app.py

import gradio as gr
import spaces
import torch

from datasets import load_dataset
from transformers import (
    AutoConfig,
    AutoTokenizer,
    AutoModelForCausalLM,
    DataCollatorForLanguageModeling,
    Trainer,
    TrainingArguments,
    pipeline,
    BitsAndBytesConfig,
)

# PEFT (LoRA / QLoRA)
from peft import LoraConfig, TaskType, get_peft_model, prepare_model_for_kbit_training, PeftModel

##############################################################################
# ZeroGPU + QLoRA Example
##############################################################################

TEXT_PIPELINE = None        # Pipeline for wuhp/myr1 (fine-tuned or base)
COMPARISON_PIPELINE = None  # Pipeline for the DeepSeek model

NUM_EXAMPLES = 50  # We'll train on 50 rows for demonstration

@spaces.GPU(duration=300)  # up to 5 min
def finetune_small_subset():
    """
    1) Loads 'wuhp/myr1' in 4-bit quantization (QLoRA style),
    2) Adds LoRA adapters (trainable),
    3) Trains on a small subset of the Magpie dataset,
    4) Saves LoRA adapter to 'finetuned_myr1',
    5) Reloads LoRA adapters for inference in a pipeline.
    """

    # --- 1) Load a small subset of the Magpie dataset ---
    ds = load_dataset(
        "Magpie-Align/Magpie-Reasoning-V2-250K-CoT-Deepseek-R1-Llama-70B", 
        split="train"
    )

    # For demonstration, pick a single conversation_id
    unique_ids = list(set(ds["conversation_id"]))
    single_id = unique_ids[0]
    ds = ds.filter(lambda x: x["conversation_id"] == single_id)

    # Then select only NUM_EXAMPLES from that subset
    ds = ds.select(range(min(NUM_EXAMPLES, len(ds))))

    # --- 2) Setup 4-bit quantization with BitsAndBytes ---
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.bfloat16,  # or torch.float16
        bnb_4bit_use_double_quant=True,
        bnb_4bit_quant_type="nf4",
    )

    config = AutoConfig.from_pretrained(
        "wuhp/myr1", 
        subfolder="myr1",
        trust_remote_code=True
    )
    tokenizer = AutoTokenizer.from_pretrained(
        "wuhp/myr1", 
        subfolder="myr1",
        trust_remote_code=True
    )

    base_model = AutoModelForCausalLM.from_pretrained(
        "wuhp/myr1",
        subfolder="myr1",
        config=config,
        quantization_config=bnb_config,  # <--- QLoRA 4-bit
        device_map="auto",
        trust_remote_code=True
    )

    # Prepare the model for k-bit training
    base_model = prepare_model_for_kbit_training(base_model)

    # --- 3) Create LoRA config & wrap the base model in LoRA ---
    lora_config = LoraConfig(
        r=16,
        lora_alpha=32,
        lora_dropout=0.05,
        bias="none",
        target_modules=["q_proj", "v_proj"],
        task_type=TaskType.CAUSAL_LM,
    )
    lora_model = get_peft_model(base_model, lora_config)

    # --- 4) Tokenize dataset ---
    def tokenize_fn(ex):
        """
        Combine instruction + response into a single text.
        You can adjust this to include more fields or different formatting.
        """
        text = (
            f"Instruction: {ex['instruction']}\n\n"
            f"Response: {ex['response']}"
        )
        return tokenizer(text, truncation=True, max_length=512)

    ds = ds.map(tokenize_fn, batched=False, remove_columns=ds.column_names)
    ds.set_format("torch")

    collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)

    # Training args
    training_args = TrainingArguments(
        output_dir="finetuned_myr1",
        num_train_epochs=1,
        per_device_train_batch_size=1,
        gradient_accumulation_steps=2,
        logging_steps=5,
        save_steps=999999,   # effectively don't save mid-epoch
        save_total_limit=1,
        fp16=False,          # rely on bfloat16 from quantization
    )

    # Trainer
    trainer = Trainer(
        model=lora_model,
        args=training_args,
        train_dataset=ds,
        data_collator=collator,
    )

    # --- 5) Train ---
    trainer.train()

    # --- 6) Save LoRA adapter + tokenizer ---
    trainer.model.save_pretrained("finetuned_myr1")
    tokenizer.save_pretrained("finetuned_myr1")

    # --- 7) Reload the base model + LoRA adapter for inference
    base_model_2 = AutoModelForCausalLM.from_pretrained(
        "wuhp/myr1",
        subfolder="myr1",
        config=config,
        quantization_config=bnb_config,
        device_map="auto",
        trust_remote_code=True
    )
    base_model_2 = prepare_model_for_kbit_training(base_model_2)

    lora_model_2 = PeftModel.from_pretrained(
        base_model_2,
        "finetuned_myr1",
    )

    global TEXT_PIPELINE
    TEXT_PIPELINE = pipeline("text-generation", model=lora_model_2, tokenizer=tokenizer)

    return "Finetuning complete. Model loaded for inference."


def ensure_pipeline():
    """
    If we haven't finetuned yet (TEXT_PIPELINE is None),
    load the base model in 4-bit with NO LoRA.
    """
    global TEXT_PIPELINE
    if TEXT_PIPELINE is None:
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_compute_dtype=torch.bfloat16,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type="nf4",
        )
        config = AutoConfig.from_pretrained("wuhp/myr1", subfolder="myr1", trust_remote_code=True)
        tokenizer = AutoTokenizer.from_pretrained("wuhp/myr1", subfolder="myr1", trust_remote_code=True)
        base_model = AutoModelForCausalLM.from_pretrained(
            "wuhp/myr1",
            subfolder="myr1",
            config=config,
            quantization_config=bnb_config,
            device_map="auto",
            trust_remote_code=True
        )
        TEXT_PIPELINE = pipeline("text-generation", model=base_model, tokenizer=tokenizer)
    return TEXT_PIPELINE


def ensure_comparison_pipeline():
    """
    Load the DeepSeek model pipeline if not already loaded.
    """
    global COMPARISON_PIPELINE
    if COMPARISON_PIPELINE is None:
        # If you prefer 4-bit, you can define BitsAndBytesConfig here, 
        # but let's keep it simpler for demonstration (fp16 or bf16).
        config = AutoConfig.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Llama-8B")
        tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Llama-8B")
        model = AutoModelForCausalLM.from_pretrained(
            "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
            config=config,
            device_map="auto"
        )
        COMPARISON_PIPELINE = pipeline(
            "text-generation",
            model=model,
            tokenizer=tokenizer
        )
    return COMPARISON_PIPELINE


@spaces.GPU(duration=120)  # up to 2 min for text generation
def predict(prompt, temperature, top_p, min_new_tokens, max_new_tokens):
    """
    Generates text from the fine-tuned (LoRA) model if present, else the base model.
    """
    pipe = ensure_pipeline()
    out = pipe(
        prompt,
        temperature=float(temperature),
        top_p=float(top_p),
        min_new_tokens=int(min_new_tokens),
        max_new_tokens=int(max_new_tokens),
        do_sample=True
    )
    return out[0]["generated_text"]


@spaces.GPU(duration=120)  # up to 2 min for text generation
def compare_models(prompt, temperature, top_p, min_new_tokens, max_new_tokens):
    """
    Generates text side-by-side from the local myr1 pipeline (fine-tuned or base)
    AND from the DeepSeek model. Returns two strings.
    """
    local_pipe = ensure_pipeline()
    comp_pipe = ensure_comparison_pipeline()

    local_out = local_pipe(
        prompt,
        temperature=float(temperature),
        top_p=float(top_p),
        min_new_tokens=int(min_new_tokens),
        max_new_tokens=int(max_new_tokens),
        do_sample=True
    )
    local_text = local_out[0]["generated_text"]

    comp_out = comp_pipe(
        prompt,
        temperature=float(temperature),
        top_p=float(top_p),
        min_new_tokens=int(min_new_tokens),
        max_new_tokens=int(max_new_tokens),
        do_sample=True
    )
    comp_text = comp_out[0]["generated_text"]

    return local_text, comp_text


# Build Gradio UI
with gr.Blocks() as demo:
    gr.Markdown("# QLoRA Fine-tuning & Comparison Demo")
    gr.Markdown("**Fine-tune wuhp/myr1** on a small subset of the Magpie dataset, then generate or compare output with the DeepSeek model.")

    finetune_btn = gr.Button("Finetune 4-bit (QLoRA) on Magpie subset (up to 5 min)")
    status_box = gr.Textbox(label="Finetune Status")
    finetune_btn.click(fn=finetune_small_subset, outputs=status_box)

    gr.Markdown("### Generate with myr1 (fine-tuned if done, else base)")

    prompt_in = gr.Textbox(lines=3, label="Prompt")
    temperature = gr.Slider(0.0, 1.5, step=0.1, value=0.7, label="Temperature")
    top_p = gr.Slider(0.0, 1.0, step=0.05, value=0.9, label="Top-p")
    min_tokens = gr.Slider(50, 1024, value=50, step=10, label="Min New Tokens")
    max_tokens = gr.Slider(50, 1024, value=200, step=50, label="Max New Tokens")

    output_box = gr.Textbox(label="myr1 Output", lines=8)
    gen_btn = gr.Button("Generate with myr1")

    gen_btn.click(
        fn=predict,
        inputs=[prompt_in, temperature, top_p, min_tokens, max_tokens],
        outputs=output_box
    )

    gr.Markdown("### Compare myr1 vs DeepSeek side-by-side")

    compare_btn = gr.Button("Compare")
    out_local = gr.Textbox(label="myr1 Output", lines=8)
    out_deepseek = gr.Textbox(label="DeepSeek Output", lines=8)

    compare_btn.click(
        fn=compare_models,
        inputs=[prompt_in, temperature, top_p, min_tokens, max_tokens],
        outputs=[out_local, out_deepseek]
    )

demo.launch()