|
| 1 | +# Copyright © 2025 Apple Inc. |
| 2 | + |
| 3 | +from dataclasses import dataclass |
| 4 | +from typing import Any, Optional |
| 5 | + |
| 6 | +import mlx.core as mx |
| 7 | +import mlx.nn as nn |
| 8 | + |
| 9 | +from .base import BaseModelArgs, create_attention_mask, scaled_dot_product_attention |
| 10 | + |
| 11 | + |
| 12 | +@dataclass |
| 13 | +class ModelArgs(BaseModelArgs): |
| 14 | + model_type: str |
| 15 | + block_size: int |
| 16 | + layer_norm_eps: float |
| 17 | + n_embd: int |
| 18 | + n_head: int |
| 19 | + n_kv_heads: int |
| 20 | + n_layer: int |
| 21 | + rope_theta: float |
| 22 | + vocab_size: int |
| 23 | + tie_word_embeddings: bool = True |
| 24 | + |
| 25 | + |
| 26 | +class Lille130mAttention(nn.Module): |
| 27 | + def __init__(self, args: ModelArgs): |
| 28 | + super().__init__() |
| 29 | + self.n_head = args.n_head |
| 30 | + self.n_kv_heads = args.n_kv_heads |
| 31 | + self.head_dim = args.n_embd // args.n_head |
| 32 | + self.scale = self.head_dim**-0.5 |
| 33 | + |
| 34 | + self.qkv_proj = nn.Linear( |
| 35 | + args.n_embd, (args.n_head + 2 * args.n_kv_heads) * self.head_dim, bias=False |
| 36 | + ) |
| 37 | + self.out_proj = nn.Linear(args.n_head * self.head_dim, args.n_embd, bias=False) |
| 38 | + |
| 39 | + self.norm = nn.RMSNorm(args.n_embd, eps=args.layer_norm_eps) |
| 40 | + |
| 41 | + self.rope = nn.RoPE(args.n_embd // args.n_head, True, args.rope_theta) |
| 42 | + |
| 43 | + def __call__( |
| 44 | + self, |
| 45 | + x: mx.array, |
| 46 | + mask: Optional[mx.array] = None, |
| 47 | + cache: Optional[Any] = None, |
| 48 | + ) -> mx.array: |
| 49 | + B, L, D = x.shape |
| 50 | + |
| 51 | + qkv = self.qkv_proj(self.norm(x)) |
| 52 | + |
| 53 | + q_size = self.n_head * self.head_dim |
| 54 | + kv_size = self.n_kv_heads * self.head_dim |
| 55 | + |
| 56 | + queries, keys, values = mx.split(qkv, [q_size, q_size + kv_size], axis=-1) |
| 57 | + |
| 58 | + queries = queries.reshape(B, L, self.n_head, -1).transpose(0, 2, 1, 3) |
| 59 | + keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3) |
| 60 | + values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3) |
| 61 | + |
| 62 | + if cache is not None: |
| 63 | + queries = self.rope(queries, offset=cache.offset) |
| 64 | + keys = self.rope(keys, offset=cache.offset) |
| 65 | + keys, values = cache.update_and_fetch(keys, values) |
| 66 | + else: |
| 67 | + queries = self.rope(queries) |
| 68 | + keys = self.rope(keys) |
| 69 | + |
| 70 | + output = scaled_dot_product_attention( |
| 71 | + queries, keys, values, cache=cache, scale=self.scale, mask=mask |
| 72 | + ) |
| 73 | + |
| 74 | + output = output.transpose(0, 2, 1, 3).reshape(B, L, -1) |
| 75 | + return self.out_proj(output) |
| 76 | + |
| 77 | + |
| 78 | +class Lille130mMLP(nn.Module): |
| 79 | + def __init__(self, args: ModelArgs): |
| 80 | + super().__init__() |
| 81 | + hidden_dim = 256 * round(int(8 * args.n_embd / 3) / 256) |
| 82 | + |
| 83 | + self.norm = nn.RMSNorm(args.n_embd, eps=args.layer_norm_eps) |
| 84 | + self.gate_proj = nn.Linear(args.n_embd, hidden_dim, bias=False) |
| 85 | + self.up_proj = nn.Linear(args.n_embd, hidden_dim, bias=False) |
| 86 | + self.down_proj = nn.Linear(hidden_dim, args.n_embd, bias=False) |
| 87 | + |
| 88 | + def __call__(self, x: mx.array) -> mx.array: |
| 89 | + h = self.norm(x) |
| 90 | + return self.down_proj(nn.silu(self.gate_proj(h)) * self.up_proj(h)) |
| 91 | + |
| 92 | + |
| 93 | +class Lille130Block(nn.Module): |
| 94 | + def __init__(self, args: ModelArgs): |
| 95 | + super().__init__() |
| 96 | + self.attention = Lille130mAttention(args) |
| 97 | + self.feed_forward = Lille130mMLP(args) |
| 98 | + |
| 99 | + def __call__( |
| 100 | + self, |
| 101 | + x: mx.array, |
| 102 | + mask: Optional[mx.array] = None, |
| 103 | + cache: Optional[Any] = None, |
| 104 | + ) -> mx.array: |
| 105 | + h = x + self.attention(x, mask, cache) |
| 106 | + out = h + self.feed_forward(h) |
| 107 | + return out |
| 108 | + |
| 109 | + |
| 110 | +class Lille130(nn.Module): |
| 111 | + def __init__(self, args: ModelArgs): |
| 112 | + super().__init__() |
| 113 | + self.tok_embeddings = nn.Embedding(args.vocab_size, args.n_embd) |
| 114 | + self.layers = [Lille130Block(args=args) for _ in range(args.n_layer)] |
| 115 | + self.norm = nn.RMSNorm(args.n_embd, eps=args.layer_norm_eps) |
| 116 | + |
| 117 | + def __call__( |
| 118 | + self, |
| 119 | + inputs: mx.array, |
| 120 | + cache: Optional[Any] = None, |
| 121 | + ) -> mx.array: |
| 122 | + h = self.tok_embeddings(inputs) |
| 123 | + |
| 124 | + if cache is None: |
| 125 | + cache = [None] * len(self.layers) |
| 126 | + |
| 127 | + mask = create_attention_mask(h, cache[0]) |
| 128 | + |
| 129 | + for layer, c in zip(self.layers, cache): |
| 130 | + h = layer(h, mask, cache=c) |
| 131 | + |
| 132 | + return self.tok_embeddings.as_linear(self.norm(h)) |
| 133 | + |
| 134 | + |
| 135 | +class Model(nn.Module): |
| 136 | + def __init__(self, args: ModelArgs): |
| 137 | + super().__init__() |
| 138 | + self.args = args |
| 139 | + self.model_type = args.model_type |
| 140 | + self.transformer = Lille130(args) |
| 141 | + |
| 142 | + def __call__( |
| 143 | + self, |
| 144 | + inputs: mx.array, |
| 145 | + cache: Optional[Any] = None, |
| 146 | + ) -> mx.array: |
| 147 | + return self.transformer(inputs, cache=cache) |
| 148 | + |
| 149 | + @property |
| 150 | + def layers(self): |
| 151 | + return self.transformer.layers |
| 152 | + |
| 153 | + def sanitize(self, weights): |
| 154 | + return {k: v for k, v in weights.items() if "rotary_emb" not in k} |
0 commit comments