Upload 6 files

README.md

@@ -0,0 +1,43 @@
+## ChatGCLM-330M
+<img src="./banner.png" alt="ChatGCLM Hero" width="600">
+<strong>A high-performance language model architecture.</strong>
+
+---
+
+## Overview
+
+**ChatGCLM** is a generative language model that deviates from the traditional Transformer architecture by utilizing a hybrid approach of **Local** and **Global Convolutions**. By leveraging Fast Fourier Transforms (FFT) for global context, ChatGCLM achieves a massive receptive field with a fraction of the computational overhead associated with standard attention mechanisms.
+
+The architecture is designed for efficiency, speed, and high-quality generation, featuring a custom vocabulary reduction system that optimizes the embedding space for specific datasets.
+
+
+## 📦 Installation
+
+Download this repository and extract it.
+
+---
+
+## Usage
+
+### 1. Training the Model
+Place your `.txt` data files in the `data/` directory and run:
+```bash
+python train_chatgclm.py
+```
+This script will build the vocabulary and train the foundation model
+
+### 2. Interactive Chat Interface
+Launch the Tkinter-based UI to interact with your model:
+```bash
+python chat_interface.py
+```
+
+---
+
+## Fine-tuning
+
+You may fine-tune the model by resuming training from a checkpoint, you may use a different dataset as long as the vocabulary is the same, you may also change parameters such as the learning rate, batch size, etc.
+
+<p align="center">
+  Built with ❤️ by AG 
+</p>

model.py

@@ -0,0 +1,94 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+D_MODEL = 1152
+N_LAYERS = 22
+MAX_SEQ_LEN = 4096
+LOCAL_KERNEL_SIZE = 5
+GLOBAL_KERNEL_SIZE = 256
+USE_GLOBAL_EVERY_N_LAYERS = 2
+FFT_SIZE = 1024
+
+class GlobalConv1D(nn.Module):
+    def __init__(self, d_model, kernel_size, fft_size):
+        super().__init__()
+        self.kernel = nn.Parameter(torch.randn(d_model, kernel_size) * 0.01)
+        self.kernel_size = kernel_size
+        self.fft_size = fft_size
+
+    def forward(self, x):
+        B, C, T = x.shape
+        K = min(self.kernel_size, T)
+        overlap = K - 1
+        block = self.fft_size - overlap
+        x = F.pad(x, (overlap, 0))
+        k = self.kernel[:, :K]
+        k = F.pad(k, (0, self.fft_size - K))
+        k_f = torch.fft.rfft(k, n=self.fft_size)
+        outs = []
+        pos = 0
+        while pos < T:
+            seg = x[..., pos:pos+self.fft_size]
+            if seg.shape[-1] < self.fft_size:
+                seg = F.pad(seg, (0, self.fft_size - seg.shape[-1]))
+            y = torch.fft.irfft(torch.fft.rfft(seg, n=self.fft_size) * k_f.unsqueeze(0), n=self.fft_size)
+            outs.append(y[..., overlap:overlap+block])
+            pos += block
+        return torch.cat(outs, dim=-1)[..., :T]
+
+class LocalConv1D(nn.Module):
+    def __init__(self, d_model, k):
+        super().__init__()
+        self.k = k
+        self.dw = nn.Conv1d(d_model, d_model, k, groups=d_model)
+        self.pw = nn.Conv1d(d_model, d_model, 1)
+
+    def forward(self, x):
+        x = F.pad(x, (self.k - 1, 0))
+        return self.pw(F.relu(self.dw(x)))
+
+class Block(nn.Module):
+    def __init__(self, d_model, use_global):
+        super().__init__()
+        self.use_global = use_global
+        self.ln1 = nn.LayerNorm(d_model)
+        self.local = LocalConv1D(d_model, LOCAL_KERNEL_SIZE)
+        if use_global:
+            self.ln2 = nn.LayerNorm(d_model)
+            self.global_conv = GlobalConv1D(d_model, GLOBAL_KERNEL_SIZE, FFT_SIZE)
+        self.ln3 = nn.LayerNorm(d_model)
+        self.ff = nn.Sequential(
+            nn.Linear(d_model, d_model*4),
+            nn.GELU(),
+            nn.Linear(d_model*4, d_model)
+        )
+
+    def forward(self, x):
+        x = x + self.local(self.ln1(x).transpose(1,2)).transpose(1,2)
+        if self.use_global:
+            x = x + self.global_conv(self.ln2(x).transpose(1,2)).transpose(1,2)
+        return x + self.ff(self.ln3(x))
+
+class ChatGCLM(nn.Module):
+    def __init__(self, vocab):
+        super().__init__()
+        self.emb = nn.Embedding(vocab, D_MODEL)
+        self.pos = nn.Embedding(MAX_SEQ_LEN, D_MODEL)
+        self.layers = nn.ModuleList([
+            Block(D_MODEL, i % USE_GLOBAL_EVERY_N_LAYERS == 0)
+            for i in range(N_LAYERS)
+        ])
+        self.ln = nn.LayerNorm(D_MODEL)
+        self.head = nn.Linear(D_MODEL, vocab)
+        self.head.weight = self.emb.weight
+
+    def forward(self, x):
+        T = x.size(1)
+        if T > MAX_SEQ_LEN:
+            x = x[:, -MAX_SEQ_LEN:]
+            T = MAX_SEQ_LEN
+        h = self.emb(x) + self.pos(torch.arange(T, device=x.device))
+        for layer in self.layers:
+            h = layer(h)
+        return self.head(self.ln(h))

sample.py

@@ -0,0 +1,103 @@
+import os
+import torch
+import torch.nn.functional as F
+import tiktoken
+from model import ChatGCLM, MAX_SEQ_LEN
+
+MODEL_PATH = None
+for f in os.listdir("."):
+    if f.startswith("ChatGCLM_") and f.endswith(".pt"):
+        MODEL_PATH = f
+        break
+
+if MODEL_PATH is None:
+    print("Error: No model checkpoint found!")
+    print("Please train the model first with: python3 train_chatgclm.py")
+    exit(1)
+
+TOKENIZER_NAME = "gpt2"
+EOS_ID = 2
+
+def load_model(device):
+    tok = tiktoken.get_encoding(TOKENIZER_NAME)
+    vocab_size = tok.n_vocab
+    
+    model = ChatGCLM(vocab_size).to(device)
+    if os.path.exists(MODEL_PATH) and os.path.getsize(MODEL_PATH) > 0:
+        print(f"Loading model from: {MODEL_PATH}")
+        model.load_state_dict(torch.load(MODEL_PATH, map_location=device))
+        model.eval()
+        return model, tok
+    else:
+        print(f"Error: Could not load model from {MODEL_PATH}")
+        return None, None
+
+@torch.no_grad()
+def generate(model, prompt, tokenizer, device, max_new_tokens=200, temperature=0.8, top_k=50):
+    model.eval()
+    input_ids = tokenizer.encode(prompt)
+    x = torch.tensor([input_ids], dtype=torch.long, device=device)
+    
+    print(f"\n{'='*70}")
+    print(f"PROMPT: {prompt}")
+    print(f"{'='*70}")
+    print("GENERATED TEXT:")
+    print(prompt, end="", flush=True)
+    
+    generated_tokens = []
+    for _ in range(max_new_tokens):
+        ctx = x[:, -MAX_SEQ_LEN:] if x.size(1) > MAX_SEQ_LEN else x
+        logits = model(ctx)
+        next_token_logits = logits[:, -1, :] / temperature
+        
+        if top_k is not None:
+            v, _ = torch.topk(next_token_logits, min(top_k, next_token_logits.size(-1)))
+            next_token_logits[next_token_logits < v[:, [-1]]] = -float('Inf')
+        
+        probs = F.softmax(next_token_logits, dim=-1)
+        next_token = torch.multinomial(probs, num_samples=1)
+        idx = next_token.item()
+        
+        if idx == EOS_ID:
+            break
+        
+        x = torch.cat((x, next_token), dim=1)
+        generated_tokens.append(idx)
+        token_text = tokenizer.decode([idx])
+        print(token_text, end="", flush=True)
+    
+    print(f"\n{'='*70}\n")
+    return tokenizer.decode(generated_tokens)
+
+if __name__ == "__main__":
+    device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+    print(f"Using device: {device}")
+    
+    model, tokenizer = load_model(device)
+    
+    if model is None:
+        exit(1)
+    
+    test_prompts = [
+        "Once upon a time",
+        "The future of AI is",
+        "In a world where",
+    ]
+    
+    print("\n" + "="*70)
+    print("ChatGCLM Text Generation Demo")
+    print("="*70)
+    
+    for prompt in test_prompts:
+        generate(model, prompt, tokenizer, device, max_new_tokens=150, temperature=0.8, top_k=50)
+    
+    print("\n" + "="*70)
+    print("Interactive Mode - Enter your own prompts!")
+    print("="*70)
+    
+    while True:
+        user_prompt = input("\nEnter prompt (or 'exit' to quit): ")
+        if user_prompt.lower() == 'exit':
+            break
+        if user_prompt.strip():
+            generate(model, user_prompt, tokenizer, device, max_new_tokens=200, temperature=0.8, top_k=50)

train_chatgclm.py

@@ -0,0 +1,174 @@
+import os
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.data import Dataset, DataLoader
+from tqdm import tqdm
+import tiktoken
+import contextlib
+from model import ChatGCLM, MAX_SEQ_LEN
+
+if os.name != "nt":
+    os.environ.setdefault("PYTORCH_CUDA_ALLOC_CONF", "expandable_segments:True")
+
+if torch.cuda.is_available():
+    torch.set_float32_matmul_precision("high")
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+
+DATA_DIR = "data"
+DATA_PCT = 0.002
+TOKENIZER_NAME = "gpt2"
+VOCAB_SAVE_PATH = "vocab_map.pt"
+
+EPOCHS = 50
+MICRO_BATCH_SIZE = 1
+GRAD_ACCUM_STEPS = 8
+LEARNING_RATE = 5e-4
+MIN_LR = 1e-5
+
+SAVE_N_EPOCHS = 1
+
+PAD_ID = 0
+SEP_ID = 1
+EOS_ID = 2
+OFFSET = 3
+
+def build_dataset_vocab(data_dir, tokenizer, save_path):
+    vocab_size = tokenizer.n_vocab
+    torch.save({
+        "vocab_size": vocab_size,
+        "PAD_ID": PAD_ID,
+        "SEP_ID": SEP_ID,
+        "EOS_ID": EOS_ID,
+    }, save_path)
+    return vocab_size
+
+class RemappedTextDataset(Dataset):
+    def __init__(self, ids, max_len):
+        self.ids = ids
+        self.max_len = max_len
+
+    def __len__(self):
+        return max(0, (len(self.ids) - 1) // self.max_len)
+
+    def __getitem__(self, i):
+        start = i * self.max_len
+        x = self.ids[start : start + self.max_len]
+        y = self.ids[start + 1 : start + self.max_len + 1]
+        
+        if len(x) < self.max_len:
+            x = x + [0] * (self.max_len - len(x))
+        if len(y) < self.max_len:
+            y = y + [0] * (self.max_len - len(y))
+            
+        return torch.tensor(x, dtype=torch.long), torch.tensor(y, dtype=torch.long)
+
+def format_params(num):
+    if num >= 1_000_000_000:
+        return f"{num/1_000_000_000:.1f}B"
+    elif num >= 1_000_000:
+        return f"{num/1_000_000:.1f}M"
+    else:
+        return f"{num/1_000:.1f}K"
+
+@torch.no_grad()
+def estimate_loss(model, dl, device, ctx):
+    model.eval()
+    losses = []
+    limit = 50
+    for i, (x, y) in enumerate(dl):
+        if i >= limit: break
+        x, y = x.to(device), y.to(device)
+        with ctx:
+            logits = model(x)
+            loss = F.cross_entropy(logits.reshape(-1, logits.size(-1)), y.reshape(-1), ignore_index=PAD_ID)
+        losses.append(loss.item())
+    model.train()
+    return sum(losses) / len(losses) if losses else 0.0
+
+def train():
+    device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+    tok = tiktoken.get_encoding(TOKENIZER_NAME)
+    vocab = build_dataset_vocab(DATA_DIR, tok, VOCAB_SAVE_PATH)
+
+    full_text = ""
+    for f in os.listdir(DATA_DIR):
+        if not f.endswith(".txt"): continue
+        fpath = os.path.join(DATA_DIR, f)
+        content = open(fpath, "r", encoding="utf-8").read()
+        full_text += content + "\n"
+
+    ids = tok.encode(full_text) + [EOS_ID]
+    
+    n = len(ids)
+    split_idx = int(n * 0.9)
+    train_ids = ids[:split_idx]
+    val_ids = ids[split_idx:]
+    
+    train_ds = RemappedTextDataset(train_ids, MAX_SEQ_LEN)
+    val_ds = RemappedTextDataset(val_ids, MAX_SEQ_LEN)
+    train_dl = DataLoader(train_ds, batch_size=MICRO_BATCH_SIZE, shuffle=True)
+    val_dl = DataLoader(val_ds, batch_size=MICRO_BATCH_SIZE, shuffle=False)
+
+    model = ChatGCLM(vocab).to(device)
+    num_params = sum(p.numel() for p in model.parameters())
+    param_str = format_params(num_params)
+    save_path = f"ChatGCLM_{param_str}.pt"
+    
+    print("-" * 30)
+    print(f"ChatGCLM TRAINING START")
+    print(f"Model ID:    {save_path}")
+    print(f"Parameters:  {num_params:,}")
+    print(f"Device:      {device}")
+    print(f"Vocab Size:  {vocab}")
+    print(f"Learning Rate: {LEARNING_RATE}")
+    print(f"Epochs:      {EPOCHS}")
+    print("-" * 30)
+
+    if os.path.exists(save_path) and os.path.getsize(save_path) > 0:
+        print(f"⏳ Found checkpoint at {save_path}, loading...")
+        model.load_state_dict(torch.load(save_path, map_location=device))
+        print("✓ Model weights loaded successfully! Resuming training.")
+    else:
+        print("ℹ No checkpoint found. Starting training from scratch.")
+
+    opt = torch.optim.AdamW(model.parameters(), lr=LEARNING_RATE)
+    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=EPOCHS, eta_min=MIN_LR)
+    loss_fn = nn.CrossEntropyLoss(ignore_index=PAD_ID)
+    ctx = torch.amp.autocast(device) if device == "cuda" else contextlib.nullcontext()
+    scaler = torch.amp.GradScaler(device) if device == "cuda" else None
+
+    for ep in range(EPOCHS):
+        opt.zero_grad(set_to_none=True)
+        pbar = tqdm(train_dl, desc=f"Epoch {ep+1}/{EPOCHS}")
+        running_loss = 0.0
+        for i, (x, y) in enumerate(pbar):
+            x, y = x.to(device), y.to(device)
+            with ctx:
+                logits = model(x)
+                loss = loss_fn(logits.reshape(-1, vocab), y.reshape(-1))
+                loss_val = loss.item()
+                loss = loss / GRAD_ACCUM_STEPS
+            if scaler:
+                scaler.scale(loss).backward()
+            else:
+                loss.backward()
+            if (i+1) % GRAD_ACCUM_STEPS == 0:
+                if scaler:
+                    scaler.step(opt)
+                    scaler.update()
+                else:
+                    opt.step()
+                opt.zero_grad(set_to_none=True)
+            running_loss = 0.9 * running_loss + 0.1 * loss_val if running_loss > 0 else loss_val
+            pbar.set_postfix(loss=f"{running_loss:.4f}")
+        val_loss = estimate_loss(model, val_dl, device, ctx)
+        current_lr = scheduler.get_last_lr()[0]
+        print(f"Epoch {ep+1} | Train Loss: {running_loss:.4f} | Val Loss: {val_loss:.4f} | LR: {current_lr:.6f}")
+        torch.save(model.state_dict(), save_path)
+        print(f"✓ Model saved successfully after epoch {ep+1} to {save_path}")
+        scheduler.step()
+
+if __name__ == "__main__":
+    train()

vocab_map.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3622ed31c3f722a9e12ae90ffdc9a51a063809d43c7aee885c1b75037161b202
+size 1337