code/model.py

"""Dense decoder-only GPT model used by train.py."""

from __future__ import annotations

import torch
import torch.nn as nn
import torch.nn.functional as F

from config import ModelConfig


class RMSNorm(nn.Module):
    def __init__(self, width: int, eps: float = 1e-5):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(width))
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        variance = x.float().square().mean(dim=-1, keepdim=True)
        return (x * torch.rsqrt(variance + self.eps)).type_as(x) * self.weight


class Attention(nn.Module):
    def __init__(self, cfg: ModelConfig):
        super().__init__()
        assert cfg.d_model % cfg.n_heads == 0
        self.n_heads = cfg.n_heads
        self.head_dim = cfg.d_model // cfg.n_heads
        self.qkv = nn.Linear(cfg.d_model, 3 * cfg.d_model, bias=False)
        self.output = nn.Linear(cfg.d_model, cfg.d_model, bias=False)
        self.dropout = cfg.dropout

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        batch, length, width = x.shape
        q, k, v = self.qkv(x).chunk(3, dim=-1)
        q = q.view(batch, length, self.n_heads, self.head_dim).transpose(1, 2)
        k = k.view(batch, length, self.n_heads, self.head_dim).transpose(1, 2)
        v = v.view(batch, length, self.n_heads, self.head_dim).transpose(1, 2)
        output = F.scaled_dot_product_attention(
            q, k, v, is_causal=True, dropout_p=self.dropout if self.training else 0.0
        )
        return self.output(output.transpose(1, 2).contiguous().view(batch, length, width))


class SwiGLU(nn.Module):
    def __init__(self, cfg: ModelConfig):
        super().__init__()
        self.gate_up = nn.Linear(cfg.d_model, 2 * cfg.mlp_hidden, bias=False)
        self.down = nn.Linear(cfg.mlp_hidden, cfg.d_model, bias=False)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        gate, up = self.gate_up(x).chunk(2, dim=-1)
        return self.down(F.silu(gate) * up)


class TransformerBlock(nn.Module):
    def __init__(self, cfg: ModelConfig):
        super().__init__()
        self.attention_norm = RMSNorm(cfg.d_model)
        self.attention = Attention(cfg)
        self.mlp_norm = RMSNorm(cfg.d_model)
        self.mlp = SwiGLU(cfg)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = x + self.attention(self.attention_norm(x))
        return x + self.mlp(self.mlp_norm(x))


class GPT(nn.Module):
    """A dense ~50M model. Tied embeddings keep vocabulary capacity efficient."""

    def __init__(self, cfg: ModelConfig):
        super().__init__()
        self.cfg = cfg
        self.token_embedding = nn.Embedding(cfg.vocab_size, cfg.d_model)
        self.position_embedding = nn.Embedding(cfg.block_size, cfg.d_model)
        self.blocks = nn.ModuleList([TransformerBlock(cfg) for _ in range(cfg.n_layers)])
        self.final_norm = RMSNorm(cfg.d_model)
        self.lm_head = nn.Linear(cfg.d_model, cfg.vocab_size, bias=False)
        self.lm_head.weight = self.token_embedding.weight
        self.gradient_checkpointing = False
        self.apply(self._init_weights)

    @staticmethod
    def _init_weights(module: nn.Module) -> None:
        if isinstance(module, (nn.Linear, nn.Embedding)):
            nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(self, input_ids: torch.Tensor, targets: torch.Tensor | None = None):
        _, length = input_ids.shape
        if length > self.cfg.block_size:
            raise ValueError(f"Sequence length {length} exceeds {self.cfg.block_size}.")
        positions = torch.arange(length, device=input_ids.device)
        x = self.token_embedding(input_ids) + self.position_embedding(positions)
        checkpointing = self.training and self.gradient_checkpointing
        for block in self.blocks:
            x = torch.utils.checkpoint.checkpoint(block, x, use_reentrant=False) if checkpointing else block(x)
        x = self.final_norm(x)
        if targets is None:
            return self.lm_head(x[:, [-1]]), None

        # Do not allocate [batch, sequence, 32k] logits at once: this is VRAM-safe on 6GB.
        hidden = x[:, :-1].contiguous().view(-1, self.cfg.d_model)
        labels = targets[:, 1:].contiguous().view(-1)
        loss_sum = hidden.new_zeros((), dtype=torch.float32)
        for start in range(0, labels.numel(), 256):
            loss_sum = loss_sum + F.cross_entropy(
                self.lm_head(hidden[start : start + 256]).float(), labels[start : start + 256], reduction="sum"
            )
        return None, loss_sum / labels.numel()

    @torch.inference_mode()
    def generate(
        self,
        input_ids: torch.Tensor,
        max_new_tokens: int,
        temperature: float = 0.8,
        top_k: int = 50,
        eos_id: int | None = None,
    ):
        for _ in range(max_new_tokens):
            logits, _ = self(input_ids[:, -self.cfg.block_size :])
            logits = logits[:, -1] / max(temperature, 1e-5)
            if top_k > 0:
                threshold = torch.topk(logits, min(top_k, logits.size(-1))).values[:, [-1]]
                logits = logits.masked_fill(logits < threshold, float("-inf"))
            next_token = torch.multinomial(F.softmax(logits, dim=-1), 1)
            input_ids = torch.cat((input_ids, next_token), dim=1)
            if eos_id is not None and (next_token == eos_id).all():
                break
        return input_ids


def parameter_count(model: nn.Module) -> int:
    return sum(parameter.numel() for parameter in model.parameters())