File size: 24,683 Bytes
4c9d0f9
46cec6c
4c9d0f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
198601d
4c9d0f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d0a90dd
 
4c9d0f9
 
 
 
 
 
 
 
 
 
 
 
 
d0a90dd
4c9d0f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d0a90dd
4c9d0f9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d0a90dd
4c9d0f9
 
 
 
 
 
 
 
 
 
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
538
539
540
541
542
543
544
545
546
547
/**
# Copyright (c) 2025-2026 CMS Manhattan
# All rights reserved.
# Author: Konstantin Vladimirovich Grabko
# Email: grabko@cmsmanhattan.com
# Phone: +1(516)777-0945
*/

package com.cbsinc.cms.llm.ml;

import ai.djl.Device;
import ai.djl.Model;
import ai.djl.engine.Engine;
import ai.djl.ndarray.NDArray;
import ai.djl.ndarray.NDList;
import ai.djl.ndarray.NDManager;
import ai.djl.ndarray.index.NDIndex;
import ai.djl.ndarray.types.DataType;
import ai.djl.ndarray.types.Shape;
import ai.djl.nn.Block;
import ai.djl.nn.Parameter;
import ai.djl.training.DefaultTrainingConfig;
import ai.djl.training.GradientCollector;
import ai.djl.training.Trainer;
import ai.djl.training.TrainingConfig;
import ai.djl.training.initializer.NormalInitializer;
import ai.djl.training.listener.TrainingListener;
import ai.djl.training.loss.Loss;
import ai.djl.training.optimizer.Optimizer;
import ai.djl.training.tracker.Tracker;
import ai.djl.util.Pair;

import java.io.DataInputStream;
import java.io.IOException;
import java.nio.file.Files;
import java.nio.file.Path;
import java.nio.file.Paths;
import java.util.Comparator;
import java.util.List;
import java.util.Random;
import java.util.regex.Matcher;
import java.util.regex.Pattern;
import java.util.stream.Collectors;
import java.util.stream.Stream;

/**
 * =============================================================================
 * STAGE 2: Ternary QAT on SlimOrca — Java on the PYTORCH ENGINE (libtorch).
 *
 * This is the "PyTorch on Java" build: DJL's Java API executing every tensor
 * op on native libtorch (the same C++ kernels CPython PyTorch uses). Ops this
 * model needs that only the PyTorch engine provides reliably: stopGradient
 * (the STE in BitLinear), stepped slicing (interleaved RoPE 0::2 / 1::2),
 * gather (masked cross-entropy), stack, and CUDA execution.
 *
 * --------------------------- DEPENDENCIES (Gradle) ---------------------------
 *   implementation platform("ai.djl:bom:0.36.0")
 *   implementation "ai.djl:api"
 *   implementation "ai.djl.pytorch:pytorch-engine"
 *   // Pick ONE native runtime:
 *   runtimeOnly "ai.djl.pytorch:pytorch-native-cu124::linux-x86_64" // NVIDIA GPU
 *   // runtimeOnly "ai.djl.pytorch:pytorch-native-cpu::linux-x86_64" // CPU only
 *   runtimeOnly "ai.djl.pytorch:pytorch-jni"
 *
 * Maven uses the same artifact IDs. If no native runtime is bundled, the
 * first run downloads libtorch automatically.
 *
 * ------------------------------- JVM FLAGS -------------------------------
 *   -Dai.djl.default_engine=PyTorch
 *   -Xmx8g   (JVM heap holds only Java objects; tensors live in NATIVE
 *             memory, so a huge -Xmx is neither needed nor helpful)
 *
 * ------------------------------ SHARD FORMAT ------------------------------
 * DJL cannot unpickle .pt files. Convert each PyTorch shard once with this
 * small script (run in the Python env that created the shards):
 *
 *   # pt_to_ndlist.py <shard.pt>
 *   import torch, numpy as np, sys
 *   d = torch.load(sys.argv[1], map_location="cpu", weights_only=False)
 *   ids = torch.nn.utils.rnn.pad_sequence(d["input_ids"], batch_first=True,
 *                                          padding_value=0).to(torch.int64)
 *   lbl = torch.nn.utils.rnn.pad_sequence(d["labels"], batch_first=True,
 *                                          padding_value=-100).to(torch.int64)
 *   np.save(sys.argv[1] + ".ids.npy", ids.numpy())
 *   np.save(sys.argv[1] + ".lbl.npy", lbl.numpy())
 *
 * The trainer reads either slimorca_data_<N>.ndlist (named NDList with
 * "input_ids"/"labels") or the .npy pair next to slimorca_data_<N>.pt.
 * =============================================================================
 */
public class JiRackJDLStage2TrainerPt {

    // ========================= SETTINGS =========================
    private static Path DATA_DIR = Paths.get("data/SlimOrca/shards");
    private static Path OUTPUT_DIR = Paths.get("build/JRock_Ternary_SlimOrca_pt");
    private static Path STAGE1_DIR = Paths.get("checkpoints/ternary_stage1");
    private static String STAGE1_NAME = "jirack_data_22";

    private static int BATCH_SIZE = 2;
    private static int GRAD_ACCUM = 5;
    private static float LR = 2e-5f;               // [S-6]
    private static final float WEIGHT_DECAY = 1e-4f;
    private static final float CLIP_GRAD = 1.0f;
    private static final double VAL_RATIO = 0.05;
    private static final int AUTOSAVE_EVERY = 1000;
    private static final long VAL_SEED = 42L;
    private static final long IGNORE_INDEX = -100L;
    private static boolean USE_SMOKE_CONFIG = false;
    // =============================================================

    private static final Pattern SHARD_NUM = Pattern.compile("data_(\\d+)");
    private static final Pattern CKPT_NUM = Pattern.compile("shard_(\\d+)");

    private long globalStep = 0;
    private Device device;

    /* --------------------- Engine & device selection --------------------- */
    private Device selectDevice() {
        Engine engine = Engine.getInstance();
        if (!"PyTorch".equals(engine.getEngineName())) {
            throw new IllegalStateException(
                    "This trainer requires the PyTorch engine (found: "
                    + engine.getEngineName() + "). Add ai.djl.pytorch:pytorch-engine "
                    + "to the classpath and set -Dai.djl.default_engine=PyTorch.");
        }
        System.out.println("Engine: PyTorch " + engine.getVersion());
        int gpus = engine.getGpuCount();
        if (gpus > 0) {
            System.out.println("CUDA devices: " + gpus + " -> using gpu(0)");
            return Device.gpu(0);
        }
        System.out.println("No CUDA device -> CPU (use --smoke for the tiny config)");
        return Device.cpu();
    }

    /* ------------------- Masked, shifted cross-entropy ------------------- */
    /** CE(logits[:, :-1, :], labels[:, 1:]) ignoring label == -100. [S-3] */
    static class MaskedShiftedCELoss extends Loss {
        MaskedShiftedCELoss() {
            super("MaskedShiftedCE");
        }

        @Override
        public NDArray evaluate(NDList labels, NDList predictions) {
            NDArray logits = predictions.singletonOrThrow();      // (B, T, V)
            NDArray target = labels.singletonOrThrow();           // (B, T) INT64

            long T = logits.getShape().get(1);
            long V = logits.getShape().get(2);

            NDArray shiftLogits = logits.get(new NDIndex(":, 0:" + (T - 1) + ", :"))
                                        .reshape(-1, V);
            NDArray shiftLabels = target.get(new NDIndex(":, 1:" + T))
                                        .reshape(-1);

            NDArray mask = shiftLabels.neq(IGNORE_INDEX);
            NDArray maskF = mask.toType(DataType.FLOAT32, false);
            // Replace -100 with 0 so gather() has a valid index; masked later.
            NDArray safeLabels = shiftLabels.mul(mask.toType(DataType.INT64, false))
                                            .reshape(-1, 1);

            // Loss math in FP32 for stability regardless of model dtype.
            NDArray logProb = shiftLogits.toType(DataType.FLOAT32, false).logSoftmax(1);
            NDArray picked = logProb.gather(safeLabels, 1).reshape(-1);
            NDArray nll = picked.neg().mul(maskF);

            return nll.sum().div(maskF.sum().maximum(1.0f));
        }
    }

    /* ----------------------------- Shard IO ----------------------------- */
    private static NDArray[] loadShard(NDManager manager, Path shardFile) throws IOException {
        String fn = shardFile.getFileName().toString();
        if (fn.endsWith(".ndlist")) {
            try (DataInputStream dis = new DataInputStream(Files.newInputStream(shardFile))) {
                NDList list = NDList.decode(manager, dis);
                NDArray ids = null;
                NDArray lbl = null;
                for (NDArray a : list) {
                    if ("input_ids".equals(a.getName())) ids = a;
                    else if ("labels".equals(a.getName())) lbl = a;
                }
                if (ids == null || lbl == null) {
                    ids = list.get(0);
                    lbl = list.get(1);
                }
                return new NDArray[]{ids.toType(DataType.INT64, false),
                                     lbl.toType(DataType.INT64, false)};
            }
        }
        // .npy pair produced by the converter in the header
        Path idsNpy = shardFile.resolveSibling(fn + ".ids.npy");
        Path lblNpy = shardFile.resolveSibling(fn + ".lbl.npy");
        if (Files.exists(idsNpy) && Files.exists(lblNpy)) {
            NDArray ids = manager.decode(Files.readAllBytes(idsNpy));
            NDArray lbl = manager.decode(Files.readAllBytes(lblNpy));
            return new NDArray[]{ids.toType(DataType.INT64, false),
                                 lbl.toType(DataType.INT64, false)};
        }
        throw new IOException("Shard not readable: " + shardFile
                + " (need .ndlist or the .npy pair — see the converter in the header)");
    }

    private static int shardNum(Path p) {
        Matcher m = SHARD_NUM.matcher(p.getFileName().toString());
        if (!m.find()) throw new IllegalArgumentException("Bad shard name: " + p);
        return Integer.parseInt(m.group(1));
    }

    /* --------------------------- Checkpoint IO --------------------------- */
    private void saveCheckpoint(Model model, Path dir, String name) throws IOException {
        Files.createDirectories(dir);
        model.save(dir, name);
        Files.writeString(dir.resolve(name + ".step"), Long.toString(globalStep));
    }

    private long readStep(Path dir, String name) {
        try {
            return Long.parseLong(Files.readString(dir.resolve(name + ".step")).trim());
        } catch (Exception e) {
            return 0L;
        }
    }

    /* ------------------------------ Training ------------------------------ */
    public void train() throws Exception {
        device = selectDevice();
        Files.createDirectories(OUTPUT_DIR);
        System.out.println("Loading JiRack Ternary — Stage 2 (PyTorch engine)...");

        try (Model model = Model.newInstance("jirack_stage2", device);
             NDManager rootManager = NDManager.newBaseManager(device)) {

            Block block = JiRackJDLTernary_10b.buildModel();
            block.setInitializer(new NormalInitializer(JiRackJDLTernary_10b.INIT_STD), Parameter.Type.WEIGHT);
            model.setBlock(block);

            TrainingConfig config = new DefaultTrainingConfig(new MaskedShiftedCELoss())
                    .optOptimizer(Optimizer.adam()
                            .optLearningRateTracker(Tracker.fixed(LR))
                            .optWeightDecays(WEIGHT_DECAY)
                            .optClipGrad(CLIP_GRAD)
                            .build())
                    .optDevices(new Device[]{device})
                    .addTrainingListeners(TrainingListener.Defaults.logging());

            try (Trainer trainer = model.newTrainer(config)) {
                trainer.initialize(new Shape(BATCH_SIZE, 64));
                JiRackJDLTernary_10b.applyDepthScaledInit(block);

                // ---------- Resume: stage-2 checkpoint > stage-1 best ----------
                List<Path> stage2 = listCheckpoints(OUTPUT_DIR);
                int lastDone = -1;
                if (!stage2.isEmpty()) {
                    Path latest = stage2.get(stage2.size() - 1);
                    String name = stripParams(latest);
                    System.out.println("Resuming stage 2 from: " + name);
                    model.load(OUTPUT_DIR, name);
                    globalStep = readStep(OUTPUT_DIR, name);
                    Matcher m = CKPT_NUM.matcher(name);
                    if (m.find()) lastDone = Integer.parseInt(m.group(1));
                } else {
                    Path s1 = STAGE1_DIR.resolve(STAGE1_NAME + "-0000.params");
                    if (Files.exists(s1)) {
                        System.out.println("Starting from stage-1 best: " + s1);
                        model.load(STAGE1_DIR, STAGE1_NAME);  // weights only [S-1]
                        globalStep = 0;   // stage-2 has its own step counter
                    } else {
                        System.out.println("WARNING: no stage-1 checkpoint at "
                                + s1.toAbsolutePath()
                                + " — random init (only sensible with --smoke).");
                    }
                }

                // [S-2] lambda constant 1.0 — warmup was done in stage 1.
                JiRackJDLTernary_10b.setLambda(block, 1.0f);
                System.out.printf(
                        "lambda=1.0 (constant) | global_step=%d | LR=%.1e | device=%s%n",
                        globalStep, LR, device);

                // -------------------- Shards & fixed val set --------------------
                List<Path> allShards = listShards(DATA_DIR);
                if (allShards.isEmpty()) {
                    throw new IOException("No shards in " + DATA_DIR.toAbsolutePath());
                }

                // [S-5] Fixed val set from shard 0, seed 42.
                System.out.println("Building fixed val set from "
                        + allShards.get(0).getFileName());
                long[] trainIdx0;
                NDArray valIds;
                NDArray valLbl;
                {
                    NDArray[] s0 = loadShard(rootManager, allShards.get(0));
                    long n0 = s0[0].getShape().get(0);
                    int valN = (int) (n0 * VAL_RATIO);
                    long[] perm = seededPermutation(n0, VAL_SEED);
                    long[] valIdx = java.util.Arrays.copyOfRange(perm, 0, valN);
                    trainIdx0 = java.util.Arrays.copyOfRange(perm, valN, (int) n0);
                    NDArray vi = rootManager.create(valIdx);
                    valIds = s0[0].get(vi).duplicate();
                    valLbl = s0[1].get(vi).duplicate();
                    s0[0].close();
                    s0[1].close();
                    vi.close();
                    System.out.println("Fixed val set: " + valN
                            + " examples (same for all shards)");
                }

                // -------------------------- Shard loop --------------------------
                for (Path shardPath : allShards) {
                    int shardIdx = shardNum(shardPath);
                    if (shardIdx <= lastDone) {
                        System.out.println("Skipping already processed: "
                                + shardPath.getFileName());
                        continue;
                    }
                    System.out.println("\nStarting shard: " + shardPath.getFileName());

                    // Shard-scoped native memory: freed when this manager closes
                    // (the GPU equivalent of del payload + empty_cache + gc).
                    try (NDManager shardManager = rootManager.newSubManager()) {
                        NDArray[] shard = loadShard(shardManager, shardPath);
                        NDArray ids = shard[0];
                        NDArray lbl = shard[1];

                        if (shardIdx == 0) {
                            NDArray tIdx = shardManager.create(trainIdx0);
                            ids = ids.get(tIdx);
                            lbl = lbl.get(tIdx);
                        }

                        trainOneShard(trainer, config, block, model,
                                ids, lbl, shardIdx, shardManager);

                        System.out.println("Validating...");
                        float valLoss = validate(trainer, config,
                                valIds, valLbl, shardManager);
                        System.out.printf(
                                "Shard %d — Fixed Val Loss: %.4f @ lambda=1.0000 (gstep=%d)%n",
                                shardIdx, valLoss, globalStep);

                        saveCheckpoint(model, OUTPUT_DIR,
                                "slimorca_ternary_shard_" + shardIdx);
                        System.out.println("Saved: slimorca_ternary_shard_" + shardIdx);
                    }
                }
            }
            System.out.println("Stage 2 (Ternary + SlimOrca, PyTorch engine) finished!");
        }
    }

    private void trainOneShard(Trainer trainer, TrainingConfig config, Block block,
                               Model model, NDArray ids, NDArray lbl,
                               int shardIdx, NDManager shardManager) throws IOException {
        long n = ids.getShape().get(0);
        long nBatches = n / BATCH_SIZE;
        long[] order = seededPermutation(n, VAL_SEED + shardIdx + 1);   // shuffle

        int micro = 0;
        GradientCollector gc = trainer.newGradientCollector();
        boolean windowHasGrads = false;
        try {
            for (long b = 0; b < nBatches; b++) {
                float lossVal;
                boolean dropped = false;

                // Per-batch scope: every intermediate tensor of this step is
                // freed when the sub-manager closes. On GPU this is what
                // prevents OOM creep that a shard-level manager can't stop.
                try (NDManager batchManager = shardManager.newSubManager()) {
                    long[] rows = java.util.Arrays.copyOfRange(
                            order, (int) (b * BATCH_SIZE), (int) ((b + 1) * BATCH_SIZE));
                    NDArray rowIdx = batchManager.create(rows);
                    NDArray batchIds = ids.get(rowIdx);
                    NDArray batchLbl = lbl.get(rowIdx);
                    batchIds.attach(batchManager);
                    batchLbl.attach(batchManager);

                    NDList preds = trainer.forward(new NDList(batchIds));
                    NDArray loss = config.getLossFunction()
                            .evaluate(new NDList(batchLbl), preds)
                            .div(GRAD_ACCUM);
                    lossVal = loss.getFloat();

                    if (Float.isNaN(lossVal) || Float.isInfinite(lossVal)) {
                        dropped = true;
                    } else {
                        gc.backward(loss);
                        windowHasGrads = true;
                        micro++;
                    }
                }

                if (dropped) {
                    // NaN/Inf: drop the whole accumulation window.
                    System.out.printf("%nNaN/Inf @ gstep=%d — window dropped%n", globalStep);
                    gc.close();
                    zeroGradients(block);
                    gc = trainer.newGradientCollector();
                    micro = 0;
                    windowHasGrads = false;
                    globalStep++;
                    continue;
                }

                if (micro == GRAD_ACCUM) {
                    trainer.step();   // clipped Adam update + zeroed grads
                    gc.close();
                    gc = trainer.newGradientCollector();
                    micro = 0;
                    windowHasGrads = false;

                    if (AUTOSAVE_EVERY > 0 && globalStep > 0
                            && globalStep % AUTOSAVE_EVERY < GRAD_ACCUM) {
                        saveCheckpoint(model, OUTPUT_DIR, "autosave_latest");
                        System.out.printf("autosave @ gstep=%d%n", globalStep);
                    }
                }

                if (b % 10 == 0) {
                    System.out.printf(
                            "Shard %d [%d/%d] loss=%.4f lambda=1.0000 gstep=%d%n",
                            shardIdx, b, nBatches, lossVal * GRAD_ACCUM, globalStep);
                }
                globalStep++;
            }
        } finally {
            if (windowHasGrads) {
                trainer.step();   // apply the leftover partial window
            }
            gc.close();
        }
    }

    private float validate(Trainer trainer, TrainingConfig config,
                           NDArray valIds, NDArray valLbl, NDManager shardManager) {
        long n = valIds.getShape().get(0);
        long nBatches = Math.max(1, n / BATCH_SIZE);
        double total = 0.0;
        int steps = 0;
        for (long b = 0; b < nBatches; b++) {
            try (NDManager batchManager = shardManager.newSubManager()) {
                long lo = b * BATCH_SIZE;
                long hi = Math.min(n, (b + 1) * BATCH_SIZE);
                NDArray batchIds = valIds.get(new NDIndex(lo + ":" + hi + ", :"));
                NDArray batchLbl = valLbl.get(new NDIndex(lo + ":" + hi + ", :"));
                batchIds.attach(batchManager);
                batchLbl.attach(batchManager);

                NDList preds = trainer.evaluate(new NDList(batchIds));
                float v = config.getLossFunction()
                        .evaluate(new NDList(batchLbl), preds).getFloat();
                if (Float.isFinite(v)) {
                    total += v;
                    steps++;
                }
            }
        }
        return steps > 0 ? (float) (total / steps) : Float.POSITIVE_INFINITY;
    }

    /* ------------------------------ Helpers ------------------------------ */
    private static void zeroGradients(Block block) {
        for (Pair<String, Parameter> p : block.getParameters()) {
            NDArray arr = p.getValue().getArray();
            if (arr.hasGradient()) {
                arr.getGradient().muli(0);
            }
        }
    }

    private static long[] seededPermutation(long n, long seed) {
        long[] idx = new long[(int) n];
        for (int i = 0; i < n; i++) idx[i] = i;
        Random rnd = new Random(seed);
        for (int i = (int) n - 1; i > 0; i--) {
            int j = rnd.nextInt(i + 1);
            long t = idx[i];
            idx[i] = idx[j];
            idx[j] = t;
        }
        return idx;
    }

    private static List<Path> listShards(Path dir) throws IOException {
        if (!Files.isDirectory(dir)) return List.of();
        try (Stream<Path> s = Files.list(dir)) {
            return s.filter(p -> {
                        String f = p.getFileName().toString();
                        return f.matches("slimorca_data_\\d+\\.ndlist")
                                || f.matches("slimorca_data_\\d+\\.pt");
                    })
                    .sorted(Comparator.comparingInt(GPTStage2TrainerPt::shardNum))
                    .collect(Collectors.toList());
        }
    }

    private static List<Path> listCheckpoints(Path dir) throws IOException {
        if (!Files.isDirectory(dir)) return List.of();
        try (Stream<Path> s = Files.list(dir)) {
            return s.filter(p -> p.getFileName().toString()
                            .matches("slimorca_ternary_shard_\\d+-\\d+\\.params"))
                    .sorted(Comparator.comparingInt(p -> {
                        Matcher m = CKPT_NUM.matcher(p.getFileName().toString());
                        m.find();
                        return Integer.parseInt(m.group(1));
                    }))
                    .collect(Collectors.toList());
        }
    }

    private static String stripParams(Path p) {
        String f = p.getFileName().toString();
        return f.substring(0, f.lastIndexOf('-'));   // drop "-0000.params"
    }

    /* -------------------------------- Main -------------------------------- */
    public static void main(String[] args) {
        System.out.println("java -Dai.djl.default_engine=PyTorch -cp Jirackkit.jar "
                + "com.cbsinc.cms.llm.ml.GPTStage2TrainerPt "
                + "[batch] [gradAccum] [lr] [dataDir] [outDir] [stage1Dir] [stage1Name] [--smoke]");

        if (args.length > 0) BATCH_SIZE = Integer.parseInt(args[0]);
        if (args.length > 1) GRAD_ACCUM = Integer.parseInt(args[1]);
        if (args.length > 2) LR = Float.parseFloat(args[2]);
        if (args.length > 3) DATA_DIR = Paths.get(args[3]);
        if (args.length > 4) OUTPUT_DIR = Paths.get(args[4]);
        if (args.length > 5) STAGE1_DIR = Paths.get(args[5]);
        if (args.length > 6) STAGE1_NAME = args[6];
        for (String a : args) {
            if ("--smoke".equals(a)) USE_SMOKE_CONFIG = true;
        }

        if (USE_SMOKE_CONFIG) {
            System.out.println("SMOKE MODE: tiny model config");
            JiRackJDLTernary_10b.smokeConfig();
        }

        try {
            new GPTStage2TrainerPt().train();
        } catch (Exception e) {
            System.err.println("Stage-2 training failed: " + e.getMessage());
            e.printStackTrace();
        }
    }
}