train.py

from unsloth import FastLanguageModel
from transformers import TrainingArguments, Trainer

# Load quantized model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="deepseek-ai/DeepSeek-V3-0324",
    dtype=torch.bfloat16,
    load_in_4bit=True,  # Or 2.71-bit
    token=os.environ["HF_TOKEN"]
)
FastLanguageModel.for_training(model)

# Training arguments
training_args = TrainingArguments(
    output_dir="/app/checkpoints",
    per_device_train_batch_size=4,  # Adjust for A100 40GB/80GB
    per_device_eval_batch_size=4,
    num_train_epochs=2,
    learning_rate=2e-5,
    save_steps=500,
    save_total_limit=2,
    evaluation_strategy="steps",
    eval_steps=500,
    logging_dir="/app/logs",
    logging_steps=100,
    fp16=False,  # bfloat16 for A100
    deepspeed="/app/ds_config.json"
)

# DeepSpeed config
with open("/app/ds_config.json", "w") as f:
    f.write('''
    {
        "fp16": {"enabled": false},
        "bf16": {"enabled": true},
        "zero_optimization": {
            "stage": 3,
            "offload_optimizer": {"device": "cpu"},
            "offload_param": {"device": "cpu"}
        },
        "train_batch_size": "auto",
        "gradient_accumulation_steps": 4
    }
    ''')

# Initialize trainer
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset["train"],
    eval_dataset=tokenized_dataset["test"]
)

# Train
trainer.train()

# Save model
model.save_pretrained("/app/fine_tuned_model")
tokenizer.save_pretrained("/app/fine_tuned_model")