Delete chat_interface.py

chat_interface.py

@@ -1,440 +0,0 @@
-import os
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import json
-import argparse
-import sys
-import sentencepiece as spm
-import math
-from dataclasses import dataclass
-
-# --- Define the CORRECT Model Architecture (copied from train_llm.py) ---
-@dataclass
-class ModelConfig:
-    vocab_size: int = 32000
-    hidden_size: int = 512
-    num_layers: int = 8
-    num_attention_heads: int = 8
-    num_key_value_heads: int = 2
-    intermediate_size: int = 1365
-    max_position_embeddings: int = 2048
-    rms_norm_eps: float = 1e-6
-    rope_theta: float = 10000.0
-
-class RMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps=1e-6):
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-
-class RotaryEmbedding(nn.Module):
-    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
-        super().__init__()
-        self.dim = dim
-        self.max_position_embeddings = max_position_embeddings
-        self.base = base
-        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim))
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-
-    def forward(self, x, seq_len=None):
-        if seq_len is None:
-            seq_len = x.shape[-2]
-        t = torch.arange(seq_len, device=x.device, dtype=self.inv_freq.dtype)
-        freqs = torch.outer(t, self.inv_freq)
-        emb = torch.cat((freqs, freqs), dim=-1)
-        cos = emb.cos()
-        sin = emb.sin()
-        return cos, sin
-
-def rotate_half(x):
-    x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None):
-    if position_ids is not None:
-        cos = cos[position_ids].unsqueeze(1)
-        sin = sin[position_ids].unsqueeze(1)
-    else:
-        cos = cos[:q.shape[-2]].unsqueeze(0).unsqueeze(0)
-        sin = sin[:q.shape[-2]].unsqueeze(0).unsqueeze(0)
-    
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
-    return q_embed, k_embed
-
-class SwiGLU(nn.Module):
-    def __init__(self, hidden_size, intermediate_size):
-        super().__init__()
-        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
-        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
-        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
-
-    def forward(self, x):
-        gate = self.gate_proj(x)
-        up = self.up_proj(x)
-        return self.down_proj(F.silu(gate) * up)
-
-class GroupedQueryAttention(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.hidden_size = config.hidden_size
-        self.num_heads = config.num_attention_heads
-        self.num_key_value_heads = config.num_key_value_heads
-        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
-        self.head_dim = self.hidden_size // self.num_heads
-        self.max_position_embeddings = config.max_position_embeddings
-        
-        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
-        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
-        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
-        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
-        
-        self.rotary_emb = RotaryEmbedding(
-            self.head_dim,
-            max_position_embeddings=self.max_position_embeddings,
-            base=config.rope_theta,
-        )
-
-    def forward(self, hidden_states, attention_mask=None, position_ids=None):
-        bsz, q_len, _ = hidden_states.size()
-        
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
-        
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-        
-        cos, sin = self.rotary_emb(value_states, seq_len=q_len)
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
-        
-        # Repeat k/v heads if n_kv_heads < n_heads
-        key_states = torch.repeat_interleave(key_states, repeats=self.num_key_value_groups, dim=1)
-        value_states = torch.repeat_interleave(value_states, repeats=self.num_key_value_groups, dim=1)
-        
-        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
-        
-        if attention_mask is not None:
-            # Convert from [batch_size, seq_len] to [batch_size, 1, 1, seq_len]
-            expanded_mask = attention_mask[:, None, None, :].to(attn_weights.dtype)
-            expanded_mask = (1.0 - expanded_mask) * torch.finfo(attn_weights.dtype).min
-            attn_weights = attn_weights + expanded_mask
-            
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
-        attn_output = torch.matmul(attn_weights, value_states)
-        
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
-        attn_output = self.o_proj(attn_output)
-        
-        return attn_output
-
-class TransformerBlock(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.self_attn = GroupedQueryAttention(config)
-        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.mlp = SwiGLU(config.hidden_size, config.intermediate_size)
-
-    def forward(self, hidden_states, attention_mask=None, position_ids=None):
-        residual = hidden_states
-        hidden_states = self.input_layernorm(hidden_states)
-        hidden_states = self.self_attn(hidden_states, attention_mask, position_ids)
-        hidden_states = residual + hidden_states
-        
-        residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-        
-        return hidden_states
-
-class LLMModel(nn.Module): # REPLACED CustomTransformer with this class
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
-        self.layers = nn.ModuleList([TransformerBlock(config) for _ in range(config.num_layers)])
-        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
-        
-    def forward(self, input_ids, attention_mask=None, position_ids=None):
-        batch_size, seq_length = input_ids.shape
-
-        if position_ids is None:
-            position_ids = torch.arange(0, seq_length, dtype=torch.long, device=input_ids.device)
-            position_ids = position_ids.unsqueeze(0).expand(batch_size, -1)
-        
-        # Create causal mask for generation
-        if attention_mask is None:
-            # Create a causal mask (lower triangular)
-            causal_mask = torch.full((seq_length, seq_length), float('-inf'), device=input_ids.device)
-            causal_mask = torch.triu(causal_mask, diagonal=1)
-            # [batch_size, 1, seq_len, seq_len]
-            attention_mask = causal_mask.unsqueeze(0).unsqueeze(0) 
-        else:
-            # If a padding mask is provided, convert it to the format we need
-            # Assuming attention_mask is [batch_size, seq_len] with 1 for valid, 0 for pad
-            attention_mask = (1.0 - attention_mask) * torch.finfo(torch.float32).min
-            attention_mask = attention_mask[:, None, None, :]
-        
-        hidden_states = self.embed_tokens(input_ids)
-        
-        for layer in self.layers:
-            hidden_states = layer(hidden_states, attention_mask, position_ids)
-            
-        hidden_states = self.norm(hidden_states)
-        logits = self.lm_head(hidden_states)
-        return logits
-
-    def generate(self, input_ids, max_new_tokens=100, do_sample=True, top_p=0.9, temperature=0.7):
-        """Simplified generation logic."""
-        self.eval()
-        generated = input_ids.clone()
-
-        for _ in range(max_new_tokens):
-            # Get logits for the last token
-            logits = self(generated)[:, -1, :] # shape: [batch_size, vocab_size]
-
-            if do_sample:
-                # Apply temperature
-                logits = logits / temperature
-                probs = torch.softmax(logits, dim=-1)
-
-                # Top-p (nucleus) sampling
-                sorted_probs, sorted_indices = torch.sort(probs, descending=True)
-                cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
-                sorted_indices_to_remove = cumulative_probs > top_p
-                sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
-                sorted_indices_to_remove[..., 0] = 0
-                indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
-                probs = probs.masked_fill(indices_to_remove, 0.0)
-                probs = probs / probs.sum(dim=-1, keepdim=True)
-
-                next_token = torch.multinomial(probs, num_samples=1)
-            else:
-                next_token = torch.argmax(logits, dim=-1, keepdim=True)
-
-            generated = torch.cat([generated, next_token], dim=-1)
-
-            # Stop if EOS is generated
-            if next_token.item() == 3:  # EOS token ID from your tokenizer
-                break
-
-        return generated
-# --- End of Model Architecture ---
-
-def load_tokenizer(tokenizer_path):
-    """Load the tokenizer and ensure it has an <UNK> token."""
-    print(f"Debug: Attempting to load tokenizer from {tokenizer_path}")
-    if not os.path.exists(tokenizer_path):
-        print(f"Error: Tokenizer file {tokenizer_path} does not exist")
-        return None
-    try:
-        sp = spm.SentencePieceProcessor()
-        sp.load(tokenizer_path)
-        if sp.unk_id() is None:
-            print("Warning: No <UNK> token in tokenizer. Using default ID 0.")
-            sp.add_unk_token(0)
-        print(f"Debug: Tokenizer loaded successfully. Vocab size: {sp.vocab_size()}")
-        return sp
-    except Exception as e:
-        print(f"Error loading tokenizer: {e}")
-        return None
-
-def load_model(model_path, config_path, device='cpu'):
-    """Load the model from checkpoint with detailed debugging."""
-    print(f"Debug: Attempting to load model from {model_path}")
-    print(f"Debug: Config path: {config_path}")
-    
-    if not os.path.exists(model_path):
-        print(f"Error: Model file {model_path} does not exist")
-        return None, None
-    if not os.path.exists(config_path):
-        print(f"Warning: Config file {config_path} not found. Using default config.")
-    
-    # Load config to get the correct parameters for the model
-    config_dict = {
-        'vocab_size': 32000,
-        'hidden_size': 512,
-        'num_layers': 8,
-        'num_attention_heads': 8,
-        'num_key_value_heads': 2,
-        'intermediate_size': 1365,
-        'max_position_embeddings': 2048,
-        'rms_norm_eps': 1e-6,
-        'rope_theta': 10000.0
-    }
-    try:
-        if os.path.exists(config_path):
-            with open(config_path, 'r') as f:
-                loaded_config = json.load(f)
-            # Update our config dict with the loaded values
-            for key in config_dict:
-                if key in loaded_config:
-                    config_dict[key] = loaded_config[key]
-            print(f"Debug: Config loaded: {config_dict}")
-    except Exception as e:
-        print(f"Warning: Failed to load config.json: {e}. Using default config.")
-    
-    # Create a ModelConfig object
-    config = ModelConfig(**config_dict)
-    
-    try:
-        print("Debug: Initializing LLMModel (correct architecture)")
-        model = LLMModel(config) # Now using the CORRECT model class
-    except Exception as e:
-        print(f"Error initializing model: {e}")
-        return None, None
-    
-    try:
-        checkpoint = torch.load(model_path, map_location=device)
-        print(f"Debug: Checkpoint type: {type(checkpoint)}")
-        if isinstance(checkpoint, dict):
-            if 'model_state_dict' in checkpoint:
-                print("Debug: Loading from full checkpoint dict")
-                model.load_state_dict(checkpoint['model_state_dict'], strict=False)
-            else:
-                print("Debug: Loading state dictionary directly")
-                model.load_state_dict(checkpoint, strict=False)
-        else:
-            print("Debug: Loading full model object (not recommended)")
-            model = checkpoint
-        model.to(device)
-        model.eval()
-        print(f"Debug: Model loaded successfully on {device}")
-        return model, config
-    except Exception as e:
-        print(f"Error loading model checkpoint: {e}")
-        return None, None
-
-def preprocess_input(text, tokenizer, max_length=512):
-    """Preprocess and tokenize input text, handling OOV tokens."""
-    print(f"Debug: Preprocessing input: {text}")
-    text = ' '.join(text.strip().split())
-    if not text:
-        return None, "Input is empty. Please provide a valid input."
-    
-    try:
-        # Use add_bos=True and add_eos=True as your training likely did
-        tokens = tokenizer.encode(text, out_type=int, add_bos=True, add_eos=True)
-        print(f"Debug: Tokenized input: {tokens}")
-        if len(tokens) > max_length:
-            # Truncate from the middle or end? Let's truncate from the end, keeping BOS
-            tokens = tokens[:max_length-1] + [tokenizer.eos_id()]
-        # Ensure the input is the right length
-        # For generation, we usually don't pad the input, we just use its actual length
-        # The model's attention mask will handle the rest.
-        return torch.tensor([tokens], dtype=torch.long), None
-    except Exception as e:
-        print(f"Tokenization error: {e}")
-        return None, f"Failed to tokenize input: {text}. Please try again."
-
-def generate_response(model, tokenizer, input_tokens, max_new_tokens=100, device='cpu'):
-    """Generate a response from the model."""
-    print(f"Debug: Generating response with input tokens shape: {input_tokens.shape}")
-    try:
-        input_tokens = input_tokens.to(device)
-        output_tokens = model.generate(input_tokens, max_new_tokens=max_new_tokens)
-        # Decode the entire sequence, then remove the input part
-        full_sequence = output_tokens[0].tolist()
-        # Find the EOS token that was originally added during preprocessing
-        input_length = input_tokens.shape[1]
-        # The response is the part after the input
-        response_tokens = full_sequence[input_length:]
-        response = tokenizer.decode(response_tokens)
-        print(f"Debug: Generated response: {response}")
-        return response, None
-    except Exception as e:
-        print(f"Inference error: {e}")
-        return None, "Failed to generate response. Please try again."
-
-def main():
-    print("🚀 Initializing CoreX AI Chat Interface...")
-    
-    default_checkpoint_path = r"D:\checkpoints"
-    default_tokenizer_path = r"D:\CoreX\tokenizer\corex_tok.model"
-    
-    parser = argparse.ArgumentParser(description="CoreX AI Chat Interface")
-    parser.add_argument('--model_path', default=default_checkpoint_path, help="Path to model checkpoints")
-    parser.add_argument('--tokenizer_path', default=default_tokenizer_path, help="Path to tokenizer")
-    args = parser.parse_args()
-    
-    print("📁 Default paths:")
-    print(f"   Model: {args.model_path}")
-    print(f"   Tokenizer: {args.tokenizer_path}")
-    print("✅ Using default paths...")
-    
-    print(f"Loading tokenizer from {args.tokenizer_path}...")
-    tokenizer = load_tokenizer(args.tokenizer_path)
-    if tokenizer is None:
-        print("Failed to load tokenizer. Exiting.")
-        return
-    
-    config_path = os.path.join(args.model_path, "config.json")
-    model_path = os.path.join(args.model_path, "final_model.pt")
-    device = 'cuda' if torch.cuda.is_available() else 'cpu'
-    
-    print(f"Loading custom model from {args.model_path}...")
-    model, config = load_model(model_path, config_path, device)
-    if model is None:
-        print("Failed to load model. Exiting.")
-        return
-    
-    print(f"Model loaded successfully on {device}")
-    print("🤖 AI Chat Interface")
-    print("=" * 50)
-    print("Type 'quit', 'exit', or 'bye' to end the conversation")
-    print("Type 'clear' to clear the conversation history")
-    print("Type 'help' for more commands")
-    print("=" * 50)
-    
-    conversation_history = []
-    
-    while True:
-        user_input = input("\n🧑 You: ").strip()
-        
-        if user_input.lower() in ['quit', 'exit', 'bye']:
-            print("👋 Goodbye!")
-            break
-        elif user_input.lower() == 'clear':
-            conversation_history = []
-            print("🗑 Conversation history cleared.")
-            continue
-        elif user_input.lower() == 'help':
-            print("Available commands:")
-            print("  quit/exit/bye: End the conversation")
-            print("  clear: Clear conversation history")
-            print("  help: Show this help message")
-            continue
-        
-        input_tokens, error = preprocess_input(user_input, tokenizer)
-        if error:
-            print(f"🤖 AI: {error}")
-            with open("rejected_inputs.log", "a") as log_file:
-                log_file.write(f"Rejected input: {user_input}\nError: {error}\n")
-            continue
-        
-        conversation_history.append({"role": "user", "content": user_input})
-        
-        response, error = generate_response(model, tokenizer, input_tokens, max_new_tokens=100, device=device)
-        if error:
-            print(f"🤖 AI: {error}")
-            continue
-        
-        conversation_history.append({"role": "assistant", "content": response})
-        print(f"\n🤖 AI: {response}")
-
-if __name__ == "__main__":
-    main()