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
// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
// SPDX-License-Identifier: MIT OR Apache-2.0
//! 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,
Fma,
ElementwiseRegion,
TransformRegion,
BatchElementwiseRegion,
MatMul,
ScaledMatMul,
ScaledQuantize,
ScaledQuantScale,
ScaledDequantize,
DotGeneral,
LayerNorm,
LayerNorm2d,
GroupNorm,
RmsNorm,
ResizeNearest2x,
Interpolate3d,
AxialRope2d,
Attention,
AttentionBackward,
AttentionBackwardAll,
RmsNormBackwardInput,
RmsNormBackwardGamma,
RmsNormBackwardBeta,
LayerNormBackwardInput,
LayerNormBackwardGamma,
GroupNormBackwardInput,
GroupNormBackwardGamma,
GroupNormBackwardBeta,
RopeBackward,
Cumsum,
CumProd,
CumMax,
CumsumBackward,
GatherBackward,
Conv2dBackwardInput,
Conv3dBackwardInput,
Conv3dBackwardWeight,
MaxPool3dBackward,
Conv2dBackwardWeight,
MaxPool2dBackward,
Rope,
Reshape,
Transpose,
Narrow,
Concat,
KvAppend,
Expand,
Gather,
Reverse,
Pad,
Slice,
Reduce,
Softmax,
SoftmaxCrossEntropy,
SoftmaxCrossEntropyWithLogits,
SoftmaxCrossEntropyBackward,
ArgMax,
ArgMin,
TopK,
Sample,
RngNormal,
RngUniform,
Conv,
Im2Col,
ConvTranspose2d,
Pool,
GroupedMatMul,
DequantGroupedMatMul,
DequantGroupedMatMulMlx,
DequantMoEWeights,
ScatterAdd,
ScatterNd,
ScatterElements,
GatherNd,
GatherElements,
DequantMatMul,
SynthMatMul,
// Native fused reconstruct (`synth_reconstruct_nk`, writes `w_bt[n,k]` in
// one dispatch). Claimed so a forward-only INFERENCE path can emit it — it
// wins the forward (~35 vs 47ms). NOT emitted by `Tensor::synth_reconstruct`
// during training, where it MEASURED net-worse: the opaque op costs ~+20ms
// in the backward (hidden from the CSE + transpose-simplification that make
// the decomposed fold's `dx` free) — more than the ~12ms forward win.
SynthReconstruct,
// NOTE: `SynthMatMulBackward` intentionally NOT claimed — the native fused
// kernels (MSL `synth_bwd_dx`/`synth_bwd_codebook`) are built and
// correctness-validated, but MEASURED slower than decomposing to Gather +
// MPS-tiled sgemm (a hand GEMM ≈ 40% of MPS, same as the forward). So the
// op decomposes via `LowerSynthMatMulBackward`; the kernels stay dormant
// for a future tiled/simdgroup implementation that could beat MPS.
SplineActivation,
SplineActivationBackwardX,
SplineActivationBackwardCoeff,
GatedDeltaNet,
SelectiveScan,
Lstm,
Gru,
Rnn,
Mamba2,
FusedSwiGLU,
FusedMatMulBiasAct,
// FusedMatMulResidual intentionally NOT claimed: decode is weight-read
// bandwidth-bound, so folding the residual into the matmul saves zero
// GPU time (measured: identical 40.5ms wait with 56 fewer dispatches),
// and its f32-only epilogue kernel would force the o_proj/down_proj
// weights to materialize F32 — blocking the F16-resident weight win
// (RLX_QWEN3_F16_WEIGHTS: 23.9→33.2 tok/s). The op + pass + kernel stay
// in-tree for a future non-bandwidth-bound path; Metal just doesn't opt in.
// (Re-confirmed on the dispatch-bound training path 2026-08: fusing all 16
// residuals gave 1.00× — removing small dispatches doesn't move wall time
// here; only removing real recompute, like the attention-bwd fusion, does.)
FusedResidualLN,
FusedResidualRmsNorm,
// Claimed so the Metal fusion pipeline may emit it;
// `MetalExecutable::compile_inner` decomposes it back to the
// primitive chain (no monolithic fused-attention MSL kernel
// yet — the per-run cost is dominated by wait_until_completed,
// not encode, so a dispatch-wrapper fusion buys nothing).
FusedAttentionBlock,
// DiT adaLN-Zero / gated residual — native MSL kernels avoid
// Expand of `[B,1,D]` modulation over the sequence axis.
AdaLayerNorm,
GatedResidual,
AdaLayerNormBackward,
GatedResidualBackward,
// User-registered custom ops dispatched through
// `rlx_metal::op_registry`. Lowering panics with a clear
// message if the named MetalKernel isn't registered;
// executor inserts a sync point + runs the host kernel
// against the unified-memory arena.
Custom,
// Op::Fft is supported via the same host-fallback pattern
// as Custom: sync the GPU, run rlx-cpu's FFT against the
// unified-memory arena, restart cmd_buf. A native Metal
// compute kernel will replace this when a workload makes
// the sync the bottleneck.
Fft,
// Op::Scan (arbitrary-body recurrence) via the same host
// fallback: compile the body once, loop it on the CPU against
// the unified-memory arena. Enables IIR (`biquad`/`sosfilt`).
Scan,
ScanBackward,
ScanBackwardXs,
LogMel,
LogMelBackward,
WelchPeaks,
// Host-fallback splat (unified-memory arena + rlx-cpu/splat).
GaussianSplatRender,
GaussianSplatRenderBackward,
GaussianSplatPrepare,
GaussianSplatRasterize,
// Core Riemannian / SPD-manifold ops. No MSL eigen kernel; they
// host-fallback to `rlx_cpu::spd` (F64) against the unified-memory
// arena via the same sync pattern as Fft/Custom (see
// `rlx_metal::spd` + `Thunk::SpdHost`). The SPD subgraph's F64
// tensors are widened to f32 for arena planning; the host step does
// the f32↔f64 conversion.
BiMap,
ReEig,
LogEig,
SpdBatchNorm,
SpdKarcherMean,
SpdKarcherMeanWeighted,
SpdLogMap,
SpdExpMap,
SpdParallelTransport,
SpdMatrixFnBatch,
ReEigBackward,
LogEigBackward,
SpdBatchNormBackwardX,
SpdBatchNormBackwardG,
SpdLogMapBackward,
SpdExpMapBackward,
SpdParallelTransportBackward,
SpdMatrixFnBatchBackward,
Eigh,
EighBackward,
EighBatch,
EighBatchBackward,
// Full OpKind coverage (host-fallback via `Thunk::HostOp` /
// `eval_single_op_f32`, or primitive expand for fused ops the
// CPU catch-all would Nop). See `lower_cpu_nop_fused_for_metal`
// + the HostOp arm in `thunk/compile.rs`.
Quantize,
Dequantize,
FakeQuantize,
FakeQuantizeLSQ,
FakeQuantizeLSQBackwardX,
FakeQuantizeLSQBackwardScale,
DenseSolve,
BatchedDenseSolve,
// Cholesky / TriangularSolve / Det / LogDet host-stage to CPU
// LAPACK via the `_other => Thunk::HostOp` catch-all (same as
// DenseSolve F64).
Cholesky,
TriangularSolve,
Det,
LogDet,
// Sort / ArgSort host-stage to CPU (stable strided sort) via the
// same `_other => Thunk::HostOp` catch-all as Det / LogDet.
Sort,
Svd,
Qr,
ArgSort,
BatchNormInference,
BatchNormInferenceBackwardInput,
BatchNormInferenceBackwardGamma,
BatchNormInferenceBackwardBeta,
Conv3d,
ConvTranspose3d,
// Native MSL ReluBackward / ActivationBackward (Fixed kinds).
ReluBackward,
ActivationBackward,
FakeQuantizeBackward,
// Native MSL C64 Wirtinger surface (`complex_norm_sq` /
// `complex_norm_sq_backward` / `conjugate_c64`).
ComplexNormSq,
ComplexNormSqBackward,
Conjugate,
// Native MSL ternary-pruned FFT butterfly (`fft_butterfly_stage`).
FftButterflyStage,
LoraMatMul,
PartitionedConv,
QMatMul,
QConv2d,
FusedConvBiasAct,
FusedTransformerLayer,
If,
While,
CustomFn,
]
};