File size: 10,510 Bytes
8932cec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
"""Standalone MLX model definition shipped with checkpoints."""

import math
from dataclasses import dataclass

import mlx.core as mx
import mlx.nn as nn
from mlx.utils import tree_flatten


@dataclass
class GPTConfig:
    sequence_len: int = 256
    vocab_size: int = 2075
    n_layer: int = 4
    n_head: int = 2
    n_kv_head: int = 2
    n_embd: int = 256
    window_pattern: str = "SSSL"


class ChessTokenizer:
    """Minimal exact-token tokenizer for metadata and atomic UCI move tokens."""

    def __init__(self, vocabulary):
        self.id_to_token = vocabulary["id_to_token"]
        self.token_to_id = {token: index for index, token in enumerate(self.id_to_token)}
        self.pad_token_id = self.token_to_id["PAD"]
        self.bos_token_id = self.token_to_id["BOS"]
        self.eos_token_id = self.token_to_id["EOS"]

    @classmethod
    def from_pretrained(cls, checkpoint_dir):
        import json
        from pathlib import Path

        vocabulary = json.loads((Path(checkpoint_dir) / "vocab.json").read_text())
        return cls(vocabulary)

    def encode_tokens(self, tokens):
        return [self.token_to_id[token] for token in tokens]

    def decode_ids(self, token_ids):
        return [self.id_to_token[int(token_id)] for token_id in token_ids]


def norm(x):
    return x * mx.rsqrt(mx.mean(x * x, axis=-1, keepdims=True) + 1e-5)


def has_ve(layer_idx, n_layer):
    return layer_idx % 2 == (n_layer - 1) % 2


def create_additive_causal_mask(seq_len, dtype=mx.float32):
    indices = mx.arange(seq_len)
    blocked = indices[None, :] > indices[:, None]
    return mx.where(blocked, mx.array(float("-inf"), dtype=dtype), mx.array(0.0, dtype=dtype))


def create_sliding_window_mask(seq_len, window_size, dtype=mx.float32):
    indices = mx.arange(seq_len)
    causal = indices[None, :] > indices[:, None]
    too_far = (indices[:, None] - indices[None, :]) >= window_size
    blocked = causal | too_far
    return mx.where(blocked, mx.array(float("-inf"), dtype=dtype), mx.array(0.0, dtype=dtype))


class CausalSelfAttention(nn.Module):
    def __init__(self, config, layer_idx):
        super().__init__()
        self.n_head = config.n_head
        self.n_kv_head = config.n_kv_head
        self.n_embd = config.n_embd
        self.head_dim = self.n_embd // self.n_head
        assert self.n_embd % self.n_head == 0
        assert self.n_kv_head <= self.n_head and self.n_head % self.n_kv_head == 0
        self.c_q = nn.Linear(self.n_embd, self.n_head * self.head_dim, bias=False)
        self.c_k = nn.Linear(self.n_embd, self.n_kv_head * self.head_dim, bias=False)
        self.c_v = nn.Linear(self.n_embd, self.n_kv_head * self.head_dim, bias=False)
        self.c_proj = nn.Linear(self.n_embd, self.n_embd, bias=False)
        self.ve_gate_channels = 32
        self.ve_gate = (
            nn.Linear(self.ve_gate_channels, self.n_kv_head, bias=False)
            if has_ve(layer_idx, config.n_layer)
            else None
        )
        self.rope = nn.RoPE(self.head_dim, traditional=True, base=10000)

    def __call__(self, x, ve, mask):
        batch_size, seq_len, _ = x.shape
        q = self.c_q(x).reshape(batch_size, seq_len, self.n_head, self.head_dim)
        k = self.c_k(x).reshape(batch_size, seq_len, self.n_kv_head, self.head_dim)
        v = self.c_v(x).reshape(batch_size, seq_len, self.n_kv_head, self.head_dim)

        if ve is not None and self.ve_gate is not None:
            ve = ve.reshape(batch_size, seq_len, self.n_kv_head, self.head_dim)
            gate = 2 * mx.sigmoid(self.ve_gate(x[..., : self.ve_gate_channels]))
            v = v + mx.expand_dims(gate, axis=-1) * ve

        q = q.transpose(0, 2, 1, 3)
        k = k.transpose(0, 2, 1, 3)
        v = v.transpose(0, 2, 1, 3)
        q = norm(self.rope(q))
        k = norm(self.rope(k))
        scale = 1.0 / math.sqrt(self.head_dim)
        y = mx.fast.scaled_dot_product_attention(q, k, v, scale=scale, mask=mask)
        y = y.transpose(0, 2, 1, 3).reshape(batch_size, seq_len, -1)
        return 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=False)
        self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=False)

    def __call__(self, x):
        x = self.c_fc(x)
        x = mx.maximum(x, 0) ** 2
        return self.c_proj(x)


class Block(nn.Module):
    def __init__(self, config, layer_idx):
        super().__init__()
        self.attn = CausalSelfAttention(config, layer_idx)
        self.mlp = MLP(config)

    def __call__(self, x, ve, mask):
        x = x + self.attn(norm(x), ve, mask)
        x = x + self.mlp(norm(x))
        return x


class GPT(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.window_sizes = self._compute_window_sizes(config)
        self.wte = nn.Embedding(config.vocab_size, config.n_embd)
        self.blocks = [Block(config, i) for i in range(config.n_layer)]
        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
        self.resid_lambdas = mx.ones((config.n_layer,), dtype=mx.float32)
        self.x0_lambdas = mx.zeros((config.n_layer,), dtype=mx.float32)
        head_dim = config.n_embd // config.n_head
        kv_dim = config.n_kv_head * head_dim
        self.value_embeds = {
            str(i): nn.Embedding(config.vocab_size, kv_dim)
            for i in range(config.n_layer)
            if has_ve(i, config.n_layer)
        }
        self._mask_cache = {}

    def init_weights(self):
        n_embd = self.config.n_embd
        scale = 3**0.5 * n_embd**-0.5
        self.wte.weight = mx.random.normal(self.wte.weight.shape).astype(mx.bfloat16)
        self.lm_head.weight = (mx.random.normal(self.lm_head.weight.shape) * 0.001).astype(mx.bfloat16)
        for block in self.blocks:
            block.attn.c_q.weight = mx.random.uniform(-scale, scale, block.attn.c_q.weight.shape).astype(mx.bfloat16)
            block.attn.c_k.weight = mx.random.uniform(-scale, scale, block.attn.c_k.weight.shape).astype(mx.bfloat16)
            block.attn.c_v.weight = mx.random.uniform(-scale, scale, block.attn.c_v.weight.shape).astype(mx.bfloat16)
            block.attn.c_proj.weight = mx.zeros_like(block.attn.c_proj.weight).astype(mx.bfloat16)
            block.mlp.c_fc.weight = mx.random.uniform(-scale, scale, block.mlp.c_fc.weight.shape).astype(mx.bfloat16)
            block.mlp.c_proj.weight = mx.zeros_like(block.mlp.c_proj.weight).astype(mx.bfloat16)
            if block.attn.ve_gate is not None:
                block.attn.ve_gate.weight = mx.zeros_like(block.attn.ve_gate.weight).astype(mx.bfloat16)
        self.resid_lambdas = mx.ones((self.config.n_layer,), dtype=mx.float32)
        self.x0_lambdas = mx.full((self.config.n_layer,), 0.1, dtype=mx.float32)
        for ve in self.value_embeds.values():
            ve.weight = mx.random.uniform(-scale, scale, ve.weight.shape).astype(mx.bfloat16)

    def _compute_window_sizes(self, config):
        pattern = config.window_pattern.upper()
        assert all(char in "SL" for char in pattern)
        mapping = {"L": config.sequence_len, "S": config.sequence_len // 2}
        windows = [mapping[pattern[i % len(pattern)]] for i in range(config.n_layer)]
        windows[-1] = config.sequence_len
        return windows

    def _get_masks(self, seq_len):
        dtype = self.wte.weight.dtype
        for window_size in set(self.window_sizes):
            key = (seq_len, window_size, dtype)
            if key not in self._mask_cache:
                self._mask_cache[key] = (
                    create_additive_causal_mask(seq_len, dtype)
                    if window_size >= seq_len
                    else create_sliding_window_mask(seq_len, window_size, dtype)
                )
        return [self._mask_cache[(seq_len, size, dtype)] for size in self.window_sizes]

    def __call__(self, idx, targets=None, reduction="mean"):
        _, seq_len = idx.shape
        masks = self._get_masks(seq_len)
        x = norm(self.wte(idx))
        x0 = x
        for i, block in enumerate(self.blocks):
            x = self.resid_lambdas[i].astype(x.dtype) * x + self.x0_lambdas[i].astype(x.dtype) * x0
            ve = self.value_embeds[str(i)](idx) if str(i) in self.value_embeds else None
            x = block(x, ve, masks[i])
        logits = self.lm_head(norm(x)).astype(mx.float32)
        logits = 15.0 * mx.tanh(logits / 15.0)
        if targets is None:
            return logits
        valid = targets != -1
        safe_targets = mx.where(valid, targets, mx.zeros_like(targets))
        ce = nn.losses.cross_entropy(logits, safe_targets, reduction="none") * valid
        if reduction == "none":
            return ce
        return mx.sum(ce) / mx.maximum(mx.sum(valid), 1)


def load_model(checkpoint_dir):
    """Load a published MLX checkpoint and its configuration."""
    import json
    from pathlib import Path

    checkpoint_dir = Path(checkpoint_dir)
    raw = json.loads((checkpoint_dir / "config.json").read_text())
    config = GPTConfig(**raw["model_config"])
    model = GPT(config)
    load_weights(model, checkpoint_dir / "model.safetensors")
    mx.eval(model.parameters())
    return model


def load_weights(model, weights_path):
    """Load flattened weights without MLX confusing numeric dict keys for list indices."""
    current = dict(tree_flatten(model.parameters()))
    loaded = mx.load(str(weights_path))
    if current.keys() != loaded.keys():
        missing = sorted(current.keys() - loaded.keys())
        extra = sorted(loaded.keys() - current.keys())
        raise ValueError(f"Checkpoint key mismatch; missing={missing}, extra={extra}")
    for path, value in loaded.items():
        if current[path].shape != value.shape:
            raise ValueError(f"Shape mismatch for {path}: {value.shape} != {current[path].shape}")
        parts = path.split(".")
        target = model
        for part in parts[:-1]:
            if isinstance(target, list):
                target = target[int(part)]
            elif isinstance(target, dict):
                target = target[part]
            else:
                target = getattr(target, part)
        final = parts[-1]
        if isinstance(target, list):
            target[int(final)] = value
        elif isinstance(target, dict):
            target[final] = value
        else:
            setattr(target, final, value)
    return model