File size: 10,153 Bytes
f6e3ff4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
258
259
260
261
262
263
264
265
#!/usr/bin/env python3
"""
leeknet_500m.py — Scaled TCF-1 architecture for ~500M params.

Same hybrid attention + Mamba SSM design as the 36M character-level model.
Differences:
  - BPE tokenizer (vocab 32k) instead of character-level
  - Wider: n_embed 1024 (vs 512)
  - Deeper: 12 hybrid pairs (vs 4)
  - Longer context: block_size 2048 (vs 512)
  - Persistent SSM state still threads through all pairs and across tokens

Architecture (per hybrid pair):
  Attention (reasons over context)
  + Mamba SSM (holds and updates persistent state)
  + FeedForward (transforms)

Usage:
  python3 leeknet_500m.py info             # show parameter count
  python3 leeknet_500m.py train_a          # Stage A pretraining
  python3 leeknet_500m.py train_b          # Stage B SFT
  python3 leeknet_500m.py train_c          # Stage C voice imprint
  python3 leeknet_500m.py chat             # interactive
"""

import math
import json
import sys
import time
from pathlib import Path

import mlx.core as mx
import mlx.nn as nn
import mlx.optimizers as optim
import mlx.utils as mlx_utils
import numpy as np
import sentencepiece as spm

# ---------------------------------------------------------------------------
# Paths
# ---------------------------------------------------------------------------
ROOT          = Path(__file__).parent
TOKENIZER_DIR = ROOT / 'tokenizer'
DATA_A        = ROOT / 'data' / 'A_knowledge'
DATA_B        = ROOT / 'data' / 'B_instruction'
VOICE_DIR     = ROOT / 'memory' / 'corpus'
CKPT_DIR      = ROOT / 'checkpoints_500m'
CKPT_DIR.mkdir(exist_ok=True)

TOKENIZER_MODEL = TOKENIZER_DIR / 'leek_bpe_32k.model'

# ---------------------------------------------------------------------------
# Config — scales from the 36M version
# ---------------------------------------------------------------------------
N_VOCAB     = 32000          # from BPE tokenizer
N_EMBED     = 1024           # was 512
N_HEAD      = 16             # was 8
N_PAIRS     = 12             # was 4
SSM_D_STATE = 16
SSM_D_CONV  = 4
SSM_EXPAND  = 2
DROPOUT     = 0.0            # disabled — relying on data diversity
BLOCK_SIZE  = 2048           # was 512

# Tools (still emitted as text — harness handles execution)
TOOLS = ['<none>', 'query_soul', 'bash', 'read_file', 'write_file', 'query_memory']

# Training defaults — adjust per stage
BATCH_SIZE   = 8
LEARN_RATE   = 3e-4
WARMUP_STEPS = 500
WEIGHT_DECAY = 0.1

# ---------------------------------------------------------------------------
# SSM block — Mamba-style selective state
# ---------------------------------------------------------------------------
class MambaBlock(nn.Module):
    def __init__(self, d_model, d_state=16, d_conv=4, expand=2):
        super().__init__()
        self.d_model = d_model
        self.d_state = d_state
        self.d_inner = int(expand * d_model)

        self.in_proj  = nn.Linear(d_model, self.d_inner * 2, bias=False)
        self.conv1d   = nn.Conv1d(
            in_channels=self.d_inner,
            out_channels=self.d_inner,
            kernel_size=d_conv,
            padding=d_conv - 1,
            bias=True,
        )
        self.x_proj   = nn.Linear(self.d_inner, d_state * 2 + 1, bias=False)
        self.dt_proj  = nn.Linear(1, self.d_inner, bias=True)
        self.out_proj = nn.Linear(self.d_inner, d_model, bias=False)
        self.norm     = nn.LayerNorm(d_model)

        A = np.arange(1, d_state + 1, dtype=np.float32)
        self.A_log = mx.array(np.log(A))
        self.D     = mx.ones(self.d_inner)

    def __call__(self, x, h_prev=None):
        B, T, D = x.shape
        x_in = self.norm(x)
        xz = self.in_proj(x_in)
        x_, z = xz[..., :self.d_inner], xz[..., self.d_inner:]

        x_conv = self.conv1d(x_)[:, :T, :]
        x_act  = mx.maximum(x_conv, 0) * mx.sigmoid(x_conv)  # silu-ish

        xproj = self.x_proj(x_act)
        dt = xproj[..., :1]
        B_  = xproj[..., 1:1+self.d_state]
        C   = xproj[..., 1+self.d_state:]

        delta = nn.softplus(self.dt_proj(dt))
        A = -mx.exp(self.A_log)

        # serial scan with persistent state
        h = h_prev if h_prev is not None else mx.zeros((B, self.d_inner, self.d_state))
        ys = []
        for t in range(T):
            dt_t = delta[:, t, :]                     # (B, d_inner)
            x_t  = x_act[:, t, :]                     # (B, d_inner)
            B_t  = B_[:, t, :]                        # (B, d_state)
            C_t  = C[:,  t, :]                        # (B, d_state)

            # discretize A and B per timestep
            dA = mx.exp(dt_t[:, :, None] * A[None, None, :])  # (B, d_inner, d_state)
            dB = dt_t[:, :, None] * B_t[:, None, :]            # (B, d_inner, d_state)

            # state update: h_t = dA * h_{t-1} + dB * x_t
            h = dA * h + dB * x_t[:, :, None]         # (B, d_inner, d_state)

            # output projection: y_t = sum_state(h_t * C_t)
            y = (h * C_t[:, None, :]).sum(axis=-1)    # (B, d_inner)
            ys.append(y[:, None, :])

        y_out = mx.concatenate(ys, axis=1)
        y_out = y_out + self.D * x_act
        y_out = y_out * mx.sigmoid(z)
        return x + self.out_proj(y_out), h

# ---------------------------------------------------------------------------
# Attention block
# ---------------------------------------------------------------------------
class AttentionBlock(nn.Module):
    def __init__(self, n_embed, n_head, dropout):
        super().__init__()
        assert n_embed % n_head == 0
        self.n_head = n_head
        self.head_dim = n_embed // n_head
        self.qkv  = nn.Linear(n_embed, 3 * n_embed, bias=False)
        self.proj = nn.Linear(n_embed, n_embed, bias=False)
        self.norm = nn.LayerNorm(n_embed)
        self.drop = nn.Dropout(dropout)

    def __call__(self, x):
        B, T, D = x.shape
        x_in = self.norm(x)
        qkv = self.qkv(x_in)
        qkv = qkv.reshape(B, T, 3, self.n_head, self.head_dim).transpose(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]
        scores = (q @ k.transpose(0, 1, 3, 2)) / math.sqrt(self.head_dim)
        mask = mx.tril(mx.ones((T, T))) == 0
        scores = mx.where(mask, -1e9, scores)
        attn   = mx.softmax(scores, axis=-1)
        out    = (attn @ v).transpose(0, 2, 1, 3).reshape(B, T, D)
        return x + self.drop(self.proj(out))

# ---------------------------------------------------------------------------
# FeedForward
# ---------------------------------------------------------------------------
class FeedForward(nn.Module):
    def __init__(self, n_embed, dropout):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(n_embed, 4 * n_embed, bias=False),
            nn.GELU(),
            nn.Linear(4 * n_embed, n_embed, bias=False),
            nn.Dropout(dropout),
        )
        self.norm = nn.LayerNorm(n_embed)

    def __call__(self, x):
        return x + self.net(self.norm(x))

# ---------------------------------------------------------------------------
# Hybrid pair: Attention + SSM + FFN
# ---------------------------------------------------------------------------
class HybridPair(nn.Module):
    def __init__(self, n_embed, n_head, dropout):
        super().__init__()
        self.attn = AttentionBlock(n_embed, n_head, dropout)
        self.ssm  = MambaBlock(n_embed, SSM_D_STATE, SSM_D_CONV, SSM_EXPAND)
        self.ff   = FeedForward(n_embed, dropout)

    def __call__(self, x, h=None):
        x      = self.attn(x)
        x, h   = self.ssm(x, h)
        x      = self.ff(x)
        return x, h

# ---------------------------------------------------------------------------
# LeekNet 500M
# ---------------------------------------------------------------------------
class LeekNet500M(nn.Module):
    def __init__(self, vocab_size=N_VOCAB, n_embed=N_EMBED, n_head=N_HEAD,
                 n_pairs=N_PAIRS, block_size=BLOCK_SIZE, dropout=DROPOUT):
        super().__init__()
        self.block_size = block_size
        self.tok_embed  = nn.Embedding(vocab_size, n_embed)
        self.pos_embed  = nn.Embedding(block_size, n_embed)
        self.drop       = nn.Dropout(dropout)
        self.pairs      = [HybridPair(n_embed, n_head, dropout) for _ in range(n_pairs)]
        self.ln_final   = nn.LayerNorm(n_embed)
        self.lm_head    = nn.Linear(n_embed, vocab_size, bias=False)

    def forward(self, idx, states=None):
        B, T = idx.shape
        pos  = mx.arange(T)
        x    = self.drop(self.tok_embed(idx) + self.pos_embed(pos))
        if states is None:
            states = [None] * len(self.pairs)
        new_states = []
        for pair, h in zip(self.pairs, states):
            x, h = pair(x, h)
            new_states.append(h)
        x = self.ln_final(x)
        return x, new_states

    def __call__(self, idx, n_think=1):
        states = None
        for _ in range(n_think):
            x, states = self.forward(idx, states)
        return self.lm_head(x)

# ---------------------------------------------------------------------------
# Quick sanity / param count
# ---------------------------------------------------------------------------
def info():
    model = LeekNet500M()
    n_params = sum(v.size for _, v in mlx_utils.tree_flatten(model.parameters()))
    print(f'\nLeekNet 500M:')
    print(f'  vocab:       {N_VOCAB:,}')
    print(f'  n_embed:     {N_EMBED}')
    print(f'  n_pairs:     {N_PAIRS}')
    print(f'  n_head:      {N_HEAD}')
    print(f'  block_size:  {BLOCK_SIZE}')
    print(f'  parameters:  {n_params/1e6:.1f}M')

    tok = spm.SentencePieceProcessor(model_file=str(TOKENIZER_MODEL))
    print(f'  tokenizer:   {TOKENIZER_MODEL.name}')
    print(f'  vocab_size:  {tok.vocab_size()}')

# ---------------------------------------------------------------------------
# Entry
# ---------------------------------------------------------------------------
if __name__ == '__main__':
    cmd = sys.argv[1] if len(sys.argv) > 1 else 'info'
    if cmd == 'info':
        info()
    else:
        print(f'training entry points (train_a/b/c) will be wired in next.')
        print(f'usage: python3 leeknet_500m.py info')