mlx-native 0.3.2

Pure-Rust Metal GPU compute library for MLX-compatible inference on Apple Silicon
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
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

//! Tests for the MoE gate (top-K expert routing with softmax) GPU kernel.

#![allow(clippy::expect_used, clippy::unwrap_used, clippy::panic)]

use mlx_native::{DType, KernelRegistry, MlxDevice, MlxBuffer};
use mlx_native::ops::moe_gate::{moe_gate, MoeGateParams};

fn setup() -> (MlxDevice, KernelRegistry) {
    let device = MlxDevice::new().expect("MlxDevice::new");
    let registry = KernelRegistry::new();
    (device, registry)
}

fn f32_to_bf16_bytes(v: f32) -> [u8; 2] {
    let bits = v.to_bits();
    let bf16 = (bits >> 16) as u16;
    bf16.to_le_bytes()
}

fn alloc_bf16(device: &MlxDevice, data: &[f32]) -> MlxBuffer {
    let n = data.len();
    let mut buf = device
        .alloc_buffer(n * 2, DType::BF16, vec![n])
        .expect("alloc bf16");
    let bytes = buf.as_mut_slice::<u8>().expect("write bf16");
    for (i, &v) in data.iter().enumerate() {
        let b = f32_to_bf16_bytes(v);
        bytes[i * 2] = b[0];
        bytes[i * 2 + 1] = b[1];
    }
    buf
}

fn alloc_f32(device: &MlxDevice, data: &[f32]) -> MlxBuffer {
    let mut buf = device
        .alloc_buffer(data.len() * 4, DType::F32, vec![data.len()])
        .expect("alloc f32");
    buf.as_mut_slice::<f32>().expect("w").copy_from_slice(data);
    buf
}

fn rms_norm(x: &[f32], weight: &[f32], eps: f32) -> Vec<f32> {
    let ss: f32 = x.iter().map(|v| v * v).sum::<f32>() / x.len() as f32;
    let rms = (ss + eps).sqrt();
    x.iter().zip(weight).map(|(&xi, &wi)| (xi / rms) * wi).collect()
}

#[allow(dead_code)]
fn softmax(values: &[f32]) -> Vec<f32> {
    let max_val = values.iter().copied().fold(f32::NEG_INFINITY, f32::max);
    let exps: Vec<f32> = values.iter().map(|&v| (v - max_val).exp()).collect();
    let sum: f32 = exps.iter().sum();
    exps.iter().map(|&e| e / sum).collect()
}

#[test]
fn test_moe_gate_basic() {
    let (device, mut registry) = setup();

    let hidden_dim = 16;
    let n_experts = 8;
    let top_k = 3;
    let rms_eps = 1e-6f32;

    let hidden_data: Vec<f32> = (0..hidden_dim).map(|i| (i as f32) * 0.1).collect();
    let norm_weight: Vec<f32> = vec![1.0; hidden_dim];
    let per_expert_scale: Vec<f32> = vec![1.0; n_experts];

    let mut router_data = vec![0.0f32; n_experts * hidden_dim];
    for e in 0..n_experts {
        for d in 0..hidden_dim {
            router_data[e * hidden_dim + d] = match e {
                5 => 2.0,
                2 => 1.5,
                7 => 1.0,
                _ => 0.1,
            };
        }
    }

    // Reference: RMS-norm the hidden state, then router matmul, then top-K + softmax
    let normed = rms_norm(&hidden_data, &norm_weight, rms_eps);
    let mut ref_logits = vec![0.0f32; n_experts];
    for e in 0..n_experts {
        for d in 0..hidden_dim {
            ref_logits[e] += normed[d] * router_data[e * hidden_dim + d];
        }
    }

    let mut indexed: Vec<(usize, f32)> = ref_logits.iter().enumerate().map(|(i, &v)| (i, v)).collect();
    indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
    let top_indices: Vec<u32> = indexed[..top_k].iter().map(|&(i, _)| i as u32).collect();

    // GPU buffers — hidden_state must be bf16
    let hidden_buf = alloc_bf16(&device, &hidden_data);
    let router_buf = alloc_f32(&device, &router_data);
    let norm_weight_buf = alloc_f32(&device, &norm_weight);
    let per_expert_scale_buf = alloc_f32(&device, &per_expert_scale);
    let expert_ids_buf = device
        .alloc_buffer(top_k * 4, DType::U32, vec![top_k])
        .expect("ids");
    let weights_buf = device
        .alloc_buffer(top_k * 4, DType::F32, vec![top_k])
        .expect("weights");

    let mut encoder = device.command_encoder().expect("encoder");
    moe_gate(
        &mut encoder, &mut registry, device.metal_device(),
        &hidden_buf, &router_buf, &norm_weight_buf, &per_expert_scale_buf,
        &expert_ids_buf, &weights_buf,
        &MoeGateParams { hidden_dim, n_experts, top_k, seq_len: 1, rms_eps },
    ).expect("moe_gate");
    encoder.commit_and_wait().expect("commit");

    let out_ids: &[u32] = expert_ids_buf.as_slice().expect("read ids");
    assert_eq!(out_ids[0], top_indices[0], "top-1 expert should be {}", top_indices[0]);
    assert_eq!(out_ids[1], top_indices[1], "top-2 expert should be {}", top_indices[1]);
    assert_eq!(out_ids[2], top_indices[2], "top-3 expert should be {}", top_indices[2]);

    let out_weights: &[f32] = weights_buf.as_slice().expect("read weights");
    let sum: f32 = out_weights.iter().sum();
    assert!(
        (sum - 1.0).abs() < 1e-4,
        "weights should sum to 1.0, got {}",
        sum
    );
    for k in 0..top_k {
        assert!(
            out_weights[k] > 0.0,
            "weight {} should be positive, got {}",
            k, out_weights[k]
        );
    }
}

#[test]
fn test_moe_gate_128_experts_top8() {
    let (device, mut registry) = setup();

    let hidden_dim = 32;
    let n_experts = 128;
    let top_k = 8;
    let rms_eps = 1e-6f32;

    let hidden_data: Vec<f32> = (0..hidden_dim).map(|i| (i as f32 - 16.0) * 0.01).collect();
    let norm_weight: Vec<f32> = vec![1.0; hidden_dim];
    let per_expert_scale: Vec<f32> = vec![1.0; n_experts];

    let mut router_data = vec![0.0f32; n_experts * hidden_dim];
    for e in 0..n_experts {
        for d in 0..hidden_dim {
            router_data[e * hidden_dim + d] = (e as f32) * 0.01;
        }
    }

    // Reference: RMS-norm + router matmul + top-K
    let normed = rms_norm(&hidden_data, &norm_weight, rms_eps);
    let mut ref_logits = vec![0.0f32; n_experts];
    for e in 0..n_experts {
        for d in 0..hidden_dim {
            ref_logits[e] += normed[d] * router_data[e * hidden_dim + d];
        }
    }
    let mut indexed: Vec<(usize, f32)> = ref_logits.iter().enumerate().map(|(i, &v)| (i, v)).collect();
    indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
    let ref_top_ids: Vec<u32> = indexed[..top_k].iter().map(|&(i, _)| i as u32).collect();

    let hidden_buf = alloc_bf16(&device, &hidden_data);
    let router_buf = alloc_f32(&device, &router_data);
    let norm_weight_buf = alloc_f32(&device, &norm_weight);
    let per_expert_scale_buf = alloc_f32(&device, &per_expert_scale);
    let expert_ids_buf = device
        .alloc_buffer(top_k * 4, DType::U32, vec![top_k])
        .expect("ids");
    let weights_buf = device
        .alloc_buffer(top_k * 4, DType::F32, vec![top_k])
        .expect("weights");

    let mut encoder = device.command_encoder().expect("encoder");
    moe_gate(
        &mut encoder, &mut registry, device.metal_device(),
        &hidden_buf, &router_buf, &norm_weight_buf, &per_expert_scale_buf,
        &expert_ids_buf, &weights_buf,
        &MoeGateParams { hidden_dim, n_experts, top_k, seq_len: 1, rms_eps },
    ).expect("moe_gate");
    encoder.commit_and_wait().expect("commit");

    let out_ids: &[u32] = expert_ids_buf.as_slice().expect("read ids");
    for k in 0..top_k {
        assert_eq!(
            out_ids[k], ref_top_ids[k],
            "top-{} expert mismatch: expected {}, got {}",
            k + 1, ref_top_ids[k], out_ids[k]
        );
    }

    let out_weights: &[f32] = weights_buf.as_slice().expect("read weights");
    let sum: f32 = out_weights.iter().sum();
    assert!(
        (sum - 1.0).abs() < 1e-4,
        "weights should sum to 1.0, got {}",
        sum
    );
    for k in 0..top_k {
        assert!(
            out_weights[k] > 0.0,
            "weight {} should be positive, got {}",
            k, out_weights[k]
        );
    }
}

// ---- Validation tests ----

#[test]
fn test_moe_gate_topk_gt_nexperts() {
    let (device, mut registry) = setup();

    let buf = device.alloc_buffer(64, DType::F32, vec![16]).expect("buf");
    let ids = device.alloc_buffer(64, DType::U32, vec![16]).expect("ids");
    let mut encoder = device.command_encoder().expect("enc");

    let result = moe_gate(
        &mut encoder, &mut registry, device.metal_device(),
        &buf, &buf, &buf, &buf, &ids, &buf,
        &MoeGateParams { hidden_dim: 4, n_experts: 4, top_k: 8, seq_len: 1, rms_eps: 1e-6 },
    );
    assert!(result.is_err(), "top_k > n_experts should error");
}

#[test]
fn test_moe_gate_too_many_experts() {
    let (device, mut registry) = setup();

    let buf = device.alloc_buffer(64, DType::F32, vec![16]).expect("buf");
    let ids = device.alloc_buffer(64, DType::U32, vec![16]).expect("ids");
    let mut encoder = device.command_encoder().expect("enc");

    let result = moe_gate(
        &mut encoder, &mut registry, device.metal_device(),
        &buf, &buf, &buf, &buf, &ids, &buf,
        &MoeGateParams { hidden_dim: 4, n_experts: 256, top_k: 1, seq_len: 1, rms_eps: 1e-6 },
    );
    assert!(result.is_err(), "n_experts > 128 should error");
}