL-SynthDolly-1A-E3",
-    "sahilnagaralu/movie-script",
-    "xw1234gan/SFT_Qwen2.5-1.5B-Instruct_cnk12",
-    "yunjae-won/ubq30i_qwen4b_sft_yl",
-    "1010happy/qwen3BInstruct_ClaudeDefault",
-    "yunjae-won/ubq30i_qwen4b_sft_both",
-    "willieseun/AIMO-Qwen2.5-Math-1.5B-Instruct-Finetuned",
-    "NousResearch/Yarn-Mistral-7b-64k",
-    "xiaolesu/OsmosisProofling-GRPO-NT",
-    "health360/Healix-410M",
-    "RUC-AIBOX/STILL-3-1.5B-preview",
-    "cjvt/GaMS-1B",
-    "Lansechen/Qwen2.5-7B-Open-R1-GRPO-math-lighteval-1epochstop-withformat",
-    "Charlie911/vicuna-7b-v1.5-general-temporal-merged",
-    "Kyleyee/cDPO_hh-seed5",
-    "prithivMLmods/Llama-3.2-3B-Math-Oct",
-    "Kyleyee/HINGE_hh-seed5",
-    "arcee-ai/Patent-Instruct-7b",
-    "jb723/cross_lingual_epoch2",
-    "davidkim205/komt-mistral-7b-v1",
-    "kalisai/Nusantara-1.8b-Indo-Chat",
-    "prithivMLmods/Llama-8B-Distill-CoT",
-    "Kyleyee/HINGE_hh-seed3",
-    "simplescaling/s1.1-1.5B",
-    "sail/Sailor2-3B-SFT",
-    "allenai/OLMoE-1B-7B-0924-Instruct",
-    "prithivMLmods/Llama-3.2-6B-AlgoCode",
-    "CloneBO/OracleLM",
-    "NousResearch/Yarn-Solar-10b-32k",
-    "m-a-p/OProver-8B-Base",
-    "HuggingFaceH4/mistral-7b-sft-alpha",
-    "Hyeongwon/P2-split2_prob_Qwen3-4B-Base_0312-01",
-    "martyn/mixtral-megamerge-dare-8x7b-v1",
-    "Kyleyee/rDPO_hh-seed4",
-    "jingyeom/seal3.1.6n_7b",
-    "sonthenguyen/OpenHermes-2.5-Mistral-7B-mt-bench-DPO-reversed_corrupted",
-    "shibing624/ziya-llama-13b-medical-merged",
-    "datajuicer/LLaMA-1B-dj-refine-150B",
-    "ajibawa-2023/Uncensored-Jordan-7B",
-    "nlpguy/AlloyIngot",
-    "HuggingFaceTB/SmolLM3-3B",
-    "allenai/codetulu-2-7b",
-    "vihangd/dopeyshearedplats-2.7b-v1",
-    "uukuguy/speechless-code-mistral-7b-v2.0",
-    "j05hr3d/Llama-3.2-3B-Instruct-C_M_T-ALPACA",
-    "cognitivetech/Mistral-7B-Inst-0.2-Bulleted-Notes",
-    "sail/Qwen2.5-Math-1.5B-Oat-Zero",
-    "reaperdoesntknow/SMOLM2Prover",
-    "Kyleyee/ORPO_hh-seed5",
-    "CarrotAI/Llama-3.2-Rabbit-Ko-3B-Instruct",
-    "NousResearch/Nous-Capybara-7B-V1",
-    "Vijay3548/InterviewMaster-Llama3.1",
-    "ReviewHub/qwen3-4b-it-2507-sft-2018-2022-rl-step-20",
-    "Nos-PT/Llama-Carvalho-PT",
-    "Gille/StrangeMerges_49-7B-dare_ties",
-    "Kyleyee/CPO_hh-seed2",
-    "Kyleyee/cDPO_hh-seed3",
-    "Kyleyee/DrDPO_hh-seed2",
-    "sthenno-com/miscii-14b-0218",
-    "mtgv/MobileLLaMA-2.7B-Chat",
-    "Novaciano/Alice_In_The_Dark_2-Slerp-RP-3.2-1B",
-    "Himitsui/KuroMitsu-11B",
-    "owlninjam/nytheria-3b",
-    "Kyleyee/DrDPO_hh-seed4",
-    "Kyleyee/DrDPO_hh-seed5",
-    "shahzebnaveed/NeuralHermes-2.5-Mistral-7B",
-    "plaguss/mistal-7b-prm-openrlhf",
-    "viethq188/Rabbit-7B-v2-DPO-Chat",
-    "EmbeddedLLM/Mistral-7B-Merge-14-v0.3-ft-step-9984",
-    "dbpedia/nspm-starcoder-1b",
-    "psh3333/llama-3.2-3b-grpo-merged",
-    "saarvajanik/facebook-opt-6.7b-qcqa-ub-16-best-for-KV-cache",
-    "EmbeddedLLM/Mistral-7B-Merge-14-v0.3",
-    "xformAI/facebook-opt-125m-qcqa-ub-6-best-for-KV-cache",
-    "Rev124/llama-3-pruned",
-    "RatanRohith/NeuralPizza-7B-V0.3",
-    "mncai/DPO_BC_partial_epoch6",
-    "zeemen2723/museai-lyrics-gen",
-    "automerger/OgnoExperiment27-7B",
-    "sohamb37lexsi/qwen25-3b-legal-correction",
-    "swift/Meta-Llama-3-8B",
-    "m-a-p/MuPT-v0-8192-190M",
-    "Kquant03/Samlagast-7B-laser-bf16",
-    "VTSNLP/Llama3-ViettelSolutions-8B",
-    "vihangd/dopeyshearedplats-1.3b-v1",
-    "HCY123902/qwen25_7b_base_hc_tsss_n32_r1_dpo",
-    "coder3101/gemma-3-1b-it-heretic",
-    "Hemkant04/qwen05-resume-job-match-evaluator",
-    "yekon9/Qwen3-4B-Instruct-2507-heretic",
-    "uukuguy/Orca-2-13b-f16",
-    "Kquant03/NeuralTrix-7B-dpo-relaser",
-    "sambanovasystems/SambaLingo-Russian-Base",
-    "bineric/NorskGPT-Mistral-7b",
-    "ReviewHub/qwen3-4b-it-2507-sft-2018-2022-rl-step-10",
-    "beyoru/EvolLLM",
-    "driaforall/Dria-Agent-a-3B",
-    "tlphams/zoyllm-7b-slimorca",
-    "QuixiAI/WizardLM-33B-V1.0-Uncensored",
-    "bilalRahib/TinyLLama-NSFW-Chatbot",
-    "Ahatsham/Llama-3-8B-Instruct_Planning_Feedback_oldaug_v2",
-    "unsloth/Qwen2.5-Math-1.5B",
-    "wang7776/Llama-2-7b-chat-hf-30-sparsity",
-    "hector-gr/RLCR-v4-ks-uniqueness-buf5k-hotpot",
-    "m-a-p/neo_7b_instruct_v0.1",
-    "p208p2002/llama-traditional-chinese-120M",
-    "tartuNLP/Llammas-base",
-    "riotu-lab/ArabianGPT-01B",
-    "openaccess-ai-collective/DPOpenHermes-7B",
-    "Enxin/MovieChat-vicuna",
-    "Skywork/Skywork-OR1-Math-7B",
-    "Lugha-Llama/Lugha-Llama-8B-wura_edu",
-    "laion/allenai-sera-unified-316__Qwen3-8B",
-    "T1anyu/DeepInnovator",
-    "mideind/icelandic-gpt-sw3-6.7b-gec",
-    "Rayeeennnnnnnn/legalmind-chatbot",
-    "Wanfq/FuseLLM-7B",
-    "mrvinph/Qwen2.5-0.5B-Instruct-Gensyn-Swarm-placid_wily_woodpecker",
-    "shivanikerai/Llama-2-7b-chat-hf-title-ner-and-title-suggestions-v2.0",
-    "Invalid-Null/PeiYangMe-0.7",
-    "open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_GradDiff_lr1e-05_alpha5_epoch5",
-    "Tesslate/Tessa-T1-3B",
-    "LorenaYannnnn/unsafe_compliance-Qwen3-0.6B-OURS_self-seed_1",
-    "HiTZ/latxa-7b-v1",
-    "huggyllama/llama-7b",
-    "FlyPig23/Llama3.2-3B_Paper_Impact_media_SFT_1ep",
-    "hongzhouyu/FineMedLM-o1",
-    "shisa-ai/shisa-v1-llama3-8b.lr-5e6",
-    "senseable/Westlake-7B",
-    "simplescaling/s1.1-3B",
-    "Borjan/finki-gpt-140M",
-    "TinyLlama/TinyLlama_v1.1_chinese",
-    "Manirajan/interview_tiny",
-    "ajibawa-2023/Young-Children-Storyteller-Mistral-7B",
-    "TinyLlama/TinyLlama_v1.1_math_code",
-    "manotham/Thai-dialogue-transalate_sft_80K",
-    "nassimjp/Maral-7B-alpha-1",
-    "MSL7/INEX16-7b",
-    "Rayeeennnnnnnn/mizan-legal-tunisian",
-    "maheshrawat18/Qwen3-4B-2507-sft1",
-    "prithivMLmods/Theta-Crucis-0.6B-Turbo1",
-    "mlabonne/llama-2-7b-miniguanaco",
-    "FreekCoolAI/privacy-gemma-qlora",
-    "lex-hue/Delexa-V0.1-7b",
-    "goldfish-models/tur_latn_100mb",
-    "shisa-ai/ablation-18-rafbestseq-shisa-v2-llama-3.1-8b-lr8e6",
-    "yash-lulla/Legal_AI_Assistant",
-    "prithivMLmods/Deepthink-Llama-3-8B-Preview",
-    "prithivMLmods/QwQ-R1-Distill-7B-CoT",
-    "sstoica12/acquisition_metamath_llama_instruct-3_1-8b-math_format_500_combined_openr1math",
-    "goldfish-models/rus_cyrl_1000mb",
-    "goldfish-models/ukr_cyrl_1000mb",
-    "eren23/dpo-binarized-NeutrixOmnibe-7B",
-    "RAANA-IA/Gheya-med",
-    "mesolitica/malaysian-tinyllama-1.1b-16k-instructions-v2",
-    "bue0912/ToolOmni-Qwen3-4B",
-    "BRlkl/distill-sft-qwen3-4b-full",
-    "prithivMLmods/PocketThinker-QwQ-3B-Instruct",
-    "prithivMLmods/Megatron-Bots-1.7B-Reasoning",
-    "kumarprince070107/geobot",
-    "ClaudioSavelli/FAME_GA_llama32-1b-instruct-qa",
-    "kairawal/Qwen3-0.6B-EL-SynthDolly-1A-E8",
-    "prithivMLmods/Poseidon-Reasoning-1.7B",
-    "PKU-Alignment/ProgressGym-HistLlama3-8B-C014-instruct-v0.2",
-    "unsloth/Qwen2-7B",
-    "prithivMLmods/Open-Xi-Math-Preview",
-    "MiniLLM/teacher-gpt2-1.5B",
-    "Rakancorle1/qwen2.5-7b_Instruct_policy_traj_30k_full",
-    "prithivMLmods/Lang-Exster-0.5B-Instruct",
-    "prithivMLmods/Nenque-MoT-0.6B-Elite14",
-    "posicube/Llama2-chat-AYT-13B",
-    "lomahony/pythia-70m-helpful-sft",
-    "kyubeen/code-grpo-checkpoint-950",
-    "nicholasKluge/TeenyTinyLlama-160m",
-    "allenai/Llama-3.1-Tulu-3-8B",
-    "BarraHome/Mistroll-7B-v2.2",
-    "posicube/Llama-chat-AY-13B",
-    "hongli-zhan/MINT-empathy-Qwen3-4B",
-    "jhhj25/qwen3-moe-neuron_structure_drop-p50-s1k-128samples-sft",
-    "Ikonz-Studios/seva-sarathi-intent-qwen3-1.7b",
-    "NECOUDBFM/Jellyfish-8B",
-    "gradients-io-tournaments/augmented-1db17e1d682d23fd",
-    "Locutusque/LocutusqueXFelladrin-TinyMistral248M-Instruct",
-    "ChuGyouk/R16",
-    "oveja1122/toolcalling-merged-demo",
-    "kmseong/llama3_2_3b-instruct-math-safedelta-scale0.8",
-    "EleutherAI/SmolLM2-1.7B-magpie-ultra-v1.0-math-431k-s",
-    "paulml/NeuralOmniBeagleMBX-v3-7B",
-    "skemessage/Qwen2.5-7B-Instruct-neuron",
+BIREN_MODELS = [
+    "zipaltrivedi/dotnet-coder-14b",
 ]
 
+CAMBRICON_MODELS = [
+]
+
+METAX_MODELS = [
+    "zipaltrivedi/dotnet-coder-14b",
+]
+
+HYGON_MODELS = [
+    "zipaltrivedi/dotnet-coder-14b",
+    "cds-jb/qwen3-14b-butterfly-subliminal-fullft",
+    "lllqaq/Qwen2.5-Coder-14B-Instruct-num11-v1-v2-v3-pairs-v3-triples-post-r2egym",
+    "deepmako/Mako-32B-Conductor",
+]
+
+KUNLUNXIN_MODELS = [
+]
+
+PPU_MODELS = [
+    "fpadovani/eng-latn-100mb-after-ppt-Dp-10mb-ckpt500_seed10",
+    "fpadovani/ita-latn-10mb-after-ppt-Dp-100mb-ckpt500_seed3407",
+    "fpadovani/ita-latn-10mb-after-ppt-Dp-10mb-ckpt500_seed3407",
+    "richardr1126/spider-skeleton-wizard-coder-merged",
+    "fpadovani/dan-latn-10mb-after-ppt-shuff-dyck-10mb-ckpt500_seed3407",
+    "fpadovani/ita-latn-10mb-after-ppt-shuff-dyck-100mb-ckpt500_seed3407",
+    "omerkaragulmez/XbyK-0.1",
+    "fpadovani/eng-latn-10mb-after-ppt-Dp-10mb-ckpt500_seed10",
+    "fpadovani/eng-latn-10mb-100mb_seed10",
+    "flax-community/gpt2-medium-indonesian",
+    "RedHatAI/QwQ-32B-Preview-quantized.w8a8",
+    "fractalego/fact-checking",
+    "KoboldAI/GPT-J-6B-Adventure",
+    "EasierAI/Falcon-3-1B",
+    "theprint/mistral-7b-cthulhu",
+    "iwalton3/phoenix",
+    "renzhenzhen/internLM2-for-triples",
+    "u2mithrandir/epsi_tmall",
+]
+
+# 本轮提交：ppu_zw_810e(18) / hygon_k100-ai(4) / MetaX_c-500(1) / Biren_166m(1)，共24个。
+#
+# 候选池是 api_verify_model_download_status_a.txt 里那 24711 个【已下载】的模型，
+# 正好满足机制A「模型必须已下载到平台存储」的前提（v1.0.37 那次157个失败就是因为没下载）。
+#
+# 本轮严格口径与放宽口径（别的卡「已验证」vs「已验证或验证中」）结果完全相同：
+# 这批已下载模型里 24007 个有验证记录的，24006 个都已至少一张卡「已验证」，
+# 放宽只多捞出 1 个；真正的瓶颈是各卡「无记录」的模型太少——
+# Kunlunxin_p-800 / Cambricon_mlu-370-x8 / Iluvatar_bi-150 均为 0，故不列入 GPU_JOBS。
+GPU_JOBS: List[Tuple[str, List[str]]] = [
+    ("ppu_zw_810e",          PPU_MODELS),
+    ("hygon_k100-ai",        HYGON_MODELS),
+    ("MetaX_c-500",          METAX_MODELS),
+    ("Biren_166m",           BIREN_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 | done | error
-    "total":         len(ALL_MODEL_IDS),
+    "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()
 
@@ -311,14 +156,124 @@ def _run_http():
     print("[http] 已关闭", flush=True)
 
 # ══════════════════════════════════════════════════════════
-#  业务逻辑
+#  各 GPU 的 config_content 模板
 # ══════════════════════════════════════════════════════════
-def _submit_task(token: str, model_id: str) -> Tuple[bool, str]:
-    headers = {
-        "Content-Type": "application/json",
-        "Authorization": f"Bearer {token}",
-    }
-    config_content = f"""gpu_type: ppu_zw_810e
+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
@@ -334,14 +289,14 @@ sut_config:
     command:
       - bash
       - /opt/t-head/entrypoint.sh
-      - python3 
-      - -m 
-      - asllm.entrypoints.api_server 
-      - --model 
-      - /model 
-      - --port 
+      - python3
+      - -m
+      - asllm.entrypoints.api_server
+      - --model
+      - /model
+      - --port
       - '30000'
-      - --host 
+      - --host
       - 0.0.0.0
       - --served-model-name
       - llm
@@ -368,21 +323,67 @@ ref_config:
       - -tp
       - '1'
 """
+    
+    elif gpu_type == "Iluvatar_bi-150":
+            return f"""docker_image: harbor-contest.4pd.io/luopingyi/enginex-iluvatar-bi150/vllm:0.8.3
+nv_docker_image: harbor.4pd.io/dooke/vllm/vllm/vllm-openai:v0.11.0
+framework: vllm
+api: completion
+temperature: 0.7
+repetition_penalty: 1.2
+top_p: 0.9
+
+max_model_len: 4096
+max_tokens: 1024
+sut_config:
+  gpu_num: 1
+  values:
+    command: ['vllm', 'serve', '/model', '--port', '80', '--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']  
+"""
+    
+    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],
+        "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,
+            "gpuType":  gpu_type,
             "taskType": TASK_TYPE,
         }],
     }
-    print(f"[payload] {json.dumps(payload, indent=2, ensure_ascii=False)}", flush=True)
+    print(f"[payload] gpu={gpu_type} model={model_id}", flush=True)
     try:
         resp = requests.post(
             BASE_URL + SUBMIT_ENDPOINT,
@@ -393,46 +394,88 @@ ref_config:
         result = resp.json()
         if result.get("code") == 0:
             task_id = result.get("data", {}).get("id", "")
-            print(f"[worker] OK  {model_id}  task_id={task_id}", flush=True)
-            return True, task_id
+            print(f"[worker] OK  {model_id} (GPU={gpu_type})  task_id={task_id}", flush=True)
+            return True, task_id, ""
         else:
-            print(f"[worker] FAIL {model_id}: {result.get('message')}", flush=True)
-            return False, ""
+            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}: {e}", flush=True)
-        return False, ""
+        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]] = []
+    successful: List[Tuple[str, str, str]] = []
     token = AUTH_TOKEN
     print("[worker] 使用预设 Token，跳过登录", flush=True)
 
-    for model_id in ALL_MODEL_IDS:
-        if _shutdown.is_set():
-            break
-        ok, task_id = _submit_task(token, model_id)
-        if ok:
-            _state["submitted"] += 1
-            successful.append((task_id, model_id))
-        else:
-            _state["failed"] += 1
+    # 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
+    ]
 
-    # 写入结果文件
-    try:
-        with open("submitted_validation_tasks.txt", "w", encoding="utf-8") as f:
-            for tid, mid in successful:
-                f.write(f"{tid}\t{mid}\n")
-    except Exception:
-        pass
+    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"[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
@@ -465,4 +508,4 @@ def main():
 
 
 if __name__ == "__main__":
-    main()
+    main()