Initial commit with model trained with loss less than 0.099999

README.md

@@ -1,13 +1,107 @@
 ---
 title: NextWordGPT
-emoji: 🦀
-colorFrom: green
-colorTo: indigo
+emoji: 🏃
+colorFrom: purple
+colorTo: yellow
 sdk: gradio
 sdk_version: 5.12.0
 app_file: app.py
 pinned: false
-short_description: Next Word Generator Which Trained On Shakespeare's Text
+short_description: 'Transformer trained on Shakespearean text '
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+
+<pre>
+Epoch 1/50: 100%|██████████| 82/82 [01:16<00:00,  1.08step/s, loss=6.2489]
+Epoch 1/50, Loss: 7.0745, Time: 76.07s
+Epoch 2/50: 100%|██████████| 82/82 [01:22<00:00,  1.00s/step, loss=5.6592]
+Epoch 2/50, Loss: 5.6716, Time: 82.14s
+Epoch 3/50: 100%|██████████| 82/82 [01:25<00:00,  1.05s/step, loss=5.2294]
+Epoch 3/50, Loss: 5.1465, Time: 85.97s
+Epoch 4/50: 100%|██████████| 82/82 [01:27<00:00,  1.07s/step, loss=4.8800]
+Epoch 4/50, Loss: 4.8121, Time: 87.40s
+Epoch 5/50: 100%|██████████| 82/82 [01:28<00:00,  1.08s/step, loss=4.6155]
+Epoch 5/50, Loss: 4.5597, Time: 88.28s
+Epoch 6/50: 100%|██████████| 82/82 [01:29<00:00,  1.10s/step, loss=4.4006]
+Epoch 6/50, Loss: 4.3344, Time: 89.88s
+Epoch 7/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=4.1696]
+Epoch 7/50, Loss: 4.1084, Time: 91.19s
+Epoch 8/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=3.9078]
+Epoch 8/50, Loss: 3.8753, Time: 91.43s
+Epoch 9/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=3.6197]
+Epoch 9/50, Loss: 3.6167, Time: 91.38s
+Epoch 10/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=3.3067]
+Epoch 10/50, Loss: 3.3436, Time: 91.24s
+Epoch 11/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=3.0890]
+Epoch 11/50, Loss: 2.9951, Time: 91.45s
+Epoch 12/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=2.7631]
+Epoch 12/50, Loss: 2.7189, Time: 91.25s
+Epoch 13/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=2.5140]
+Epoch 13/50, Loss: 2.4935, Time: 91.21s
+Epoch 14/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=2.3475]
+Epoch 14/50, Loss: 2.3095, Time: 91.42s
+Epoch 15/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=2.1527]
+Epoch 15/50, Loss: 2.1343, Time: 91.61s
+Epoch 16/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=1.9820]
+Epoch 16/50, Loss: 1.9522, Time: 91.35s
+Epoch 17/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=1.7411]
+Epoch 17/50, Loss: 1.7585, Time: 91.53s
+Epoch 18/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=1.5516]
+Epoch 18/50, Loss: 1.5744, Time: 91.77s
+Epoch 19/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=1.3633]
+Epoch 19/50, Loss: 1.4087, Time: 91.45s
+Epoch 20/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=1.2165]
+Epoch 20/50, Loss: 1.2397, Time: 91.37s
+Epoch 21/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=1.1129]
+Epoch 21/50, Loss: 1.0790, Time: 91.69s
+Epoch 22/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.9431]
+Epoch 22/50, Loss: 0.9302, Time: 91.61s
+Epoch 23/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.8262]
+Epoch 23/50, Loss: 0.8121, Time: 91.39s
+Epoch 24/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.7406]
+Epoch 24/50, Loss: 0.7170, Time: 91.36s
+Epoch 25/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.6618]
+Epoch 25/50, Loss: 0.6387, Time: 91.58s
+Epoch 26/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.5878]
+Epoch 26/50, Loss: 0.5709, Time: 91.55s
+Epoch 27/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.5246]
+Epoch 27/50, Loss: 0.5079, Time: 91.23s
+Epoch 28/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.4453]
+Epoch 28/50, Loss: 0.4472, Time: 91.39s
+Epoch 29/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.3966]
+Epoch 29/50, Loss: 0.3912, Time: 91.58s
+Epoch 30/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.3454]
+Epoch 30/50, Loss: 0.3401, Time: 91.14s
+Epoch 31/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.3288]
+Epoch 31/50, Loss: 0.3059, Time: 91.06s
+Epoch 32/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.2900]
+Epoch 32/50, Loss: 0.2712, Time: 91.22s
+Epoch 33/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.2608]
+Epoch 33/50, Loss: 0.2438, Time: 91.44s
+Epoch 34/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.2365]
+Epoch 34/50, Loss: 0.2215, Time: 91.02s
+Epoch 35/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.2159]
+Epoch 35/50, Loss: 0.2017, Time: 91.14s
+Epoch 36/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1979]
+Epoch 36/50, Loss: 0.1840, Time: 91.59s
+Epoch 37/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1814]
+Epoch 37/50, Loss: 0.1681, Time: 91.70s
+Epoch 38/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1661]
+Epoch 38/50, Loss: 0.1539, Time: 91.46s
+Epoch 39/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1522]
+Epoch 39/50, Loss: 0.1410, Time: 91.53s
+Epoch 40/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1390]
+Epoch 40/50, Loss: 0.1295, Time: 91.60s
+Epoch 41/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1350]
+Epoch 41/50, Loss: 0.1215, Time: 91.51s
+Epoch 42/50: 100%|██████████| 82/82 [01:31<00:00,  1.11s/step, loss=0.1304]
+Epoch 42/50, Loss: 0.1156, Time: 91.43s
+Epoch 43/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1247]
+Epoch 43/50, Loss: 0.1099, Time: 91.80s
+Epoch 44/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1162]
+Epoch 44/50, Loss: 0.1047, Time: 91.56s
+Epoch 45/50: 100%|██████████| 82/82 [01:31<00:00,  1.12s/step, loss=0.1122]
+Epoch 45/50, Loss: 0.0998, Time: 91.53s
+</pre>

app.py

@@ -0,0 +1,56 @@
+import torch
+import torch.nn.functional as F
+from transformers import GPT2Tokenizer
+import gradio as gr
+
+# Load tokenizer
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2")  # Using GPT-2 tokenizer for compatibility
+
+# Load model
+from train_get2_8_init import GPT, GPTConfig
+
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+
+# Initialize the model
+config = GPTConfig()
+model = GPT(config)
+model.load_state_dict(torch.load("decoder_only_transformer.pth", map_location=torch.device(device)))
+model.eval()
+model.to(device)
+
+# Prediction function
+def generate_text(input_text, max_length=50, top_k=50):
+    with torch.no_grad():
+        # Tokenize input
+        tokens = tokenizer.encode(input_text, return_tensors="pt").to(device)
+        x = tokens
+        while x.size(1) < max_length:
+            # Forward pass to get logits
+            logits = model(x)[0]  # (B, T, vocab_size)
+            logits = logits[:, -1, :]  # Take the logits at the last position
+            
+            # Get probabilities and do top-k sampling
+            probs = F.softmax(logits, dim=-1)
+            topk_probs, topk_indices = torch.topk(probs, top_k, dim=-1)
+            ix = torch.multinomial(topk_probs, 1)  # Sample token
+            xcol = torch.gather(topk_indices, -1, ix)  # Gather indices
+            x = torch.cat((x, xcol), dim=1)  # Append to sequence
+
+        # Decode tokens into text
+        generated_text = tokenizer.decode(x[0])
+        return generated_text
+
+# Gradio Interface
+def gradio_interface(input_text):
+    return generate_text(input_text)
+
+interface = gr.Interface(
+    fn=gradio_interface,
+    inputs=gr.Textbox(lines=2, placeholder="Enter your text here..."),
+    outputs=gr.Textbox(lines=2, placeholder="Generated text will appear here..."),
+    title="Text Prediction App",
+    description="Enter a text prompt to generate the next sequence of words.",
+)
+
+# Launch the app
+interface.launch()

decoder_only_transformer.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d66ba9508c76b5b60af5713845ebe0528ea2e034d4fb015243bc6f62e764e144
+size 548151190

lr_finder.py

@@ -0,0 +1,90 @@
+from torch_lr_finder import LRFinder
+from torch.nn import CrossEntropyLoss
+import torch.optim as optim
+import torch
+from transformer import Config, DecoderOnlyTransformer
+
+class DataLoaderLite:
+    def __init__(self, B, T):
+        self.B = B
+        self.T = T
+
+        # at init load tokens from disk and store them in memory
+        with open('input.txt', 'r') as f:
+            text = f.read()
+        enc = tiktoken.get_encoding('gpt2') 
+        tokens = enc.encode(text)
+        self.tokens = torch.tensor(tokens)
+        print(f'loaded {len(self.tokens)} tokens')
+        print(f'1 epoch = {len(self.tokens) // (B * T)} batches')
+
+        # state
+        self.current_position = 0
+    
+    def next_batch(self):
+        B, T = self.B, self.T
+        buf = self.tokens[self.current_position: self.current_position + B * T + 1]
+        x = (buf[:-1]).view(B, T) # inputs
+        y = (buf[1:]).view(B, T) # targets
+        # advance the position in the tensor
+        self.current_position += B*T
+        # if loading the next batch would be out of bounds, reset
+        if self.current_position + (B * T + 1) > len(self.tokens):
+            self.current_position = 0
+        return x, y
+
+
+batches, no_of_tokens = 16, 128
+train_loader = DataLoaderLite(B=batches, T=no_of_tokens)
+steps_per_epoch = len(train_loader.tokens) // (batches * no_of_tokens)
+
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+# Model configuration
+config = Config()
+
+# Tokenizer
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2")  # Use GPT-2 tokenizer for compatibility
+
+# Load trained model
+model = DecoderOnlyTransformer(config)
+model.load_state_dict(torch.load("decoder_only_transformer.pth", map_location=torch.device('cpu')))
+model.eval()
+model.to(device)
+
+amp_config = {
+    'device_type': 'cuda',
+    'dtype': torch.float16,
+}
+criterion = CrossEntropyLoss()
+grad_scaler = torch.cuda.amp.GradScaler()
+optimizer = optim.Adam(model.parameters(), lr=1e-3)
+
+# Define a custom batch fetching wrapper
+class CustomDataLoader:
+    def __init__(self, next_batch_func, num_batches):
+        self.next_batch_func = next_batch_func
+        self.num_batches = num_batches
+        self.current_batch = 0
+
+    def __iter__(self):
+        self.current_batch = 0
+        return self
+
+    def __next__(self):
+        if self.current_batch < self.num_batches:
+            self.current_batch += 1
+            return self.next_batch_func()
+        else:
+            raise StopIteration
+
+# Create a custom data loader using next_batch
+custom_train_loader = CustomDataLoader(train_loader.next_batch(), num_batches=steps_per_epoch)
+
+# Use the custom data loader with LRFinder
+lr_finder = LRFinder(
+    model, optimizer, criterion, device='cuda',
+    amp_backend='torch', amp_config=amp_config, grad_scaler=grad_scaler
+)
+lr_finder.range_test(custom_train_loader, end_lr=5, num_iter=1000, step_mode='exp')
+lr_finder.plot()
+lr_finder.reset()

requirements.txt

@@ -0,0 +1,3 @@
+torch
+transformers
+gradio

train.py

@@ -0,0 +1,275 @@
+import os
+import math
+import time
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from dataclasses import dataclass
+from torch.optim.lr_scheduler import StepLR
+from torch.cuda.amp import GradScaler, autocast
+import tiktoken
+from tqdm import tqdm
+
+class CausalSelfAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
+        self.c_proj.NANGPT_SCALE_INIT = 1
+        # regularization
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size)).view(1, 1, config.block_size, config.block_size))
+
+    def forward(self, x):
+        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        # nh is "number of heads", hs is "head size", and C (number of channels) = nh * hs
+        # e.g. in GPT-2 (124M), n_head=12, hs=64, so nh*hs=C=768 channels in the Transformer
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+        att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float('-inf'))
+        att = F.softmax(att, dim=-1)
+        y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+        # output projection
+        y = self.c_proj(y)
+        return y
+
+
+class MLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd)
+        self.gelu    = nn.GELU(approximate='tanh')
+        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd)
+        self.c_proj.NANOGPT_SCALE_INIT = 1
+
+    def forward(self, x):
+        x = self.c_fc(x)
+        x = self.gelu(x)
+        x = self.c_proj(x)
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.n_embd)
+        self.attn = CausalSelfAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd)
+        self.mlp = MLP(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+@dataclass
+class GPTConfig:
+    block_size: int = 1024 # max sequence length
+    vocab_size: int = 50257 # number of tokens: 50,000 BPE merges + 256 bytes tokens + 1 <|endoftext|> token
+    n_layer: int = 12 # number of layers
+    n_head: int = 12 # number of heads
+    n_embd: int = 768 # embedding dimension
+
+
+class GPT(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        self.transformer = nn.ModuleDict(dict(
+            wte = nn.Embedding(config.vocab_size, config.n_embd),
+            wpe = nn.Embedding(config.block_size, config.n_embd),
+            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
+            ln_f = nn.LayerNorm(config.n_embd),
+        ))
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        # weight sharing
+        self.transformer.wte.weight = self.lm_head.weight
+
+        # weight initialization
+        self.apply(self._init_weights)
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            std = 0.02
+            if hasattr(module, 'NANGPT_SCALE_INIT'):
+                std *= (2 * self.config.n_layer) ** -0.5
+            torch.nn.init.normal_(module.weight, mean = 0.0, std = std)
+            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 forward(self, idx, targets=None):
+        # idx is of shape (B, T)
+        B, T = idx.size()
+        assert T <= self.config.block_size, f"Cannot forward sequence of length {T}, block size is only {self.config.block_size}"
+        # forward the token and posisition embeddings
+        pos = torch.arange(0, T, dtype=torch.long, device=idx.device) # shape (T)
+        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (T, n_embd)
+        tok_emb = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)
+        x = tok_emb + pos_emb
+        # forward the blocks of the transformer
+        for block in self.transformer.h:
+            x = block(x)
+        # forward the final layernorm and the classifier
+        x = self.transformer.ln_f(x)
+        logits = self.lm_head(x) # (B, T, vocab_size)
+        loss = None
+        if targets is not None:
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+        return logits, loss
+
+    @classmethod
+    def from_pretrained(cls, model_type):
+        """Loads pretrained GPT-2 model weights from huggingface"""
+        assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}
+        from transformers import GPT2LMHeadModel
+        print("loading weights from pretrained gpt: %s" % model_type)
+
+        # n_layer, n_head and n_embd are determined from model_type
+        config_args = {
+            'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params
+            'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params
+            'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params
+            'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params
+        }[model_type]
+        config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints
+        config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints
+        # create a from-scratch initialized minGPT model
+        config = GPTConfig(**config_args)
+        model = GPT(config)
+        sd = model.state_dict()
+        sd_keys = sd.keys()
+        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param
+
+        # init a huggingface/transformers model
+        model_hf = GPT2LMHeadModel.from_pretrained(model_type)
+        sd_hf = model_hf.state_dict()
+
+        # copy while ensuring all of the parameters are aligned and match in names and shapes
+        sd_keys_hf = sd_hf.keys()
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)
+        transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']
+        # basically the openai checkpoints use a "Conv1D" module, but we only want to use a vanilla Linear
+        # this means that we have to transpose these weights when we import them
+        assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}"
+        for k in sd_keys_hf:
+            if any(k.endswith(w) for w in transposed):
+                # special treatment for the Conv1D weights we need to transpose
+                assert sd_hf[k].shape[::-1] == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k].t())
+            else:
+                # vanilla copy over the other parameters
+                assert sd_hf[k].shape == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k])
+
+        return model
+
+# model = GPT.from_pretrained('gpt2')
+
+device = 'cpu'
+if torch.cuda.is_available():
+    device = 'cuda'
+elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+    device = "mps"
+print(f"using device: {device}")
+
+# SEED
+torch.manual_seed(1337)
+if torch.cuda.is_available():
+    torch.cuda.manual_seed(1337)
+
+# STOP
+num_return_sequences = 5
+max_length = 30
+
+
+
+import tiktoken
+
+class DataLoaderLite:
+    def __init__(self, B, T):
+        self.B = B
+        self.T = T
+
+        # at init load tokens from disk and store them in memory
+        with open('/kaggle/input/input-txt/input.txt', 'r') as f:
+            text = f.read()
+        enc = tiktoken.get_encoding('gpt2') 
+        tokens = enc.encode(text)
+        self.tokens = torch.tensor(tokens)
+        print(f'loaded {len(self.tokens)} tokens')
+        print(f'1 epoch = {len(self.tokens) // (B * T)} batches')
+
+        # state
+        self.current_position = 0
+    
+    def next_batch(self):
+        B, T = self.B, self.T
+        buf = self.tokens[self.current_position: self.current_position + B * T + 1]
+        x = (buf[:-1]).view(B, T) # inputs
+        y = (buf[1:]).view(B, T) # targets
+        # advance the position in the tensor
+        self.current_position += B*T
+        # if loading the next batch would be out of bounds, reset
+        if self.current_position + (B * T + 1) > len(self.tokens):
+            self.current_position = 0
+        return x, y
+
+model = GPT(GPTConfig())
+model.to(device)
+
+batches, no_of_tokens = 16, 256
+train_loader = DataLoaderLite(B=batches, T=no_of_tokens)
+optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4, weight_decay=0.01)
+scheduler = StepLR(optimizer, step_size=10, gamma=0.5)
+
+# Training Loop
+steps_per_epoch = len(train_loader.tokens) // (batches * no_of_tokens)
+print(steps_per_epoch)
+EPOCHS = 50
+for epoch in range(EPOCHS):
+    loss_list = []
+    train_loader_temp = train_loader
+    start_time = time.time()
+
+    with tqdm(total=steps_per_epoch, desc=f"Epoch {epoch + 1}/{EPOCHS}", unit="step") as pbar:
+        for step in range(steps_per_epoch):
+            x, y = train_loader.next_batch()
+            x, y = x.to(device), y.to(device)
+            optimizer.zero_grad()
+            logits, loss = model(x, y)
+            loss.backward()
+            optimizer.step()
+            loss_list.append(loss.item())
+            pbar.update(1)
+            pbar.set_postfix(loss=f"{loss.item():.4f}")
+
+    scheduler.step()
+    epoch_loss = sum(loss_list) / len(loss_list)
+    print(f"Epoch {epoch + 1}/{EPOCHS}, Loss: {epoch_loss:.4f}, Time: {time.time() - start_time:.2f}s")
+    if(epoch_loss < 0.099999):
+        break

train_get2-8-init.py

@@ -0,0 +1,287 @@
+# Solving for residual std scaling issue
+import os
+import math
+import time
+import inspect
+from dataclasses import dataclass
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+
+class CausalSelfAttention(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        assert config.n_embd % config.n_head == 0
+        # key, query, value projections for all heads, but in a batch
+        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
+        # output projection
+        self.c_proj = nn.Linear(config.n_embd, config.n_embd)
+        self.c_proj.NANGPT_SCALE_INIT = 1
+        # regularization
+        self.n_head = config.n_head
+        self.n_embd = config.n_embd
+        self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size)).view(1, 1, config.block_size, config.block_size))
+
+    def forward(self, x):
+        B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
+        # calculate query, key, values for all heads in batch and move head forward to be the batch dim
+        # nh is "number of heads", hs is "head size", and C (number of channels) = nh * hs
+        # e.g. in GPT-2 (124M), n_head=12, hs=64, so nh*hs=C=768 channels in the Transformer
+        qkv = self.c_attn(x)
+        q, k, v = qkv.split(self.n_embd, dim=2)
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # (B, nh, T, hs)
+
+        att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
+        att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float('-inf'))
+        att = F.softmax(att, dim=-1)
+        y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
+
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
+        # output projection
+        y = self.c_proj(y)
+        return y
+
+
+class MLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc    = nn.Linear(config.n_embd, 4 * config.n_embd)
+        self.gelu    = nn.GELU(approximate='tanh')
+        self.c_proj  = nn.Linear(4 * config.n_embd, config.n_embd)
+        self.c_proj.NANOGPT_SCALE_INIT = 1
+
+    def forward(self, x):
+        x = self.c_fc(x)
+        x = self.gelu(x)
+        x = self.c_proj(x)
+        return x
+
+class Block(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.n_embd)
+        self.attn = CausalSelfAttention(config)
+        self.ln_2 = nn.LayerNorm(config.n_embd)
+        self.mlp = MLP(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+@dataclass
+class GPTConfig:
+    block_size: int = 1024 # max sequence length
+    vocab_size: int = 50257 # number of tokens: 50,000 BPE merges + 256 bytes tokens + 1 <|endoftext|> token
+    n_layer: int = 12 # number of layers
+    n_head: int = 12 # number of heads
+    n_embd: int = 768 # embedding dimension
+
+
+class GPT(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        self.transformer = nn.ModuleDict(dict(
+            wte = nn.Embedding(config.vocab_size, config.n_embd),
+            wpe = nn.Embedding(config.block_size, config.n_embd),
+            h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
+            ln_f = nn.LayerNorm(config.n_embd),
+        ))
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
+
+        # weight sharing
+        self.transformer.wte.weight = self.lm_head.weight
+
+        # weight initialization
+        self.apply(self._init_weights)
+
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            std = 0.02
+            if hasattr(module, 'NANGPT_SCALE_INIT'):
+                std *= (2 * self.config.n_layer) ** -0.5
+            torch.nn.init.normal_(module.weight, mean = 0.0, std = std)
+            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 forward(self, idx, targets=None):
+        # idx is of shape (B, T)
+        B, T = idx.size()
+        assert T <= self.config.block_size, f"Cannot forward sequence of length {T}, block size is only {self.config.block_size}"
+        # forward the token and posisition embeddings
+        pos = torch.arange(0, T, dtype=torch.long, device=idx.device) # shape (T)
+        pos_emb = self.transformer.wpe(pos) # position embeddings of shape (T, n_embd)
+        tok_emb = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)
+        x = tok_emb + pos_emb
+        # forward the blocks of the transformer
+        for block in self.transformer.h:
+            x = block(x)
+        # forward the final layernorm and the classifier
+        x = self.transformer.ln_f(x)
+        logits = self.lm_head(x) # (B, T, vocab_size)
+        loss = None
+        if targets is not None:
+            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
+        return logits, loss
+
+    @classmethod
+    def from_pretrained(cls, model_type):
+        """Loads pretrained GPT-2 model weights from huggingface"""
+        assert model_type in {'gpt2', 'gpt2-medium', 'gpt2-large', 'gpt2-xl'}
+        from transformers import GPT2LMHeadModel
+        print("loading weights from pretrained gpt: %s" % model_type)
+
+        # n_layer, n_head and n_embd are determined from model_type
+        config_args = {
+            'gpt2':         dict(n_layer=12, n_head=12, n_embd=768),  # 124M params
+            'gpt2-medium':  dict(n_layer=24, n_head=16, n_embd=1024), # 350M params
+            'gpt2-large':   dict(n_layer=36, n_head=20, n_embd=1280), # 774M params
+            'gpt2-xl':      dict(n_layer=48, n_head=25, n_embd=1600), # 1558M params
+        }[model_type]
+        config_args['vocab_size'] = 50257 # always 50257 for GPT model checkpoints
+        config_args['block_size'] = 1024 # always 1024 for GPT model checkpoints
+        # create a from-scratch initialized minGPT model
+        config = GPTConfig(**config_args)
+        model = GPT(config)
+        sd = model.state_dict()
+        sd_keys = sd.keys()
+        sd_keys = [k for k in sd_keys if not k.endswith('.attn.bias')] # discard this mask / buffer, not a param
+
+        # init a huggingface/transformers model
+        model_hf = GPT2LMHeadModel.from_pretrained(model_type)
+        sd_hf = model_hf.state_dict()
+
+        # copy while ensuring all of the parameters are aligned and match in names and shapes
+        sd_keys_hf = sd_hf.keys()
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.masked_bias')] # ignore these, just a buffer
+        sd_keys_hf = [k for k in sd_keys_hf if not k.endswith('.attn.bias')] # same, just the mask (buffer)
+        transposed = ['attn.c_attn.weight', 'attn.c_proj.weight', 'mlp.c_fc.weight', 'mlp.c_proj.weight']
+        # basically the openai checkpoints use a "Conv1D" module, but we only want to use a vanilla Linear
+        # this means that we have to transpose these weights when we import them
+        assert len(sd_keys_hf) == len(sd_keys), f"mismatched keys: {len(sd_keys_hf)} != {len(sd_keys)}"
+        for k in sd_keys_hf:
+            if any(k.endswith(w) for w in transposed):
+                # special treatment for the Conv1D weights we need to transpose
+                assert sd_hf[k].shape[::-1] == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k].t())
+            else:
+                # vanilla copy over the other parameters
+                assert sd_hf[k].shape == sd[k].shape
+                with torch.no_grad():
+                    sd[k].copy_(sd_hf[k])
+
+        return model
+
+# model = GPT.from_pretrained('gpt2')
+
+device = 'cpu'
+if torch.cuda.is_available():
+    device = 'cuda'
+elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+    device = "mps"
+print(f"using device: {device}")
+
+# SEED
+torch.manual_seed(1337)
+if torch.cuda.is_available():
+    torch.cuda.manual_seed(1337)
+
+# STOP
+num_return_sequences = 5
+max_length = 30
+
+
+
+import tiktoken
+
+class DataLoaderLite:
+    def __init__(self, B, T):
+        self.B = B
+        self.T = T
+
+        # at init load tokens from disk and store them in memory
+        with open('input.txt', 'r') as f:
+            text = f.read()
+        enc = tiktoken.get_encoding('gpt2') 
+        tokens = enc.encode(text)
+        self.tokens = torch.tensor(tokens)
+        print(f'loaded {len(self.tokens)} tokens')
+        print(f'1 epoch = {len(self.tokens) // (B * T)} batches')
+
+        # state
+        self.current_position = 0
+    
+    def next_batch(self):
+        B, T = self.B, self.T
+        buf = self.tokens[self.current_position: self.current_position + B * T + 1]
+        x = (buf[:-1]).view(B, T) # inputs
+        y = (buf[1:]).view(B, T) # targets
+        # advance the position in the tensor
+        self.current_position += B*T
+        # if loading the next batch would be out of bounds, reset
+        if self.current_position + (B * T + 1) > len(self.tokens):
+            self.current_position = 0
+        return x, y
+
+
+model = GPT(GPTConfig())
+model.to(device)
+
+train_loader = DataLoaderLite(B = 4, T = 32)
+
+# NEW CODE
+optimizer = torch.optim.AdamW(model.parameters(), lr = 3e-4)
+for i in range(50):
+    x, y = train_loader.next_batch()
+    x, y = x.to(device), y.to(device)
+    optimizer.zero_grad()
+    logits, loss = model(x, y)
+    loss.backward()
+    optimizer.step()
+    print(f'step{i}, loss: {loss.item()}')
+
+
+print(loss)
+import sys; sys.exit(0)
+
+torch.manual_seed(42)
+torch.cuda.manual_seed(42)
+while x.size(1) < max_length:
+    # forward the model to get the logits
+    with torch.no_grad():
+        logits = model(x)[0] # (B, T, vocab_size)
+        # take the logits at the last position
+        logits = logits[:, -1, :] # (B, vocab_size)
+        # get the probabilities
+        probs = F.softmax(logits, dim=-1)
+        # do top-k sampling of 50 (huggingface pipeline default)
+        # topk_probs here becomes (5, 50), topk_indices is (5, 50)
+        topk_probs, topk_indices = torch.topk(probs, 50, dim=-1)
+        # select a token from the top-k probabilities
+        # note: multinomial does not demand the input to sum to 1
+        ix = torch.multinomial(topk_probs, 1) # (B, 1)
+        # gather the corresponding indices
+        xcol = torch.gather(topk_indices, -1, ix) # (B, 1)
+        # append to the sequence
+        x = torch.cat((x, xcol), dim=1)
+
+# print the generated text
+for i in range(num_return_sequences):
+    tokens = x[i, :max_length].tolist()
+    decoded = enc.decode(tokens)
+    print(">", decoded)

transformer.py

@@ -0,0 +1,125 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from dataclasses import dataclass
+
+@dataclass
+class Config:
+    vocab_size: int = 50257
+    max_seq_len: int = 2048
+    dim: int = 768
+    num_layers: int = 12
+    num_heads: int = 12
+    dropout: float = 0.1
+
+class MultiHeadAttention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.n_head = config.num_heads
+        self.n_embd = config.dim
+        
+        # Linear projections for Q, K, V
+        self.c_attn = nn.Linear(config.dim, 3 * config.dim) # [n_embd, 3 * n_embd]
+        self.c_proj = nn.Linear(config.dim, config.dim) # [n_embd, n_embd]
+        
+        self.attn_dropout = nn.Dropout(config.dropout)
+        self.resid_dropout = nn.Dropout(config.dropout)
+
+    def forward(self, x):
+        B, T, C = x.size() # [B, T, n_embd]
+        
+        # Linear projection and split into Q, K, V
+        q, k, v = self.c_attn(x).split(self.n_embd, dim=2) # [B, T, n_embd] each
+        
+        # Reshape for multi-head attention
+        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # [B, n_head, T, n_embd/n_head]
+        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # [B, n_head, T, n_embd/n_head]
+        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2) # [B, n_head, T, n_embd/n_head]
+        
+        # Attention scores
+        att = (q @ k.transpose(-2, -1)) * (1.0 / (k.size(-1) ** 0.5)) # [B, n_head, T, T]
+        att = F.softmax(att, dim=-1) # [B, n_head, T, T]
+        att = self.attn_dropout(att) # [B, n_head, T, T]
+        
+        # Weighted sum of values
+        y = att @ v # [B, n_head, T, n_embd/n_head]
+        
+        # Reshape and project
+        y = y.transpose(1, 2).contiguous().view(B, T, C) # [B, T, n_embd]
+        y = self.c_proj(y) # [B, T, n_embd]
+        y = self.resid_dropout(y) # [B, T, n_embd]
+        
+        return y
+
+class FeedForward(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.c_fc = nn.Linear(config.dim, 4 * config.dim) # [n_embd, 4 * n_embd]
+        self.c_proj = nn.Linear(4 * config.dim, config.dim) # [4 * n_embd, n_embd]
+        self.dropout = nn.Dropout(config.dropout)
+
+    def forward(self, x):
+        x = self.c_fc(x) # [B, T, 4 * n_embd]
+        x = F.gelu(x) # [B, T, 4 * n_embd]
+        x = self.c_proj(x) # [B, T, n_embd]
+        x = self.dropout(x) # [B, T, n_embd]
+        return x
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.dim) # [n_embd]
+        self.attn = MultiHeadAttention(config)
+        self.ln_2 = nn.LayerNorm(config.dim) # [n_embd]
+        self.mlp = FeedForward(config)
+
+    def forward(self, x):
+        x = x + self.attn(self.ln_1(x)) # [B, T, n_embd]
+        x = x + self.mlp(self.ln_2(x)) # [B, T, n_embd]
+        return x
+
+class DecoderOnlyTransformer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.wte = nn.Embedding(config.vocab_size, config.dim) # [vocab_size, n_embd]
+        self.wpe = nn.Embedding(config.max_seq_len, config.dim) # [max_seq_len, n_embd]
+        self.drop = nn.Dropout(config.dropout)
+        self.blocks = nn.ModuleList([TransformerBlock(config) for _ in range(config.num_layers)])
+        self.ln_f = nn.LayerNorm(config.dim) # [n_embd]
+        self.lm_head = nn.Linear(config.dim, config.vocab_size, bias=False) # [n_embd, vocab_size]
+        
+        self.apply(self._init_weights)
+
+    def _init_weights(self, module):
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            module.weight.data.normal_(mean=0.0, std=0.02)
+            if isinstance(module, nn.Linear) and module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    def forward(self, idx):
+        B, T = idx.size() # [B, T]
+        
+        # Positional embeddings
+        pos = torch.arange(0, T, dtype=torch.long, device=idx.device).unsqueeze(0) # [1, T]
+        
+        # Token and position embeddings
+        tok_emb = self.wte(idx) # [B, T, n_embd]
+        pos_emb = self.wpe(pos) # [1, T, n_embd]
+        
+        # Combine embeddings and apply dropout
+        x = self.drop(tok_emb + pos_emb) # [B, T, n_embd]
+        
+        # Transformer blocks
+        for block in self.blocks:
+            x = block(x) # [B, T, n_embd]
+        
+        # Final layer norm and linear projection
+        x = self.ln_f(x) # [B, T, n_embd]
+        logits = self.lm_head(x) # [B, T, vocab_size]
+        
+        return logits