model.py

"""
model.py -- standalone architecture definition for GTM-3-base.

This is a plain PyTorch nanoGPT-style GPT model with RoPE (rotary position
embeddings), NOT a HuggingFace `transformers` AutoModel. To load the
released weights:

    pip install torch safetensors tiktoken

    import json, torch
    from safetensors.torch import load_file
    from model import GPT, GPTConfig

    with open("config.json") as f:
        config = GPTConfig(**json.load(f))
    model = GPT(config)
    state_dict = load_file("model.safetensors")
    model.load_state_dict(state_dict)
    model.eval()

    import tiktoken
    enc = tiktoken.get_encoding("gpt2")
    ids = enc.encode_ordinary("Once upon a time,")
    x = torch.tensor([ids], dtype=torch.long)
    out = model.generate(x, max_new_tokens=100, temperature=0.8, top_k=50,
                          eot_token=enc.eot_token, repetition_penalty=1.3)
    print(enc.decode(out[0].tolist()))
"""

import math
from dataclasses import dataclass

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


@dataclass
class GPTConfig:
    vocab_size: int = 50257
    block_size: int = 1024
    n_layer: int = 10
    n_head: int = 8
    n_embd: int = 608
    dropout: float = 0.0
    bias: bool = True
    rope_theta: float = 10000.0  # standard RoPE base frequency


def precompute_rope_freqs(head_dim, max_seq_len, theta=10000.0, device="cpu"):
    """Precompute the complex rotation frequencies used by RoPE, one pair
    per (position, frequency-band). Standard formula: theta_i = theta^(-2i/dim)."""
    assert head_dim % 2 == 0, "RoPE requires an even head_dim"
    freqs = 1.0 / (theta ** (torch.arange(0, head_dim, 2, device=device).float() / head_dim))
    t = torch.arange(max_seq_len, device=device).float()
    freqs = torch.outer(t, freqs)  # (max_seq_len, head_dim/2)
    return torch.polar(torch.ones_like(freqs), freqs)  # complex64, (max_seq_len, head_dim/2)


def apply_rope(x, freqs_cis):
    """Apply rotary position embeddings to a (B, n_head, T, head_dim) tensor."""
    B, n_head, T, head_dim = x.shape
    x_complex = torch.view_as_complex(x.float().reshape(B, n_head, T, head_dim // 2, 2))
    freqs_cis = freqs_cis[:T].view(1, 1, T, head_dim // 2)
    x_rotated = x_complex * freqs_cis
    x_out = torch.view_as_real(x_rotated).reshape(B, n_head, T, head_dim)
    return x_out.type_as(x)


class CausalSelfAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        assert config.n_embd % config.n_head == 0
        self.n_head = config.n_head
        self.n_embd = config.n_embd
        self.head_dim = config.n_embd // config.n_head
        self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
        self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
        self.attn_dropout = nn.Dropout(config.dropout)
        self.resid_dropout = nn.Dropout(config.dropout)
        self.dropout = config.dropout

    def forward(self, x, freqs_cis):
        B, T, C = x.shape
        q, k, v = self.c_attn(x).split(self.n_embd, dim=2)
        q = q.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)
        k = k.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)
        v = v.view(B, T, self.n_head, C // self.n_head).transpose(1, 2)
        # RoPE is applied to queries and keys only, not values -- this is what
        # makes attention scores depend on *relative* position between tokens
        q = apply_rope(q, freqs_cis)
        k = apply_rope(k, freqs_cis)
        y = F.scaled_dot_product_attention(
            q, k, v, is_causal=True,
            dropout_p=self.dropout if self.training else 0.0,
        )
        y = y.transpose(1, 2).contiguous().view(B, T, C)
        return self.resid_dropout(self.c_proj(y))


class MLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
        self.gelu = nn.GELU()
        self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
        self.dropout = nn.Dropout(config.dropout)

    def forward(self, x):
        return self.dropout(self.c_proj(self.gelu(self.c_fc(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, freqs_cis):
        x = x + self.attn(self.ln_1(x), freqs_cis)
        x = x + self.mlp(self.ln_2(x))
        return x


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),
            # NOTE: no wpe (learned position embedding) -- RoPE replaces it entirely,
            # applied inside attention rather than added to the input embeddings
            drop=nn.Dropout(config.dropout),
            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)
        self.transformer.wte.weight = self.lm_head.weight
        head_dim = config.n_embd // config.n_head
        freqs_cis = precompute_rope_freqs(head_dim, config.block_size, theta=config.rope_theta)
        self.register_buffer("freqs_cis", freqs_cis, persistent=False)
        self.apply(self._init_weights)
        for pn, p in self.named_parameters():
            if pn.endswith("c_proj.weight"):
                nn.init.normal_(p, mean=0.0, std=0.02 / math.sqrt(2 * config.n_layer))

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            nn.init.normal_(module.weight, mean=0.0, std=0.02)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
        elif isinstance(module, nn.Embedding):
            nn.init.normal_(module.weight, mean=0.0, std=0.02)

    def forward(self, idx, targets=None):
        B, T = idx.shape
        assert T <= self.config.block_size, "sequence longer than block_size"
        x = self.transformer.drop(self.transformer.wte(idx))
        freqs_cis = self.freqs_cis.to(x.device)
        for block in self.transformer.h:
            x = block(x, freqs_cis)
        x = self.transformer.ln_f(x)
        logits = self.lm_head(x)
        loss = None
        if targets is not None:
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
        return logits, loss

    @torch.no_grad()
    def generate(self, idx, max_new_tokens, temperature=1.0, top_k=None, eot_token=None,
                 repetition_penalty=1.0):
        for _ in range(max_new_tokens):
            idx_cond = idx if idx.size(1) <= self.config.block_size else idx[:, -self.config.block_size:]
            logits, _ = self(idx_cond)
            logits = logits[:, -1, :] / temperature
            if repetition_penalty != 1.0:
                for seen_id in set(idx[0].tolist()):
                    if logits[0, seen_id] > 0:
                        logits[0, seen_id] /= repetition_penalty
                    else:
                        logits[0, seen_id] *= repetition_penalty
            if top_k is not None:
                v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
                logits[logits < v[:, [-1]]] = float("-inf")
            probs = F.softmax(logits, dim=-1)
            idx_next = torch.multinomial(probs, num_samples=1)
            idx = torch.cat((idx, idx_next), dim=1)
            if eot_token is not None and idx_next.item() == eot_token:
                break
        return idx