app_flash1.py

import gc
import os
import torch
import torch.nn as nn
import torch.optim as optim
import tempfile
import gradio as gr
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModel
from flashpack import FlashPackMixin
from huggingface_hub import Repository
from typing import Tuple
from sklearn.model_selection import train_test_split

device = torch.device("cpu")
torch.set_num_threads(4)
print(f"🔧 Using device: {device} (CPU-only)")

# ============================================================
# 1️⃣ Model
# ============================================================
class GemmaTrainer(nn.Module, FlashPackMixin):
    def __init__(self, input_dim: int, hidden_dim: int = 1024, output_dim: int = 1536):
        super().__init__()
        self.fc1 = nn.Linear(input_dim, hidden_dim)
        self.relu = nn.ReLU()
        self.fc2 = nn.Linear(hidden_dim, hidden_dim)
        self.fc3 = nn.Linear(hidden_dim, output_dim)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        x = self.relu(x)
        x = self.fc3(x)
        return x

# ============================================================
# 2️⃣ Encoder with batch mean+max pooling
# ============================================================
def build_encoder(model_name="gpt2", max_length=128):
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    embed_model = AutoModel.from_pretrained(model_name).to(device)
    embed_model.eval()

    @torch.no_grad()
    def encode_batch(prompts: list, batch_size=16) -> torch.Tensor:
        embeddings = []
        for i in range(0, len(prompts), batch_size):
            batch = prompts[i:i+batch_size]
            inputs = tokenizer(batch, return_tensors="pt", truncation=True,
                               padding="max_length", max_length=max_length).to(device)
            last_hidden = embed_model(**inputs).last_hidden_state
            mean_pool = last_hidden.mean(dim=1)
            max_pool, _ = last_hidden.max(dim=1)
            batch_emb = torch.cat([mean_pool, max_pool], dim=1)
            embeddings.append(batch_emb.cpu())
        return torch.vstack(embeddings)

    return tokenizer, embed_model, encode_batch

# ============================================================
# 3️⃣ Push model to HF
# ============================================================
def push_flashpack_model_to_hf(model, hf_repo: str):
    logs = []
    with tempfile.TemporaryDirectory() as tmp_dir:
        logs.append(f"📂 Using temporary directory: {tmp_dir}")
        repo = Repository(local_dir=tmp_dir, clone_from=hf_repo, use_auth_token=True)
        pack_path = os.path.join(tmp_dir, "model.flashpack")
        model.save_flashpack(pack_path, target_dtype=torch.float32)
        readme_path = os.path.join(tmp_dir, "README.md")
        with open(readme_path, "w") as f:
            f.write("# FlashPack Model\nThis repo contains a FlashPack model trained for short→long prompt mapping.")
        repo.push_to_hub()
        logs.append(f"✅ Model pushed to Hugging Face repo: {hf_repo}")
    return logs

# ============================================================
# 4️⃣ Train with train/test split & detailed logging
# ============================================================
def train_flashpack_model(
    dataset_name="rahul7star/prompt-enhancer-dataset",
    max_encode=1000,
    hidden_dim=1024,
    hf_repo="rahul7star/FlashPack",
    push_to_hub=True,
    test_split=0.1,
    batch_size=32,
    max_epochs=50,
    target_test_loss=0.01
) -> Tuple[GemmaTrainer, object, object, object, torch.Tensor]:

    print("📦 Loading dataset...")
    dataset = load_dataset(dataset_name, split="train")
    limit = min(max_encode, len(dataset))
    dataset = dataset.select(range(limit))
    print(f"⚡ Using {len(dataset)} prompts for training")

    short_prompts = [item["short_prompt"] for item in dataset]
    long_prompts = [item["long_prompt"] for item in dataset]

    # Split
    train_short, test_short, train_long, test_long = train_test_split(
        short_prompts, long_prompts, test_size=test_split, random_state=42
    )
    print(f"🔹 Train size: {len(train_short)}, Test size: {len(test_short)}")

    tokenizer, embed_model, encode_batch = build_encoder("gpt2", max_length=128)

    # Encode
    print("⚡ Encoding training prompts...")
    train_short_emb = encode_batch(train_short)
    train_long_emb = encode_batch(train_long)
    print(f"✅ Train embeddings shape: {train_short_emb.shape}, {train_long_emb.shape}")

    print("⚡ Encoding test prompts...")
    test_short_emb = encode_batch(test_short)
    test_long_emb = encode_batch(test_long)
    print(f"✅ Test embeddings shape: {test_short_emb.shape}, {test_long_emb.shape}")

    input_dim = train_short_emb.shape[1]
    output_dim = train_long_emb.shape[1]

    model = GemmaTrainer(input_dim=input_dim, hidden_dim=hidden_dim, output_dim=output_dim).to(device)

    criterion = nn.CosineSimilarity(dim=1)
    optimizer = optim.Adam(model.parameters(), lr=1e-3)

    n_train = train_short_emb.shape[0]

    print("🚀 Training model...")
    for epoch in range(max_epochs):
        model.train()
        epoch_loss = 0.0
        perm = torch.randperm(n_train)
        for start in range(0, n_train, batch_size):
            idx = perm[start:start+batch_size]
            inputs = train_short_emb[idx].to(device)
            targets = train_long_emb[idx].to(device)

            optimizer.zero_grad()
            outputs = model(inputs)
            loss = 1 - criterion(outputs, targets).mean()
            loss.backward()
            optimizer.step()
            epoch_loss += loss.item() * inputs.size(0)

        epoch_loss /= n_train

        # Evaluate on test
        model.eval()
        with torch.no_grad():
            test_outputs = model(test_short_emb.to(device))
            test_loss = (1 - criterion(test_outputs, test_long_emb.to(device)).mean()).item()

        print(f"Epoch {epoch+1}/{max_epochs} → Train loss: {epoch_loss:.6f}, Test loss: {test_loss:.6f}")

        # Check if model is perfect enough
        if test_loss <= target_test_loss:
            print(f"✅ Target test loss reached ({test_loss:.6f}) – stopping training early.")
            break

    # Push to HF if trained well
    logs = []
    if push_to_hub and test_loss <= target_test_loss:
        logs = push_flashpack_model_to_hf(model, hf_repo)
        for log in logs:
            print(log)
    elif push_to_hub:
        print(f"⚠️ Test loss too high ({test_loss:.6f}); skipping HF upload.")

    return model, dataset, embed_model, tokenizer, train_long_emb

# ============================================================
# 5️⃣ Load or train
# ============================================================
def get_flashpack_model(hf_repo="rahul7star/FlashPack"):
    try:
        print(f"🔁 Attempting to load FlashPack model from {hf_repo}")
        model = GemmaTrainer.from_flashpack(hf_repo)
        model.eval()
        tokenizer, embed_model, encode_batch = build_encoder("gpt2", max_length=128)
        return model, tokenizer, embed_model
    except Exception as e:
        print(f"⚠️ Load failed: {e}")
        print("⏬ Training a new FlashPack model locally...")
        return train_flashpack_model(hf_repo=hf_repo)

# ============================================================
# 6️⃣ Load or train
# ============================================================
model, tokenizer, embed_model, dataset, long_embeddings = get_flashpack_model()

# ============================================================
# 7️⃣ Inference helpers
# ============================================================
@torch.no_grad()
def encode_for_inference(prompt: str) -> torch.Tensor:
    inputs = tokenizer(prompt, return_tensors="pt", truncation=True,
                       padding="max_length", max_length=128).to(device)
    last_hidden = embed_model(**inputs).last_hidden_state
    mean_pool = last_hidden.mean(dim=1)
    max_pool, _ = last_hidden.max(dim=1)
    return torch.cat([mean_pool, max_pool], dim=1).cpu()

def enhance_prompt(user_prompt: str, temperature: float, max_tokens: int, chat_history):
    chat_history = chat_history or []
    short_emb = encode_for_inference(user_prompt)
    mapped = model(short_emb.to(device)).cpu()

    sims = (long_embeddings @ mapped.t()).squeeze(1)
    long_norms = long_embeddings.norm(dim=1)
    mapped_norm = mapped.norm()
    sims = sims / (long_norms * (mapped_norm + 1e-12))

    best_idx = int(sims.argmax().item())
    enhanced_prompt = dataset[best_idx]["long_prompt"]

    chat_history.append({"role": "user", "content": user_prompt})
    chat_history.append({"role": "assistant", "content": enhanced_prompt})
    return chat_history

# ============================================================
# 8️⃣ Gradio UI
# ============================================================
with gr.Blocks(title="Prompt Enhancer – FlashPack (CPU)", theme=gr.themes.Soft()) as demo:
    gr.Markdown(
        """
        # ✨ Prompt Enhancer (FlashPack mapper)
        Enter a short prompt, and the model will **expand it with details and creative context**.
        (CPU-only mode.)
        """
    )

    with gr.Row():
        chatbot = gr.Chatbot(height=400, label="Enhanced Prompts", type="messages")
        with gr.Column(scale=1):
            user_prompt = gr.Textbox(placeholder="Enter a short prompt...", label="Your Prompt", lines=3)
            temperature = gr.Slider(0.0, 1.0, value=0.7, step=0.05, label="Temperature")
            max_tokens = gr.Slider(32, 256, value=128, step=16, label="Max Tokens")
            send_btn = gr.Button("🚀 Enhance Prompt", variant="primary")
            clear_btn = gr.Button("🧹 Clear Chat")

    send_btn.click(enhance_prompt, [user_prompt, temperature, max_tokens, chatbot], chatbot)
    user_prompt.submit(enhance_prompt, [user_prompt, temperature, max_tokens, chatbot], chatbot)
    clear_btn.click(lambda: [], None, chatbot)

if __name__ == "__main__":
    demo.launch(show_error=True)