remote-gpu-client/examples/deep_nanogpt_resumable.py

import sys
import traceback
import os
import time

print("🔹 DEBUG: Script Wrapper Started")

try:
    import torch
    import torch.nn as nn
    from torch.nn import functional as F
    import requests
    import matplotlib.pyplot as plt

    print("🔹 DEBUG: Imports successful")

    # --- Hyperparameters ---
    batch_size = 32  # Reduced to fit 2x 72-layer models in 16GB
    block_size = 256
    learning_rate = 3e-4
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    print(f"🔹 DEBUG: Device selected: {device}")
    
    # HARD MODE CONFIG
    n_embd = 128
    n_head = 4
    n_layer = 72       # Deep!
    dropout = 0.1
    
    # Training Loop Config
    TOTAL_ITERS = 200  # Target
    CHUNK_ITERS = 20   # Steps per execution (safe < 300s)
    eval_interval = 20 # Log / Plot every 20 steps
    eval_iters = 10    # Fast eval

    torch.manual_seed(1337)
    
    # --- Persistence Paths ---
    storage_dir = "/home/user/app/storage/deep_experiment_v2"
    os.makedirs(storage_dir, exist_ok=True)
    ckpt_path_a = os.path.join(storage_dir, 'ckpt_a.pt')
    ckpt_path_b = os.path.join(storage_dir, 'ckpt_b.pt')
    history_path = os.path.join(storage_dir, 'history.pt')

    # --- Data Loading ---
    def get_data():
        if not os.path.exists('input.txt'):
            url = 'https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt'
            # print(f"Downloading {url}...")
            data = requests.get(url).text
            with open('input.txt', 'w') as f:
                f.write(data)
        else:
            with open('input.txt', 'r') as f:
                data = f.read()
        return data

    data = get_data()
    chars = sorted(list(set(data)))
    vocab_size = len(chars)
    stoi = { ch:i for i,ch in enumerate(chars) }
    itos = { i:ch for i,ch in enumerate(chars) }
    encode = lambda s: [stoi[c] for c in s]
    decode = lambda l: ''.join([itos[i] for i in l])

    data_train = torch.tensor(encode(data), dtype=torch.long)
    n = int(0.9 * len(data_train))
    train_data = data_train[:n]
    val_data = data_train[n:]

    def get_batch(split):
        data = train_data if split == 'train' else val_data
        ix = torch.randint(len(data) - block_size, (batch_size,))
        x = torch.stack([data[i:i+block_size] for i in ix])
        y = torch.stack([data[i+1:i+block_size+1] for i in ix])
        x, y = x.to(device), y.to(device)
        return x, y

    @torch.no_grad()
    def estimate_loss(model):
        out = {}
        model.eval()
        for split in ['train', 'val']:
            losses = torch.zeros(eval_iters)
            for k in range(eval_iters):
                X, Y = get_batch(split)
                logits, loss = model(X, Y)
                losses[k] = loss.item()
            out[split] = losses.mean()
        model.train()
        return out

    # --- Components ---
    class Head(nn.Module):
        def __init__(self, head_size):
            super().__init__()
            self.key = nn.Linear(n_embd, head_size, bias=False)
            self.query = nn.Linear(n_embd, head_size, bias=False)
            self.value = nn.Linear(n_embd, head_size, bias=False)
            self.register_buffer('tril', torch.tril(torch.ones(block_size, block_size)))
            self.dropout = nn.Dropout(dropout)

        def forward(self, x):
            B,T,C = x.shape
            k = self.key(x)   
            q = self.query(x) 
            wei = q @ k.transpose(-2, -1) * C**-0.5
            wei = wei.masked_fill(self.tril[:T, :T] == 0, float('-inf'))
            wei = F.softmax(wei, dim=-1)
            wei = self.dropout(wei)
            v = self.value(x)
            out = wei @ v
            return out

    class MultiHeadAttention(nn.Module):
        def __init__(self, num_heads, head_size):
            super().__init__()
            self.heads = nn.ModuleList([Head(head_size) for _ in range(num_heads)])
            self.proj = nn.Linear(n_embd, n_embd)
            self.dropout = nn.Dropout(dropout)

        def forward(self, x):
            out = torch.cat([h(x) for h in self.heads], dim=-1)
            out = self.dropout(self.proj(out))
            return out

    class FeedForward(nn.Module):
        def __init__(self, n_embd):
            super().__init__()
            self.net = nn.Sequential(
                nn.Linear(n_embd, 4 * n_embd),
                nn.ReLU(),
                nn.Linear(4 * n_embd, n_embd),
                nn.Dropout(dropout),
            )

        def forward(self, x):
            return self.net(x)

    # --- Standard Block ---
    class BlockStandard(nn.Module):
        def __init__(self, n_embd, n_head):
            super().__init__()
            head_size = n_embd // n_head
            self.sa = MultiHeadAttention(n_head, head_size)
            self.ffwd = FeedForward(n_embd)
            self.ln1 = nn.LayerNorm(n_embd)
            self.ln2 = nn.LayerNorm(n_embd)

        def forward(self, x):
            x = x + self.sa(self.ln1(x))
            x = x + self.ffwd(self.ln2(x))
            return x

    # --- mHC Block ---
    class RMSNorm(nn.Module):
        def __init__(self, dim, eps=1e-6):
            super().__init__()
            self.eps = eps
            self.weight = nn.Parameter(torch.ones(dim))

        def _norm(self, x):
            return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

        def forward(self, x):
            return self._norm(x.float()).type_as(x) * self.weight

    class BlockMHC(nn.Module):
        def __init__(self, n_embd, n_head):
            super().__init__()
            head_size = n_embd // n_head
            self.sa = MultiHeadAttention(n_head, head_size)
            self.ffwd = FeedForward(n_embd)
            self.alpha1 = nn.Parameter(torch.tensor(0.9))
            self.beta1 = nn.Parameter(torch.tensor(0.1))
            self.ln1 = RMSNorm(n_embd)
            self.alpha2 = nn.Parameter(torch.tensor(0.9))
            self.beta2 = nn.Parameter(torch.tensor(0.1))
            self.ln2 = RMSNorm(n_embd)

        def forward(self, x):
            mix1 = self.alpha1 * x + self.beta1 * self.sa(x)
            x = self.ln1(mix1)
            mix2 = self.alpha2 * x + self.beta2 * self.ffwd(x)
            x = self.ln2(mix2)
            return x

    class GPT(nn.Module):
        def __init__(self, arch_type='standard'):
            super().__init__()
            self.arch_type = arch_type
            self.token_embedding_table = nn.Embedding(vocab_size, n_embd)
            self.position_embedding_table = nn.Embedding(block_size, n_embd)
            if arch_type == 'standard':
                self.blocks = nn.Sequential(*[BlockStandard(n_embd, n_head) for _ in range(n_layer)])
                self.ln_f = nn.LayerNorm(n_embd)
            elif arch_type == 'mhc':
                self.blocks = nn.Sequential(*[BlockMHC(n_embd, n_head) for _ in range(n_layer)])
                self.ln_f = RMSNorm(n_embd)
            self.lm_head = nn.Linear(n_embd, vocab_size)

        def forward(self, idx, targets=None):
            B, T = idx.shape
            tok_emb = self.token_embedding_table(idx)
            pos_emb = self.position_embedding_table(torch.arange(T, device=device))
            x = tok_emb + pos_emb
            x = self.blocks(x)
            x = self.ln_f(x)
            logits = self.lm_head(x)
            if targets is None:
                loss = None
            else:
                B, T, C = logits.shape
                logits = logits.view(B*T, C)
                targets = targets.view(B*T)
                loss = F.cross_entropy(logits, targets)
            return logits, loss
            
        def generate(self, idx, max_new_tokens):
            for _ in range(max_new_tokens):
                idx_cond = idx[:, -block_size:]
                logits, loss = self(idx_cond)
                logits = logits[:, -1, :]
                probs = F.softmax(logits, dim=-1)
                idx_next = torch.multinomial(probs, num_samples=1)
                idx = torch.cat((idx, idx_next), dim=1)
            return idx

    # --- Resumable Logic ---
    history = {'a': [], 'b': [], 'steps': []}
    current_iter = 0

    if os.path.exists(history_path):
        print(f"🔄 Resuming from history: {history_path}")
        try:
            history = torch.load(history_path)
            if history['steps']:
                current_iter = history['steps'][-1]
            print(f"   Last Checkpoint: Step {current_iter}")
        except:
            print("   ⚠️ Error loading history, starting fresh.")
            current_iter = 0

    model_a = GPT(arch_type='standard').to(device)
    opt_a = torch.optim.AdamW(model_a.parameters(), lr=learning_rate)
    
    model_b = GPT(arch_type='mhc').to(device)
    opt_b = torch.optim.AdamW(model_b.parameters(), lr=learning_rate)
    
    # Load Weights if exist
    if os.path.exists(ckpt_path_a):
        model_a.load_state_dict(torch.load(ckpt_path_a))
        opt_a.load_state_dict(torch.load(os.path.join(storage_dir, 'opt_a.pt')))
    
    if os.path.exists(ckpt_path_b):
        model_b.load_state_dict(torch.load(ckpt_path_b))
        opt_b.load_state_dict(torch.load(os.path.join(storage_dir, 'opt_b.pt')))

    # --- Training Chunk ---
    print(f"🚀 Training Chunk: Steps {current_iter} -> {current_iter + CHUNK_ITERS} (Target: {TOTAL_ITERS})")
    
    start_chunk = time.time()
    
    for i in range(CHUNK_ITERS):
        global_step = current_iter + i + 1
        if global_step > TOTAL_ITERS:
            break
            
        # Eval & Log
        if global_step % eval_interval == 0 or global_step == 1:
            la = estimate_loss(model_a)
            lb = estimate_loss(model_b)
            print(f"Step {global_step}: Loss A={la['train']:.4f}, Loss B={lb['train']:.4f}")
            history['steps'].append(global_step)
            history['a'].append(la['train'].item())
            history['b'].append(lb['train'].item())
            
            # Save History
            torch.save(history, history_path)
            
            # Update Plots inside the chunk to give realtime feedback
            plt.figure(figsize=(10, 6))
            plt.plot(history['steps'], history['a'], label='Standard GPT', marker='o')
            plt.plot(history['steps'], history['b'], label='DeepMHC', marker='x')
            plt.xlabel('Iterations')
            plt.ylabel('Loss')
            plt.title(f'Training Stability: Deep-NanoGPT (Layers={n_layer})')
            plt.legend()
            plt.grid(True)
            plt.savefig(os.path.join(storage_dir, 'comparison_loss_v2.png'))
            plt.close()

        # Step A
        xb, yb = get_batch('train')
        logits, loss = model_a(xb, yb)
        opt_a.zero_grad(set_to_none=True)
        loss.backward()
        # Optional: Clip grad to prevent immediate explosion, but let's see it fail naturally
        # torch.nn.utils.clip_grad_norm_(model_a.parameters(), 1.0)
        opt_a.step()
        
        # Step B
        xb, yb = get_batch('train')
        logits, loss = model_b(xb, yb)
        opt_b.zero_grad(set_to_none=True)
        loss.backward()
        opt_b.step()

    print(f"🏁 Chunk Done in {time.time()-start_chunk:.2f}s. Global Step: {global_step}")

    # --- Save Checkpoints ---
    torch.save(model_a.state_dict(), ckpt_path_a)
    torch.save(opt_a.state_dict(), os.path.join(storage_dir, 'opt_a.pt'))
    torch.save(model_b.state_dict(), ckpt_path_b)
    torch.save(opt_b.state_dict(), os.path.join(storage_dir, 'opt_b.pt'))
    torch.save(history, history_path)
    
    print("💾 Checkpoints Saved.")
    
    # Return 1 if more work needed, 0 if done
    if global_step < TOTAL_ITERS:
        print("CONTINUE_TRAINING") # Signal to client
    else:
        print("TRAINING_COMPLETE")
        # Final Generation
        context = torch.zeros((1, 1), dtype=torch.long, device=device)
        print("Generating final samples...")
        out_a = decode(model_a.generate(context, max_new_tokens=200)[0].tolist())
        out_b = decode(model_b.generate(context, max_new_tokens=200)[0].tolist())
        
        sp = os.path.join(storage_dir, 'generation_sample_v2.txt')
        with open(sp, 'w') as f:
            f.write(f"--- MODEL A (Standard, L={n_layer}) ---\n{out_a}\n\n")
            f.write(f"--- MODEL B (mHC, L={n_layer}) ---\n{out_b}\n\n")
            
        # Copy to cwd for download
        os.system(f"cp {os.path.join(storage_dir, 'comparison_loss_v2.png')} .")
        os.system(f"cp {sp} .")
        print("✅ Final Artifacts prepared for download.")

except Exception as e:
    print(f"\n❌ FATAL SCRIPT ERROR: {e}")
    traceback.print_exc()