File size: 10,752 Bytes
198b1d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
592dcd6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
198b1d8
 
 
592dcd6
198b1d8
 
592dcd6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
198b1d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Vortex-Embed v4.5 — Standalone Single-File Native 4-Bit Embedding Engine.

Lightweight, 4.72 MB RAM native 4-bit embedding engine with Matryoshka support.
Repo: https://huggingface.co/VTXAI/Vortex-Embed-v4-5-sentence
"""
from __future__ import annotations

import json
from dataclasses import dataclass, field
from pathlib import Path
from typing import List, Optional, Sequence, Tuple, Union

import numpy as np
from safetensors.numpy import load_file, save_file

try:
    from tokenizers import Tokenizer
except ImportError:
    Tokenizer = None


class VortexEmbedConfig:
    def __init__(
        self,
        vocab_size: int = 29528,
        embedding_dim: int = 256,
        block_size: int = 32,
        num_blocks: int = 8,
        model_type: str = "vortex-embed",
        architectures: Optional[List[str]] = None,
        quantization: str = "lf4",
        bits: int = 4,
        sif_a: float = 0.05,
        sif_pc: float = 1.0,
        pc_k: int = 1,
        matryoshka_dim: Optional[int] = None,
        **kwargs,
    ):
        self.vocab_size = vocab_size
        self.embedding_dim = embedding_dim
        self.block_size = block_size
        self.num_blocks = num_blocks
        self.model_type = model_type
        self.architectures = architectures or ["VortexEmbedV4_5"]
        self.quantization = quantization
        self.bits = bits
        self.sif_a = sif_a
        self.sif_pc = sif_pc
        self.pc_k = pc_k
        self.matryoshka_dim = matryoshka_dim

    @classmethod
    def from_dict(cls, d: dict) -> "VortexEmbedConfig":
        return cls(**d)

    def to_dict(self) -> dict:
        return {
            "vocab_size": self.vocab_size,
            "embedding_dim": self.embedding_dim,
            "block_size": self.block_size,
            "num_blocks": self.num_blocks,
            "model_type": self.model_type,
            "architectures": self.architectures,
            "quantization": self.quantization,
            "bits": self.bits,
            "sif_a": self.sif_a,
            "sif_pc": self.sif_pc,
            "pc_k": self.pc_k,
            "matryoshka_dim": self.matryoshka_dim,
        }



class VortexEmbedV4_5:
    """Vortex-Embed v4.5 — Native 4-Bit Sentence Embedding Model.

    Features:
      - 4.72 MB RAM Footprint (Zero FP32 matrix in RAM).
      - On-the-fly dequantization per batch.
      - Matryoshka Representation Learning (truncation to 256, 128, 64 dims).
    """

    def __init__(
        self,
        packed: np.ndarray,
        scales: np.ndarray,
        zeros: np.ndarray,
        tokenizer_data: Union[str, Path],
        config: Union[dict, VortexEmbedConfig],
        *,
        matryoshka_dim: Optional[int] = None,
    ) -> None:
        self.packed = np.asarray(packed, dtype=np.uint8)
        self.scales = np.asarray(scales, dtype=np.float16)
        self.zeros = np.asarray(zeros, dtype=np.float16)
        self.tokenizer_data = str(tokenizer_data)
        self.config = config if isinstance(config, VortexEmbedConfig) else VortexEmbedConfig.from_dict(config)
        self.vocab_size = int(self.config.vocab_size)
        self.dim = int(self.config.embedding_dim)
        self.block_size = int(self.config.block_size)
        self.num_blocks = int(self.config.num_blocks)
        self.sif_a = float(self.config.sif_a)
        self.sif_pc = float(self.config.sif_pc)
        self.pc_k = int(self.config.pc_k)
        self.matryoshka_dim = matryoshka_dim or self.config.matryoshka_dim

        self._tokenizer: Optional[Tokenizer] = None
        self._sif_weights: Optional[np.ndarray] = None
        self._pc_directions: Optional[np.ndarray] = None

    @property
    def tokenizer(self) -> Tokenizer:
        if self._tokenizer is None:
            if Tokenizer is None:
                raise RuntimeError("tokenizers required: pip install tokenizers")
            self._tokenizer = Tokenizer.from_file(self.tokenizer_data)
        return self._tokenizer

    @property
    def model_size_mb(self) -> float:
        """Returns actual in-RAM size of stored parameters (4.72 MB)."""
        return (self.packed.nbytes + self.scales.nbytes + self.zeros.nbytes) / 1e6

    @property
    def on_disk_size_mb(self) -> float:
        return (self.packed.nbytes + self.scales.nbytes + self.zeros.nbytes) / 1e6

    @classmethod
    def from_pretrained(
        cls,
        path_or_id: Union[str, Path],
        matryoshka_dim: Optional[int] = None,
        **overrides,
    ) -> "VortexEmbedV4_5":
        path = Path(path_or_id)
        if not path.is_dir():
            from huggingface_hub import snapshot_download
            path = Path(snapshot_download(str(path_or_id)))
        tensors = load_file(str(path / "model.safetensors"))
        config = json.loads((path / "config.json").read_text())
        for k, v in overrides.items():
            if k in VortexEmbedConfig.__dataclass_fields__:
                config[k] = v
        return cls(
            packed=tensors["embedding_packed"],
            scales=tensors["embedding_scales"],
            zeros=tensors["embedding_zeros"],
            tokenizer_data=str(path / "tokenizer.json"),
            config=config,
            matryoshka_dim=matryoshka_dim,
        )

    def save_pretrained(self, path: Union[str, Path]) -> None:
        out = Path(path)
        out.mkdir(parents=True, exist_ok=True)
        save_file(
            {
                "embedding_packed": self.packed,
                "embedding_scales": self.scales,
                "embedding_zeros": self.zeros,
            },
            str(out / "model.safetensors"),
        )
        (out / "config.json").write_text(json.dumps(self.config.to_dict(), indent=2))
        if not (out / "tokenizer.json").exists():
            (out / "tokenizer.json").write_text(Path(self.tokenizer_data).read_text())

    def _dequantize_ids_on_the_fly(self, token_ids: np.ndarray) -> np.ndarray:
        if token_ids.size == 0:
            return np.empty((0, self.dim), dtype=np.float32)

        p = self.packed[token_ids]
        s = self.scales[token_ids].astype(np.float32)[:, :, None]
        z = self.zeros[token_ids].astype(np.float32)[:, :, None]

        low = (p & 0x0F).astype(np.float32)
        high = ((p >> 4) & 0x0F).astype(np.float32)

        n = len(token_ids)
        padded = p.shape[1] * 2
        unpacked = np.empty((n, padded), dtype=np.float32)
        unpacked[:, 0::2] = low
        unpacked[:, 1::2] = high

        blocked = unpacked.reshape(n, self.num_blocks, self.block_size)
        out = (blocked * s + z).reshape(n, padded)
        return out[:, : self.dim]

    def fit_idf(self, corpus_token_lists: Sequence[Sequence[int]]) -> "VortexEmbedV4_5":
        flat = (np.concatenate(corpus_token_lists) if corpus_token_lists else np.empty(0, dtype=np.int64))
        total = max(int(flat.size), 1)
        counts = np.bincount(flat, minlength=self.vocab_size).astype(np.float64)
        p = counts / total
        denom = self.sif_a + p
        with np.errstate(divide="ignore", invalid="ignore"):
            weights = np.where(p > 0, self.sif_a / denom, 1.0)
        self._sif_weights = weights.astype(np.float32)
        return self

    def fit_pc(self, corpus_embeddings: np.ndarray, k: Optional[int] = None) -> "VortexEmbedV4_5":
        if k is None:
            k = self.pc_k
        if corpus_embeddings.size == 0 or k <= 0:
            return self
        x = corpus_embeddings.astype(np.float32)
        x = x - x.mean(axis=0, keepdims=True)
        try:
            _, _, vt = np.linalg.svd(x, full_matrices=False)
            pcs = vt[:k].astype(np.float32)
            pcs = pcs / (np.linalg.norm(pcs, axis=1, keepdims=True) + 1e-12)
            self._pc_directions = pcs
        except np.linalg.LinAlgError:
            self._pc_directions = None
        return self

    def _apply_pc(self, x: np.ndarray) -> np.ndarray:
        if self.sif_pc <= 0 or self._pc_directions is None:
            return x
        out = x
        for pc in self._pc_directions:
            proj = (out @ pc)[:, None] * pc[None, :]
            out = out - self.sif_pc * proj
        return out

    def _tokenize_batch(self, texts: Sequence[str]) -> List[List[int]]:
        encoded = self.tokenizer.encode_batch(list(texts))
        return [[tid for tid in item.ids if 0 <= int(tid) < self.vocab_size] for item in encoded]

    def encode_batch(
        self,
        texts: Sequence[str],
        *,
        normalize: bool = True,
        truncate_dim: Optional[int] = None,
    ) -> np.ndarray:
        if not texts:
            return np.zeros((0, self.dim), dtype=np.float32)

        token_lists = self._tokenize_batch(list(texts))
        n = len(token_lists)
        flat = (np.concatenate(token_lists) if token_lists else np.empty(0, dtype=np.int64))

        if flat.size == 0:
            return np.zeros((n, self.dim), dtype=np.float32)

        unique_ids, inverse_indices = np.unique(flat, return_inverse=True)
        unique_embs = self._dequantize_ids_on_the_fly(unique_ids)
        token_embs = unique_embs[inverse_indices]

        if self._sif_weights is not None:
            w = self._sif_weights[flat].astype(np.float32)[:, None]
            token_embs = token_embs * w

        chunk_lens = np.array([len(ids) for ids in token_lists], dtype=np.int64)
        chunk_ends = np.cumsum(chunk_lens)
        boundaries = np.empty(n + 1, dtype=np.int64)
        boundaries[0] = 0
        boundaries[1:] = chunk_ends

        sums = np.add.reduceat(token_embs, boundaries[:-1], axis=0)

        if self._sif_weights is not None:
            w_full = self._sif_weights[flat].astype(np.float32)
            w_per_row = np.add.reduceat(w_full, boundaries[:-1])
            w_per_row = np.maximum(w_per_row, 1e-12)
        else:
            w_per_row = np.maximum(chunk_lens.astype(np.float32), 1.0)

        embeddings = sums / w_per_row[:, None]
        embeddings = self._apply_pc(embeddings)

        dim = truncate_dim if truncate_dim is not None else self.matryoshka_dim
        if dim is not None and 0 < dim < self.dim:
            embeddings = embeddings[:, :dim]

        if normalize and embeddings.shape[0] > 0:
            norms = np.linalg.norm(embeddings, axis=1, keepdims=True)
            np.divide(embeddings, np.maximum(norms, 1e-12), out=embeddings)

        return embeddings

    def encode(
        self,
        texts: Union[str, Sequence[str]],
        *,
        normalize: bool = True,
        truncate_dim: Optional[int] = None,
    ) -> np.ndarray:
        if isinstance(texts, str):
            return self.encode_batch([texts], normalize=normalize, truncate_dim=truncate_dim)[0]
        return self.encode_batch(list(texts), normalize=normalize, truncate_dim=truncate_dim)