Fine-tuning/backend/app/services/model_test_service.py

from pathlib import Path
from typing import Any

from app.config import get_settings
from app.core.logging import logger

settings = get_settings()


async def test_model(model_id: str, prompt: str, max_new_tokens: int = 128, temperature: float = 0.8, top_p: float = 0.95) -> dict[str, Any]:
    """加载已缓存模型并生成测试响应。"""
    if settings.use_remote_compute:
        return _test_model_remote(model_id, prompt, max_new_tokens, temperature, top_p)
    return _test_model_local(model_id, prompt, max_new_tokens, temperature, top_p)


def _test_model_remote(model_id: str, prompt: str, max_new_tokens: int, temperature: float, top_p: float) -> dict[str, Any]:
    """通过 SSH 在算力节点执行模型测试。

    通过环境变量传递参数，base64 编码脚本通过 stdin 管道传给 docker exec -i python。
    """
    import base64
    import json
    from app.core.remote_executor import ssh_exec

    container = settings.compute_node_docker_container
    python = settings.compute_node_python
    workdir = settings.compute_node_workdir

    # 参数通过 base64 编码，脚本内通过 os.environ 读取，完全避免引号/转义问题
    prompt_b64 = base64.b64encode(prompt.encode('utf-8')).decode()
    do_sample = str(temperature > 0).lower()

    # 独立的 Python 脚本（参数通过环境变量传入）
    script = rf"""\
import json, os, base64
from pathlib import Path
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModel

def find_model_path(model_id):
    for base in ['/root/.cache/huggingface/hub', '/root/.cache/modelscope/hub', '/root/models']:
        bp = Path(base)
        if not bp.is_dir():
            continue
        try:
            for child in bp.rglob('config.json'):
                if child.parent.is_dir():
                    return str(child.parent)
        except Exception:
            pass
    return None

model_id = os.environ.get('MODEL_ID', '')
prompt = base64.b64decode(os.environ.get('PROMPT_B64', '')).decode('utf-8')
max_new_tokens = int(os.environ.get('MAX_TOKENS', '128'))
temperature = float(os.environ.get('TEMPERATURE', '0.8'))
top_p = float(os.environ.get('TOP_P', '0.95'))
do_sample = os.environ.get('DO_SAMPLE', 'true').lower() == 'true'

model_path = find_model_path(model_id)
if model_path is None:
    print(json.dumps({{'error': f'Model not found in cache: {{model_id}}'}}))
    exit(1)

t = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
t.pad_token = t.pad_token or t.eos_token

m = None
for cls, kw in [(AutoModelForCausalLM, {{'trust_remote_code': True}}), (AutoModel, {{'trust_remote_code': True}})]:
    try:
        m = cls.from_pretrained(model_path, torch_dtype=torch.float16, device_map='auto', **kw)
        break
    except Exception:
        pass

if m is None:
    print(json.dumps({{'error': 'Unable to load model'}}))
    exit(1)

m.eval()
inp = t(prompt, return_tensors='pt').to(m.device)
out = m.generate(**inp, max_new_tokens=max_new_tokens, temperature=temperature, top_p=top_p, do_sample=do_sample, pad_token_id=t.eos_token_id)
gen = t.decode(out[0][inp['input_ids'].shape[1]:], skip_special_tokens=True)
print(json.dumps({{'generated_text': gen}}))
"""

    script_b64 = base64.b64encode(script.encode()).decode()

    # 环境变量通过 docker exec -e 传入容器，脚本通过 stdin 管道传入
    remote_cmd = (
        f"echo {script_b64} | base64 -d | "
        f"docker exec -i -w {workdir} "
        f"-e MODEL_ID={model_id} "
        f"-e PROMPT_B64={prompt_b64} "
        f"-e MAX_TOKENS={max_new_tokens} "
        f"-e TEMPERATURE={temperature} "
        f"-e TOP_P={top_p} "
        f"-e DO_SAMPLE={do_sample} "
        f"{container} {python}"
    )

    code, stdout, stderr = ssh_exec(remote_cmd, timeout=600)

    logger.info(f"Remote test result: code={code}, stdout_len={len(stdout)}, stderr_len={len(stderr)}")
    if stdout:
        logger.info(f"stdout (first 500): {stdout[:500]}")
    if stderr:
        logger.info(f"stderr (first 500): {stderr[:500]}")

    if code != 0:
        logger.error(f"Remote model test failed: {stderr}")
        return {"error": stderr.strip() or "Remote test failed"}

    # 提取最后一行 JSON
    for line in reversed(stdout.strip().split("\n")):
        line = line.strip()
        if line.startswith("{"):
            try:
                result = json.loads(line)
                result["model_id"] = model_id
                result["prompt"] = prompt
                return result
            except json.JSONDecodeError:
                continue

    return {"error": f"Invalid response: {stdout[:500]}"}


async def _test_model_local(model_id: str, prompt: str, max_new_tokens: int, temperature: float, top_p: float) -> dict[str, Any]:
    """本地执行模型测试（仅用于开发环境）。"""
    import torch
    from transformers import AutoModel, AutoModelForCausalLM, AutoTokenizer, AutoConfig

    from app.services.model_service import resolve_model_path
    model_path = await resolve_model_path(model_id)
    if not model_path:
        return {"error": f"Model not found in cache: {model_id}"}

    model_dir = Path(model_path)
    if not (model_dir / "config.json").exists():
        return {"error": f"Model directory not found: {model_dir}"}

    tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    # 通用加载策略：尝试多种加载方式，自动兼容各种新架构
    model = None
    for loader_cls, kwargs in [
        (AutoModelForCausalLM, {"trust_remote_code": True}),
        (AutoModel, {"trust_remote_code": True}),
    ]:
        try:
            model = loader_cls.from_pretrained(
                model_dir,
                torch_dtype=torch.float16,
                device_map="auto",
                **kwargs,
            )
            break
        except Exception:
            continue

    if model is None:
        return {"error": f"Unable to load model with any available loader. Model type may not be supported yet."}
    model.eval()

    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            top_p=top_p,
            do_sample=temperature > 0,
            pad_token_id=tokenizer.eos_token_id,
        )

    generated_text = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)

    return {
        "model_id": model_id,
        "prompt": prompt,
        "generated_text": generated_text,
    }