YG_FT/compute/engines/llama_factory/adapter.py

from __future__ import annotations

import re
from dataclasses import dataclass
from pathlib import Path
from typing import Any


@dataclass(frozen=True)
class LlamaFactoryCommand:
    command: list[str]
    work_dir: str
    env: dict[str, str]


def validate_config(config: dict[str, Any]) -> list[str]:
    errors: list[str] = []
    if not config.get("base_model") and not config.get("model_name_or_path"):
        errors.append("base_model or model_name_or_path is required")
    if not config.get("dataset") and not config.get("dataset_dir"):
        errors.append("dataset or dataset_dir is required")
    try:
        learning_rate = float(config.get("learning_rate", 0.0002))
    except (TypeError, ValueError):
        learning_rate = 0
    if learning_rate <= 0:
        errors.append("learning_rate must be greater than zero")
    try:
        epochs = int(config.get("n_epochs", config.get("num_train_epochs", 1)))
    except (TypeError, ValueError):
        epochs = 0
    if epochs <= 0:
        errors.append("n_epochs must be greater than zero")
    return errors


def _optional_arg(config: dict[str, Any], command: list[str], option: str, *keys: str) -> None:
    for key in keys:
        value = config.get(key)
        if value is not None and value != "":
            command.extend([option, str(value)])
            return


def build_command(config: dict[str, Any], llama_factory_home: str = "/app/LLaMA-Factory") -> LlamaFactoryCommand:
    errors = validate_config(config)
    if errors:
        raise ValueError("; ".join(errors))

    engine = str(config.get("engine") or config.get("training_engine") or "llama_factory")
    if engine == "smoke":
        output_dir = config.get("output_dir") or f"/data/yg-ft/outputs/{config.get('name', 'training-smoke')}"
        script = (
            "import json, os, time; "
            f"out={str(output_dir)!r}; "
            "os.makedirs(out, exist_ok=True); "
            "print('[INFO] smoke training started', flush=True); "
            "\nfor step in range(1, 7):\n"
            "    loss=round(1.8/(step+1), 4)\n"
            "    lr=round(0.0002*(1-step/10), 8)\n"
            "    print({'loss': loss, 'grad_norm': round(0.4 + step*0.03, 4), 'learning_rate': lr, 'epoch': round(step/6, 4)}, flush=True)\n"
            "    time.sleep(0.4)\n"
            "\nopen(os.path.join(out, 'adapter_config.json'), 'w', encoding='utf-8').write(json.dumps({'engine':'smoke','status':'completed'})); "
            "print('***** train metrics *****', flush=True); "
            "print('train_loss = 0.12', flush=True); "
            "print('***** train metrics end *****', flush=True)"
        )
        return LlamaFactoryCommand(command=["python", "-u", "-c", script], work_dir="/app", env={})

    model_path = config.get("base_model") or config.get("model_name_or_path")
    dataset = config.get("dataset") or config.get("dataset_name")
    dataset_dir = config.get("dataset_dir")
    output_dir = config.get("output_dir") or f"/data/yg-ft/outputs/{config.get('name', 'training-job')}"
    command = [
        "llamafactory-cli",
        "train",
        "--stage",
        str(config.get("stage", "sft")).lower(),
        "--do_train",
        "true",
        "--model_name_or_path",
        str(model_path),
        "--dataset",
        str(dataset or "default"),
        "--template",
        str(config.get("template", "qwen")),
        "--finetuning_type",
        str(config.get("train_method", config.get("finetuning_type", "lora"))),
        "--output_dir",
        str(output_dir),
        "--per_device_train_batch_size",
        str(config.get("batch_size", 2)),
        "--learning_rate",
        str(config.get("learning_rate", 0.0002)),
        "--num_train_epochs",
        str(config.get("n_epochs", 3)),
        "--save_steps",
        str(config.get("save_steps", 50)),
        "--logging_steps",
        str(config.get("logging_steps", 10)),
        "--overwrite_output_dir",
        "true",
        "--plot_loss",
        "true",
    ]
    if dataset_dir:
        command.extend(["--dataset_dir", str(dataset_dir)])
    _optional_arg(config, command, "--cutoff_len", "max_length", "cutoff_len")
    _optional_arg(config, command, "--lr_scheduler_type", "lr_scheduler_type")
    _optional_arg(config, command, "--warmup_ratio", "warmup_ratio")
    _optional_arg(config, command, "--weight_decay", "weight_decay")
    _optional_arg(config, command, "--lora_rank", "lora_rank", "rank")
    _optional_arg(config, command, "--lora_alpha", "lora_alpha")
    _optional_arg(config, command, "--lora_dropout", "lora_dropout")
    quantization_bit = int(config.get("quantization_bit", 0) or 0)
    if quantization_bit in {4, 8}:
        command.extend(["--quantization_bit", str(quantization_bit)])
    return LlamaFactoryCommand(command=command, work_dir=str(Path(llama_factory_home)), env={})


def parse_log_line(line: str) -> dict[str, float] | None:
    if "loss" not in line or "learning_rate" not in line:
        return None
    result: dict[str, float] = {}
    for key in ["loss", "grad_norm", "learning_rate", "epoch"]:
        match = re.search(rf"['\"]?{key}['\"]?\s*:\s*([-+]?\d+(?:\.\d+)?(?:[eE][-+]?\d+)?)", line)
        if match:
            result[key] = float(match.group(1))
    return result or None