app.py

import torch
import torch.nn as nn
from torch.nn import functional as F
import tiktoken
import gradio as gr
import math
import os

device = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'

# Tokenizer setup
enc = tiktoken.get_encoding("gpt2")
vocab_size = enc.n_vocab + 1 # +1 for mask token
mask_token_id = enc.n_vocab

def encode(s):
    return enc.encode(s)

def decode(l):
    return enc.decode([t for t in l if t != mask_token_id])

def format_masked_text(l):
    chunks = []
    current_chunk = []
    for t in l:
        if t == mask_token_id:
            if current_chunk:
                chunks.append(enc.decode(current_chunk))
                current_chunk = []
            chunks.append(" [MASK] ")
        else:
            current_chunk.append(t)
            
    if current_chunk:
        chunks.append(enc.decode(current_chunk))
        
    return "".join(chunks)

def norm(x):
    return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + 1e-5)

def apply_rotary_emb(x, cos, sin):
    assert x.ndim == 4
    d = x.shape[3] // 2
    x1, x2 = x[..., :d], x[..., d:]
    y1 = x1 * cos + x2 * sin
    y2 = x1 * (-sin) + x2 * cos
    out = torch.cat([y1, y2], 3)
    return out.to(x.dtype)

class MultiHeadAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.c_q = nn.Linear(config.n_embd, config.n_embd, bias=False)
        self.c_k = nn.Linear(config.n_embd, config.n_embd, bias=False)
        self.c_v = nn.Linear(config.n_embd, config.n_embd, bias=False)
        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=False)

    def forward(self, x, cos_sin):
        B, T, C = x.size()
        q = self.c_q(x).view(B, T, self.config.n_head, self.config.head_dim)
        k = self.c_k(x).view(B, T, self.config.n_head, self.config.head_dim)
        v = self.c_v(x).view(B, T, self.config.n_head, self.config.head_dim)
        cos, sin = cos_sin
        q, k = apply_rotary_emb(q, cos, sin), apply_rotary_emb(k, cos, sin)
        q, k = norm(q), norm(k)
        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
        y = F.scaled_dot_product_attention(q, k, v, is_causal=False)
        y = y.transpose(1, 2).contiguous().view(B, T, -1)
        y = self.c_proj(y)
        return y

class MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        hidden_dim = int(8 * config.n_embd / 3) 
        self.w1 = nn.Linear(config.n_embd, hidden_dim, bias=False)
        self.w2 = nn.Linear(config.n_embd, hidden_dim, bias=False)
        self.c_proj = nn.Linear(hidden_dim, config.n_embd, bias=False)

    def forward(self, x):
        return self.c_proj(F.silu(self.w1(x)) * self.w2(x))

class Block(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.attn = MultiHeadAttention(config)
        self.mlp = MLP(config)

    def forward(self, x, cos_sin):
        x = x + self.attn(norm(x), cos_sin)
        x = x + self.mlp(norm(x))
        return x

class Model(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.token_emb = nn.Embedding(vocab_size, config.n_embd)
        self.time_emb = nn.Sequential(
            nn.Linear(1, config.n_embd),
            nn.SiLU(),
            nn.Linear(config.n_embd, config.n_embd),
        )
        self.rotary_seq_len = config.block_size * 2
        cos, sin = self._precompute_rotary_embeddings(self.rotary_seq_len)
        self.register_buffer("cos", cos, persistent=False)
        self.register_buffer("sin", sin, persistent=False)
        self.blocks = nn.ModuleList([Block(config) for _ in range(config.n_layer)])
        self.lm_head = nn.Linear(config.n_embd, vocab_size, bias=False)
        self.lm_head.weight = self.token_emb.weight  # tie weights
        self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                torch.nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def _precompute_rotary_embeddings(self, seq_len, base=10000, device=None):
        if device is None:
            device = self.token_emb.weight.device
        channel_range = torch.arange(0, self.config.head_dim, 2, dtype=torch.float32, device=device)
        inv_freq = 1.0 / (base ** (channel_range / self.config.head_dim))
        t = torch.arange(seq_len, dtype=torch.float32, device=device)
        freqs = torch.outer(t, inv_freq)
        cos, sin = freqs.cos(), freqs.sin()
        cos, sin = cos[None, :, None, :], sin[None, :, None, :]
        return cos, sin

    def forward(self, idx, targets=None, mask=None, mask_rate=None):
        B, T = idx.size()
        x = self.token_emb(idx)
        if mask_rate is not None:
            t = mask_rate.float().unsqueeze(-1)  # (B, 1, 1)
            x = x + self.time_emb(t)
        x = norm(x)
        cos_sin = (self.cos[:, :T], self.sin[:, :T])
        for block in self.blocks:
            x = block(x, cos_sin)
        x = norm(x)
        logits = self.lm_head(x)
        
        if targets is None:
            loss = None
        else:
            B, T, C = logits.shape
            logits_flat = logits.view(B * T, C)
            targets_flat = targets.view(B * T)
            if mask is not None:
                mask_flat = mask.view(B * T)
                loss = F.cross_entropy(logits_flat, targets_flat, reduction="none")
                loss = (loss * mask_flat).sum() / mask_flat.sum()
            else:
                loss = F.cross_entropy(logits_flat, targets_flat)
        return logits, loss

class Config:
    def __init__(self, model_type):
        self.block_size = 512
        if model_type == 'medium':
            self.n_embd = 512
            self.n_head = 8
            self.n_layer = 8
            self.weights_path = "tinystories_diffusion_med_dual.pt"
        elif model_type == 'gpt2':
            self.n_embd = 768
            self.n_head = 12
            self.n_layer = 12
            self.weights_path = "tinystories_diffusion_GPT2_dual.pt"
        else:
            raise ValueError("model_type must be 'medium' or 'gpt2'")
        self.head_dim = self.n_embd // self.n_head

# Dynamic loading
loaded_model_type = None
loaded_model = None

def get_model(model_type):
    global loaded_model_type, loaded_model
    if loaded_model_type == model_type and loaded_model is not None:
        return loaded_model, Config(model_type)
    
    print(f"Loading {model_type} model...")
    config = Config(model_type)
    model = Model(config)
    weights_path = config.weights_path
    
    if os.path.exists(weights_path):
        state_dict = torch.load(weights_path, map_location=device, weights_only=True)
        unwrapped_state_dict = {}
        for k, v in state_dict.items():
            # Handle 'module.' prefix from DataParallel if present
            if k.startswith("module."):
                unwrapped_state_dict[k[7:]] = v
            else:
                unwrapped_state_dict[k] = v
        model.load_state_dict(unwrapped_state_dict)
        print("Model loaded successfully!")
    else:
        print(f"Warning: {weights_path} not found. Running with uninitialized random parameters.")
    
    model.to(device)
    loaded_model = model
    loaded_model_type = model_type
    return model, config

@torch.no_grad()
def generate_diffusion(prompt, max_new_tokens=100, mode="Direct Output", model_type="medium"):
    model, config = get_model(model_type)
    prompt_tokens = encode(prompt)
    model.eval()
    prompt_len = len(prompt_tokens)
    all_tokens = prompt_tokens.copy()
    temp = 1.0
    confidence_threshold = 0.95
    top_k = 3

    while len(all_tokens) - len(prompt_tokens) < max_new_tokens:
        curr_prompt_len = len(all_tokens)
        block_len = min(config.block_size - curr_prompt_len, len(prompt_tokens) + max_new_tokens - len(all_tokens))
        if block_len <= 0: break
        
        x = torch.full((1, config.block_size), mask_token_id, dtype=torch.long, device=device)
        x[0, :curr_prompt_len] = torch.tensor(all_tokens[-curr_prompt_len:], device=device)
        
        masked = torch.zeros(1, config.block_size, dtype=torch.bool, device=device)
        masked[0, curr_prompt_len : curr_prompt_len + block_len] = True

        while masked.any():
            logits, _ = model(x)
            probs = F.softmax(logits / temp, dim=-1)
            top_k_probs, top_k_indices = torch.topk(probs, k=top_k, dim=-1)
            confidences = top_k_probs.sum(dim=-1)

            decode_mask = (confidences >= confidence_threshold) & masked
            if not decode_mask.any():
                masked_confidences = torch.where(masked, confidences, torch.tensor(-float('inf')).to(device))
                decode_mask.view(-1)[masked_confidences.argmax()] = True

            top_k_probs_norm = top_k_probs / top_k_probs.sum(dim=-1, keepdim=True)
            sampled_k = torch.multinomial(top_k_probs_norm.view(-1, top_k), 1).view(1, config.block_size)
            sampled_tokens = torch.gather(top_k_indices, -1, sampled_k.unsqueeze(-1)).squeeze(-1)

            x = torch.where(decode_mask, sampled_tokens, x)
            masked = masked & ~decode_mask
            
            if mode == "Show Generation Process":
                current_block = x[0, curr_prompt_len : curr_prompt_len + block_len].tolist()
                yield format_masked_text(all_tokens + current_block)
            
        all_tokens.extend(x[0, curr_prompt_len : curr_prompt_len + block_len].tolist())
        
    full_output = decode(all_tokens)
    yield full_output

def gradio_fn(prompt, display_mode, max_tokens, model_type):
    for text in generate_diffusion(prompt, max_new_tokens=max_tokens, mode=display_mode, model_type=model_type):
        yield text

# Gradio
with gr.Blocks(theme=gr.themes.Monochrome()) as demo:
    gr.Markdown("# TinyStories Diffusion LM")
    gr.Markdown("A non-autoregressive language model leveraging parallel block-decoding and SwiGLU networks.")
    
    with gr.Row():
        with gr.Column():
            prompt_in = gr.Textbox(lines=2, placeholder="Once upon a time, there was a little girl who", label="Prompt (approx 10 words)")
            
            model_type_in = gr.Radio(["medium", "gpt2"], value="medium", label="Model Architecture")
            mode = gr.Radio(["Direct Output", "Show Generation Process"], value="Direct Output", label="Display Mode")
            max_tokens = gr.Slider(minimum=20, maximum=1000, value=100, step=1, label="Max Tokens")
            
            generate_btn = gr.Button("Generate Story", variant='primary')
        
        with gr.Column():
            output = gr.Textbox(lines=10, label="Output")
            
    generate_btn.click(fn=gradio_fn, inputs=[prompt_in, mode, max_tokens, model_type_in], outputs=output)

if __name__ == "__main__":
    demo.queue().launch()