File size: 10,489 Bytes
0e3d4b8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
"""Neural network layers for SplitBit LLM — pure NumPy implementation.

Layers:
- Embedding (token ID → dense vector)
- Multi-head self-attention with RoPE
- Feed-forward network (MLP with GELU)
- Layer normalization (pre-norm)
- KV cache for fast autoregressive generation

All weights stored as NumPy arrays, quantized via SplitBitQuantizer.
"""

from __future__ import annotations

import logging
import math
from typing import Any

import numpy as np

from .quantization import SplitBitQuantizer

logger = logging.getLogger(__name__)


def gelu(x: np.ndarray) -> np.ndarray:
    """GELU activation — Gaussian Error Linear Unit."""
    return 0.5 * x * (1.0 + np.tanh(math.sqrt(2.0 / math.pi) * (x + 0.044715 * x ** 3)))


def softmax(x: np.ndarray, axis: int = -1) -> np.ndarray:
    """Numerically stable softmax."""
    x_max = np.max(x, axis=axis, keepdims=True)
    exp_x = np.exp(x - x_max)
    return exp_x / np.sum(exp_x, axis=axis, keepdims=True)


def layer_norm(x: np.ndarray, gamma: np.ndarray, beta: np.ndarray, eps: float = 1e-5) -> np.ndarray:
    """Layer normalization."""
    mean = np.mean(x, axis=-1, keepdims=True)
    var = np.var(x, axis=-1, keepdims=True)
    return gamma * (x - mean) / np.sqrt(var + eps) + beta


def rope(pos: np.ndarray, d_head: int, base: float = 10000.0) -> tuple[np.ndarray, np.ndarray]:
    """Rotary Position Embedding (RoPE).

    Returns cos and sin tensors for rotating Q and K.
    """
    inv_freq = 1.0 / (base ** (np.arange(0, d_head, 2) / d_head))
    # pos: [seq_len], inv_freq: [d_head/2]
    freqs = np.outer(pos, inv_freq)  # [seq_len, d_head/2]
    cos = np.cos(freqs)
    sin = np.sin(freqs)
    # Repeat to match d_head
    cos = np.repeat(cos, 2, axis=-1)  # [seq_len, d_head]
    sin = np.repeat(sin, 2, axis=-1)
    return cos, sin


def apply_rope(x: np.ndarray, cos: np.ndarray, sin: np.ndarray) -> np.ndarray:
    """Apply rotary embedding to tensor x.

    x: [batch, n_heads, seq_len, d_head]
    cos/sin: [seq_len, d_head]
    """
    x1 = x[..., 0::2]  # even indices
    x2 = x[..., 1::2]  # odd indices
    # Rotate
    cos = cos[None, None, :, :]  # [1, 1, seq_len, d_head]
    sin = sin[None, None, :, :]
    rotated = np.empty_like(x)
    rotated[..., 0::2] = x1 * cos[..., 0::2] - x2 * sin[..., 0::2]
    rotated[..., 1::2] = x1 * sin[..., 1::2] + x2 * cos[..., 1::2]
    return rotated


class Embedding:
    """Token embedding layer."""

    def __init__(self, vocab_size: int, d_model: int) -> None:
        # Xavier/Glorot initialization
        std = math.sqrt(2.0 / (vocab_size + d_model))
        self.weight = np.random.randn(vocab_size, d_model).astype(np.float32) * std
        self.d_model = d_model
        self.vocab_size = vocab_size

    def forward(self, token_ids: np.ndarray) -> np.ndarray:
        """token_ids: [batch, seq_len] → [batch, seq_len, d_model]"""
        return self.weight[token_ids]

    def backward(self, grad: np.ndarray, token_ids: np.ndarray) -> np.ndarray:
        """Gradient w.r.t. embedding weights."""
        grad_weight = np.zeros_like(self.weight)
        np.add.at(grad_weight, token_ids, grad)
        return grad_weight


class Linear:
    """Linear layer: y = x @ W^T + b, with SplitBit quantization support."""

    def __init__(self, in_features: int, out_features: int, bias: bool = True) -> None:
        std = math.sqrt(2.0 / (in_features + out_features))
        self.weight = np.random.randn(out_features, in_features).astype(np.float32) * std
        self.bias = np.zeros(out_features, dtype=np.float32) if bias else None
        self.in_features = in_features
        self.out_features = out_features
        self.use_bias = bias
        self._quantized = None

    def quantize(self, quantizer: SplitBitQuantizer) -> None:
        """Quantize weights for storage/inference."""
        self._quantized = {
            "weight": quantizer.quantize(self.weight),
            "bias": self.bias.copy() if self.bias is not None else None,
        }

    def dequantize(self) -> None:
        """Restore full-precision weights."""
        self._quantized = None

    def forward(self, x: np.ndarray) -> np.ndarray:
        """x: [..., in_features] → [..., out_features]"""
        w = self.weight
        out = x @ w.T
        if self.bias is not None:
            out = out + self.bias
        return out

    def forward_with_cache(self, x: np.ndarray, kv_cache: dict | None = None, layer_idx: int = 0,
                           is_kv: bool = False) -> np.ndarray:
        """Forward pass that optionally uses/appends to KV cache."""
        return self.forward(x)


class MultiHeadAttention:
    """Multi-head self-attention with RoPE and KV cache."""

    def __init__(self, d_model: int, n_heads: int, max_seq_len: int = 512) -> None:
        self.d_model = d_model
        self.n_heads = n_heads
        self.d_head = d_model // n_heads
        self.max_seq_len = max_seq_len

        self.wq = Linear(d_model, d_model, bias=False)
        self.wk = Linear(d_model, d_model, bias=False)
        self.wv = Linear(d_model, d_model, bias=False)
        self.wo = Linear(d_model, d_model, bias=False)

        # Precompute RoPE
        pos = np.arange(max_seq_len, dtype=np.float32)
        self._cos, self._sin = rope(pos, self.d_head)

        # KV cache: {layer_idx: (k, v)}
        self._kv_cache: dict[int, tuple[np.ndarray, np.ndarray]] = {}

    def forward(
        self,
        x: np.ndarray,
        layer_idx: int = 0,
        use_cache: bool = False,
        past_len: int = 0,
    ) -> np.ndarray:
        """
        x: [batch, seq_len, d_model]
        Returns: [batch, seq_len, d_model]
        """
        batch, seq_len, _ = x.shape

        # Project to Q, K, V
        q = self.wq.forward(x)  # [batch, seq_len, d_model]
        k = self.wk.forward(x)
        v = self.wv.forward(x)

        # Reshape to [batch, n_heads, seq_len, d_head]
        q = q.reshape(batch, seq_len, self.n_heads, self.d_head).transpose(0, 2, 1, 3)
        k = k.reshape(batch, seq_len, self.n_heads, self.d_head).transpose(0, 2, 1, 3)
        v = v.reshape(batch, seq_len, self.n_heads, self.d_head).transpose(0, 2, 1, 3)

        # Apply RoPE to Q and K
        pos_start = past_len
        pos_end = past_len + seq_len
        if pos_end > self.max_seq_len:
            # Extend RoPE tables dynamically
            pos = np.arange(pos_end, dtype=np.float32)
            cos_ext, sin_ext = rope(pos, self.d_head)
            cos = cos_ext[pos_start:pos_end]
            sin = sin_ext[pos_start:pos_end]
            self._cos = cos_ext
            self._sin = sin_ext
        else:
            cos = self._cos[pos_start:pos_end]
            sin = self._sin[pos_start:pos_end]
        q = apply_rope(q, cos, sin)
        k = apply_rope(k, cos, sin)

        # KV cache
        if use_cache:
            if layer_idx in self._kv_cache:
                past_k, past_v = self._kv_cache[layer_idx]
                k = np.concatenate([past_k, k], axis=2)
                v = np.concatenate([past_v, v], axis=2)
            self._kv_cache[layer_idx] = (k, v)

        # Scaled dot-product attention
        # q: [batch, n_heads, seq_len, d_head]
        # k: [batch, n_heads, total_len, d_head]
        scores = q @ k.transpose(0, 1, 3, 2) / math.sqrt(self.d_head)
        # Causal mask
        total_len = k.shape[2]
        causal = np.triu(np.ones((seq_len, total_len), dtype=bool), k=total_len - seq_len)
        scores = np.where(causal[None, None, :, :], -1e9, scores)
        attn = softmax(scores, axis=-1)

        # Apply attention to V
        out = attn @ v  # [batch, n_heads, seq_len, d_head]
        out = out.transpose(0, 2, 1, 3).reshape(batch, seq_len, self.d_model)

        return self.wo.forward(out)

    def reset_cache(self) -> None:
        self._kv_cache.clear()


class FeedForward:
    """Feed-forward network: 2-layer MLP with GELU."""

    def __init__(self, d_model: int, d_ff: int) -> None:
        self.w1 = Linear(d_model, d_ff, bias=False)
        self.w2 = Linear(d_ff, d_model, bias=False)

    def forward(self, x: np.ndarray) -> np.ndarray:
        """x: [..., d_model] → [..., d_model]"""
        return self.w2.forward(gelu(self.w1.forward(x)))


class TransformerLayer:
    """Single transformer layer: pre-norm attention + pre-norm FFN."""

    def __init__(self, d_model: int, n_heads: int, d_ff: int, max_seq_len: int = 512) -> None:
        self.attn = MultiHeadAttention(d_model, n_heads, max_seq_len)
        self.ffn = FeedForward(d_model, d_ff)

        # Layer norm parameters
        self.ln1_gamma = np.ones(d_model, dtype=np.float32)
        self.ln1_beta = np.zeros(d_model, dtype=np.float32)
        self.ln2_gamma = np.ones(d_model, dtype=np.float32)
        self.ln2_beta = np.zeros(d_model, dtype=np.float32)

    def forward(
        self,
        x: np.ndarray,
        layer_idx: int = 0,
        use_cache: bool = False,
        past_len: int = 0,
    ) -> np.ndarray:
        """Pre-norm transformer layer."""
        # Attention with residual
        normed = layer_norm(x, self.ln1_gamma, self.ln1_beta)
        attn_out = self.attn.forward(normed, layer_idx=layer_idx, use_cache=use_cache, past_len=past_len)
        x = x + attn_out

        # FFN with residual
        normed = layer_norm(x, self.ln2_gamma, self.ln2_beta)
        ffn_out = self.ffn.forward(normed)
        x = x + ffn_out

        return x

    def get_params(self) -> dict[str, Any]:
        """Get all parameters as a dict (for saving/quantization)."""
        return {
            "wq": self.attn.wq.weight,
            "wk": self.attn.wk.weight,
            "wv": self.attn.wv.weight,
            "wo": self.attn.wo.weight,
            "w1": self.ffn.w1.weight,
            "w2": self.ffn.w2.weight,
            "ln1_gamma": self.ln1_gamma,
            "ln1_beta": self.ln1_beta,
            "ln2_gamma": self.ln2_gamma,
            "ln2_beta": self.ln2_beta,
        }

    def set_params(self, params: dict[str, Any]) -> None:
        """Set parameters from a dict."""
        self.attn.wq.weight = params["wq"]
        self.attn.wk.weight = params["wk"]
        self.attn.wv.weight = params["wv"]
        self.attn.wo.weight = params["wo"]
        self.ffn.w1.weight = params["w1"]
        self.ffn.w2.weight = params["w2"]
        self.ln1_gamma = params["ln1_gamma"]
        self.ln1_beta = params["ln1_beta"]
        self.ln2_gamma = params["ln2_gamma"]
        self.ln2_beta = params["ln2_beta"]