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
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314


#[cfg(test)]
#[cfg(feature = "cuda")]
mod tests {
    use super::*;
    use crate::cuda::executor::test_fixtures::{
        generate_q4_0_weights, generate_q5_0_weights, generate_q8_0_weights,
    };

    /// Helper to create CudaExecutor for tests
    fn create_executor() -> Option<CudaExecutor> {
        CudaExecutor::new(0).ok()
    }

    // ========================================================================
    // Q8_0 GEMV Tests
    // ========================================================================

    #[test]
    fn test_q8_0_gemv_into_basic() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        // K=256, N=64: 64 output rows, 8 blocks per row (32 elements/block)
        let k = 256u32;
        let n = 64u32;
        let blocks = (n as usize) * (k as usize / 32);
        let weights = generate_q8_0_weights(blocks);

        // Upload weights to GPU
        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let weight_ptr = weight_buf.as_ptr();

        // Create input/output buffers
        let input: Vec<f32> = (0..k as usize).map(|i| (i as f32) * 0.01).collect();
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        // Execute (may fail on PTX generation but exercises path)
        let result = exec.q8_0_gemv_into(weight_ptr, &input_buf, &output_buf, n, k);
        let _ = result;
    }

    #[test]
    fn test_q8_0_gemv_into_large() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        let k = 512u32;
        let n = 128u32;
        let blocks = (n as usize) * (k as usize / 32);
        let weights = generate_q8_0_weights(blocks);

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = vec![0.5f32; k as usize];
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result = exec.q8_0_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    // ========================================================================
    // Q5_0 GEMV Tests
    // ========================================================================

    #[test]
    fn test_q5_0_gemv_into_basic() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        let k = 256u32;
        let n = 64u32;
        let blocks = (n as usize) * (k as usize / 32);
        let weights = generate_q5_0_weights(blocks);

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = (0..k as usize).map(|i| (i as f32) * 0.01).collect();
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result = exec.q5_0_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    #[test]
    fn test_q5_0_gemv_into_large() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        let k = 512u32;
        let n = 128u32;
        let blocks = (n as usize) * (k as usize / 32);
        let weights = generate_q5_0_weights(blocks);

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = vec![0.5f32; k as usize];
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result = exec.q5_0_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    // ========================================================================
    // Q4_0 GEMV Tests
    // ========================================================================

    #[test]
    fn test_q4_0_gemv_into_basic() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        let k = 256u32;
        let n = 64u32;
        let blocks = (n as usize) * (k as usize / 32);
        let weights = generate_q4_0_weights(blocks);

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = (0..k as usize).map(|i| (i as f32) * 0.01).collect();
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result = exec.q4_0_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    #[test]
    fn test_q4_0_gemv_into_single_row() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        let k = 256u32;
        let n = 1u32;
        let blocks = k as usize / 32;
        let weights = generate_q4_0_weights(blocks);

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = vec![1.0f32; k as usize];
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result = exec.q4_0_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    // ========================================================================
    // Q4_1 GEMV Tests
    // ========================================================================

    #[test]
    fn test_q4_1_gemv_into_basic() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        // Q4_1 has same block count as Q4_0 but 20 bytes per block instead of 18
        let k = 256u32;
        let n = 64u32;
        let blocks = (n as usize) * (k as usize / 32);
        // Use Q4_0 weights (18 bytes/block), Q4_1 expects 20 bytes/block
        // The kernel will interpret incorrectly, but we're testing the path
        let weights = generate_q4_0_weights(blocks);

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = (0..k as usize).map(|i| (i as f32) * 0.01).collect();
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result = exec.q4_1_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    // ========================================================================
    // Q5_K GEMV Tests
    // ========================================================================

    #[test]
    fn test_q5k_gemv_into_basic() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        // Q5_K: 176 bytes per 256 elements (super-block format)
        let k = 256u32;
        let n = 64u32;
        // Q5_K needs k to be multiple of 256
        let superblocks = (n as usize) * (k as usize / 256);
        // Simulate Q5K weight data
        let weights = vec![0u8; superblocks * 176];

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = (0..k as usize).map(|i| (i as f32) * 0.01).collect();
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result = exec.q5k_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    // ========================================================================
    // Q6_K GEMV Tests
    // ========================================================================

    #[test]
    fn test_q6k_gemv_into_basic() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        // Q6_K: 210 bytes per 256 elements
        let k = 256u32;
        let n = 64u32;
        let superblocks = (n as usize) * (k as usize / 256);
        let weights = vec![0u8; superblocks * 210];

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = (0..k as usize).map(|i| (i as f32) * 0.01).collect();
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result = exec.q6k_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    #[test]
    fn test_coalesced_q6k_gemv_into_basic() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        let k = 256u32;
        let n = 64u32;
        let superblocks = (n as usize) * (k as usize / 256);
        let weights = vec![0u8; superblocks * 210];

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = (0..k as usize).map(|i| (i as f32) * 0.01).collect();
        let output = vec![0.0f32; n as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result =
            exec.coalesced_q6k_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, n, k);
        let _ = result;
    }

    // ========================================================================
    // Batched Q6K GEMV Tests
    // ========================================================================

    #[test]
    fn test_batched_q6k_gemv_into_basic() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        let m = 4u32; // batch size
        let k = 256u32;
        let n = 64u32;
        let superblocks = (n as usize) * (k as usize / 256);
        let weights = vec![0u8; superblocks * 210];

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        // Batched input: m * k elements
        let input: Vec<f32> = (0..(m * k) as usize).map(|i| (i as f32) * 0.001).collect();
        // Batched output: m * n elements
        let output = vec![0.0f32; (m * n) as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result =
            exec.batched_q6k_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, m, n, k);
        let _ = result;
    }

    #[test]
    fn test_batched_q6k_gemv_into_m8() {
        let Some(mut exec) = create_executor() else {
            return;
        };

        let m = 8u32;
        let k = 256u32;
        let n = 32u32;
        let superblocks = (n as usize) * (k as usize / 256);
        let weights = vec![0u8; superblocks * 210];

        let weight_buf = GpuBuffer::from_host(&exec.context, &weights).unwrap();
        let input: Vec<f32> = vec![0.5f32; (m * k) as usize];
        let output = vec![0.0f32; (m * n) as usize];
        let input_buf = GpuBuffer::from_host(&exec.context, &input).unwrap();
        let output_buf = GpuBuffer::from_host(&exec.context, &output).unwrap();

        let result =
            exec.batched_q6k_gemv_into(weight_buf.as_ptr(), &input_buf, &output_buf, m, n, k);
        let _ = result;
    }
}