431 lines
17 KiB
Python
431 lines
17 KiB
Python
"""
|
||
xc_validation_strategy — 主入口
|
||
|
||
启动后针对 4 张 GPU 卡（Biren_166m / Cambricon_mlu-370-x8 / MetaX_c-500 /
|
||
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，
|
||
Bearer Token 认证），之后保持 HTTP 服务存活。
|
||
同时暴露 /health（K8s 探活）和 /status（运行状态）。
|
||
"""
|
||
|
||
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.eyJ1c2VyQWNjb3VudCI6Inpob3VzaGFzaGEiLCJpZCI6MTQsInVzZXJSb2xlIjoibGVhZGVyYm9hcmQiLCJleHAiOjE3ODU3NDY3NTMsImlhdCI6MTc4NTE0MTk1M30.KwUuefNAFSNwq3_Pnaw2nef8ZC6WgsECQ_LMeQnKk2c"
|
||
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 = [
|
||
"BigRatz/LOL-AI-2026",
|
||
"aimeri/spoomplesmaxx-cardmaker-v1",
|
||
"Alibaba-DT/Logics-STEM-8B-SFT",
|
||
"Muneebmn123/insurance-voice-qwen25-1_5b",
|
||
"AnkitBirGurung/NEMO-12B-SFT-Further",
|
||
"danilarudenko/editorai-mini",
|
||
"EphemeralYou/Prompt-Refine-MiniCPM5-1B",
|
||
"mtepe01/mentorx-mistral-7b-automata-merged",
|
||
"DarkArtsForge/Helix-SCE-12B-jh",
|
||
"rpant/iolai26-solve",
|
||
"NithinAI12/NithinX-Omni-LLM-v1",
|
||
"ConvexAI/Luminex-34B-v0.2",
|
||
"codellama/CodeLlama-34b-hf",
|
||
"Lipas007/iol-ai-2026-qwen14b-awq",
|
||
"D-Z-W/finetuned-teacher",
|
||
"huan1999/ziya-llama-13b-medical-merged",
|
||
"codellama/CodeLlama-34b-Python-hf",
|
||
"ld4ad/gemma-2-9b-dunhuang",
|
||
"harindhar10/Olmo-7b_1M_Smiles_lora",
|
||
"allenai/Olmo-3-7B-Think-DPO",
|
||
"allenai/Olmo-3-32B-Think-DPO",
|
||
"RedHatAI/gemma-2-9b-it",
|
||
"prashanthsura/gemma-2-2b-legal-financial-sft",
|
||
"pfnet/plamo-2-8b",
|
||
"sail/Sailor2-20B-128K-SFT",
|
||
"facebook/layerskip-llama3.2-1B",
|
||
"Qwen/Qwen2.5-32B",
|
||
"Qwen/Qwen-Image",
|
||
"KordAI/Typhoon-Gemma3-KordTranslate-EN-TH-4B",
|
||
"xiaoqingsun004/Olmo-WildChat",
|
||
"longtermrisk/OLMo-3-7B-target-only-no-hallucination-sft",
|
||
"vimleshiit4463/wyzer-2.0-smollm2-135m",
|
||
"trl-lib/pythia-1b-deduped-tldr-sft",
|
||
]
|
||
|
||
CAMBRICON_MODELS = [
|
||
"clear-blue-sky/evolai-reborn-tfm-008",
|
||
"clear-blue-sky/evolai-reborn-tfm-010",
|
||
"BigRatz/LOL-AI-2026",
|
||
"clear-blue-sky/evolai-reborn-tfm-001",
|
||
"clear-blue-sky/evolai-reborn-tfm-019",
|
||
"clear-blue-sky/evolai-reborn-tfm-007",
|
||
"aimeri/spoomplesmaxx-cardmaker-v1",
|
||
"aipatseer/inst_ft_qwen_0.6b_summ",
|
||
"clear-blue-sky/evolai-reborn-tfm-003",
|
||
"clear-blue-sky/evolai-reborn-tfm-004",
|
||
"Alibaba-DT/Logics-STEM-8B-SFT",
|
||
"clear-blue-sky/evolai-reborn-tfm-002",
|
||
"prism-ml/Ternary-Bonsai-4B-unpacked",
|
||
"Muneebmn123/insurance-voice-qwen25-1_5b",
|
||
"AnkitBirGurung/NEMO-12B-SFT-Further",
|
||
"danilarudenko/editorai-mini",
|
||
"rpant/iolai26-solve",
|
||
]
|
||
|
||
METAX_MODELS = [
|
||
"Phoenix9781/evolai-tf-model-105",
|
||
"clear-blue-sky/evolai-reborn-tfm-008",
|
||
"clear-blue-sky/evolai-reborn-tfm-010",
|
||
"BigRatz/LOL-AI-2026",
|
||
"clear-blue-sky/evolai-reborn-tfm-001",
|
||
"clear-blue-sky/evolai-reborn-tfm-019",
|
||
"clear-blue-sky/evolai-reborn-tfm-007",
|
||
"clear-blue-sky/evolai-reborn-tfm-005",
|
||
"Lin2es/evolai-tfm-03o",
|
||
"clear-blue-sky/evolai-reborn-tfm-009",
|
||
"aimeri/spoomplesmaxx-cardmaker-v1",
|
||
"aipatseer/inst_ft_qwen_0.6b_summ",
|
||
"reaperdoesntknow/DualMind-TKD-Agentic-1.7B",
|
||
"clear-blue-sky/evolai-reborn-tfm-011",
|
||
"clear-blue-sky/evolai-reborn-tfm-003",
|
||
"clear-blue-sky/evolai-reborn-tfm-004",
|
||
"clear-blue-sky/evolai-reborn-tfm-002",
|
||
"amd/ReasonLite-0.6B",
|
||
"prism-ml/Ternary-Bonsai-4B-unpacked",
|
||
"Muneebmn123/insurance-voice-qwen25-1_5b",
|
||
"danilarudenko/editorai-mini",
|
||
"rpant/iolai26-solve",
|
||
]
|
||
|
||
KUNLUNXIN_MODELS = [
|
||
"Patriae/patriae-cuban-dialect-model",
|
||
"Phoenix9781/evolai-tf-model-105",
|
||
"clear-blue-sky/evolai-reborn-tfm-008",
|
||
"clear-blue-sky/evolai-reborn-tfm-010",
|
||
"clear-blue-sky/evolai-reborn-tfm-001",
|
||
"clear-blue-sky/evolai-reborn-tfm-019",
|
||
"clear-blue-sky/evolai-reborn-tfm-007",
|
||
"clear-blue-sky/evolai-reborn-tfm-005",
|
||
"Lin2es/evolai-tfm-03o",
|
||
"clear-blue-sky/evolai-reborn-tfm-009",
|
||
"aimeri/spoomplesmaxx-cardmaker-v1",
|
||
"aipatseer/inst_ft_qwen_0.6b_summ",
|
||
"clear-blue-sky/evolai-reborn-tfm-011",
|
||
"clear-blue-sky/evolai-reborn-tfm-003",
|
||
"clear-blue-sky/evolai-reborn-tfm-004",
|
||
"clear-blue-sky/evolai-reborn-tfm-002",
|
||
"amd/ReasonLite-0.6B",
|
||
"prism-ml/Ternary-Bonsai-4B-unpacked",
|
||
"EthanGao123/CellHermes-v1.0",
|
||
"RecursiveMAS/Mixture-Science-BioMistral-7B",
|
||
"chenyitian-shanshu/SIRL-Gurobi",
|
||
"RedHatAI/gemma-2-9b-it",
|
||
"gaunernst/gemma-3-27b-it-qat-autoawq",
|
||
"promotion/qwen3-8b-simpo-avg-b2p5-g1p0-s42",
|
||
"llamaindex/vdr-2b-multi-v1",
|
||
"KordAI/Typhoon-Gemma3-KordTranslate-EN-TH-4B",
|
||
"TheDrummer/UnslopNemo-12B-v3",
|
||
"gradients-io-tournaments/tournament-tourn_c5d86c82ce819a79_20260706-b78a01d4-0a6a-49e1-9190-5e88ae329937-5DS6XMVr",
|
||
"ArliAI/Mistral-Nemo-12B-ArliAI-RPMax-v1.1",
|
||
]
|
||
|
||
# 按顺序处理：Biren → Cambricon → MetaX → Kunlunxin
|
||
GPU_JOBS: List[Tuple[str, List[str]]] = [
|
||
("Biren_166m", BIREN_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)
|
||
|
||
# ══════════════════════════════════════════════════════════
|
||
# 全局状态（供 /status 展示）
|
||
# ══════════════════════════════════════════════════════════
|
||
_state = {
|
||
"strategy_id": STRATEGY_ID,
|
||
"phase": "starting", # starting | submitting | done | error
|
||
"total": TOTAL_MODELS,
|
||
"submitted": 0,
|
||
"failed": 0,
|
||
"per_gpu": {gpu: 0 for gpu, _ in GPU_JOBS},
|
||
"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)
|
||
|
||
# ══════════════════════════════════════════════════════════
|
||
# 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']
|
||
"""
|
||
else:
|
||
raise ValueError(f"未知的 GPU_TYPE: {gpu_type}")
|
||
|
||
|
||
|
||
|
||
|
||
# ══════════════════════════════════════════════════════════
|
||
# 业务逻辑
|
||
# ══════════════════════════════════════════════════════════
|
||
def _submit_task(token: str, gpu_type: str, model_id: str) -> Tuple[bool, 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:
|
||
print(f"[worker] FAIL {model_id} (GPU={gpu_type}): {result.get('message')}", flush=True)
|
||
return False, ""
|
||
except Exception as e:
|
||
print(f"[worker] ERROR {model_id} (GPU={gpu_type}): {e}", flush=True)
|
||
return False, ""
|
||
|
||
|
||
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)
|
||
|
||
for gpu_type, model_list in GPU_JOBS:
|
||
if _shutdown.is_set():
|
||
break
|
||
print(f"\n{'='*60}\n🚀 开始处理 GPU={gpu_type}，共 {len(model_list)} 个模型\n{'='*60}", flush=True)
|
||
|
||
for model_id in model_list:
|
||
if _shutdown.is_set():
|
||
break
|
||
ok, task_id = _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))
|
||
else:
|
||
_state["failed"] += 1
|
||
|
||
# 写入结果文件
|
||
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
|
||
|
||
_state["finished_at"] = datetime.utcnow().isoformat()
|
||
_state["phase"] = "done"
|
||
print(
|
||
f"[worker] 完成 submitted={_state['submitted']} failed={_state['failed']} "
|
||
f"total={_state['total']} per_gpu={_state['per_gpu']}",
|
||
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()
|