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
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480


impl DispatchMetrics {
    /// Histogram bucket boundaries in microseconds
    pub const BUCKET_BOUNDARIES: [u64; 4] = [100, 500, 1000, 5000];

    /// Create new metrics tracker
    #[must_use]
    pub fn new() -> Self {
        Self {
            cpu_dispatches: std::sync::atomic::AtomicUsize::new(0),
            gpu_dispatches: std::sync::atomic::AtomicUsize::new(0),
            cpu_latency_count: std::sync::atomic::AtomicUsize::new(0),
            cpu_latency_sum_us: std::sync::atomic::AtomicU64::new(0),
            gpu_latency_count: std::sync::atomic::AtomicUsize::new(0),
            gpu_latency_sum_us: std::sync::atomic::AtomicU64::new(0),
            cpu_latency_buckets: [
                std::sync::atomic::AtomicUsize::new(0),
                std::sync::atomic::AtomicUsize::new(0),
                std::sync::atomic::AtomicUsize::new(0),
                std::sync::atomic::AtomicUsize::new(0),
                std::sync::atomic::AtomicUsize::new(0),
            ],
            gpu_latency_buckets: [
                std::sync::atomic::AtomicUsize::new(0),
                std::sync::atomic::AtomicUsize::new(0),
                std::sync::atomic::AtomicUsize::new(0),
                std::sync::atomic::AtomicUsize::new(0),
                std::sync::atomic::AtomicUsize::new(0),
            ],
            cpu_latency_min_us: std::sync::atomic::AtomicU64::new(u64::MAX),
            cpu_latency_max_us: std::sync::atomic::AtomicU64::new(0),
            gpu_latency_min_us: std::sync::atomic::AtomicU64::new(u64::MAX),
            gpu_latency_max_us: std::sync::atomic::AtomicU64::new(0),
            cpu_latency_sum_sq_us: std::sync::atomic::AtomicU64::new(0),
            gpu_latency_sum_sq_us: std::sync::atomic::AtomicU64::new(0),
            start_time_ms: std::sync::atomic::AtomicU64::new(
                std::time::SystemTime::now()
                    .duration_since(std::time::UNIX_EPOCH)
                    .map(|d| d.as_millis() as u64)
                    .unwrap_or(0),
            ),
        }
    }

    fn bucket_index(latency_us: u64) -> usize {
        for (i, &boundary) in Self::BUCKET_BOUNDARIES.iter().enumerate() {
            if latency_us < boundary {
                return i;
            }
        }
        4
    }

    /// Record a CPU dispatch
    pub fn record_cpu_dispatch(&self) {
        self.cpu_dispatches
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    /// Record a GPU dispatch
    pub fn record_gpu_dispatch(&self) {
        self.gpu_dispatches
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
    }

    /// Record CPU dispatch latency
    pub fn record_cpu_latency(&self, latency: std::time::Duration) {
        let latency_us = latency.as_micros() as u64;
        self.cpu_latency_count
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_sum_us
            .fetch_add(latency_us, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_buckets[Self::bucket_index(latency_us)]
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_min_us
            .fetch_min(latency_us, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_max_us
            .fetch_max(latency_us, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_sum_sq_us.fetch_add(
            latency_us * latency_us,
            std::sync::atomic::Ordering::Relaxed,
        );
    }

    /// Record GPU dispatch latency
    pub fn record_gpu_latency(&self, latency: std::time::Duration) {
        let latency_us = latency.as_micros() as u64;
        self.gpu_latency_count
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_sum_us
            .fetch_add(latency_us, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_buckets[Self::bucket_index(latency_us)]
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_min_us
            .fetch_min(latency_us, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_max_us
            .fetch_max(latency_us, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_sum_sq_us.fetch_add(
            latency_us * latency_us,
            std::sync::atomic::Ordering::Relaxed,
        );
    }

    /// Get CPU dispatch count
    #[must_use]
    pub fn cpu_dispatches(&self) -> usize {
        self.cpu_dispatches
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get GPU dispatch count
    #[must_use]
    pub fn gpu_dispatches(&self) -> usize {
        self.gpu_dispatches
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get total dispatches
    #[must_use]
    pub fn total_dispatches(&self) -> usize {
        self.cpu_dispatches() + self.gpu_dispatches()
    }

    /// Get GPU dispatch ratio
    #[must_use]
    pub fn gpu_ratio(&self) -> f64 {
        let total = self.total_dispatches();
        if total == 0 {
            0.0
        } else {
            self.gpu_dispatches() as f64 / total as f64
        }
    }

    /// Get CPU latency count
    #[must_use]
    pub fn cpu_latency_count(&self) -> usize {
        self.cpu_latency_count
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get GPU latency count
    #[must_use]
    pub fn gpu_latency_count(&self) -> usize {
        self.gpu_latency_count
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get mean CPU latency in microseconds
    #[must_use]
    pub fn cpu_latency_mean_us(&self) -> f64 {
        let count = self.cpu_latency_count();
        if count == 0 {
            0.0
        } else {
            self.cpu_latency_sum_us
                .load(std::sync::atomic::Ordering::Relaxed) as f64
                / count as f64
        }
    }

    /// Get mean GPU latency in microseconds
    #[must_use]
    pub fn gpu_latency_mean_us(&self) -> f64 {
        let count = self.gpu_latency_count();
        if count == 0 {
            0.0
        } else {
            self.gpu_latency_sum_us
                .load(std::sync::atomic::Ordering::Relaxed) as f64
                / count as f64
        }
    }

    /// Get CPU latency sum
    #[must_use]
    pub fn cpu_latency_sum_us(&self) -> u64 {
        self.cpu_latency_sum_us
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get GPU latency sum
    #[must_use]
    pub fn gpu_latency_sum_us(&self) -> u64 {
        self.gpu_latency_sum_us
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get CPU latency min
    #[must_use]
    pub fn cpu_latency_min_us(&self) -> u64 {
        if self.cpu_latency_count() == 0 {
            0
        } else {
            self.cpu_latency_min_us
                .load(std::sync::atomic::Ordering::Relaxed)
        }
    }

    /// Get CPU latency max
    #[must_use]
    pub fn cpu_latency_max_us(&self) -> u64 {
        self.cpu_latency_max_us
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get GPU latency min
    #[must_use]
    pub fn gpu_latency_min_us(&self) -> u64 {
        if self.gpu_latency_count() == 0 {
            0
        } else {
            self.gpu_latency_min_us
                .load(std::sync::atomic::Ordering::Relaxed)
        }
    }

    /// Get GPU latency max
    #[must_use]
    pub fn gpu_latency_max_us(&self) -> u64 {
        self.gpu_latency_max_us
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get CPU latency variance
    #[must_use]
    pub fn cpu_latency_variance_us(&self) -> f64 {
        let count = self.cpu_latency_count();
        if count < 2 {
            return 0.0;
        }
        let sum = self
            .cpu_latency_sum_us
            .load(std::sync::atomic::Ordering::Relaxed) as f64;
        let sum_sq = self
            .cpu_latency_sum_sq_us
            .load(std::sync::atomic::Ordering::Relaxed) as f64;
        let n = count as f64;
        (sum_sq / n) - (sum / n).powi(2)
    }

    /// Get CPU latency stddev
    #[must_use]
    pub fn cpu_latency_stddev_us(&self) -> f64 {
        self.cpu_latency_variance_us().sqrt()
    }

    /// Get GPU latency variance
    #[must_use]
    pub fn gpu_latency_variance_us(&self) -> f64 {
        let count = self.gpu_latency_count();
        if count < 2 {
            return 0.0;
        }
        let sum = self
            .gpu_latency_sum_us
            .load(std::sync::atomic::Ordering::Relaxed) as f64;
        let sum_sq = self
            .gpu_latency_sum_sq_us
            .load(std::sync::atomic::Ordering::Relaxed) as f64;
        let n = count as f64;
        (sum_sq / n) - (sum / n).powi(2)
    }

    /// Get GPU latency stddev
    #[must_use]
    pub fn gpu_latency_stddev_us(&self) -> f64 {
        self.gpu_latency_variance_us().sqrt()
    }

    /// Get CPU latency histogram buckets
    #[must_use]
    pub fn cpu_latency_buckets(&self) -> [usize; 5] {
        [
            self.cpu_latency_buckets[0].load(std::sync::atomic::Ordering::Relaxed),
            self.cpu_latency_buckets[1].load(std::sync::atomic::Ordering::Relaxed),
            self.cpu_latency_buckets[2].load(std::sync::atomic::Ordering::Relaxed),
            self.cpu_latency_buckets[3].load(std::sync::atomic::Ordering::Relaxed),
            self.cpu_latency_buckets[4].load(std::sync::atomic::Ordering::Relaxed),
        ]
    }

    /// Get GPU latency histogram buckets
    #[must_use]
    pub fn gpu_latency_buckets(&self) -> [usize; 5] {
        [
            self.gpu_latency_buckets[0].load(std::sync::atomic::Ordering::Relaxed),
            self.gpu_latency_buckets[1].load(std::sync::atomic::Ordering::Relaxed),
            self.gpu_latency_buckets[2].load(std::sync::atomic::Ordering::Relaxed),
            self.gpu_latency_buckets[3].load(std::sync::atomic::Ordering::Relaxed),
            self.gpu_latency_buckets[4].load(std::sync::atomic::Ordering::Relaxed),
        ]
    }

    fn estimate_percentile_from_buckets(buckets: &[usize; 5], percentile: f64) -> f64 {
        const BUCKET_UPPER_BOUNDS: [f64; 5] = [100.0, 500.0, 1000.0, 5000.0, 10000.0];
        const BUCKET_LOWER_BOUNDS: [f64; 5] = [0.0, 100.0, 500.0, 1000.0, 5000.0];
        let total: usize = buckets.iter().sum();
        if total == 0 {
            return 0.0;
        }
        let target_rank = (percentile / 100.0) * total as f64;
        let mut cumulative: f64 = 0.0;
        for (i, &count) in buckets.iter().enumerate() {
            let prev_cumulative = cumulative;
            cumulative += count as f64;
            if cumulative >= target_rank {
                if count == 0 {
                    return BUCKET_LOWER_BOUNDS[i];
                }
                let fraction = (target_rank - prev_cumulative) / count as f64;
                return BUCKET_LOWER_BOUNDS[i]
                    + fraction * (BUCKET_UPPER_BOUNDS[i] - BUCKET_LOWER_BOUNDS[i]);
            }
        }
        BUCKET_UPPER_BOUNDS[4]
    }

    /// Get CPU p50 latency
    #[must_use]
    pub fn cpu_latency_p50_us(&self) -> f64 {
        Self::estimate_percentile_from_buckets(&self.cpu_latency_buckets(), 50.0)
    }

    /// Get CPU p95 latency
    #[must_use]
    pub fn cpu_latency_p95_us(&self) -> f64 {
        Self::estimate_percentile_from_buckets(&self.cpu_latency_buckets(), 95.0)
    }

    /// Get CPU p99 latency
    #[must_use]
    pub fn cpu_latency_p99_us(&self) -> f64 {
        Self::estimate_percentile_from_buckets(&self.cpu_latency_buckets(), 99.0)
    }

    /// Get GPU p50 latency
    #[must_use]
    pub fn gpu_latency_p50_us(&self) -> f64 {
        Self::estimate_percentile_from_buckets(&self.gpu_latency_buckets(), 50.0)
    }

    /// Get GPU p95 latency
    #[must_use]
    pub fn gpu_latency_p95_us(&self) -> f64 {
        Self::estimate_percentile_from_buckets(&self.gpu_latency_buckets(), 95.0)
    }

    /// Get GPU p99 latency
    #[must_use]
    pub fn gpu_latency_p99_us(&self) -> f64 {
        Self::estimate_percentile_from_buckets(&self.gpu_latency_buckets(), 99.0)
    }

    /// Get bucket boundaries as strings
    #[must_use]
    pub fn bucket_boundaries_us(&self) -> Vec<String> {
        vec![
            format!("0-{}", Self::BUCKET_BOUNDARIES[0]),
            format!(
                "{}-{}",
                Self::BUCKET_BOUNDARIES[0],
                Self::BUCKET_BOUNDARIES[1]
            ),
            format!(
                "{}-{}",
                Self::BUCKET_BOUNDARIES[1],
                Self::BUCKET_BOUNDARIES[2]
            ),
            format!(
                "{}-{}",
                Self::BUCKET_BOUNDARIES[2],
                Self::BUCKET_BOUNDARIES[3]
            ),
            format!("{}+", Self::BUCKET_BOUNDARIES[3]),
        ]
    }

    /// Get start time
    #[must_use]
    pub fn start_time_ms(&self) -> u64 {
        self.start_time_ms
            .load(std::sync::atomic::Ordering::Relaxed)
    }

    /// Get elapsed seconds
    #[must_use]
    pub fn elapsed_seconds(&self) -> f64 {
        let start = self.start_time_ms();
        let now = std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .map(|d| d.as_millis() as u64)
            .unwrap_or(0);
        (now.saturating_sub(start)) as f64 / 1000.0
    }

    /// Get throughput
    #[must_use]
    pub fn throughput_rps(&self) -> f64 {
        let elapsed = self.elapsed_seconds();
        if elapsed < 0.001 {
            0.0
        } else {
            self.total_dispatches() as f64 / elapsed
        }
    }

    /// Get CPU latency CV
    #[must_use]
    pub fn cpu_latency_cv(&self) -> f64 {
        let mean = self.cpu_latency_mean_us();
        if mean < 0.001 {
            0.0
        } else {
            (self.cpu_latency_stddev_us() / mean) * 100.0
        }
    }

    /// Get GPU latency CV
    #[must_use]
    pub fn gpu_latency_cv(&self) -> f64 {
        let mean = self.gpu_latency_mean_us();
        if mean < 0.001 {
            0.0
        } else {
            (self.gpu_latency_stddev_us() / mean) * 100.0
        }
    }

    /// Get CPU/GPU speedup
    #[must_use]
    pub fn cpu_gpu_speedup(&self) -> f64 {
        let gpu_mean = self.gpu_latency_mean_us();
        if gpu_mean < 0.001 {
            0.0
        } else {
            self.cpu_latency_mean_us() / gpu_mean
        }
    }

    /// Reset all metrics
    pub fn reset(&self) {
        self.cpu_dispatches
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.gpu_dispatches
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_count
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_sum_us
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_count
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_sum_us
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_min_us
            .store(u64::MAX, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_max_us
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_min_us
            .store(u64::MAX, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_max_us
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.cpu_latency_sum_sq_us
            .store(0, std::sync::atomic::Ordering::Relaxed);
        self.gpu_latency_sum_sq_us
            .store(0, std::sync::atomic::Ordering::Relaxed);
        for bucket in &self.cpu_latency_buckets {
            bucket.store(0, std::sync::atomic::Ordering::Relaxed);
        }
        for bucket in &self.gpu_latency_buckets {
            bucket.store(0, std::sync::atomic::Ordering::Relaxed);
        }
        let now = std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .map(|d| d.as_millis() as u64)
            .unwrap_or(0);
        self.start_time_ms
            .store(now, std::sync::atomic::Ordering::Relaxed);
    }
}