jonmabe/tiny-llm-cli-sft

1"""2Tiny Transformer with modern components:3- RoPE (Rotary Position Embeddings)4- RMSNorm5- SwiGLU activation6- Weight tying7"""8import torch9import torch.nn as nn10import torch.nn.functional as F11import math12 13class RMSNorm(nn.Module):14    def __init__(self, dim: int, eps: float = 1e-6):15        super().__init__()16        self.eps = eps17        self.weight = nn.Parameter(torch.ones(dim))18 19    def forward(self, x):20        norm = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)21        return x * norm * self.weight22 23 24class RotaryEmbedding(nn.Module):25    def __init__(self, dim: int, max_seq_len: int = 512, base: int = 10000):26        super().__init__()27        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))28        self.register_buffer("inv_freq", inv_freq)29        self.max_seq_len = max_seq_len30        31    def forward(self, x, seq_len: int):32        t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq)33        freqs = torch.einsum("i,j->ij", t, self.inv_freq)34        emb = torch.cat((freqs, freqs), dim=-1)35        return emb.cos(), emb.sin()36 37 38def rotate_half(x):39    x1, x2 = x.chunk(2, dim=-1)40    return torch.cat((-x2, x1), dim=-1)41 42 43def apply_rotary_pos_emb(q, k, cos, sin):44    cos = cos.unsqueeze(0).unsqueeze(0)  # [1, 1, seq_len, dim]45    sin = sin.unsqueeze(0).unsqueeze(0)46    q_embed = (q * cos) + (rotate_half(q) * sin)47    k_embed = (k * cos) + (rotate_half(k) * sin)48    return q_embed, k_embed49 50 51class SwiGLU(nn.Module):52    def __init__(self, hidden_size: int, intermediate_size: int):53        super().__init__()54        self.w1 = nn.Linear(hidden_size, intermediate_size, bias=False)55        self.w2 = nn.Linear(intermediate_size, hidden_size, bias=False)56        self.w3 = nn.Linear(hidden_size, intermediate_size, bias=False)57 58    def forward(self, x):59        return self.w2(F.silu(self.w1(x)) * self.w3(x))60 61 62class Attention(nn.Module):63    def __init__(self, hidden_size: int, num_heads: int, dropout: float = 0.0):64        super().__init__()65        self.num_heads = num_heads66        self.head_dim = hidden_size // num_heads67        68        self.q_proj = nn.Linear(hidden_size, hidden_size, bias=False)69        self.k_proj = nn.Linear(hidden_size, hidden_size, bias=False)70        self.v_proj = nn.Linear(hidden_size, hidden_size, bias=False)71        self.o_proj = nn.Linear(hidden_size, hidden_size, bias=False)72        73        self.rotary = RotaryEmbedding(self.head_dim)74        self.dropout = nn.Dropout(dropout)75 76    def forward(self, x, mask=None):77        B, T, C = x.shape78        79        q = self.q_proj(x).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)80        k = self.k_proj(x).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)81        v = self.v_proj(x).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)82        83        cos, sin = self.rotary(x, T)84        q, k = apply_rotary_pos_emb(q, k, cos, sin)85        86        # Scaled dot-product attention87        scale = 1.0 / math.sqrt(self.head_dim)88        attn = torch.matmul(q, k.transpose(-2, -1)) * scale89        90        if mask is not None:91            attn = attn.masked_fill(mask == 0, float('-inf'))92        93        attn = F.softmax(attn, dim=-1)94        attn = self.dropout(attn)95        96        out = torch.matmul(attn, v)97        out = out.transpose(1, 2).contiguous().view(B, T, C)98        return self.o_proj(out)99 100 101class TransformerBlock(nn.Module):102    def __init__(self, hidden_size: int, num_heads: int, intermediate_size: int, dropout: float = 0.0):103        super().__init__()104        self.norm1 = RMSNorm(hidden_size)105        self.attn = Attention(hidden_size, num_heads, dropout)106        self.norm2 = RMSNorm(hidden_size)107        self.ffn = SwiGLU(hidden_size, intermediate_size)108 109    def forward(self, x, mask=None):110        x = x + self.attn(self.norm1(x), mask)111        x = x + self.ffn(self.norm2(x))112        return x113 114 115class TinyLLM(nn.Module):116    def __init__(117        self,118        vocab_size: int = 32000,119        hidden_size: int = 512,120        num_layers: int = 12,121        num_heads: int = 8,122        intermediate_size: int = 1408,123        max_position_embeddings: int = 512,124        dropout: float = 0.0,125        tie_weights: bool = True,126    ):127        super().__init__()128        self.vocab_size = vocab_size129        self.hidden_size = hidden_size130        131        self.embed_tokens = nn.Embedding(vocab_size, hidden_size)132        self.layers = nn.ModuleList([133            TransformerBlock(hidden_size, num_heads, intermediate_size, dropout)134            for _ in range(num_layers)135        ])136        self.norm = RMSNorm(hidden_size)137        self.lm_head = nn.Linear(hidden_size, vocab_size, bias=False)138        139        if tie_weights:140            self.lm_head.weight = self.embed_tokens.weight141        142        # Causal mask143        self.register_buffer(144            "causal_mask",145            torch.tril(torch.ones(max_position_embeddings, max_position_embeddings))146        )147        148        self._init_weights()149    150    def _init_weights(self):151        for module in self.modules():152            if isinstance(module, nn.Linear):153                torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)154            elif isinstance(module, nn.Embedding):155                torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)156    157    def forward(self, input_ids, labels=None):158        B, T = input_ids.shape159        160        x = self.embed_tokens(input_ids)161        mask = self.causal_mask[:T, :T]162        163        for layer in self.layers:164            x = layer(x, mask)165        166        x = self.norm(x)167        logits = self.lm_head(x)168        169        loss = None170        if labels is not None:171            shift_logits = logits[..., :-1, :].contiguous()172            shift_labels = labels[..., 1:].contiguous()173            loss = F.cross_entropy(174                shift_logits.view(-1, self.vocab_size),175                shift_labels.view(-1),176                ignore_index=-100177            )178        179        return {"loss": loss, "logits": logits}180    181    def count_parameters(self):182        return sum(p.numel() for p in self.parameters())183 184 185if __name__ == "__main__":186    # Test model187    model = TinyLLM()188    print(f"Parameters: {model.count_parameters() / 1e6:.2f}M")189    190    x = torch.randint(0, 32000, (2, 128))191    out = model(x, labels=x)192    print(f"Loss: {out['loss'].item():.4f}")193    print(f"Logits shape: {out['logits'].shape}")194