rlx-models 0.2.5

Model loading for RLX — config parsing, safetensors weights, graph builders
Documentation 
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

#!/usr/bin/env python3
# RLX — versatile ML compiler + runtime.
# Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation, version 3.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
#
# You should have received a copy of the GNU General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.

"""Emit JSON lines for RLX ↔ TIDE PyTorch component parity (standalone torch refs)."""

from __future__ import annotations

import json
import math

import torch
import torch.nn.functional as F


def emit(obj: dict) -> None:
    print(json.dumps(obj), flush=True)


def ramp(n: int, scale: float) -> list[float]:
    return [0.001 + scale * i * 0.01 for i in range(n)]


def get_num_transfer_tokens(block_length: int, steps: int) -> list[int]:
    if steps == 0:
        return []
    base = block_length // steps
    remainder = block_length % steps
    out = [base] * steps
    for i in range(remainder):
        out[i] += 1
    return out


def group_limited_topk(
    scores: torch.Tensor,
    n_group: int,
    topk_group: int,
    top_k: int,
) -> tuple[torch.Tensor, torch.Tensor]:
    num_tokens, num_experts = scores.size()
    epg = num_experts // n_group
    group_scores = scores.view(num_tokens, n_group, -1).topk(2, dim=-1)[0].sum(dim=-1)
    group_idx = torch.topk(group_scores, k=topk_group, dim=-1, sorted=False)[1]
    group_mask = torch.zeros_like(group_scores)
    group_mask.scatter_(1, group_idx, 1)
    score_mask = (
        group_mask.unsqueeze(-1)
        .expand(num_tokens, n_group, epg)
        .reshape(num_tokens, -1)
    )
    masked_scores = scores.masked_fill(~score_mask.bool(), float("-inf"))
    return torch.topk(masked_scores, k=top_k, dim=-1)


def gate_forward(
    hidden: torch.Tensor,
    weight: torch.Tensor,
    expert_bias: torch.Tensor,
    n_group: int,
    topk_group: int,
    top_k: int,
    routed_scaling: float,
) -> tuple[torch.Tensor, torch.Tensor]:
    logits = F.linear(hidden.float(), weight.float())
    scores = torch.sigmoid(logits.float())
    scores_for_routing = scores + expert_bias
    _, topk_idx = group_limited_topk(scores_for_routing, n_group, topk_group, top_k)
    gathered = torch.gather(scores, dim=1, index=topk_idx).type_as(logits)
    if top_k > 1:
        topk_weight = gathered / (gathered.sum(dim=-1, keepdim=True) + 1e-20)
    else:
        topk_weight = gathered
    topk_weight = topk_weight * routed_scaling
    return topk_idx, topk_weight


def main() -> int:
    for block_length, steps in [(32, 32), (10, 3), (7, 4)]:
        emit(
            {
                "test": "transfer_schedule",
                "block_length": block_length,
                "steps": steps,
                "schedule": get_num_transfer_tokens(block_length, steps),
            }
        )

    seq_len, block_length = 8, 4
    num_blocks = (seq_len + block_length - 1) // block_length
    block_mask = torch.tril(torch.ones(num_blocks, num_blocks, dtype=torch.float32))
    attn = (
        block_mask.repeat_interleave(block_length, dim=0)
        .repeat_interleave(block_length, dim=1)
        .unsqueeze(0)
        .unsqueeze(0)
        .log()[:, :, :seq_len, :seq_len]
    )
    flat = attn.reshape(-1)
    emit(
        {
            "test": "block_mask",
            "seq_len": seq_len,
            "block_length": block_length,
            "mask": ["-inf" if math.isinf(v) else float(v) for v in flat.tolist()],
        }
    )

    scores = torch.tensor(
        [[0.1, 0.9, 0.2, 0.8], [0.5, 0.5, 0.5, 0.5]],
        dtype=torch.float32,
    )
    probs, idx = group_limited_topk(scores, n_group=2, topk_group=1, top_k=2)
    emit(
        {
            "test": "group_limited_topk",
            "indices": [int(x) for x in idx.reshape(-1).tolist()],
            "probs": [float(x) for x in probs.reshape(-1).tolist()],
        }
    )

    h, e, rows = 16, 4, 4
    hidden = torch.tensor(
        [[0.01 * i for i in range(h)] for _ in range(rows)], dtype=torch.float32
    )
    # RLX stores router row-major [hidden, expert]; F.linear weight is [expert, hidden].
    weight = torch.tensor(ramp(h * e, 1.1), dtype=torch.float32).reshape(h, e).t()
    bias = torch.tensor(ramp(e, 0.01), dtype=torch.float32)
    top_idx, top_weight = gate_forward(
        hidden, weight, bias, n_group=2, topk_group=1, top_k=2, routed_scaling=2.5
    )
    emit(
        {
            "test": "gate_forward",
            "indices": [int(x) for x in top_idx.reshape(-1).tolist()],
            "weights": [float(x) for x in top_weight.reshape(-1).tolist()],
        }
    )

    for step in range(4):
        refresh = (1 == 0) or (step % 2 == 0)
        emit({"test": "refresh", "num_block": 1, "prefill": 0, "step": step, "refresh": refresh})

    emit({"test": "done"})
    return 0


if __name__ == "__main__":
    raise SystemExit(main())