spark-scripts/finetune.py

#!/usr/bin/env python3
"""Fine-tune TAIDE 12B on Audrey Tang transcripts using SFT.

On DGX Spark (128GB unified memory), this runs full-parameter SFT
on a 12B model. For memory safety, we use gradient checkpointing
and bf16 mixed precision.

Expected time: ~4-8 hours for 2 epochs on 56K training examples.
"""

import torch
from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    TrainingArguments,
)
from trl import SFTTrainer, SFTConfig

# Model and data paths
MODEL_PATH = "./models/taide-12b"
TRAIN_PATH = "./training_data/train.jsonl"
EVAL_PATH = "./training_data/eval.jsonl"
OUTPUT_DIR = "./models/taide-12b-audrey"

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Load model — full precision on DGX Spark's 128GB unified memory
model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",  # Flash attention via SDPA
)
model.config.use_cache = False  # Required for gradient checkpointing

# Load datasets
train_dataset = load_dataset("json", data_files=TRAIN_PATH, split="train")
eval_dataset = load_dataset("json", data_files=EVAL_PATH, split="train")

print(f"Train: {len(train_dataset)}, Eval: {len(eval_dataset)}")


def format_chat(example):
    """Format messages into the Gemma chat template."""
    return tokenizer.apply_chat_template(
        example["messages"],
        tokenize=False,
        add_generation_prompt=False,
    )


# Training config
training_args = SFTConfig(
    output_dir=OUTPUT_DIR,
    num_train_epochs=2,
    per_device_train_batch_size=2,
    per_device_eval_batch_size=2,
    gradient_accumulation_steps=8,  # effective batch size = 16
    gradient_checkpointing=True,
    gradient_checkpointing_kwargs={"use_reentrant": False},
    learning_rate=2e-5,
    lr_scheduler_type="cosine",
    warmup_ratio=0.05,
    weight_decay=0.01,
    bf16=True,
    logging_steps=10,
    eval_strategy="steps",
    eval_steps=500,
    save_strategy="steps",
    save_steps=500,
    save_total_limit=3,
    max_seq_length=4096,
    packing=True,  # Pack short examples together for efficiency
    dataset_text_field="text",
    report_to="none",
)

# Preprocess: apply chat template
train_dataset = train_dataset.map(
    lambda x: {"text": format_chat(x)}, remove_columns=train_dataset.column_names
)
eval_dataset = eval_dataset.map(
    lambda x: {"text": format_chat(x)}, remove_columns=eval_dataset.column_names
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    processing_class=tokenizer,
)

print("Starting training...")
trainer.train()

# Save final model
trainer.save_model(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)
print(f"Model saved to {OUTPUT_DIR}")