model.py

import torch
from torch import nn
from torch.nn import functional as F
from model_utils import Block

device = "cuda" if torch.cuda.is_available() else "cpu"


class Gpt(nn.Module):
    def __init__(self, vocab_size, n_embd, n_layer, n_head, block_size, dropout=None):
        super().__init__()
        self.vocab_size = vocab_size
        self.n_embd = n_embd
        self.n_layer = n_layer
        self.n_head = n_head
        self.block_size = block_size
        self.dropout = dropout
        # each token directly reads off the logits for the next token from a lookup table
        self.token_embedding_table = nn.Embedding(self.vocab_size, self.n_embd)
        self.position_embedding_table = nn.Embedding(self.block_size, self.n_embd)
        self.blocks = nn.Sequential(
            *[Block(self.n_embd, self.n_head, self.block_size, self.dropout) for _ in range(self.n_layer)]
        )
        self.ln_f = nn.LayerNorm(self.n_embd)  # final layer norm
        self.lm_head = nn.Linear(self.n_embd, self.vocab_size)

    def forward(self, idx, targets=None):
        B, T = idx.shape

        # idx and targets are both (B,T) tensor of integers
        tok_emb = self.token_embedding_table(idx)  # (B,T,C)
        pos_emb = self.position_embedding_table(torch.arange(T, device=device))  # (T,C)
        x = tok_emb + pos_emb  # (B,T,C)
        x = self.blocks(x)  # (B,T,C)
        x = self.ln_f(x)  # (B,T,C)
        logits = self.lm_head(x)  # (B,T,vocab_size)

        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, block_size):
        # idx is (B, T) array of indices in the current context
        for _ in range(max_new_tokens):
            # crop idx to the last block_size tokens
            idx_cond = idx[:, -block_size:]
            # get the predictions
            logits, loss = self(idx_cond)
            # focus only on the last time step
            logits = logits[:, -1, :]  # becomes (B, C)
            # apply softmax to get probabilities
            probs = F.softmax(logits, dim=-1)  # (B, C)
            # sample from the distribution
            idx_next = torch.multinomial(probs, num_samples=1)  # (B, 1)
            # append sampled index to the running sequence
            idx = torch.cat((idx, idx_next), dim=1)  # (B, T+1)
        return idx