realizar 0.8.5

Pure Rust ML inference engine built from scratch - model serving for GGUF and safetensors
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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274


#[test]
#[serial]
fn test_cov020_gemm_fused_size_mismatch() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let m = 32u32;
    let n = 32u32;
    let k = 64u32;

    // A is wrong size
    let a = vec![0.1f32; 100];
    let b = vec![0.1f32; (k * n) as usize];
    let mut c = vec![0.0f32; (m * n) as usize];

    let result = executor.gemm_fused(&a, &b, None, &mut c, m, n, k, 0);

    assert!(result.is_err(), "Should fail with size mismatch");
}

#[test]
#[serial]
fn test_cov020_gemm_fused_bias_size_mismatch() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let m = 32u32;
    let n = 32u32;
    let k = 64u32;

    let a = vec![0.1f32; (m * k) as usize];
    let b = vec![0.1f32; (k * n) as usize];
    let mut c = vec![0.0f32; (m * n) as usize];
    let bias = vec![0.1f32; 10]; // Wrong size, should be n = 32

    let result = executor.gemm_fused(&a, &b, Some(&bias), &mut c, m, n, k, 0);

    assert!(result.is_err(), "Should fail with bias size mismatch");
    let err_msg = format!("{:?}", result.err().unwrap());
    assert!(
        err_msg.contains("Bias") || err_msg.contains("mismatch"),
        "Error should mention bias mismatch: {}",
        err_msg
    );
}

#[test]
#[serial]
fn test_cov020_gemm_fused_valid_no_bias() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let m = 32u32;
    let n = 32u32;
    let k = 64u32;

    let a = vec![0.1f32; (m * k) as usize];
    let b = vec![0.1f32; (k * n) as usize];
    let mut c = vec![0.0f32; (m * n) as usize];

    let result = executor.gemm_fused(&a, &b, None, &mut c, m, n, k, 0);

    assert!(
        result.is_ok(),
        "gemm_fused (no bias) should succeed: {:?}",
        result.err()
    );
}

#[test]
#[serial]
fn test_cov020_gemm_fused_with_bias_and_relu() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let m = 32u32;
    let n = 32u32;
    let k = 64u32;

    let a = vec![0.1f32; (m * k) as usize];
    let b = vec![0.1f32; (k * n) as usize];
    let mut c = vec![0.0f32; (m * n) as usize];
    let bias = vec![0.1f32; n as usize]; // Correct size

    // activation = 1 (ReLU)
    let result = executor.gemm_fused(&a, &b, Some(&bias), &mut c, m, n, k, 1);

    assert!(
        result.is_ok(),
        "gemm_fused with bias+relu should succeed: {:?}",
        result.err()
    );
}

#[test]
#[serial]
fn test_cov020_softmax_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let mut data = vec![1.0f32, 2.0, 3.0, 4.0];

    let result = executor.softmax(&mut data);

    assert!(result.is_ok(), "softmax should succeed: {:?}", result.err());

    // Check output sums to 1
    let sum: f32 = data.iter().sum();
    assert!(
        (sum - 1.0).abs() < 0.01,
        "Softmax should sum to 1, got {}",
        sum
    );

    // Check values are in (0, 1)
    for val in &data {
        assert!(
            *val > 0.0 && *val < 1.0,
            "Softmax values should be in (0, 1): {}",
            val
        );
    }
}

#[test]
#[serial]
fn test_cov020_gemm_cached_async_weight_not_found() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let m = 32u32;
    let n = 32u32;
    let k = 64u32;

    let input_buf = GpuBuffer::<f32>::new(&executor.context, (k * n) as usize).expect("input");
    let output_buf = GpuBuffer::<f32>::new(&executor.context, (m * n) as usize).expect("output");

    let result =
        executor.gemm_cached_async("nonexistent_async_weight", &input_buf, &output_buf, m, n, k);

    assert!(result.is_err(), "Should fail when weight not cached");
    let err_msg = format!("{:?}", result.err().unwrap());
    assert!(
        err_msg.contains("not cached"),
        "Error should mention not cached: {}",
        err_msg
    );
}

#[test]
#[serial]
fn test_cov020_q4k_matvec_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let m = 32u32; // Output dimension
    let k = 256u32; // Input dimension (must be divisible by 256 for Q4K)

    // Q4K format: 144 bytes per 256 values
    let num_superblocks = (m as usize * k as usize + 255) / 256;
    let weight_bytes = num_superblocks * 144;
    let weights = vec![0u8; weight_bytes];

    let input = vec![0.1f32; k as usize];
    let mut output = vec![0.0f32; m as usize];

    let result = executor.q4k_matvec(&weights, &input, &mut output, m, k);

    assert!(
        result.is_ok(),
        "q4k_matvec should succeed: {:?}",
        result.err()
    );
}

#[test]
#[serial]
fn test_cov020_q4k_gemv_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 32u32; // Output dimension
    let k = 256u32; // Input dimension (must be divisible by 256)

    // Q4K format: 144 bytes per 256 values
    let num_superblocks = (n as usize * k as usize + 255) / 256;
    let weight_bytes = num_superblocks * 144;
    let weights = vec![0u8; weight_bytes];

    let input = vec![0.1f32; k as usize];
    let mut output = vec![0.0f32; n as usize];

    let result = executor.q4k_gemv(&weights, &input, &mut output, n, k);

    assert!(
        result.is_ok(),
        "q4k_gemv should succeed: {:?}",
        result.err()
    );
}

#[test]
#[serial]
fn test_cov020_q5k_gemv_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 32u32;
    let k = 256u32;

    // Q5K format: 176 bytes per 256 values
    let num_superblocks = (n as usize * k as usize + 255) / 256;
    let weight_bytes = num_superblocks * 176;
    let weights = vec![0u8; weight_bytes];

    let input = vec![0.1f32; k as usize];
    let mut output = vec![0.0f32; n as usize];

    let result = executor.q5k_gemv(&weights, &input, &mut output, n, k);

    assert!(
        result.is_ok(),
        "q5k_gemv should succeed: {:?}",
        result.err()
    );
}

#[test]
#[serial]
fn test_cov020_q6k_gemv_basic() {
    if !CudaExecutor::is_available() {
        return;
    }
    let mut executor = CudaExecutor::new(0).expect("CUDA executor");

    let n = 32u32;
    let k = 256u32;

    // Q6K format: 210 bytes per 256 values
    let num_superblocks = (n as usize * k as usize + 255) / 256;
    let weight_bytes = num_superblocks * 210;
    let weights = vec![0u8; weight_bytes];

    let input = vec![0.1f32; k as usize];
    let mut output = vec![0.0f32; n as usize];

    let result = executor.q6k_gemv(&weights, &input, &mut output, n, k);

    assert!(
        result.is_ok(),
        "q6k_gemv should succeed: {:?}",
        result.err()
    );
}