train.py

import os
from datasets import load_from_disk
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
import torch

model_name = "mergekit-community/Qwen-2.5-Coder"  # ваш base_model
out_dir = "D:\\out_peft"

os.environ['HF_HOME'] = 'D:\\huggingface_cache'

# load dataset
ds = load_from_disk("processed_ds")
train_ds = ds["train"]
eval_ds = ds["test"]

# tokenizer & model
tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_4bit=True)

# prepare and apply LoRA
model = prepare_model_for_kbit_training(model)
lora_config = LoraConfig(
    r=8, lora_alpha=32, target_modules=["q_proj","v_proj","k_proj","o_proj"],
    lora_dropout=0.1, bias="none", task_type="CAUSAL_LM"
)
model = get_peft_model(model, lora_config)

# tokenization fn
def tokenize_fn(batch):
    inputs = [a + tokenizer.eos_token + b for a,b in zip(batch["input_text"], batch["target_text"])]
    out = tokenizer(inputs, truncation=True, padding="max_length", max_length=1024)
    out["labels"] = out["input_ids"].copy()
    return out

train_ds = train_ds.map(tokenize_fn, batched=True, remove_columns=train_ds.column_names)
eval_ds = eval_ds.map(tokenize_fn, batched=True, remove_columns=eval_ds.column_names)

training_args = TrainingArguments(
    output_dir=out_dir,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=8,
    num_train_epochs=3,
    learning_rate=2e-4,
    fp16=True,
    logging_steps=50,
    save_total_limit=2,
    optim="paged_adamw_8bit"
)

trainer = Trainer(model=model, args=training_args, train_dataset=train_ds, eval_dataset=eval_ds)
trainer.train()

# save PEFT weights (small)
model.save_pretrained(out_dir)
tokenizer.save_pretrained(out_dir)
print("Saved PEFT to", out_dir)