app.py

# ==============================================================================
# HuggingFace Space - Sam Model Chat Interface with Streaming
# ==============================================================================
# Loads model directly from HuggingFace Hub: Smilyai-labs/Sam-1-large
# ==============================================================================

import gradio as gr
import tensorflow as tf
import keras
import numpy as np
from tokenizers import Tokenizer
from huggingface_hub import hf_hub_download
import os

# ==============================================================================
# Model Configuration
# ==============================================================================

MODEL_REPO = "Smilyai-labs/Sam-1-large"  # Your HuggingFace model repo
MAX_NEW_TOKENS = 512
TEMPERATURE = 0.8
TOP_P = 0.9
TOP_K = 50

# ==============================================================================
# Custom Keras Layers (Must match training code)
# ==============================================================================

@keras.saving.register_keras_serializable()
class RotaryEmbedding(keras.layers.Layer):
    def __init__(self, dim, max_len=2048, theta=10000, **kwargs):
        super().__init__(**kwargs)
        self.dim = dim
        self.max_len = max_len
        self.theta = theta
        self.built_cache = False
    
    def build(self, input_shape):
        if not self.built_cache:
            inv_freq = 1.0 / (self.theta ** (tf.range(0, self.dim, 2, dtype=tf.float32) / self.dim))
            t = tf.range(self.max_len, dtype=tf.float32)
            freqs = tf.einsum("i,j->ij", t, inv_freq)
            emb = tf.concat([freqs, freqs], axis=-1)
            
            self.cos_cached = tf.constant(tf.cos(emb), dtype=tf.float32)
            self.sin_cached = tf.constant(tf.sin(emb), dtype=tf.float32)
            self.built_cache = True
        
        super().build(input_shape)
    
    def rotate_half(self, x):
        x1, x2 = tf.split(x, 2, axis=-1)
        return tf.concat([-x2, x1], axis=-1)
    
    def call(self, q, k):
        seq_len = tf.shape(q)[2]
        dtype = q.dtype
        cos = tf.cast(self.cos_cached[:seq_len, :], dtype)[None, None, :, :]
        sin = tf.cast(self.sin_cached[:seq_len, :], dtype)[None, None, :, :]
        
        q_rotated = (q * cos) + (self.rotate_half(q) * sin)
        k_rotated = (k * cos) + (self.rotate_half(k) * sin)
        
        return q_rotated, k_rotated
    
    def get_config(self):
        config = super().get_config()
        config.update({"dim": self.dim, "max_len": self.max_len, "theta": self.theta})
        return config


@keras.saving.register_keras_serializable()
class RMSNorm(keras.layers.Layer):
    def __init__(self, epsilon=1e-5, **kwargs):
        super().__init__(**kwargs)
        self.epsilon = epsilon
    
    def build(self, input_shape):
        self.scale = self.add_weight(name="scale", shape=(input_shape[-1],), initializer="ones")
    
    def call(self, x):
        variance = tf.reduce_mean(tf.square(x), axis=-1, keepdims=True)
        return x * tf.math.rsqrt(variance + self.epsilon) * self.scale
    
    def get_config(self):
        config = super().get_config()
        config.update({"epsilon": self.epsilon})
        return config


@keras.saving.register_keras_serializable()
class TransformerBlock(keras.layers.Layer):
    def __init__(self, d_model, n_heads, ff_dim, dropout, max_len, rope_theta, layer_idx=0, **kwargs):
        super().__init__(**kwargs)
        self.d_model = d_model
        self.n_heads = n_heads
        self.ff_dim = ff_dim
        self.dropout_rate = dropout
        self.max_len = max_len
        self.rope_theta = rope_theta
        self.head_dim = d_model // n_heads
        self.layer_idx = layer_idx
        
        self.pre_attn_norm = RMSNorm()
        self.pre_ffn_norm = RMSNorm()
        
        self.q_proj = keras.layers.Dense(d_model, use_bias=False, name="q_proj")
        self.k_proj = keras.layers.Dense(d_model, use_bias=False, name="k_proj")
        self.v_proj = keras.layers.Dense(d_model, use_bias=False, name="v_proj")
        self.out_proj = keras.layers.Dense(d_model, use_bias=False, name="o_proj")
        
        self.rope = RotaryEmbedding(self.head_dim, max_len=max_len, theta=rope_theta)
        
        self.gate_proj = keras.layers.Dense(ff_dim, use_bias=False, name="gate_proj")
        self.up_proj = keras.layers.Dense(ff_dim, use_bias=False, name="up_proj")
        self.down_proj = keras.layers.Dense(d_model, use_bias=False, name="down_proj")
        
        self.dropout = keras.layers.Dropout(dropout)
    
    def call(self, x, training=None):
        B, T, D = tf.shape(x)[0], tf.shape(x)[1], self.d_model
        dtype = x.dtype
        
        res = x
        y = self.pre_attn_norm(x)
        
        q = tf.transpose(tf.reshape(self.q_proj(y), [B, T, self.n_heads, self.head_dim]), [0, 2, 1, 3])
        k = tf.transpose(tf.reshape(self.k_proj(y), [B, T, self.n_heads, self.head_dim]), [0, 2, 1, 3])
        v = tf.transpose(tf.reshape(self.v_proj(y), [B, T, self.n_heads, self.head_dim]), [0, 2, 1, 3])
        
        q, k = self.rope(q, k)
        
        scores = tf.matmul(q, k, transpose_b=True) / tf.sqrt(tf.cast(self.head_dim, dtype))
        
        mask = tf.where(
            tf.linalg.band_part(tf.ones([T, T], dtype=dtype), -1, 0) == 0,
            tf.constant(-1e9, dtype=dtype),
            tf.constant(0.0, dtype=dtype)
        )
        scores += mask
        attn = tf.matmul(tf.nn.softmax(scores, axis=-1), v)
        
        attn = tf.reshape(tf.transpose(attn, [0, 2, 1, 3]), [B, T, D])
        x = res + self.dropout(self.out_proj(attn), training=training)
        
        res = x
        y = self.pre_ffn_norm(x)
        ffn = self.down_proj(keras.activations.silu(self.gate_proj(y)) * self.up_proj(y))
        
        return res + self.dropout(ffn, training=training)
    
    def get_config(self):
        config = super().get_config()
        config.update({
            "d_model": self.d_model,
            "n_heads": self.n_heads,
            "ff_dim": self.ff_dim,
            "dropout": self.dropout_rate,
            "max_len": self.max_len,
            "rope_theta": self.rope_theta,
            "layer_idx": self.layer_idx
        })
        return config


@keras.saving.register_keras_serializable()
class SAM1Model(keras.Model):
    def __init__(self, **kwargs):
        super().__init__()
        if 'config' in kwargs and isinstance(kwargs['config'], dict):
            self.cfg = kwargs['config']
        elif 'vocab_size' in kwargs:
            self.cfg = kwargs
        else:
            self.cfg = kwargs.get('cfg', kwargs)
        
        self.embed = keras.layers.Embedding(self.cfg['vocab_size'], self.cfg['d_model'], name="embed_tokens")
        
        ff_dim = int(self.cfg['d_model'] * self.cfg['ff_mult'])
        block_args = {
            'd_model': self.cfg['d_model'],
            'n_heads': self.cfg['n_heads'],
            'ff_dim': ff_dim,
            'dropout': self.cfg['dropout'],
            'max_len': self.cfg['max_len'],
            'rope_theta': self.cfg['rope_theta']
        }
        
        self.blocks = []
        for i in range(self.cfg['n_layers']):
            block = TransformerBlock(name=f"block_{i}", layer_idx=i, **block_args)
            self.blocks.append(block)
        
        self.norm = RMSNorm(name="final_norm")
        self.lm_head = keras.layers.Dense(self.cfg['vocab_size'], use_bias=False, name="lm_head")
    
    def call(self, input_ids, training=None):
        x = self.embed(input_ids)
        
        for block in self.blocks:
            x = block(x, training=training)
        
        return self.lm_head(self.norm(x))
    
    def get_config(self):
        base_config = super().get_config()
        base_config['config'] = self.cfg
        return base_config

# ==============================================================================
# Load Model and Tokenizer from HuggingFace Hub
# ==============================================================================

print("🔥 Loading Sam model from HuggingFace Hub...")
print(f"   Repository: {MODEL_REPO}")

try:
    # Download model file
    print("📥 Downloading model.keras...")
    model_path = hf_hub_download(
        repo_id=MODEL_REPO,
        filename="model.keras",
        cache_dir="./model_cache"
    )
    print(f"✅ Model downloaded to: {model_path}")
    
    # Download tokenizer
    print("📥 Downloading tokenizer.json...")
    tokenizer_path = hf_hub_download(
        repo_id=MODEL_REPO,
        filename="tokenizer.json",
        cache_dir="./model_cache"
    )
    print(f"✅ Tokenizer downloaded to: {tokenizer_path}")
    
    # Load tokenizer
    tokenizer = Tokenizer.from_file(tokenizer_path)
    eos_token = "<|endoftext|>"
    eos_token_id = tokenizer.token_to_id(eos_token)
    print(f"✅ Tokenizer loaded (vocab_size={tokenizer.get_vocab_size()})")
    
    # Load model
    print("🔄 Loading model into memory...")
    model = keras.models.load_model(model_path)
    print(f"✅ Model loaded successfully!")
    
except Exception as e:
    print(f"❌ Error loading model: {e}")
    print("\n💡 Troubleshooting:")
    print("1. Make sure the model repo exists: https://huggingface.co/Smilyai-labs/Sam-1-large")
    print("2. Check that model.keras and tokenizer.json are in the repo")
    print("3. If repo is private, you may need to login: huggingface-cli login")
    raise

# ==============================================================================
# Generation Functions
# ==============================================================================

def sample_token(logits, temperature=1.0, top_p=0.9, top_k=50):
    """Sample next token with temperature, top-p, and top-k"""
    logits = logits / temperature
    
    # Top-k filtering
    if top_k > 0:
        top_k_logits, top_k_indices = tf.nn.top_k(logits, k=min(top_k, logits.shape[-1]))
        logits = tf.where(
            tf.reduce_any(tf.equal(tf.expand_dims(tf.range(logits.shape[-1]), 0), 
                                   tf.expand_dims(top_k_indices, -1)), axis=1),
            logits,
            tf.fill(logits.shape, -1e10)
        )
    
    # Top-p (nucleus) filtering
    if top_p < 1.0:
        sorted_logits = tf.sort(logits, direction='DESCENDING')
        sorted_probs = tf.nn.softmax(sorted_logits)
        cumsum_probs = tf.cumsum(sorted_probs)
        
        sorted_indices_to_remove = cumsum_probs > top_p
        sorted_indices_to_remove = tf.concat([
            [False],
            sorted_indices_to_remove[:-1]
        ], axis=0)
        
        sorted_indices = tf.argsort(logits, direction='DESCENDING')
        indices_to_remove = tf.gather(sorted_indices_to_remove, tf.argsort(sorted_indices))
        
        logits = tf.where(indices_to_remove, -1e10, logits)
    
    # Sample
    probs = tf.nn.softmax(logits)
    next_token = tf.random.categorical(tf.math.log(probs[None, :]), num_samples=1)[0, 0]
    
    return next_token.numpy()


def generate_stream(prompt, max_new_tokens=512, temperature=0.8, top_p=0.9, top_k=50):
    """Generate text with streaming (yields tokens as they're generated)"""
    
    # Format prompt
    formatted_prompt = f"User: {prompt}\nSam:"
    
    # Tokenize
    encoding = tokenizer.encode(formatted_prompt)
    input_ids = np.array([encoding.ids], dtype=np.int32)
    
    # Check if prompt is too long
    if input_ids.shape[1] > 1000:
        yield "❌ Error: Prompt is too long (max 1000 tokens)"
        return
    
    generated_text = ""
    
    for _ in range(max_new_tokens):
        # Get logits
        logits = model(input_ids, training=False)
        next_token_logits = logits[0, -1, :].numpy()
        
        # Sample next token
        next_token = sample_token(next_token_logits, temperature, top_p, top_k)
        
        # Stop if EOS
        if next_token == eos_token_id:
            break
        
        # Decode token
        token_text = tokenizer.decode([next_token])
        generated_text += token_text
        
        # Yield for streaming
        yield generated_text
        
        # Append to input
        input_ids = np.concatenate([input_ids, [[next_token]]], axis=1)
        
        # Stop if we hit max length
        if input_ids.shape[1] >= 1024:
            break


def chat_interface(message, history, temperature, top_p, top_k, max_tokens):
    """Gradio chat interface with streaming"""
    
    if not message.strip():
        yield history
        return
    
    # Build conversation context from history (last 3 turns to save tokens)
    conversation = ""
    recent_history = history[-3:] if len(history) > 3 else history
    
    for user_msg, bot_msg in recent_history:
        conversation += f"User: {user_msg}\nSam: {bot_msg}\n"
    
    # Add current message
    full_prompt = conversation + message if conversation else message
    
    # Add user message to history immediately
    history.append([message, ""])
    
    # Generate with streaming
    for response_chunk in generate_stream(
        full_prompt, 
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        top_k=top_k
    ):
        # Update the bot's response in history
        history[-1][1] = response_chunk
        yield history


# ==============================================================================
# Gradio Interface
# ==============================================================================

with gr.Blocks(theme=gr.themes.Soft(), title="Chat with Sam") as demo:
    gr.Markdown("""
    # 🤖 Chat with Sam
    
    **Sam** is a fine-tuned language model trained on math, code, reasoning, and conversational data.
    
    ### ✨ Capabilities:
    - 🧮 **Math**: Solve arithmetic and word problems (trained on GSM8K)
    - 💻 **Code**: Write Python, JavaScript, and more (trained on CodeAlpaca)
    - 🤔 **Reasoning**: Show step-by-step thinking with `<think>` tags
    - 💬 **Chat**: Natural conversations on any topic
    
    ### 📊 Model Info:
    - **Architecture**: 768d, 16 layers, 12 heads (~100M parameters)
    - **Context**: 1024 tokens
    - **Training**: TPU v5e-8 on multi-dataset mix
    """)
    
    chatbot = gr.Chatbot(
        label="💬 Conversation",
        height=450,
        show_copy_button=True,
        avatar_images=(None, "🤖"),
    )
    
    with gr.Row():
        msg = gr.Textbox(
            label="Your message",
            placeholder="Ask Sam anything... (e.g., 'What is 127 * 43?' or 'Write a function to sort a list')",
            lines=2,
            scale=4,
            autofocus=True
        )
        submit = gr.Button("Send 🚀", scale=1, variant="primary")
    
    with gr.Accordion("⚙️ Generation Settings", open=False):
        with gr.Row():
            temperature = gr.Slider(
                minimum=0.1,
                maximum=2.0,
                value=TEMPERATURE,
                step=0.1,
                label="Temperature",
                info="Higher = more creative/random"
            )
            top_p = gr.Slider(
                minimum=0.1,
                maximum=1.0,
                value=TOP_P,
                step=0.05,
                label="Top-p",
                info="Nucleus sampling threshold"
            )
        with gr.Row():
            top_k = gr.Slider(
                minimum=1,
                maximum=100,
                value=TOP_K,
                step=1,
                label="Top-k",
                info="Vocabulary size limit"
            )
            max_tokens = gr.Slider(
                minimum=50,
                maximum=512,
                value=MAX_NEW_TOKENS,
                step=50,
                label="Max tokens",
                info="Maximum response length"
            )
    
    with gr.Row():
        clear = gr.Button("🗑️ Clear Chat")
        
    with gr.Accordion("💡 Example Prompts", open=False):
        gr.Examples(
            examples=[
                ["What is 127 * 43?"],
                ["Write a Python function to reverse a string"],
                ["Explain how photosynthesis works"],
                ["What's the capital of France?"],
                ["Write a haiku about coding"],
                ["How do I sort a list in Python?"],
            ],
            inputs=msg,
            label="Click to try:"
        )
    
    gr.Markdown("""
    ---
    ### 📝 Tips:
    - Sam uses conversational format: `User: ... Sam: ...`
    - Watch for `<think>` tags showing reasoning process
    - Adjust temperature for more creative (higher) or focused (lower) responses
    - Model remembers last 3 conversation turns for context
    
    ### 🔗 Links:
    - Model: [Smilyai-labs/Sam-1-large](https://huggingface.co/Smilyai-labs/Sam-1-large)
    - Training: TPU v5e-8 on Kaggle
    - Framework: TensorFlow/Keras
    """)
    
    # Event handlers
    def respond(message, chat_history, temperature, top_p, top_k, max_tokens):
        """Handle message and generate response"""
        # Add user message to history
        chat_history.append([message, None])
        
        # Build conversation context from history (last 3 turns to save tokens)
        conversation = ""
        recent_history = chat_history[:-1][-3:] if len(chat_history) > 1 else []
        
        for user_msg, bot_msg in recent_history:
            if bot_msg:  # Only include completed turns
                conversation += f"User: {user_msg}\nSam: {bot_msg}\n"
        
        # Add current message
        full_prompt = conversation + message if conversation else message
        
        # Generate with streaming
        chat_history[-1][1] = ""
        for response_chunk in generate_stream(
            full_prompt, 
            max_new_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            top_k=top_k
        ):
            chat_history[-1][1] = response_chunk
            yield chat_history
    
    msg.submit(
        respond,
        [msg, chatbot, temperature, top_p, top_k, max_tokens],
        chatbot
    ).then(
        lambda: gr.Textbox(value=""),
        None,
        msg
    )
    
    submit.click(
        respond,
        [msg, chatbot, temperature, top_p, top_k, max_tokens],
        chatbot
    ).then(
        lambda: gr.Textbox(value=""),
        None,
        msg
    )
    
    clear.click(lambda: None, None, chatbot, queue=False)

# Launch
if __name__ == "__main__":
    print("\n" + "="*70)
    print("🚀 STARTING SAM CHAT INTERFACE".center(70))
    print("="*70)
    print(f"\n✅ Model loaded from: {MODEL_REPO}")
    print(f"✅ Vocab size: {tokenizer.get_vocab_size()}")
    print(f"✅ Ready to chat!\n")
    
    demo.queue()  # Enable streaming
    demo.launch()