mlex 0.1.3

Safe, idiomatic Rust runtime for Apple MLX: quantized LLM inference (Qwen, Gemma4, NemotronH, DharaAR, ...) with multi-modal support
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

// Copyright © 2023 Apple Inc.

#include "doctest/doctest.h"

#include "mlx/mlx.h"
#include "mlx/primitives.h"

using namespace mlx::core;

void test_arg_reduce_small(
    Device d,
    const array& x,
    ArgReduce::ReduceType r,
    Shape out_shape,
    int axis,
    std::vector<int> expected_output) {
  auto s = default_stream(d);
  auto y =
      array(out_shape, uint32, std::make_shared<ArgReduce>(s, r, axis), {x});
  y.eval();
  const uint32_t* ydata = y.data<uint32_t>();
  for (int i = 0; i < y.size(); i++) {
    CHECK_EQ(expected_output[i], ydata[i]);
  }
}

void test_arg_reduce_against_cpu(
    const array& x,
    ArgReduce::ReduceType r,
    Shape out_shape,
    int axis) {
  auto y1 = array(
      out_shape,
      uint32,
      std::make_shared<ArgReduce>(default_stream(Device::cpu), r, axis),
      {x});
  auto y2 = array(
      out_shape,
      uint32,
      std::make_shared<ArgReduce>(default_stream(Device::gpu), r, axis),
      {x});
  y1.eval();
  y2.eval();
  CHECK(array_equal(y1, y2).item<bool>());
}

TEST_CASE("test arg reduce small") {
  auto x = array(
      {0, 2, 1, 7, 5, -5, 0, 2, 1, 7, 5, -5,
       0, 2, 1, 7, 5, -5, 0, 2, 1, 7, 5, -5},
      {2, 3, 4});
  test_arg_reduce_small(
      Device::cpu, x, ArgReduce::ArgMin, {2, 3}, 2, {0, 1, 3, 0, 1, 3});
  test_arg_reduce_small(
      Device::cpu, x, ArgReduce::ArgMin, {2, 4}, 1, {0, 1, 1, 2, 0, 1, 1, 2});
  test_arg_reduce_small(
      Device::cpu,
      x,
      ArgReduce::ArgMin,
      {3, 4},
      0,
      {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0});
  test_arg_reduce_small(
      Device::cpu, x, ArgReduce::ArgMax, {2, 3}, 2, {3, 0, 1, 3, 0, 1});
  test_arg_reduce_small(
      Device::cpu, x, ArgReduce::ArgMax, {2, 4}, 1, {1, 2, 2, 0, 1, 2, 2, 0});
  test_arg_reduce_small(
      Device::cpu,
      x,
      ArgReduce::ArgMax,
      {3, 4},
      0,
      {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0});

  if (!metal::is_available()) {
    INFO("Skipping arg reduction gpu tests");
    return;
  }

  test_arg_reduce_small(
      Device::gpu, x, ArgReduce::ArgMin, {2, 3}, 2, {0, 1, 3, 0, 1, 3});
  test_arg_reduce_small(
      Device::gpu, x, ArgReduce::ArgMin, {2, 4}, 1, {0, 1, 1, 2, 0, 1, 1, 2});
  test_arg_reduce_small(
      Device::gpu,
      x,
      ArgReduce::ArgMin,
      {3, 4},
      0,
      {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0});
  test_arg_reduce_small(
      Device::gpu, x, ArgReduce::ArgMax, {2, 3}, 2, {3, 0, 1, 3, 0, 1});
  test_arg_reduce_small(
      Device::gpu, x, ArgReduce::ArgMax, {2, 4}, 1, {1, 2, 2, 0, 1, 2, 2, 0});
  test_arg_reduce_small(
      Device::gpu,
      x,
      ArgReduce::ArgMax,
      {3, 4},
      0,
      {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0});
}

TEST_CASE("test arg reduce against cpu") {
  if (!metal::is_available()) {
    INFO("Skipping arg reduction gpu tests");
    return;
  }

  auto x = random::uniform(array(0.0), array(1.0), {127, 92, 55});
  x.eval();
  test_arg_reduce_against_cpu(x, ArgReduce::ArgMin, {127, 92}, 2);
  test_arg_reduce_against_cpu(x, ArgReduce::ArgMin, {127, 55}, 1);
  test_arg_reduce_against_cpu(x, ArgReduce::ArgMin, {92, 55}, 0);
  test_arg_reduce_against_cpu(x, ArgReduce::ArgMax, {127, 92}, 2);
  test_arg_reduce_against_cpu(x, ArgReduce::ArgMax, {127, 55}, 1);
  test_arg_reduce_against_cpu(x, ArgReduce::ArgMax, {92, 55}, 0);

  auto y = random::uniform(array(0.0), array(1.0), {1234});
  y.eval();
  test_arg_reduce_against_cpu(y, ArgReduce::ArgMin, {}, 0);
  test_arg_reduce_against_cpu(y, ArgReduce::ArgMax, {}, 0);
}

void test_arg_reduce_small_bool(
    Device d,
    ArgReduce::ReduceType r,
    Shape out_shape,
    int axis,
    std::vector<int> expected_output) {
  auto s = default_stream(d);
  auto x = array(
      {0, 2, 1, 7, 5, -5, 0, 2, 1, 7, 5, -5,
       0, 2, 1, 7, 5, -5, 0, 2, 1, 7, 5, -5},
      {2, 3, 4});
  x.eval();
  auto y =
      array(out_shape, uint32, std::make_shared<ArgReduce>(s, r, axis), {x});
  y.eval();
  const uint32_t* ydata = y.data<uint32_t>();
  for (int i = 0; i < y.size(); i++) {
    CHECK_EQ(expected_output[i], ydata[i]);
  }
}

TEST_CASE("test arg reduce bool") {
  if (!metal::is_available()) {
    INFO("Skipping arg reduction gpu tests");
    return;
  }
  auto x = array(
      {false, true,  true,  false, false, false, false, true,
       true,  false, true,  true,  false, true,  true,  false,
       false, false, false, true,  true,  false, true,  true},
      {2, 3, 4});
  x.eval();
  test_arg_reduce_small(
      Device::gpu, x, ArgReduce::ArgMin, {2, 3}, 2, {0, 0, 1, 0, 0, 1});
  test_arg_reduce_small(
      Device::gpu, x, ArgReduce::ArgMin, {2, 4}, 1, {0, 1, 1, 0, 0, 1, 1, 0});
  test_arg_reduce_small(
      Device::gpu,
      x,
      ArgReduce::ArgMin,
      {3, 4},
      0,
      {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0});
  test_arg_reduce_small(
      Device::gpu, x, ArgReduce::ArgMax, {2, 3}, 2, {1, 3, 0, 1, 3, 0});
  test_arg_reduce_small(
      Device::gpu, x, ArgReduce::ArgMax, {2, 4}, 1, {2, 0, 0, 1, 2, 0, 0, 1});
  test_arg_reduce_small(
      Device::gpu,
      x,
      ArgReduce::ArgMax,
      {3, 4},
      0,
      {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0});
}

TEST_CASE("test arg reduce edge cases") {
  auto a = argmin(array(1.0));
  CHECK_EQ(a.item<uint32_t>(), 0);
  auto b = argmax(array(1.0));
  CHECK_EQ(b.item<uint32_t>(), 0);
  CHECK_THROWS(argmin(array({})));
  CHECK_THROWS(argmax(array({})));
}

TEST_CASE("test arg reduce irregular strides") {
  auto x = array(
      {0, 2, 1, 7, 5, -5, 0, 2, 1, 7, 5, -5,
       0, 2, 1, 7, 5, -5, 0, 2, 1, 7, 5, -5},
      {2, 3, 4});
  x = transpose(x, {2, 0, 1});
  x.eval();
  test_arg_reduce_small(
      Device::cpu, x, ArgReduce::ArgMin, {4, 2}, 2, {0, 0, 1, 1, 1, 1, 2, 2});

  if (!metal::is_available()) {
    INFO("Skipping arg reduction gpu tests");
    return;
  }
}