@mlx-node/trl 0.0.12 → 0.0.15
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/package.json +6 -4
- package/src/data/dataset.ts +193 -0
- package/src/data/sft-dataset.ts +562 -0
- package/src/index.ts +139 -0
- package/src/trainers/grpo-trainer.ts +2151 -0
- package/src/trainers/sft-config.ts +250 -0
- package/src/trainers/sft-trainer.ts +679 -0
- package/src/trainers/training-logger.ts +830 -0
- package/src/types.ts +70 -0
- package/src/utils/path-security.ts +88 -0
- package/src/utils/xml-parser.ts +209 -0
|
@@ -0,0 +1,830 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Unified Training Logger
|
|
3
|
+
*
|
|
4
|
+
* A high-level logger abstraction that handles TUI/console mode automatically.
|
|
5
|
+
* Eliminates verbose conditional checks throughout the codebase.
|
|
6
|
+
*
|
|
7
|
+
* @example
|
|
8
|
+
* ```typescript
|
|
9
|
+
* const logger = createTrainingLogger();
|
|
10
|
+
*
|
|
11
|
+
* logger.info('Loading model...'); // Console OR TUI log
|
|
12
|
+
* logger.status('loading', 'Loading...'); // TUI header update
|
|
13
|
+
* logger.step(metrics); // Training step
|
|
14
|
+
* logger.checkpoint(path, step); // Checkpoint saved
|
|
15
|
+
* ```
|
|
16
|
+
*/
|
|
17
|
+
|
|
18
|
+
import { appendFileSync, mkdirSync } from 'node:fs';
|
|
19
|
+
import { join } from 'node:path';
|
|
20
|
+
|
|
21
|
+
// ============================================================================
|
|
22
|
+
// Types
|
|
23
|
+
// ============================================================================
|
|
24
|
+
|
|
25
|
+
/**
|
|
26
|
+
* Training metrics from a training step
|
|
27
|
+
*
|
|
28
|
+
* Supports both SFT and GRPO training:
|
|
29
|
+
* - SFT: loss, perplexity, tokenAccuracy (no reward/advantage)
|
|
30
|
+
* - GRPO: loss, meanReward, stdReward, stdAdvantage
|
|
31
|
+
*/
|
|
32
|
+
export interface TrainingMetrics {
|
|
33
|
+
step: number;
|
|
34
|
+
loss: number;
|
|
35
|
+
totalTokens: number;
|
|
36
|
+
// GRPO-specific (optional for SFT)
|
|
37
|
+
meanReward?: number;
|
|
38
|
+
stdReward?: number;
|
|
39
|
+
meanAdvantage?: number;
|
|
40
|
+
/** Std of advantages - indicates reward variance within groups */
|
|
41
|
+
stdAdvantage?: number;
|
|
42
|
+
// SFT-specific (optional for GRPO)
|
|
43
|
+
perplexity?: number;
|
|
44
|
+
tokenAccuracy?: number;
|
|
45
|
+
// Timing (optional)
|
|
46
|
+
generationTimeMs?: number;
|
|
47
|
+
trainingTimeMs?: number;
|
|
48
|
+
// Memory (optional)
|
|
49
|
+
peakMemoryMb?: number;
|
|
50
|
+
activeMemoryMb?: number;
|
|
51
|
+
}
|
|
52
|
+
|
|
53
|
+
/**
|
|
54
|
+
* Generation sample for TUI display
|
|
55
|
+
*/
|
|
56
|
+
export interface GenerationSample {
|
|
57
|
+
index: number;
|
|
58
|
+
prompt: string;
|
|
59
|
+
completion: string;
|
|
60
|
+
reward: number;
|
|
61
|
+
tokens: number;
|
|
62
|
+
rewardDetails?: Record<string, number>;
|
|
63
|
+
}
|
|
64
|
+
|
|
65
|
+
/**
|
|
66
|
+
* Training configuration fields for logging
|
|
67
|
+
*/
|
|
68
|
+
export interface TrainingConfigFields {
|
|
69
|
+
numEpochs: number;
|
|
70
|
+
batchSize: number;
|
|
71
|
+
groupSize: number;
|
|
72
|
+
learningRate: number;
|
|
73
|
+
/** Training type: "sft" or "grpo" (default: "grpo") */
|
|
74
|
+
trainingType?: 'sft' | 'grpo';
|
|
75
|
+
[key: string]: unknown;
|
|
76
|
+
}
|
|
77
|
+
|
|
78
|
+
/**
|
|
79
|
+
* Configuration for the training logger
|
|
80
|
+
*/
|
|
81
|
+
export interface TrainingLoggerConfig {
|
|
82
|
+
/** Enable TUI mode (JSONL to stdout) */
|
|
83
|
+
tuiMode: boolean;
|
|
84
|
+
/** Enable console logging (non-TUI) */
|
|
85
|
+
logConsole: boolean;
|
|
86
|
+
/** Enable JSONL file logging */
|
|
87
|
+
logJsonl: boolean;
|
|
88
|
+
/** Output directory for JSONL files */
|
|
89
|
+
outputDir?: string;
|
|
90
|
+
/** Run name for log files */
|
|
91
|
+
runName?: string;
|
|
92
|
+
/** Logging frequency (every N steps) */
|
|
93
|
+
logInterval: number;
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
/**
|
|
97
|
+
* TUI message types for structured stdout output
|
|
98
|
+
*/
|
|
99
|
+
export type TuiMessage =
|
|
100
|
+
| { type: 'init'; model: string; config: Record<string, unknown> }
|
|
101
|
+
| { type: 'epoch_start'; epoch: number; totalEpochs: number; numBatches: number }
|
|
102
|
+
| {
|
|
103
|
+
type: 'step';
|
|
104
|
+
step: number;
|
|
105
|
+
loss: number;
|
|
106
|
+
totalTokens: number;
|
|
107
|
+
// GRPO-specific (optional for SFT)
|
|
108
|
+
meanReward?: number;
|
|
109
|
+
stdReward?: number;
|
|
110
|
+
stdAdvantage?: number;
|
|
111
|
+
// SFT-specific (optional for GRPO)
|
|
112
|
+
perplexity?: number;
|
|
113
|
+
tokenAccuracy?: number;
|
|
114
|
+
// Timing (optional)
|
|
115
|
+
generationTimeMs?: number;
|
|
116
|
+
trainingTimeMs?: number;
|
|
117
|
+
// Memory (optional)
|
|
118
|
+
peakMemoryMb?: number;
|
|
119
|
+
activeMemoryMb?: number;
|
|
120
|
+
}
|
|
121
|
+
| {
|
|
122
|
+
type: 'generation';
|
|
123
|
+
index: number;
|
|
124
|
+
prompt: string;
|
|
125
|
+
completion: string;
|
|
126
|
+
reward: number;
|
|
127
|
+
tokens: number;
|
|
128
|
+
rewardDetails?: Record<string, number>;
|
|
129
|
+
}
|
|
130
|
+
| { type: 'checkpoint'; path: string; step: number }
|
|
131
|
+
| { type: 'epoch_end'; epoch: number; avgLoss: number; avgReward: number; epochTimeSecs: number }
|
|
132
|
+
| { type: 'complete'; totalSteps: number; totalTimeSecs: number }
|
|
133
|
+
| { type: 'log'; level: 'info' | 'warn' | 'error' | 'debug'; message: string }
|
|
134
|
+
| { type: 'paused'; step: number }
|
|
135
|
+
| { type: 'resumed'; step: number }
|
|
136
|
+
| { type: 'status'; phase: string; message: string }
|
|
137
|
+
| { type: 'database_path'; path: string; runId: string; runName?: string }
|
|
138
|
+
| {
|
|
139
|
+
type: 'prompt';
|
|
140
|
+
id: string;
|
|
141
|
+
message: string;
|
|
142
|
+
choices: PromptChoice[];
|
|
143
|
+
default?: number[];
|
|
144
|
+
multiSelect: boolean;
|
|
145
|
+
}
|
|
146
|
+
| {
|
|
147
|
+
type: 'resume_state';
|
|
148
|
+
step: number;
|
|
149
|
+
epoch: number;
|
|
150
|
+
totalEpochs: number;
|
|
151
|
+
stepInEpoch: number;
|
|
152
|
+
totalStepsInEpoch: number;
|
|
153
|
+
metricsHistory: ResumeMetric[];
|
|
154
|
+
aggregates: ResumeAggregates;
|
|
155
|
+
};
|
|
156
|
+
|
|
157
|
+
/**
|
|
158
|
+
* Choice for interactive TUI prompt
|
|
159
|
+
*/
|
|
160
|
+
export interface PromptChoice {
|
|
161
|
+
/** Value to return when selected */
|
|
162
|
+
value: string;
|
|
163
|
+
/** Display label */
|
|
164
|
+
label: string;
|
|
165
|
+
/** Optional description */
|
|
166
|
+
description?: string;
|
|
167
|
+
}
|
|
168
|
+
|
|
169
|
+
/**
|
|
170
|
+
* Options for TUI prompts
|
|
171
|
+
*/
|
|
172
|
+
export interface PromptOptions {
|
|
173
|
+
/** Default selection index (single) or indices (multi-select) */
|
|
174
|
+
default?: number | number[];
|
|
175
|
+
/** Enable multi-select mode (checkbox style) */
|
|
176
|
+
multiSelect?: boolean;
|
|
177
|
+
}
|
|
178
|
+
|
|
179
|
+
/**
|
|
180
|
+
* Historical metric for sparkline restoration
|
|
181
|
+
*/
|
|
182
|
+
export interface ResumeMetric {
|
|
183
|
+
step: number;
|
|
184
|
+
loss: number;
|
|
185
|
+
meanReward: number;
|
|
186
|
+
/** Std of advantages - indicates reward variance within groups */
|
|
187
|
+
stdAdvantage: number;
|
|
188
|
+
perplexity?: number;
|
|
189
|
+
tokenAccuracy?: number;
|
|
190
|
+
/** Time for generation phase (milliseconds) */
|
|
191
|
+
generationTimeMs?: number;
|
|
192
|
+
/** Time for training phase (milliseconds) */
|
|
193
|
+
trainingTimeMs?: number;
|
|
194
|
+
}
|
|
195
|
+
|
|
196
|
+
/**
|
|
197
|
+
* Aggregate statistics for resume state
|
|
198
|
+
*/
|
|
199
|
+
export interface ResumeAggregates {
|
|
200
|
+
bestReward: number;
|
|
201
|
+
avgReward: number;
|
|
202
|
+
rewardCount: number;
|
|
203
|
+
bestLoss: number;
|
|
204
|
+
avgLoss: number;
|
|
205
|
+
lossCount: number;
|
|
206
|
+
totalTokens: number;
|
|
207
|
+
/** Average generation time (milliseconds) */
|
|
208
|
+
avgGenerationTimeMs: number;
|
|
209
|
+
/** Average training time (milliseconds) */
|
|
210
|
+
avgTrainingTimeMs: number;
|
|
211
|
+
}
|
|
212
|
+
|
|
213
|
+
/**
|
|
214
|
+
* Resume state data sent when resuming from checkpoint
|
|
215
|
+
*/
|
|
216
|
+
export interface ResumeState {
|
|
217
|
+
step: number;
|
|
218
|
+
epoch: number;
|
|
219
|
+
totalEpochs: number;
|
|
220
|
+
/** Current batch within epoch */
|
|
221
|
+
stepInEpoch: number;
|
|
222
|
+
/** Total batches per epoch */
|
|
223
|
+
totalStepsInEpoch: number;
|
|
224
|
+
metricsHistory: ResumeMetric[];
|
|
225
|
+
aggregates: ResumeAggregates;
|
|
226
|
+
}
|
|
227
|
+
|
|
228
|
+
/**
|
|
229
|
+
* JSONL log event types
|
|
230
|
+
*/
|
|
231
|
+
export type LogEvent =
|
|
232
|
+
| { event: 'training_start'; timestamp: string; config: TrainingLoggerConfig }
|
|
233
|
+
| { event: 'training_config'; num_examples: number; config: Record<string, unknown>; timestamp: string }
|
|
234
|
+
| {
|
|
235
|
+
event: 'step';
|
|
236
|
+
step: number;
|
|
237
|
+
loss: number;
|
|
238
|
+
mean_reward: number;
|
|
239
|
+
std_reward: number;
|
|
240
|
+
mean_advantage: number;
|
|
241
|
+
total_tokens: number;
|
|
242
|
+
timestamp: string;
|
|
243
|
+
}
|
|
244
|
+
| {
|
|
245
|
+
event: 'epoch';
|
|
246
|
+
epoch: number;
|
|
247
|
+
avg_loss: number;
|
|
248
|
+
avg_reward: number;
|
|
249
|
+
avg_advantage: number;
|
|
250
|
+
total_tokens: number;
|
|
251
|
+
timestamp: string;
|
|
252
|
+
}
|
|
253
|
+
| { event: 'epoch_start'; epoch: number; num_batches: number; timestamp: string }
|
|
254
|
+
| { event: 'training_complete'; final_step: number; total_time_ms: number; timestamp: string }
|
|
255
|
+
| { event: 'checkpoint'; step: number; path: string; timestamp: string };
|
|
256
|
+
|
|
257
|
+
// ============================================================================
|
|
258
|
+
// Metrics Aggregator
|
|
259
|
+
// ============================================================================
|
|
260
|
+
|
|
261
|
+
/**
|
|
262
|
+
* Aggregator for computing running statistics over an epoch
|
|
263
|
+
*/
|
|
264
|
+
class MetricsAggregator {
|
|
265
|
+
private values: number[] = [];
|
|
266
|
+
|
|
267
|
+
add(value: number): void {
|
|
268
|
+
this.values.push(value);
|
|
269
|
+
}
|
|
270
|
+
|
|
271
|
+
mean(): number {
|
|
272
|
+
if (this.values.length === 0) return 0;
|
|
273
|
+
return this.values.reduce((a, b) => a + b, 0) / this.values.length;
|
|
274
|
+
}
|
|
275
|
+
|
|
276
|
+
sum(): number {
|
|
277
|
+
return this.values.reduce((a, b) => a + b, 0);
|
|
278
|
+
}
|
|
279
|
+
|
|
280
|
+
reset(): void {
|
|
281
|
+
this.values = [];
|
|
282
|
+
}
|
|
283
|
+
}
|
|
284
|
+
|
|
285
|
+
// ============================================================================
|
|
286
|
+
// Training Logger
|
|
287
|
+
// ============================================================================
|
|
288
|
+
|
|
289
|
+
/**
|
|
290
|
+
* Unified training logger that handles TUI/console mode automatically
|
|
291
|
+
*/
|
|
292
|
+
export class TrainingLogger {
|
|
293
|
+
private config: TrainingLoggerConfig;
|
|
294
|
+
private jsonlPath?: string;
|
|
295
|
+
private startTime: number;
|
|
296
|
+
private lastLogTime: number;
|
|
297
|
+
|
|
298
|
+
// Epoch-level aggregators
|
|
299
|
+
private epochLoss = new MetricsAggregator();
|
|
300
|
+
private epochReward = new MetricsAggregator();
|
|
301
|
+
private epochAdvantage = new MetricsAggregator();
|
|
302
|
+
private epochTokens = new MetricsAggregator();
|
|
303
|
+
|
|
304
|
+
constructor(config: Partial<TrainingLoggerConfig> = {}) {
|
|
305
|
+
// TUI mode is ONLY enabled via environment variable (set by mlx-tui)
|
|
306
|
+
// This ensures users don't accidentally miss output in CLI mode
|
|
307
|
+
const tuiMode = process.env.MLX_TUI_MODE === '1';
|
|
308
|
+
|
|
309
|
+
this.config = {
|
|
310
|
+
tuiMode,
|
|
311
|
+
logConsole: config.logConsole ?? true,
|
|
312
|
+
logJsonl: config.logJsonl ?? false,
|
|
313
|
+
logInterval: config.logInterval ?? 1,
|
|
314
|
+
outputDir: config.outputDir,
|
|
315
|
+
runName: config.runName,
|
|
316
|
+
};
|
|
317
|
+
this.startTime = Date.now();
|
|
318
|
+
this.lastLogTime = this.startTime;
|
|
319
|
+
|
|
320
|
+
this.setupJsonl();
|
|
321
|
+
}
|
|
322
|
+
|
|
323
|
+
private setupJsonl(): void {
|
|
324
|
+
if (!this.config.logJsonl || this.config.tuiMode) return;
|
|
325
|
+
if (!this.config.outputDir) return;
|
|
326
|
+
|
|
327
|
+
try {
|
|
328
|
+
mkdirSync(this.config.outputDir, { recursive: true });
|
|
329
|
+
const runName = this.config.runName || 'grpo-training';
|
|
330
|
+
this.jsonlPath = join(this.config.outputDir, `${runName}.jsonl`);
|
|
331
|
+
|
|
332
|
+
this.writeJsonl({
|
|
333
|
+
event: 'training_start',
|
|
334
|
+
timestamp: new Date().toISOString(),
|
|
335
|
+
config: this.config,
|
|
336
|
+
});
|
|
337
|
+
} catch {
|
|
338
|
+
// Silently ignore - logging shouldn't crash training
|
|
339
|
+
}
|
|
340
|
+
}
|
|
341
|
+
|
|
342
|
+
// ==========================================================================
|
|
343
|
+
// Core Logging Methods
|
|
344
|
+
// ==========================================================================
|
|
345
|
+
|
|
346
|
+
/** Log info message */
|
|
347
|
+
info(message: string): void {
|
|
348
|
+
if (this.config.tuiMode) {
|
|
349
|
+
this.writeTui({ type: 'log', level: 'info', message });
|
|
350
|
+
} else if (this.config.logConsole) {
|
|
351
|
+
console.log(message);
|
|
352
|
+
}
|
|
353
|
+
}
|
|
354
|
+
|
|
355
|
+
/** Log warning message */
|
|
356
|
+
warn(message: string): void {
|
|
357
|
+
if (this.config.tuiMode) {
|
|
358
|
+
this.writeTui({ type: 'log', level: 'warn', message });
|
|
359
|
+
} else {
|
|
360
|
+
console.warn(message);
|
|
361
|
+
}
|
|
362
|
+
}
|
|
363
|
+
|
|
364
|
+
/** Log error message */
|
|
365
|
+
error(message: string): void {
|
|
366
|
+
if (this.config.tuiMode) {
|
|
367
|
+
this.writeTui({ type: 'log', level: 'error', message });
|
|
368
|
+
} else {
|
|
369
|
+
console.error(message);
|
|
370
|
+
}
|
|
371
|
+
}
|
|
372
|
+
|
|
373
|
+
/** Log debug message (only in verbose mode) */
|
|
374
|
+
debug(message: string): void {
|
|
375
|
+
if (this.config.tuiMode) {
|
|
376
|
+
this.writeTui({ type: 'log', level: 'debug', message });
|
|
377
|
+
} else if (this.config.logConsole && process.env.DEBUG) {
|
|
378
|
+
console.log(`[DEBUG] ${message}`);
|
|
379
|
+
}
|
|
380
|
+
}
|
|
381
|
+
|
|
382
|
+
/** Update status (updates TUI header, shows in console) */
|
|
383
|
+
status(phase: string, message: string): void {
|
|
384
|
+
if (this.config.tuiMode) {
|
|
385
|
+
this.writeTui({ type: 'status', phase, message });
|
|
386
|
+
} else if (this.config.logConsole) {
|
|
387
|
+
console.log(message);
|
|
388
|
+
}
|
|
389
|
+
}
|
|
390
|
+
|
|
391
|
+
/** Print decorative banner (console only, suppressed in TUI mode) */
|
|
392
|
+
banner(...lines: string[]): void {
|
|
393
|
+
if (this.config.tuiMode) return;
|
|
394
|
+
if (!this.config.logConsole) return;
|
|
395
|
+
for (const line of lines) {
|
|
396
|
+
console.log(line);
|
|
397
|
+
}
|
|
398
|
+
}
|
|
399
|
+
|
|
400
|
+
// ==========================================================================
|
|
401
|
+
// Training Event Methods
|
|
402
|
+
// ==========================================================================
|
|
403
|
+
|
|
404
|
+
/** Log training initialization */
|
|
405
|
+
init(model: string, config: TrainingConfigFields, numExamples?: number): void {
|
|
406
|
+
const trainingType = config.trainingType ?? 'grpo';
|
|
407
|
+
|
|
408
|
+
if (this.config.tuiMode) {
|
|
409
|
+
this.writeTui({ type: 'init', model, config: { ...config, trainingType } });
|
|
410
|
+
} else if (this.config.logConsole) {
|
|
411
|
+
const trainingLabel = trainingType === 'sft' ? 'SFT' : 'GRPO';
|
|
412
|
+
console.log(`\n Starting ${trainingLabel} training`);
|
|
413
|
+
if (numExamples) console.log(` Examples: ${numExamples}`);
|
|
414
|
+
console.log(` Epochs: ${config.numEpochs}`);
|
|
415
|
+
console.log(` Batch size: ${config.batchSize}`);
|
|
416
|
+
if (trainingType !== 'sft') {
|
|
417
|
+
console.log(` Group size: ${config.groupSize}`);
|
|
418
|
+
}
|
|
419
|
+
console.log(` Learning rate: ${config.learningRate}`);
|
|
420
|
+
}
|
|
421
|
+
|
|
422
|
+
if (this.jsonlPath && numExamples !== undefined) {
|
|
423
|
+
this.writeJsonl({
|
|
424
|
+
event: 'training_config',
|
|
425
|
+
num_examples: numExamples,
|
|
426
|
+
config: { ...config, trainingType },
|
|
427
|
+
timestamp: new Date().toISOString(),
|
|
428
|
+
});
|
|
429
|
+
}
|
|
430
|
+
}
|
|
431
|
+
|
|
432
|
+
/** Log epoch start */
|
|
433
|
+
epochStart(epoch: number, totalEpochs: number, numBatches: number): void {
|
|
434
|
+
if (this.config.tuiMode) {
|
|
435
|
+
this.writeTui({
|
|
436
|
+
type: 'epoch_start',
|
|
437
|
+
epoch: epoch + 1,
|
|
438
|
+
totalEpochs,
|
|
439
|
+
numBatches,
|
|
440
|
+
});
|
|
441
|
+
} else if (this.config.logConsole) {
|
|
442
|
+
console.log(`\n=== Epoch ${epoch + 1}/${totalEpochs} (${numBatches} batches) ===`);
|
|
443
|
+
}
|
|
444
|
+
|
|
445
|
+
if (this.jsonlPath) {
|
|
446
|
+
this.writeJsonl({
|
|
447
|
+
event: 'epoch_start',
|
|
448
|
+
epoch: epoch + 1,
|
|
449
|
+
num_batches: numBatches,
|
|
450
|
+
timestamp: new Date().toISOString(),
|
|
451
|
+
});
|
|
452
|
+
}
|
|
453
|
+
}
|
|
454
|
+
|
|
455
|
+
/** Log training step */
|
|
456
|
+
step(metrics: TrainingMetrics, batchIdx?: number, numBatches?: number): void {
|
|
457
|
+
// Aggregate for epoch (use available metrics)
|
|
458
|
+
this.epochLoss.add(metrics.loss);
|
|
459
|
+
if (metrics.meanReward !== undefined) {
|
|
460
|
+
this.epochReward.add(metrics.meanReward);
|
|
461
|
+
}
|
|
462
|
+
if (metrics.meanAdvantage !== undefined) {
|
|
463
|
+
this.epochAdvantage.add(metrics.meanAdvantage);
|
|
464
|
+
}
|
|
465
|
+
this.epochTokens.add(metrics.totalTokens);
|
|
466
|
+
|
|
467
|
+
if (this.config.tuiMode) {
|
|
468
|
+
// Build step message with only available fields (no fake values!)
|
|
469
|
+
const stepMsg: TuiMessage & { type: 'step' } = {
|
|
470
|
+
type: 'step',
|
|
471
|
+
step: metrics.step,
|
|
472
|
+
loss: metrics.loss,
|
|
473
|
+
totalTokens: metrics.totalTokens,
|
|
474
|
+
};
|
|
475
|
+
|
|
476
|
+
// Add GRPO-specific fields if present
|
|
477
|
+
if (metrics.meanReward !== undefined) {
|
|
478
|
+
stepMsg.meanReward = metrics.meanReward;
|
|
479
|
+
}
|
|
480
|
+
if (metrics.stdReward !== undefined) {
|
|
481
|
+
stepMsg.stdReward = metrics.stdReward;
|
|
482
|
+
}
|
|
483
|
+
if (metrics.stdAdvantage !== undefined) {
|
|
484
|
+
stepMsg.stdAdvantage = metrics.stdAdvantage;
|
|
485
|
+
}
|
|
486
|
+
|
|
487
|
+
// Add SFT-specific fields if present
|
|
488
|
+
if (metrics.perplexity !== undefined) {
|
|
489
|
+
stepMsg.perplexity = metrics.perplexity;
|
|
490
|
+
}
|
|
491
|
+
if (metrics.tokenAccuracy !== undefined) {
|
|
492
|
+
stepMsg.tokenAccuracy = metrics.tokenAccuracy;
|
|
493
|
+
}
|
|
494
|
+
|
|
495
|
+
// Add timing if present
|
|
496
|
+
if (metrics.generationTimeMs !== undefined) {
|
|
497
|
+
stepMsg.generationTimeMs = metrics.generationTimeMs;
|
|
498
|
+
}
|
|
499
|
+
if (metrics.trainingTimeMs !== undefined) {
|
|
500
|
+
stepMsg.trainingTimeMs = metrics.trainingTimeMs;
|
|
501
|
+
}
|
|
502
|
+
|
|
503
|
+
// Add memory if present
|
|
504
|
+
if (metrics.peakMemoryMb !== undefined) {
|
|
505
|
+
stepMsg.peakMemoryMb = metrics.peakMemoryMb;
|
|
506
|
+
}
|
|
507
|
+
if (metrics.activeMemoryMb !== undefined) {
|
|
508
|
+
stepMsg.activeMemoryMb = metrics.activeMemoryMb;
|
|
509
|
+
}
|
|
510
|
+
|
|
511
|
+
this.writeTui(stepMsg);
|
|
512
|
+
} else if (this.config.logConsole && metrics.step % this.config.logInterval === 0) {
|
|
513
|
+
const now = Date.now();
|
|
514
|
+
const stepTime = (now - this.lastLogTime) / this.config.logInterval;
|
|
515
|
+
this.lastLogTime = now;
|
|
516
|
+
|
|
517
|
+
const batchInfo =
|
|
518
|
+
batchIdx !== undefined && numBatches !== undefined ? ` | Batch ${batchIdx + 1}/${numBatches}` : '';
|
|
519
|
+
|
|
520
|
+
// Build log message based on available metrics (SFT vs GRPO)
|
|
521
|
+
let logMsg = `Step ${metrics.step}${batchInfo} | Loss: ${metrics.loss.toFixed(4)}`;
|
|
522
|
+
|
|
523
|
+
if (metrics.perplexity !== undefined) {
|
|
524
|
+
// SFT format
|
|
525
|
+
logMsg += ` | Perplexity: ${metrics.perplexity.toFixed(2)}`;
|
|
526
|
+
}
|
|
527
|
+
if (metrics.tokenAccuracy !== undefined) {
|
|
528
|
+
logMsg += ` | Acc: ${(metrics.tokenAccuracy * 100).toFixed(1)}%`;
|
|
529
|
+
}
|
|
530
|
+
if (metrics.meanReward !== undefined) {
|
|
531
|
+
// GRPO format
|
|
532
|
+
logMsg += ` | Reward: ${metrics.meanReward.toFixed(4)}`;
|
|
533
|
+
}
|
|
534
|
+
if (metrics.meanAdvantage !== undefined) {
|
|
535
|
+
logMsg += ` | Adv: ${metrics.meanAdvantage.toFixed(4)}`;
|
|
536
|
+
}
|
|
537
|
+
|
|
538
|
+
logMsg += ` | Tokens: ${metrics.totalTokens} | Time: ${stepTime.toFixed(0)}ms/step`;
|
|
539
|
+
console.log(logMsg);
|
|
540
|
+
}
|
|
541
|
+
|
|
542
|
+
if (this.jsonlPath && metrics.step % this.config.logInterval === 0) {
|
|
543
|
+
this.writeJsonl({
|
|
544
|
+
event: 'step',
|
|
545
|
+
step: metrics.step,
|
|
546
|
+
loss: metrics.loss,
|
|
547
|
+
mean_reward: metrics.meanReward ?? 0,
|
|
548
|
+
std_reward: metrics.stdReward ?? 0,
|
|
549
|
+
mean_advantage: metrics.meanAdvantage ?? 0,
|
|
550
|
+
total_tokens: metrics.totalTokens,
|
|
551
|
+
timestamp: new Date().toISOString(),
|
|
552
|
+
});
|
|
553
|
+
}
|
|
554
|
+
}
|
|
555
|
+
|
|
556
|
+
/** Log epoch end/summary */
|
|
557
|
+
epochEnd(epoch: number, totalEpochs: number, epochTimeSecs?: number): void {
|
|
558
|
+
const avgLoss = this.epochLoss.mean();
|
|
559
|
+
const avgReward = this.epochReward.mean();
|
|
560
|
+
const avgAdvantage = this.epochAdvantage.mean();
|
|
561
|
+
const totalTokens = this.epochTokens.sum();
|
|
562
|
+
|
|
563
|
+
if (this.config.tuiMode) {
|
|
564
|
+
this.writeTui({
|
|
565
|
+
type: 'epoch_end',
|
|
566
|
+
epoch: epoch + 1,
|
|
567
|
+
avgLoss,
|
|
568
|
+
avgReward,
|
|
569
|
+
epochTimeSecs: epochTimeSecs ?? 0,
|
|
570
|
+
});
|
|
571
|
+
} else if (this.config.logConsole) {
|
|
572
|
+
console.log(
|
|
573
|
+
`\nEpoch ${epoch + 1}/${totalEpochs} Summary | ` +
|
|
574
|
+
`Avg Loss: ${avgLoss.toFixed(4)} | ` +
|
|
575
|
+
`Avg Reward: ${avgReward.toFixed(4)} | ` +
|
|
576
|
+
`Avg Advantage: ${avgAdvantage.toFixed(4)} | ` +
|
|
577
|
+
`Total Tokens: ${totalTokens.toFixed(0)}`,
|
|
578
|
+
);
|
|
579
|
+
}
|
|
580
|
+
|
|
581
|
+
if (this.jsonlPath) {
|
|
582
|
+
this.writeJsonl({
|
|
583
|
+
event: 'epoch',
|
|
584
|
+
epoch: epoch + 1,
|
|
585
|
+
avg_loss: avgLoss,
|
|
586
|
+
avg_reward: avgReward,
|
|
587
|
+
avg_advantage: avgAdvantage,
|
|
588
|
+
total_tokens: totalTokens,
|
|
589
|
+
timestamp: new Date().toISOString(),
|
|
590
|
+
});
|
|
591
|
+
}
|
|
592
|
+
|
|
593
|
+
// Reset aggregators
|
|
594
|
+
this.epochLoss.reset();
|
|
595
|
+
this.epochReward.reset();
|
|
596
|
+
this.epochAdvantage.reset();
|
|
597
|
+
this.epochTokens.reset();
|
|
598
|
+
}
|
|
599
|
+
|
|
600
|
+
/** Log checkpoint saved */
|
|
601
|
+
checkpoint(path: string, step: number): void {
|
|
602
|
+
if (this.config.tuiMode) {
|
|
603
|
+
this.writeTui({ type: 'checkpoint', path, step });
|
|
604
|
+
} else if (this.config.logConsole) {
|
|
605
|
+
console.log(`💾 Checkpoint saved: ${path}`);
|
|
606
|
+
}
|
|
607
|
+
|
|
608
|
+
if (this.jsonlPath) {
|
|
609
|
+
this.writeJsonl({
|
|
610
|
+
event: 'checkpoint',
|
|
611
|
+
step,
|
|
612
|
+
path,
|
|
613
|
+
timestamp: new Date().toISOString(),
|
|
614
|
+
});
|
|
615
|
+
}
|
|
616
|
+
}
|
|
617
|
+
|
|
618
|
+
/** Log training completion */
|
|
619
|
+
complete(totalSteps: number): void {
|
|
620
|
+
const totalTime = Date.now() - this.startTime;
|
|
621
|
+
const totalMinutes = totalTime / 60000;
|
|
622
|
+
const totalTimeSecs = totalTime / 1000;
|
|
623
|
+
|
|
624
|
+
if (this.config.tuiMode) {
|
|
625
|
+
this.writeTui({ type: 'complete', totalSteps, totalTimeSecs });
|
|
626
|
+
} else if (this.config.logConsole) {
|
|
627
|
+
console.log(`\n✓ Training complete! Final step: ${totalSteps} | Total time: ${totalMinutes.toFixed(2)} minutes`);
|
|
628
|
+
}
|
|
629
|
+
|
|
630
|
+
if (this.jsonlPath) {
|
|
631
|
+
this.writeJsonl({
|
|
632
|
+
event: 'training_complete',
|
|
633
|
+
final_step: totalSteps,
|
|
634
|
+
total_time_ms: totalTime,
|
|
635
|
+
timestamp: new Date().toISOString(),
|
|
636
|
+
});
|
|
637
|
+
}
|
|
638
|
+
}
|
|
639
|
+
|
|
640
|
+
/** Log generation sample (TUI only) */
|
|
641
|
+
generation(sample: GenerationSample): void {
|
|
642
|
+
if (!this.config.tuiMode) return;
|
|
643
|
+
this.writeTui({
|
|
644
|
+
type: 'generation',
|
|
645
|
+
index: sample.index,
|
|
646
|
+
prompt: sample.prompt,
|
|
647
|
+
completion: sample.completion,
|
|
648
|
+
reward: sample.reward,
|
|
649
|
+
tokens: sample.tokens,
|
|
650
|
+
rewardDetails: sample.rewardDetails,
|
|
651
|
+
});
|
|
652
|
+
}
|
|
653
|
+
|
|
654
|
+
/** Log training paused (TUI only) */
|
|
655
|
+
paused(step: number): void {
|
|
656
|
+
if (!this.config.tuiMode) return;
|
|
657
|
+
this.writeTui({ type: 'paused', step });
|
|
658
|
+
}
|
|
659
|
+
|
|
660
|
+
/** Log training resumed (TUI only) */
|
|
661
|
+
resumed(step: number): void {
|
|
662
|
+
if (!this.config.tuiMode) return;
|
|
663
|
+
this.writeTui({ type: 'resumed', step });
|
|
664
|
+
}
|
|
665
|
+
|
|
666
|
+
/** Log database path for TUI DB tab */
|
|
667
|
+
databasePath(path: string, runId: string, runName?: string): void {
|
|
668
|
+
if (!this.config.tuiMode) return;
|
|
669
|
+
this.writeTui({ type: 'database_path', path, runId, runName });
|
|
670
|
+
}
|
|
671
|
+
|
|
672
|
+
/**
|
|
673
|
+
* Send resume state to TUI for sparkline and aggregate restoration.
|
|
674
|
+
* Called when resuming from checkpoint to restore TUI history.
|
|
675
|
+
*/
|
|
676
|
+
resumeState(state: ResumeState): void {
|
|
677
|
+
if (!this.config.tuiMode) return;
|
|
678
|
+
this.writeTui({
|
|
679
|
+
type: 'resume_state',
|
|
680
|
+
step: state.step,
|
|
681
|
+
epoch: state.epoch,
|
|
682
|
+
totalEpochs: state.totalEpochs,
|
|
683
|
+
stepInEpoch: state.stepInEpoch,
|
|
684
|
+
totalStepsInEpoch: state.totalStepsInEpoch,
|
|
685
|
+
metricsHistory: state.metricsHistory,
|
|
686
|
+
aggregates: state.aggregates,
|
|
687
|
+
});
|
|
688
|
+
}
|
|
689
|
+
|
|
690
|
+
/**
|
|
691
|
+
* Send an interactive prompt to the TUI and wait for response.
|
|
692
|
+
* Only works in TUI mode - returns null in non-TUI mode.
|
|
693
|
+
*
|
|
694
|
+
* @param id - Unique ID for this prompt
|
|
695
|
+
* @param message - Message to display
|
|
696
|
+
* @param choices - Available choices
|
|
697
|
+
* @param options - Prompt options
|
|
698
|
+
* @returns The selected value(s), or null if not in TUI mode.
|
|
699
|
+
* For multi-select, returns comma-separated values (use promptMulti for array).
|
|
700
|
+
*/
|
|
701
|
+
async prompt(id: string, message: string, choices: PromptChoice[], options?: PromptOptions): Promise<string | null> {
|
|
702
|
+
if (!this.config.tuiMode) {
|
|
703
|
+
return null; // Caller should handle non-TUI mode
|
|
704
|
+
}
|
|
705
|
+
|
|
706
|
+
const defaultIndices =
|
|
707
|
+
options?.default !== undefined
|
|
708
|
+
? Array.isArray(options.default)
|
|
709
|
+
? options.default
|
|
710
|
+
: [options.default]
|
|
711
|
+
: undefined;
|
|
712
|
+
|
|
713
|
+
// Send prompt to TUI
|
|
714
|
+
this.writeTui({
|
|
715
|
+
type: 'prompt',
|
|
716
|
+
id,
|
|
717
|
+
message,
|
|
718
|
+
choices,
|
|
719
|
+
default: defaultIndices,
|
|
720
|
+
multiSelect: options?.multiSelect ?? false,
|
|
721
|
+
});
|
|
722
|
+
|
|
723
|
+
// Wait for response via stdin
|
|
724
|
+
return new Promise((resolve) => {
|
|
725
|
+
const onData = (data: Buffer) => {
|
|
726
|
+
const line = data.toString().trim();
|
|
727
|
+
// Expected format: PROMPT:<id>:<value>
|
|
728
|
+
// For multi-select, value is comma-separated
|
|
729
|
+
if (line.startsWith('PROMPT:')) {
|
|
730
|
+
const parts = line.split(':');
|
|
731
|
+
if (parts.length >= 3 && parts[1] === id) {
|
|
732
|
+
const value = parts.slice(2).join(':'); // Handle values with colons
|
|
733
|
+
process.stdin.removeListener('data', onData);
|
|
734
|
+
process.stdin.pause();
|
|
735
|
+
resolve(value);
|
|
736
|
+
}
|
|
737
|
+
}
|
|
738
|
+
};
|
|
739
|
+
|
|
740
|
+
process.stdin.resume();
|
|
741
|
+
process.stdin.on('data', onData);
|
|
742
|
+
});
|
|
743
|
+
}
|
|
744
|
+
|
|
745
|
+
/**
|
|
746
|
+
* Send a multi-select prompt to the TUI and wait for response.
|
|
747
|
+
* Convenience wrapper that returns an array of selected values.
|
|
748
|
+
*
|
|
749
|
+
* @param id - Unique ID for this prompt
|
|
750
|
+
* @param message - Message to display
|
|
751
|
+
* @param choices - Available choices
|
|
752
|
+
* @param defaultIndices - Optional default selection indices
|
|
753
|
+
* @returns Array of selected values, or null if not in TUI mode
|
|
754
|
+
*/
|
|
755
|
+
async promptMulti(
|
|
756
|
+
id: string,
|
|
757
|
+
message: string,
|
|
758
|
+
choices: PromptChoice[],
|
|
759
|
+
defaultIndices?: number[],
|
|
760
|
+
): Promise<string[] | null> {
|
|
761
|
+
const result = await this.prompt(id, message, choices, {
|
|
762
|
+
multiSelect: true,
|
|
763
|
+
default: defaultIndices,
|
|
764
|
+
});
|
|
765
|
+
|
|
766
|
+
if (result === null) return null;
|
|
767
|
+
if (result === '') return []; // No selections
|
|
768
|
+
return result.split(',');
|
|
769
|
+
}
|
|
770
|
+
|
|
771
|
+
// ==========================================================================
|
|
772
|
+
// Accessors
|
|
773
|
+
// ==========================================================================
|
|
774
|
+
|
|
775
|
+
/** Check if TUI mode is enabled */
|
|
776
|
+
get isTuiMode(): boolean {
|
|
777
|
+
return this.config.tuiMode;
|
|
778
|
+
}
|
|
779
|
+
|
|
780
|
+
/** Get the log interval */
|
|
781
|
+
get logInterval(): number {
|
|
782
|
+
return this.config.logInterval;
|
|
783
|
+
}
|
|
784
|
+
|
|
785
|
+
// ==========================================================================
|
|
786
|
+
// Internal Methods
|
|
787
|
+
// ==========================================================================
|
|
788
|
+
|
|
789
|
+
private writeTui(msg: TuiMessage): void {
|
|
790
|
+
if (!this.config.tuiMode) return;
|
|
791
|
+
process.stdout.write(JSON.stringify(msg) + '\n');
|
|
792
|
+
}
|
|
793
|
+
|
|
794
|
+
private writeJsonl(data: LogEvent): void {
|
|
795
|
+
if (!this.jsonlPath) return;
|
|
796
|
+
|
|
797
|
+
try {
|
|
798
|
+
const line = JSON.stringify(data) + '\n';
|
|
799
|
+
appendFileSync(this.jsonlPath, line, 'utf8');
|
|
800
|
+
} catch {
|
|
801
|
+
// Silently ignore - logging shouldn't crash training
|
|
802
|
+
}
|
|
803
|
+
}
|
|
804
|
+
}
|
|
805
|
+
|
|
806
|
+
// ============================================================================
|
|
807
|
+
// Factory Function
|
|
808
|
+
// ============================================================================
|
|
809
|
+
|
|
810
|
+
/**
|
|
811
|
+
* Create a training logger instance
|
|
812
|
+
*
|
|
813
|
+
* @example
|
|
814
|
+
* ```typescript
|
|
815
|
+
* // Auto-detect TUI mode from environment
|
|
816
|
+
* const logger = createTrainingLogger();
|
|
817
|
+
*
|
|
818
|
+
* // Explicit configuration
|
|
819
|
+
* const logger = createTrainingLogger({
|
|
820
|
+
* tuiMode: false,
|
|
821
|
+
* logConsole: true,
|
|
822
|
+
* logJsonl: true,
|
|
823
|
+
* outputDir: './outputs',
|
|
824
|
+
* logInterval: 10,
|
|
825
|
+
* });
|
|
826
|
+
* ```
|
|
827
|
+
*/
|
|
828
|
+
export function createTrainingLogger(config?: Partial<TrainingLoggerConfig>): TrainingLogger {
|
|
829
|
+
return new TrainingLogger(config);
|
|
830
|
+
}
|