oxidelm 0.2.0

Bare-metal pure-Rust Transformer engine with reverse-mode autograd, BPE tokenization, and wgpu acceleration
Documentation
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
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
//! Offline checkpoint and model inspection without training data.

use crate::memory::{estimate_training_memory, format_bytes, format_memory_report, MemoryEstimate};
use crate::model::{CausalLM, CausalLmConfig};
use crate::optimizer::AdamWConfig;
use crate::tokenizer::{BpeTokenizer, SpecialTokens, BYTE_VOCAB_SIZE, FIRST_MERGE_ID};
use crate::trainer::{
    load_checkpoint, LearningRateSchedule, TrainerConfig, TrainerError,
};
use crate::Tensor;
use std::fmt::Write as _;
use std::path::{Path, PathBuf};

pub const TOKENIZER_SAMPLE: &str = "The quick brown fox jumps over the lazy dog. 0123456789";

#[derive(Clone, Debug)]
pub struct ParameterInfo {
    pub name: String,
    pub shape: Vec<usize>,
    pub numel: usize,
    pub dtype: &'static str,
}

#[derive(Clone, Debug)]
pub struct TokenizerStats {
    pub vocab_size: usize,
    pub merge_count: usize,
    pub special_tokens: SpecialTokens,
    pub byte_token_count: usize,
    pub special_token_count: usize,
    pub merge_token_count: usize,
    pub min_token_bytes: usize,
    pub max_token_bytes: usize,
    pub average_token_bytes: f64,
    pub sample_characters: usize,
    pub sample_bytes: usize,
    pub sample_tokens: usize,
    pub sample_tokens_per_character: f64,
    pub sample_bytes_per_token: f64,
}

#[derive(Clone, Debug)]
pub struct CheckpointReport {
    pub path: PathBuf,
    pub file_bytes: u64,
    pub version: u32,
    pub model_config: CausalLmConfig,
    pub parameters: Vec<ParameterInfo>,
    pub total_parameters: usize,
    pub tie_word_embeddings: bool,
    pub memory: MemoryEstimate,
    pub tokenizer: TokenizerStats,
    pub optimizer_config: AdamWConfig,
    pub optimizer_step: usize,
    pub schedule: LearningRateSchedule,
    pub trainer_config: TrainerConfig,
    pub training_step: usize,
    pub cursor: usize,
    pub tokens_processed: usize,
    pub best_validation_loss: Option<f32>,
    pub rng_model_seed: u64,
    pub rng_dropout_stream: usize,
    pub data_loader_seed: u64,
    pub data_loader_batch_size: usize,
    pub data_loader_sequence_length: usize,
}

/// Collect named parameter tensors in the same order as [`CausalLM::parameters`].
pub fn named_parameters(model: &CausalLM) -> Vec<(String, Tensor)> {
    model.named_parameters()
}

pub fn parameter_infos(model: &CausalLM) -> Vec<ParameterInfo> {
    named_parameters(model)
        .into_iter()
        .map(|(name, tensor)| ParameterInfo {
            name,
            shape: tensor.shape(),
            numel: tensor.numel(),
            dtype: "F32",
        })
        .collect()
}

pub fn estimate_memory(
    _total_parameters: usize,
    config: &CausalLmConfig,
    trainer_config: &TrainerConfig,
) -> MemoryEstimate {
    estimate_training_memory(config, trainer_config)
}

pub fn tokenizer_stats(tokenizer: &BpeTokenizer) -> TokenizerStats {
    let mut min_token_bytes = usize::MAX;
    let mut max_token_bytes = 0usize;
    let mut total_token_bytes = 0usize;
    let mut byte_token_count = 0usize;
    let mut special_token_count = 0usize;
    let mut merge_token_count = 0usize;

    for (token_id, bytes) in tokenizer.vocab.iter().enumerate() {
        let length = bytes.len();
        min_token_bytes = min_token_bytes.min(length);
        max_token_bytes = max_token_bytes.max(length);
        total_token_bytes = total_token_bytes.saturating_add(length);
        if token_id < BYTE_VOCAB_SIZE {
            byte_token_count += 1;
        } else if token_id < FIRST_MERGE_ID {
            special_token_count += 1;
        } else {
            merge_token_count += 1;
        }
    }
    if tokenizer.vocab.is_empty() {
        min_token_bytes = 0;
    }

    let vocab_size = tokenizer.vocab_size();
    let average_token_bytes = if vocab_size == 0 {
        0.0
    } else {
        total_token_bytes as f64 / vocab_size as f64
    };

    let sample_tokens = tokenizer.encode(TOKENIZER_SAMPLE);
    let sample_characters = TOKENIZER_SAMPLE.chars().count();
    let sample_bytes = TOKENIZER_SAMPLE.len();
    let sample_token_count = sample_tokens.len();
    let sample_tokens_per_character = if sample_characters == 0 {
        0.0
    } else {
        sample_token_count as f64 / sample_characters as f64
    };
    let sample_bytes_per_token = if sample_token_count == 0 {
        0.0
    } else {
        sample_bytes as f64 / sample_token_count as f64
    };

    TokenizerStats {
        vocab_size,
        merge_count: tokenizer.merges.len(),
        special_tokens: tokenizer.special_tokens.clone(),
        byte_token_count,
        special_token_count,
        merge_token_count,
        min_token_bytes,
        max_token_bytes,
        average_token_bytes,
        sample_characters,
        sample_bytes,
        sample_tokens: sample_token_count,
        sample_tokens_per_character,
        sample_bytes_per_token,
    }
}

pub fn inspect_checkpoint(path: impl AsRef<Path>) -> Result<CheckpointReport, TrainerError> {
    let path = path.as_ref();
    let file_bytes = std::fs::metadata(path)?.len();
    let checkpoint = load_checkpoint(path)?;
    let parameters = parameter_infos(&checkpoint.model);
    let total_parameters = parameters.iter().map(|parameter| parameter.numel).sum();
    let memory = estimate_training_memory(&checkpoint.model.config, &checkpoint.trainer_config);
    let tokenizer = tokenizer_stats(&checkpoint.tokenizer);
    Ok(CheckpointReport {
        path: path.to_path_buf(),
        file_bytes,
        version: checkpoint.version,
        model_config: checkpoint.model.config.clone(),
        parameters,
        total_parameters,
        tie_word_embeddings: checkpoint.model.config.tie_word_embeddings,
        memory,
        tokenizer,
        optimizer_config: checkpoint.optimizer.config().clone(),
        optimizer_step: checkpoint.optimizer.state().step,
        schedule: checkpoint.schedule,
        trainer_config: checkpoint.trainer_config,
        training_step: checkpoint.training_step,
        cursor: checkpoint.cursor,
        tokens_processed: checkpoint.tokens_processed,
        best_validation_loss: checkpoint.best_validation_loss,
        rng_model_seed: checkpoint.rng.model_seed,
        rng_dropout_stream: checkpoint.rng.dropout_stream,
        data_loader_seed: checkpoint.data_loader.seed,
        data_loader_batch_size: checkpoint.data_loader.batch_size,
        data_loader_sequence_length: checkpoint.data_loader.sequence_length,
    })
}

pub fn format_report(report: &CheckpointReport) -> String {
    let mut output = String::new();
    let _ = writeln!(output, "checkpoint path={}", report.path.display());
    let _ = writeln!(output, "checkpoint file_bytes={}", report.file_bytes);
    let _ = writeln!(output, "checkpoint version={}", report.version);
    let _ = writeln!(
        output,
        "checkpoint training_step={} cursor={} tokens_processed={}",
        report.training_step, report.cursor, report.tokens_processed
    );
    let _ = writeln!(
        output,
        "checkpoint rng model_seed={} dropout_stream={}",
        report.rng_model_seed, report.rng_dropout_stream
    );
    let _ = writeln!(
        output,
        "checkpoint data_loader seed={} cursor={} batch_size={} sequence_length={}",
        report.data_loader_seed,
        report.cursor,
        report.data_loader_batch_size,
        report.data_loader_sequence_length
    );
    match report.best_validation_loss {
        Some(loss) => {
            let _ = writeln!(output, "checkpoint best_validation_loss={loss:.6}");
        }
        None => {
            let _ = writeln!(output, "checkpoint best_validation_loss=none");
        }
    }

    let config = &report.model_config;
    let _ = writeln!(
        output,
        "model vocab_size={} hidden_size={} num_layers={} num_heads={} num_key_value_heads={} intermediate_size={}",
        config.vocab_size,
        config.hidden_size,
        config.num_layers,
        config.num_heads,
        config.num_key_value_heads,
        config.intermediate_size
    );
    let _ = writeln!(
        output,
        "model max_sequence_length={} rms_norm_epsilon={} rope_theta={} eos_token_id={} seed={} tie_word_embeddings={}",
        config.max_sequence_length,
        config.rms_norm_epsilon,
        config.rope_theta,
        config.eos_token_id,
        config.seed,
        config.tie_word_embeddings
    );
    let _ = writeln!(
        output,
        "model attention_dropout={} residual_dropout={} embedding_dropout={} use_bias={}",
        config.attention_dropout,
        config.residual_dropout,
        config.embedding_dropout,
        config.use_bias
    );
    if let Some(scaling) = &config.rope_scaling {
        let _ = writeln!(
            output,
            "model rope_scaling={:?} factor={} original_max_position_embeddings={} attention_factor={:?} beta_fast={} beta_slow={}",
            scaling.kind,
            scaling.factor,
            scaling.original_max_position_embeddings,
            scaling.attention_factor,
            scaling.beta_fast,
            scaling.beta_slow
        );
    }
    let _ = writeln!(output, "parameters total={}", report.total_parameters);
    if report.tie_word_embeddings {
        let _ = writeln!(
            output,
            "parameters note=lm_head.weight tied to tok_embeddings.weight"
        );
    }
    for parameter in &report.parameters {
        let _ = writeln!(
            output,
            "parameter name={} shape={} dtype={} numel={}",
            parameter.name,
            format_shape(&parameter.shape),
            parameter.dtype,
            parameter.numel
        );
    }

    let _ = writeln!(output, "{}", format_memory_report(&report.memory, None));
    let _ = writeln!(
        output,
        "memory gradients={} (included in optimizer)",
        format_bytes(report.memory.gradient_bytes)
    );

    let tokenizer = &report.tokenizer;
    let _ = writeln!(
        output,
        "tokenizer vocab_size={} merges={} byte_tokens={} special_tokens={} merge_tokens={}",
        tokenizer.vocab_size,
        tokenizer.merge_count,
        tokenizer.byte_token_count,
        tokenizer.special_token_count,
        tokenizer.merge_token_count
    );
    let _ = writeln!(
        output,
        "tokenizer specials pad={} unk={} bos={} eos={}",
        tokenizer.special_tokens.pad,
        tokenizer.special_tokens.unk,
        tokenizer.special_tokens.bos,
        tokenizer.special_tokens.eos
    );
    let _ = writeln!(
        output,
        "tokenizer token_bytes min={} max={} avg={:.3}",
        tokenizer.min_token_bytes, tokenizer.max_token_bytes, tokenizer.average_token_bytes
    );
    let _ = writeln!(
        output,
        "tokenizer sample_bytes={} sample_chars={} sample_tokens={} tokens_per_char={:.3} bytes_per_token={:.3}",
        tokenizer.sample_bytes,
        tokenizer.sample_characters,
        tokenizer.sample_tokens,
        tokenizer.sample_tokens_per_character,
        tokenizer.sample_bytes_per_token
    );

    let optimizer = &report.optimizer_config;
    let _ = writeln!(
        output,
        "optimizer step={} beta1={} beta2={} epsilon={} weight_decay={} no_decay_rms_norm={} no_decay_embeddings=[{}] rms_norm_lr_scale={} embedding_lr_scale={}",
        report.optimizer_step,
        optimizer.beta1,
        optimizer.beta2,
        optimizer.epsilon,
        optimizer.weight_decay,
        optimizer.no_decay_rms_norm,
        optimizer.no_decay_embeddings.join(","),
        optimizer.rms_norm_lr_scale,
        optimizer.embedding_lr_scale
    );
    let schedule = &report.schedule;
    let _ = writeln!(
        output,
        "schedule base_lr={} min_lr={} warmup_steps={} total_steps={}",
        schedule.base_learning_rate,
        schedule.minimum_learning_rate,
        schedule.warmup_steps,
        schedule.total_steps
    );
    let trainer = &report.trainer_config;
    let _ = writeln!(
        output,
        "trainer batch_size={} sequence_length={} max_gradient_norm={} log_every={} gradient_accumulation_steps={} activation_checkpointing={} mixed_precision={} amp_format={} initial_loss_scale={}",
        trainer.batch_size,
        trainer.sequence_length,
        trainer.max_gradient_norm,
        trainer.log_every,
        trainer.gradient_accumulation_steps,
        trainer.activation_checkpointing,
        trainer.mixed_precision.enabled,
        trainer.mixed_precision.format.as_str(),
        trainer.mixed_precision.initial_loss_scale
    );
    output
}

fn format_shape(shape: &[usize]) -> String {
    format!(
        "[{}]",
        shape
            .iter()
            .map(ToString::to_string)
            .collect::<Vec<_>>()
            .join(", ")
    )
}

#[cfg(test)]
mod tests {
    use super::{
        estimate_memory, format_report, inspect_checkpoint, named_parameters, parameter_infos,
        tokenizer_stats,
    };
    use crate::memory::count_parameters;
    use crate::model::{CausalLM, CausalLmConfig};
    use crate::optimizer::{AdamW, AdamWConfig};
    use crate::tokenizer::{BpeTokenizer, FIRST_MERGE_ID};
    use crate::trainer::{save_checkpoint, LearningRateSchedule, TrainerConfig};

    fn tiny_config(tie_word_embeddings: bool) -> CausalLmConfig {
        CausalLmConfig {
            vocab_size: 16,
            hidden_size: 8,
            num_layers: 1,
            num_heads: 2,
            num_key_value_heads: 1,
            intermediate_size: 16,
            max_sequence_length: 8,
            rms_norm_epsilon: 1e-5,
            rope_theta: 10_000.0,
            rope_scaling: None,
            eos_token_id: 15,
            seed: 11,
            initializer_std: 0.02,
            tie_word_embeddings,
            attention_dropout: 0.0,
            residual_dropout: 0.0,
            embedding_dropout: 0.0,
            use_bias: false,
        }
    }

    #[test]
    fn named_parameters_match_parameter_list() {
        for tie in [true, false] {
            let model = CausalLM::new(tiny_config(tie)).unwrap();
            let named = named_parameters(&model);
            let parameters = model.parameters();
            assert_eq!(named.len(), parameters.len());
            let total: usize = named.iter().map(|(_, tensor)| tensor.numel()).sum();
            let expected: usize = parameters.iter().map(|parameter| parameter.numel()).sum();
            assert_eq!(total, expected);
            for ((_, named_tensor), parameter) in named.iter().zip(parameters.iter()) {
                assert_eq!(named_tensor.shape(), parameter.shape());
                assert_eq!(named_tensor.numel(), parameter.numel());
            }
        }
    }

    #[test]
    fn memory_estimate_uses_fp32_footprint() {
        let config = tiny_config(true);
        let trainer = TrainerConfig {
            batch_size: 2,
            sequence_length: 4,
            max_gradient_norm: 1.0,
            log_every: 1,
            gradient_accumulation_steps: 1,
            mixed_precision: crate::amp::MixedPrecisionConfig::disabled(),
            activation_checkpointing: false,
            data_parallel_workers: 1,
            low_memory: Default::default(),
        };
        let memory = estimate_memory(100, &config, &trainer);
        assert_eq!(memory.parameter_bytes, count_parameters(&config) * 4);
        assert_eq!(memory.gradient_bytes, memory.parameter_bytes);
        assert_eq!(memory.optimizer_bytes, memory.parameter_bytes * 3);
        let head_dim = config.hidden_size / config.num_heads;
        assert_eq!(
            memory.cache_bytes,
            2 * config.num_layers
                * trainer.batch_size
                * trainer.sequence_length
                * config.num_key_value_heads
                * head_dim
                * 4
        );
        assert!(memory.activation_bytes > 0);
        assert!(memory.staging_bytes > 0);
        assert!(memory.total_bytes >= memory.parameter_bytes);
    }

    #[test]
    fn inspect_checkpoint_reports_totals_without_training_corpus() {
        let directory = std::env::temp_dir().join(format!(
            "oxidelm-inspect-{}",
            std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap()
                .as_nanos()
        ));
        std::fs::create_dir_all(&directory).unwrap();
        let path = directory.join("model.bin");

        let tokenizer = BpeTokenizer::train_from_bytes_with_min_frequency(
            b"hello world hello rust",
            FIRST_MERGE_ID + 8,
            1,
        )
        .unwrap();
        let mut config = tiny_config(true);
        config.vocab_size = tokenizer.vocab_size();
        config.eos_token_id = tokenizer.special_tokens.eos;
        let model = CausalLM::new(config).unwrap();
        let parameters = model.parameters();
        let expected_total: usize = parameters.iter().map(|parameter| parameter.numel()).sum();
        let optimizer = AdamW::new(&parameters, AdamWConfig::default()).unwrap();
        let schedule = LearningRateSchedule {
            base_learning_rate: 3e-4,
            minimum_learning_rate: 3e-5,
            warmup_steps: 1,
            total_steps: 10,
        };
        let trainer_config = TrainerConfig {
            batch_size: 1,
            sequence_length: 4,
            max_gradient_norm: 1.0,
            log_every: 1,
            gradient_accumulation_steps: 1,
            mixed_precision: crate::amp::MixedPrecisionConfig::disabled(),
            activation_checkpointing: false,
            data_parallel_workers: 1,
            low_memory: Default::default(),
        };
        save_checkpoint(
            &path,
            &model,
            &optimizer,
            &tokenizer,
            &schedule,
            &trainer_config,
            7,
            3,
            42,
        )
        .unwrap();

        let report = inspect_checkpoint(&path).unwrap();
        assert_eq!(report.total_parameters, expected_total);
        assert_eq!(report.parameters.len(), parameter_infos(&model).len());
        assert_eq!(report.training_step, 7);
        assert_eq!(report.cursor, 3);
        assert_eq!(report.tokens_processed, 42);
        assert_eq!(report.tokenizer.vocab_size, tokenizer.vocab_size());
        assert_eq!(
            report.tokenizer.merge_count,
            tokenizer_stats(&tokenizer).merge_count
        );
        assert!(report.file_bytes > 0);

        let formatted = format_report(&report);
        assert!(formatted.contains("parameters total="));
        assert!(formatted.contains("tok_embeddings.weight"));
        assert!(formatted.contains("tokenizer vocab_size="));
        assert!(formatted.contains("memory parameters="));
        assert!(formatted.contains("activations="));
        assert!(formatted.contains("staging="));

        let _ = std::fs::remove_dir_all(directory);
    }
}