submit 338 nonquantized candidates on PPU and Hygon

This commit is contained in:
2026-09-24 22:17:16 +08:00
parent efee86e49c
commit a3d9109764
4 changed files with 371 additions and 45 deletions

66
main.py
View File

@@ -1,9 +1,9 @@
"""
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)
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。
本轮仅提交 ppu_zw_810e 和 hygon_k100-ai 的 2026-09-24 非量化候选清单;
其他 GPU 的 config_content 模板仍保留,但未列入 GPU_JOBS。
(/adminApi/async/task/create-contest-task,
Bearer Token 认证),之后保持 HTTP 服务存活。
@@ -22,6 +22,7 @@ import signal
import threading
from datetime import datetime
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
from pathlib import Path
from typing import List, Tuple
import requests
@@ -33,7 +34,7 @@ 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"
AUTH_TOKEN = os.environ.get("AUTH_TOKEN", "eyJ0eXAiOiJKV1QiLCJhbGciOiJIUzI1NiJ9.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3OTA4NjQwMDMsImlhdCI6MTc5MDI1OTIwM30.T23Tp3xcI8kkKIOwRCXmZlpe3Qo3sOIxZ8n6NbzJJ2M")
CONTEST_API_TOKEN = "ef1ef82f3c9efee413d602345fbe224d"
CONTRIBUTORS = "zhoushasha"
TASK_TYPE = "text-generation"
@@ -43,8 +44,22 @@ 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)
# 各 GPU 的模型列表
# ══════════════════════════════════════════════════════════
MODEL_DIR = Path(__file__).resolve().parent / "model_ids"
def load_model_ids(filename: str) -> List[str]:
model_ids = [
line.strip()
for line in (MODEL_DIR / filename).read_text(encoding="utf-8").splitlines()
if line.strip()
]
if not model_ids or len(model_ids) != len(set(model_ids)):
raise ValueError(f"模型列表为空或含重复 ID: {filename}")
return model_ids
BIREN_MODELS = [
"zipaltrivedi/dotnet-coder-14b",
]
@@ -56,51 +71,18 @@ 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",
]
HYGON_MODELS = load_model_ids("hygon_k100-ai_2026-09-24.txt")
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",
]
PPU_MODELS = load_model_ids("ppu_zw_810e_2026-09-24.txt")
# 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。
# 本轮提交:2026-09-24 已成功下载、非量化、目标卡无验证记录,且其他卡至少一张「已验证」的模型。
# 仅 ppu_zw_810e(17) 和 hygon_k100-ai(321),共 338 个;不重复提交上一轮的 24 个任务。
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)