File size: 7,016 Bytes
72c0672
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import List, Tuple
import itertools
import math

def _pack_rank_ids(buf: List[int], rank_bitlength: int) -> List[int]:
    per_b   = 8 // rank_bitlength
    mask    = (1 << rank_bitlength) - 1
    out_b   = []
    it      = iter(buf)

    while True:
        chunk = list(itertools.islice(it, per_b))
        if not chunk:
            break
        byte_val = 0
        for p, idx in enumerate(chunk):
            byte_val |= (idx & mask) << (p * rank_bitlength)
        out_b.append(byte_val)
    return out_b

def _unpack_rank_ids(payload: List[int], run_len: int, rank_bitlength: int):
    mask   = (1 << rank_bitlength) - 1

    byte_iter = iter(payload)
    cur_byte  = next(byte_iter)
    filled    = 8

    for _ in range(run_len):
        if filled == 0:
            cur_byte = next(byte_iter)
            filled = 8
        rank_id = cur_byte & mask
        cur_byte >>= rank_bitlength
        filled   -= rank_bitlength
        yield rank_id

class SimpleAdaptiveRankCodec:
    def __init__(
        self, 
        top_k: int = 4, 
        tau: float = 0.5, 
        min_run: int = 3,
        max_run: int = 255,
        sentinel_rle: int = 256,
        sentinel_rank_run: int = 257,
    ):
        self.top_k = top_k
        self.tau = tau
        self.min_run = min_run
        self.max_run = max_run
        self.raw_byte_offset = 256

        self.rank_bitlength = max(1, (top_k - 1).bit_length())
        assert self.rank_bitlength <= 8 and 8 % self.rank_bitlength == 0, (
            f"rank_bitlength must be between 1 and 8 and must divide 8, got {self.rank_bitlength}"
            f"top_k: {top_k}"
        )
        self.ranks_per_byte = 8 // self.rank_bitlength
        self.sentinel_rle = sentinel_rle
        self.sentinel_rank_run = sentinel_rank_run

    def encode_window(
        self,
        tokens:       List[int],
        repeat_probs: List[float],
        ranks:        List[int],
    ) -> List[int]:
        """Return a list of ints: raw bytes 0-255 and sentinel events ≥256."""
        assert len(tokens) == len(repeat_probs) == len(ranks)

        rank_buf: List[int] = []
        out: List[int]      = [tokens[0]]
        i, n = 1, len(tokens)

        def flush_rank_buf():
            if not rank_buf:
                return
            out.append(self.sentinel_rank_run)
            out.append(len(rank_buf))
            out.extend(_pack_rank_ids(rank_buf, self.rank_bitlength))
            rank_buf.clear()

        while i < n:
            tok = tokens[i]

            # --- RLE probe (uses *current* token prob) --------------------
            run = 1
            while (i + run < n and
                tokens[i + run] == tok and
                repeat_probs[i + run] >= self.tau):
                run += 1
            if run >= self.min_run:
                flush_rank_buf()
                out.extend([self.sentinel_rle, run, tok])
                i += run
                continue

            if ranks[i] < self.top_k:
                rank_buf.append(ranks[i])
            else:
                # the current token is not in top-K,
                # so we escape to a raw byte
                flush_rank_buf()
                out.append(tok)
            i += 1

        flush_rank_buf()
        return out

    def encoding_to_pseudo_bytes(self, enc: list[int]) -> list[int]:
        # NOTE: this function is not expected to be lossless, that is, 
        # we cannot reconstruct the original encoding from the pseudo-bytes
        out: list[int] = []
        i = 0
        while i < len(enc):
            tok = enc[i]
            i += 1

            if tok < self.raw_byte_offset:
                out.append(tok)

            elif tok == self.sentinel_rle:
                run = enc[i]
                raw = enc[i+1]
                i += 2
                run = min(run, self.max_run)
                # we mark the run length from 512 to 256
                out.extend([self.raw_byte_offset + self.raw_byte_offset - run, raw])
            elif tok == self.sentinel_rank_run:
                length = enc[i]
                i += 1
                n_bytes = math.ceil(length / self.ranks_per_byte)
                for _ in range(n_bytes):
                    pb = enc[i] + self.raw_byte_offset
                    out.append(pb)
                    i += 1
            else:
                raise ValueError(f"unknown token {tok}")
        return out

    def pseudo_bytes_to_encoding(self, pb: list[int], original_encoding: list[int]) -> list[int]:
        # NOTE: we do not expect the encoding-to-pseudo-bytes conversion to be lossless,
        # so we need to pass the original encoding to reconstruct the original encoding
        # this function is just for sanity check
        raise NotImplementedError("Not implemented")

    def decode_window(
        self,
        stream:       List[int],
        original_len: int,
        topk_symbols: List[List[int]],
    ) -> List[int]:
        """
        `topk_symbols[pos][idx]` must give the byte value (0-255) that
        corresponds to rank `idx` at position `pos`, e.g. recomputed from
        the helper LM during decoding.
        """
        out: List[int] = []
        # position in input stream
        i   = 0
        # position in output tokens
        pos = 0

        while pos < original_len:
            tok = stream[i]
            i += 1

            if tok < 256:
                out.append(tok)
                pos += 1

            elif tok == self.sentinel_rle:
                run_len = stream[i]
                raw = stream[i+1]
                i += 2
                out.extend([raw] * run_len)
                pos += run_len

            elif tok == self.sentinel_rank_run:
                run_len = stream[i]
                i += 1
                bytes_needed = math.ceil(run_len / self.ranks_per_byte)
                payload = stream[i: i + bytes_needed]
                i += bytes_needed
                for rank_id in _unpack_rank_ids(payload, run_len, self.rank_bitlength):
                    sym = topk_symbols[pos][rank_id]
                    out.append(sym)
                    pos += 1
            else:
                raise ValueError(f"Unknown sentinel {tok}")

        return out[:original_len]

if __name__ == "__main__":
    import torch, random
    random.seed(0)
    T, K = 384, 13                        # demonstrate non-power-of-two K
    tokens = torch.randint(0, 32, (T,)).tolist()
    repeat_probs = torch.rand(T).tolist()
    ranks = torch.randint(0, K + 5, (T,)).tolist()   # some ranks ≥K → raw
    ranks = [r if r < K else K for r in ranks]

    # fake LM top-K table for demo: identity mapping
    topk = [[tokens[t]] * K for t in range(T)]

    codec = SimpleAdaptiveRankCodec(top_k=K, tau=0.00)

    enc = codec.encode_window(tokens, repeat_probs, ranks)
    dec = codec.decode_window(enc, T, topk)

    print(f"raw={T}  encoded={len(enc)}  ratio={len(enc)/T:.2f}")
    assert dec == tokens
    print("✓ window-enc-dec round-trip passes")