rlx-mlx 0.2.13

MLX backend for RLX — Apple's array framework via hand-rolled C++ shim, eager + lazy execution
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

// 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/>.

//! OpKinds this backend claims for legalization (`Backend::supported_ops`).
//!
//! Source of truth for the coverage matrix in `docs/op-coverage.md`.
//! Kept in the backend crate so adding an op is a local edit, not a change
//! to `rlx-runtime`'s mega-`backend.rs`.

pub const SUPPORTED_OPS: &[rlx_ir::OpKind] = {
    use rlx_ir::OpKind::*;
    &[
        Input,
        Param,
        Constant,
        Activation,
        Cast,
        StopGradient,
        Binary,
        Compare,
        Where,
        ElementwiseRegion,
        TransformRegion,
        BatchElementwiseRegion,
        MatMul,
        DotGeneral,
        DenseSolve,
        BatchedDenseSolve,
        LayerNorm,
        LayerNorm2d,
        GroupNorm,
        ResizeNearest2x,
        RmsNorm,
        Attention,
        Rope,
        Reshape,
        Transpose,
        Narrow,
        Concat,
        Expand,
        Gather,
        Reverse,
        Reduce,
        Softmax,
        Cumsum,
        ArgMax,
        ArgMin,
        TopK,
        RngNormal,
        RngUniform,
        Sample,
        Conv,
        Im2Col,
        ConvTranspose2d,
        Pool,
        GroupedMatMul,
        DequantGroupedMatMul,
        DequantMoEWeights,
        ScatterAdd,
        ScatterNd,
        ScatterElements,
        GatherNd,
        GatherElements,
        LoraMatMul,
        DequantMatMul,
        SelectiveScan,
        GatedDeltaNet,
        FusedSwiGLU,
        FusedMatMulBiasAct,
        FusedResidualLN,
        FusedResidualRmsNorm,
        AdaLayerNorm,
        GatedResidual,
        // Packed DiT reverse — native composed lowering in `rlx-mlx/src/lower/`.
        AdaLayerNormBackward,
        GatedResidualBackward,
        FusedAttentionBlock,
        FusedTransformerLayer,
        If,
        While,
        // Loop-unrolled scan (Op::Scan body is statically unrolled
        // `length` times into MLX ops; mirror of Op::While's
        // bounded-unroll lowering). ScanBackward is the AD
        // companion — handled the same way.
        Scan,
        ScanBackward,
        ScanBackwardXs,
        // Recurrent ops are kept intact (NOT unfused into a giant
        // unrolled graph — that makes MLX compile crawl on long
        // sequences, e.g. a 256-step LSTM over a spatial field). MLX
        // host-evals them via the shared CPU kernel (see
        // `first_host_eval_op` + the `Op::Lstm` arm in rlx-mlx/lower.rs),
        // matching the ROCm/CoreML host-fallback pattern.
        Lstm,
        // GRU / Elman-RNN: native on-device unrolls in `rlx-mlx/lower.rs`
        // (`native_gru` / `native_rnn`), same story as `Lstm`. Without these
        // the runtime would decompose them via `unfuse` and the native
        // (fp16-capable) unrolls would never run.
        Gru,
        Rnn,
        // Tier 1 autodiff backward ops — lowered as primitive
        // compositions in `rlx-mlx/src/lower/`.
        ReluBackward,
        ActivationBackward,
        SoftmaxCrossEntropy,
        SoftmaxCrossEntropyWithLogits,
        SoftmaxCrossEntropyBackward,
        AttentionBackward,
        LayerNormBackwardInput,
        LayerNormBackwardGamma,
        // GroupNorm backward — native MLX lowering in `lower.rs`
        // (group-reshape + reduce, mirrors GroupNormBackwardInput).
        GroupNormBackwardInput,
        GroupNormBackwardGamma,
        GroupNormBackwardBeta,
        // Tier 2 — conv backward via `mc::conv_general` with the
        // same parameter-mapping MLX uses inside its built-in vjp.
        // Currently groups=1 only; grouped conv backward will
        // surface as a clear error from `lower.rs`.
        Conv2dBackwardInput,
        Conv2dBackwardWeight,
        // Tier 3 — max-pool backward via slice-strided argmax over
        // pool windows + a per-kernel-slot scatter-add, matching
        // the CPU thunk's "first-hit-wins" tiebreaking.
        MaxPool2dBackward,
        // QAT — `FakeQuantize` (PerBatch + Fixed scale modes;
        // EMA returns a clear error from `lower.rs`) and the
        // `FakeQuantizeBackward` family covering all 4 STE
        // variants. Closes the last gap vs `rlx_cpu::SUPPORTED_OPS`.
        FakeQuantize,
        FakeQuantizeBackward,
        // User-registered custom ops dispatched through
        // `rlx_mlx::op_registry`. Lowering looks up the
        // registered `MlxKernel` and calls its `execute` method
        // to produce the lazy MLX `Array` for this node.
        Custom,
        Fft,
        LogMel,
        LogMelBackward,
        WelchPeaks,
        GaussianSplatRender,
        GaussianSplatRenderBackward,
        // Op::Fft on MLX: native `mlx::fft::fft` via rlx_mlx_op_fft shim.
        // 2N real-block f32/f64 and complex64 inputs supported.
    ]
};