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
315
316
317
318
319
320
321
322
323
324
325
326
327

#[cfg(test)]
mod tests {
    use crate::brick::*;

    // T030: BrickAssertion bounds check - edge values
    #[test]
    fn t030_assertion_bounds_edge() {
        let assertion = BrickAssertion::bounds(-1.0, 1.0);

        // Exact boundary values should pass
        assert!(assertion.check_f32(&[-1.0, 0.0, 1.0], true).is_ok());

        // Just outside should fail (below)
        let result = assertion.check_f32(&[-1.01], true);
        assert!(result.is_err());

        // Just outside should fail (above)
        let result = assertion.check_f32(&[1.01], true);
        assert!(result.is_err());
    }

    // T031: FlashAttentionBrick with single sequence position
    #[test]
    fn t031_flash_attention_single_pos() {
        let brick = FlashAttentionBrick::new(2, 1, 4);
        let seq_len = 1;

        let query = vec![1.0f32; 2 * 4]; // 2 heads * 4 dim
        let keys = vec![1.0f32; 4]; // 1 seq * 1 kv_head * 4 dim
        let values = vec![1.0f32; 4];

        let output = brick
            .forward(&query, &keys, &values, seq_len)
            .expect("forward failed");
        assert_eq!(output.len(), 8);

        // With uniform inputs, output should be uniform
        for &v in &output {
            assert!(!v.is_nan());
            assert!(v.is_finite());
        }
    }

    // T032: CoalescedDp4aBrick FLOPS calculation
    #[test]
    #[cfg(feature = "cuda")]
    fn t032_coalesced_dp4a_flops() {
        let brick = CoalescedDp4aBrick::new(1024, 256);
        let flops = brick.flops();

        // GEMV: 2 * K * N
        assert_eq!(flops, 2 * 1024 * 256);
    }

    // T033: FusedFfnBrick with negative values
    #[test]
    #[cfg(feature = "cuda")]
    fn t033_fused_ffn_negative_values() {
        let brick = FusedFfnBrick::new(2, 4);

        let input = vec![-1.0f32; 2];
        let gate_proj = vec![1.0f32; 8];
        let up_proj = vec![1.0f32; 8];
        let down_proj = vec![1.0f32; 8];

        let output = brick
            .forward(&input, &gate_proj, &up_proj, &down_proj)
            .expect("forward failed");

        // SiLU of negative value should be small but not NaN
        for &v in &output {
            assert!(!v.is_nan());
            assert!(v.is_finite());
        }
    }

    // T034: CoalescedDp4aBrick with all-zero weights
    #[test]
    #[cfg(feature = "cuda")]
    fn t034_coalesced_dp4a_zero_weights() {
        let brick = CoalescedDp4aBrick::new(256, 4);

        let input_q8 = vec![1i8; 256];
        let input_scale = 1.0;
        // All 0x88 = nibbles (8,8) centered to (0,0) - effectively zero weights
        let weights_q4 = vec![0x88u8; 512];
        let weight_scales = vec![0.1f32; 4];

        let output = brick
            .forward(&input_q8, input_scale, &weights_q4, &weight_scales)
            .expect("forward failed");

        // With centered zero weights, output should be zero
        for &v in &output {
            assert_eq!(v, 0.0);
        }
    }

    // T035: ActivationQuantBrick per-channel mode
    #[test]
    fn t035_activation_quant_per_channel() {
        let per_tensor = ActivationQuantBrick::new(1024);
        let per_channel = ActivationQuantBrick::with_per_channel(1024);

        assert!(!per_tensor.per_channel);
        assert!(per_channel.per_channel);

        // Per-channel should have different (tighter) budget
        assert!(per_channel.budget().us_per_token > per_tensor.budget().us_per_token);
    }

    // T036: TransformerLayerBrick total_budget calculation
    #[test]
    fn t036_transformer_layer_total_budget() {
        let layer = TransformerLayerBrick::from_config(0, 64, 4, 2, 256, 1e-5, 10000.0, 0);

        let total = layer.total_budget_us();

        // Total should equal sum of components
        let sum = layer.attn_norm.budget().us_per_token
            + layer.qkv.budget().us_per_token
            + layer.rope.budget().us_per_token
            + layer.attention.budget().us_per_token
            + layer.o_proj.budget().us_per_token
            + layer.ffn_norm.budget().us_per_token
            + layer.ffn.budget().us_per_token;

        assert!((total - sum).abs() < 0.001);
    }

    // T037: BenchmarkReport display format
    #[test]
    fn t037_benchmark_report_display() {
        let report = BenchmarkReport {
            brick_name: "test_brick".to_string(),
            mean_us: 50.5,
            std_us: 5.2,
            cv: 0.103,
            p50_us: 50.0,
            p99_us: 62.0,
            tokens_per_sec: 19802.0,
            budget_us: 100.0,
            budget_met: true,
            statistically_valid: true,
        };

        let display = format!("{}", report);
        assert!(display.contains("test_brick"));
        assert!(display.contains("50.5"));
        assert!(display.contains("PASS"));
    }

    // T038: BenchmarkReport display FAIL case
    #[test]
    fn t038_benchmark_report_display_fail() {
        let report = BenchmarkReport {
            brick_name: "test".to_string(),
            mean_us: 150.0,
            std_us: 10.0,
            cv: 0.067,
            p50_us: 150.0,
            p99_us: 180.0,
            tokens_per_sec: 6667.0,
            budget_us: 100.0,
            budget_met: false,
            statistically_valid: true,
        };

        let display = format!("{}", report);
        assert!(display.contains("FAIL"));
    }

    // T039: FlashAttentionBrick GQA with different group sizes
    #[test]
    fn t039_flash_attention_gqa_groups() {
        // 8 query heads, 2 kv heads = group size 4
        let brick1 = FlashAttentionBrick::new(8, 2, 64);
        assert_eq!(brick1.group_size(), 4);

        // 8 query heads, 8 kv heads = group size 1 (MHA)
        let brick2 = FlashAttentionBrick::new(8, 8, 64);
        assert_eq!(brick2.group_size(), 1);

        // 8 query heads, 1 kv head = group size 8 (MQA)
        let brick3 = FlashAttentionBrick::new(8, 1, 64);
        assert_eq!(brick3.group_size(), 8);
    }

    // T040: ActivationQuantBrick large dimension
    #[test]
    fn t040_activation_quant_large_dim() {
        let brick = ActivationQuantBrick::new(4096);
        let input: Vec<f32> = (0..4096).map(|i| (i as f32 - 2048.0) / 100.0).collect();

        let (quants, scales) = brick.quantize(&input).expect("quantize failed");
        assert_eq!(quants.len(), 4096);
        assert_eq!(scales.len(), 128); // 4096 / 32 = 128 blocks

        let output = brick
            .dequantize(&quants, &scales)
            .expect("dequantize failed");
        assert_eq!(output.len(), 4096);

        // Verify roundtrip error is reasonable
        let error = brick
            .measure_error(&input, &quants, &scales)
            .expect("measure failed");
        assert!(error < 0.05, "Error {} should be < 5%", error);
    }

    // T041: QkvBrick with custom budget
    #[test]
    fn t041_qkv_brick_with_budget() {
        let custom_budget = TokenBudget::from_latency(20.0);
        let brick = QkvBrick::new(64, 64, 32, 32).with_budget(custom_budget);

        assert!((brick.budget().us_per_token - 20.0).abs() < 0.001);
    }

    // T042: RopeBrick with custom budget
    #[test]
    fn t042_rope_brick_with_budget() {
        let custom_budget = TokenBudget::from_latency(5.0);
        let brick = RopeBrick::new(64, 8, 10000.0, 0).with_budget(custom_budget);

        assert!((brick.budget().us_per_token - 5.0).abs() < 0.001);
    }

    // T043: AttentionBrick with custom budget
    #[test]
    fn t043_attention_brick_with_budget() {
        let custom_budget = TokenBudget::from_latency(25.0);
        let brick = AttentionBrick::new(8, 2, 64).with_budget(custom_budget);

        assert!((brick.budget().us_per_token - 25.0).abs() < 0.001);
    }

    // T044: OProjBrick with custom budget
    #[test]
    fn t044_oproj_brick_with_budget() {
        let custom_budget = TokenBudget::from_latency(8.0);
        let brick = OProjBrick::new(512, 64).with_budget(custom_budget);

        assert!((brick.budget().us_per_token - 8.0).abs() < 0.001);
    }

    // T045: FfnBrick with custom budget
    #[test]
    fn t045_ffn_brick_with_budget() {
        let custom_budget = TokenBudget::from_latency(30.0);
        let brick = FfnBrick::new(64, 256).with_budget(custom_budget);

        assert!((brick.budget().us_per_token - 30.0).abs() < 0.001);
    }

    // T047: TokenBudget edge case - very small latency
    #[test]
    fn t047_token_budget_small_latency() {
        let budget = TokenBudget::from_latency(0.001); // 1 nanosecond
        assert_eq!(budget.tokens_per_sec, 1_000_000_000.0);
        assert!(budget.is_met(0.0005));
        assert!(!budget.is_met(0.002));
    }

    // T048: TokenBudget edge case - very high throughput
    #[test]
    fn t048_token_budget_high_throughput() {
        let budget = TokenBudget::from_throughput(1_000_000.0); // 1M tokens/sec
        assert_eq!(budget.us_per_token, 1.0);
    }

    // T049: BrickVerification accumulate multiple failures
    #[test]
    fn t049_brick_verification_multiple_failures() {
        let mut v = BrickVerification::pass();
        assert!(v.is_valid);

        v.add("check1", false, "failed 1");
        assert!(!v.is_valid);

        v.add("check2", true, "passed");
        assert!(!v.is_valid); // Still invalid

        v.add("check3", false, "failed 3");
        assert!(!v.is_valid);

        assert_eq!(v.results.len(), 3);
    }

    // T050: ComputeBrick verify with zero budget
    #[test]
    fn t050_compute_brick_verify_zero_budget() {
        // Create a brick and manually set zero budget
        let mut brick = RmsNormBrick::new(vec![1.0; 4], 1e-5);
        brick.budget = TokenBudget::from_latency(0.0);

        // Verify should fail due to zero budget
        // Note: This creates a 0/0 situation in throughput
        // but the verify() method checks us_per_token <= 0.0
        let verification = brick.verify();
        assert!(!verification.is_valid);
    }

    // T051: FlashAttentionBrick with varying sequence lengths
    #[test]
    fn t051_flash_attention_varying_seq() {
        let brick = FlashAttentionBrick::new(4, 2, 8);

        for seq_len in [1, 2, 4, 8, 16, 32] {
            let query = vec![1.0f32; 4 * 8];
            let keys = vec![0.5f32; seq_len * 2 * 8];
            let values = vec![0.25f32; seq_len * 2 * 8];

            let output = brick
                .forward(&query, &keys, &values, seq_len)
                .unwrap_or_else(|_| panic!("forward failed for seq_len={}", seq_len));

            assert_eq!(output.len(), 4 * 8);
            assert!(output.iter().all(|&v| !v.is_nan()));
        }
    }
include!("tests_f001_brick_f002.rs");
include!("tests_f056_fused_f057.rs");
include!("tests_r006_fused.rs");
include!("tests_f083_timing_f084.rs");
include!("tests_layer_timing.rs");
}