realizar 0.8.4

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
315
316


// ============================================================================
// Tests (PMAT-802: T-COV-95)
// ============================================================================

#[cfg(all(test, feature = "cuda"))]
mod tests {
    use super::*;

    // =========================================================================
    // Constructor and Builder Tests
    // =========================================================================

    #[test]
    fn test_cuda_backend_new() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert_eq!(backend.m, 1);
        assert_eq!(backend.n, 4096);
        assert_eq!(backend.k, 4096);
        assert_eq!(backend.head_dim, 64);
        assert_eq!(backend.num_heads, 32); // default
        assert_eq!(backend.max_seq_len, 2048); // default
    }

    #[test]
    fn test_cuda_backend_with_num_heads() {
        let backend = CudaBackend::new(1, 4096, 4096, 64).with_num_heads(16);
        assert_eq!(backend.num_heads, 16);
    }

    #[test]
    fn test_cuda_backend_with_max_seq_len() {
        let backend = CudaBackend::new(1, 4096, 4096, 64).with_max_seq_len(4096);
        assert_eq!(backend.max_seq_len, 4096);
    }

    #[test]
    fn test_cuda_backend_builder_chain() {
        let backend = CudaBackend::new(1, 4096, 4096, 128)
            .with_num_heads(8)
            .with_max_seq_len(1024);
        assert_eq!(backend.num_heads, 8);
        assert_eq!(backend.max_seq_len, 1024);
        assert_eq!(backend.head_dim, 128);
    }

    #[test]
    fn test_cuda_backend_clone() {
        let backend = CudaBackend::new(2, 1024, 2048, 64);
        let cloned = backend.clone();
        assert_eq!(cloned.m, 2);
        assert_eq!(cloned.n, 1024);
        assert_eq!(cloned.k, 2048);
    }

    #[test]
    fn test_cuda_backend_debug() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        let debug_str = format!("{:?}", backend);
        assert!(debug_str.contains("CudaBackend"));
        assert!(debug_str.contains("4096"));
    }

    // =========================================================================
    // Q4_K Kernel Tests
    // =========================================================================

    #[test]
    fn test_q4k_gemm_kernel_name() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert_eq!(backend.q4k_gemm_kernel_name(), "q4k_gemm_fused");
    }

    #[test]
    fn test_q4k_blocks_per_row() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert_eq!(backend.q4k_blocks_per_row(), 128); // 4096 / 32
    }

    #[test]
    fn test_q4k_blocks_per_row_small() {
        let backend = CudaBackend::new(1, 1024, 256, 64);
        assert_eq!(backend.q4k_blocks_per_row(), 8); // 256 / 32
    }

    #[test]
    fn test_q4k_weight_bytes() {
        let backend = CudaBackend::new(1, 1024, 1024, 64);
        // blocks_per_row = 1024 / 32 = 32
        // bytes_per_row = 32 * 18 = 576
        // total = 1024 * 576 = 589,824
        assert_eq!(backend.q4k_weight_bytes(), 589_824);
    }

    #[test]
    fn test_q4k_weight_bytes_large() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        // blocks_per_row = 4096 / 32 = 128
        // bytes_per_row = 128 * 18 = 2304
        // total = 4096 * 2304 = 9,437,184
        assert_eq!(backend.q4k_weight_bytes(), 9_437_184);
    }

    #[test]
    fn test_q4k_gemm_ptx_generation() {
        let backend = CudaBackend::new(1, 1024, 256, 64);
        let ptx = backend.q4k_gemm_ptx();
        assert!(!ptx.is_empty());
        assert!(ptx.contains(".version"));
    }

    #[test]
    fn test_q4k_gemm_ptx_caching() {
        let backend = CudaBackend::new(1, 1024, 256, 64);
        let ptx1 = backend.q4k_gemm_ptx();
        let ptx2 = backend.q4k_gemm_ptx(); // Should use cache
        assert_eq!(ptx1, ptx2);
    }

    // =========================================================================
    // FlashAttention Tests
    // =========================================================================

    #[test]
    fn test_flash_attention_kernel_name_causal() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert_eq!(
            backend.flash_attention_kernel_name(true),
            "flash_attention_causal"
        );
    }

    #[test]
    fn test_flash_attention_kernel_name_non_causal() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert_eq!(
            backend.flash_attention_kernel_name(false),
            "flash_attention"
        );
    }

    #[test]
    fn test_flash_attention_smem_bytes() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        // tile_q = 64, tile_kv = 64, d = 64
        // smem = (64*64 + 64*64*2) * 4 = (4096 + 8192) * 4 = 49152
        assert_eq!(backend.flash_attention_smem_bytes(), 49152);
    }

    #[test]
    fn test_flash_attention_smem_bytes_large_head_dim() {
        let backend = CudaBackend::new(1, 4096, 4096, 128);
        // tile_q = 64, tile_kv = 64, d = 128
        // smem = (64*128 + 64*128*2) * 4 = (8192 + 16384) * 4 = 98304
        assert_eq!(backend.flash_attention_smem_bytes(), 98304);
    }

    #[test]
    fn test_flash_attention_ptx_generation() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        let ptx = backend.flash_attention_ptx(512, 64, true);
        assert!(!ptx.is_empty());
        assert!(ptx.contains(".version"));
    }

    #[test]
    fn test_flash_attention_causal_ptx_caching() {
        let backend = CudaBackend::new(1, 4096, 4096, 64).with_max_seq_len(512);
        let ptx1 = backend.flash_attention_causal_ptx();
        let ptx2 = backend.flash_attention_causal_ptx(); // Should use cache
        assert_eq!(ptx1, ptx2);
    }

    // =========================================================================
    // KV Cache Tests
    // =========================================================================

    #[test]
    fn test_kv_cache_bytes_per_layer() {
        let backend = CudaBackend::new(1, 4096, 4096, 64)
            .with_num_heads(32)
            .with_max_seq_len(2048);
        // K: 32 * 2048 * 64 * 4 = 16,777,216
        // V: 32 * 2048 * 64 * 4 = 16,777,216
        // Total: 33,554,432
        assert_eq!(backend.kv_cache_bytes_per_layer(), 33_554_432);
    }

    #[test]
    fn test_kv_cache_total_bytes() {
        let backend = CudaBackend::new(1, 4096, 4096, 64)
            .with_num_heads(32)
            .with_max_seq_len(2048);
        let per_layer = backend.kv_cache_bytes_per_layer();
        assert_eq!(backend.kv_cache_total_bytes(22), per_layer * 22);
    }

    #[test]
    fn test_kv_cache_page_tokens() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert_eq!(backend.kv_cache_page_tokens(), 64);
    }

    #[test]
    fn test_kv_cache_pages_needed_exact() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        // 64 tokens = 1 page
        assert_eq!(backend.kv_cache_pages_needed(64), 1);
        // 128 tokens = 2 pages
        assert_eq!(backend.kv_cache_pages_needed(128), 2);
    }

    #[test]
    fn test_kv_cache_pages_needed_round_up() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        // 65 tokens = 2 pages (rounds up)
        assert_eq!(backend.kv_cache_pages_needed(65), 2);
        // 100 tokens = 2 pages
        assert_eq!(backend.kv_cache_pages_needed(100), 2);
        // 129 tokens = 3 pages
        assert_eq!(backend.kv_cache_pages_needed(129), 3);
    }

    #[test]
    fn test_kv_cache_pages_needed_large() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        // 2048 tokens = 32 pages
        assert_eq!(backend.kv_cache_pages_needed(2048), 32);
    }

    // =========================================================================
    // Launch Configuration Tests
    // =========================================================================

    #[test]
    fn test_q4k_gemm_launch_config() {
        let backend = CudaBackend::new(1, 1024, 4096, 64);
        let (grid, block) = backend.q4k_gemm_launch_config();
        // grid = (n/32, m/32, 1) = (1024/32, 1/32 rounded up, 1) = (32, 1, 1)
        assert_eq!(grid, (32, 1, 1));
        // block = (32*32, 1, 1) = (1024, 1, 1)
        assert_eq!(block, (1024, 1, 1));
    }

    #[test]
    fn test_flash_attention_launch_config() {
        let backend = CudaBackend::new(1, 4096, 4096, 64).with_num_heads(8);
        let (grid, block) = backend.flash_attention_launch_config(256);
        // grid = (256/64, 8, 1) = (4, 8, 1)
        assert_eq!(grid, (4, 8, 1));
        // block = (64 * 64, 1, 1) = (4096, 1, 1)
        assert_eq!(block, (4096, 1, 1));
    }

    // =========================================================================
    // Validation Tests
    // =========================================================================

    #[test]
    fn test_validate_dimensions_valid() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert!(backend.validate_dimensions());
    }

    #[test]
    fn test_validate_dimensions_k_not_divisible() {
        let backend = CudaBackend::new(1, 4096, 4095, 64); // k not divisible by 32
        assert!(!backend.validate_dimensions());
    }

    #[test]
    fn test_validate_dimensions_head_dim_not_power_of_two() {
        let backend = CudaBackend::new(1, 4096, 4096, 80); // 80 is not power of 2
        assert!(!backend.validate_dimensions());
    }

    #[test]
    fn test_validate_dimensions_zero_m() {
        let backend = CudaBackend::new(0, 4096, 4096, 64);
        assert!(!backend.validate_dimensions());
    }

    #[test]
    fn test_validate_dimensions_zero_n() {
        let backend = CudaBackend::new(1, 0, 4096, 64);
        assert!(!backend.validate_dimensions());
    }

    #[test]
    fn test_validate_dimensions_zero_k() {
        let backend = CudaBackend::new(1, 4096, 0, 64);
        assert!(!backend.validate_dimensions());
    }

    #[test]
    fn test_validate_dimensions_zero_head_dim() {
        let backend = CudaBackend::new(1, 4096, 4096, 0);
        assert!(!backend.validate_dimensions());
    }

    // =========================================================================
    // PTX Target Tests
    // =========================================================================

    #[test]
    fn test_ptx_target() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert_eq!(backend.ptx_target(), "sm_89");
    }

    #[test]
    fn test_ptx_version() {
        let backend = CudaBackend::new(1, 4096, 4096, 64);
        assert_eq!(backend.ptx_version(), (8, 0));
    }
}