File size: 29,511 Bytes
c8ba9c6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
"""Execution reconfiguration and distributed full-state continuation checks."""
from copy import deepcopy
import importlib.util
import json
import math
import os
from pathlib import Path
import subprocess
import sys
import types

import numpy as np
import pytest
import torch


ROOT = Path(__file__).resolve().parent
spec = importlib.util.spec_from_file_location("execution_pipeline", ROOT / "run_40b.py")
pipeline = importlib.util.module_from_spec(spec)
spec.loader.exec_module(pipeline)
sys.path.insert(0, str(ROOT))
from warpstate import WarpStateConfig, WarpStateLM


def signature():
    return {
        "settings": {
            "seq_len": 8, "micro_batch": 2, "grad_accum": 2,
            "world_size": 1, "train_tokens": 73, "lr": 0.002,
            "min_lr": 0.0002, "warmup_steps": 2, "weight_decay": 0.1,
            "seed": 1337, "activation_checkpointing": True,
            "loss_chunk_size": 5, "compile": False, "compile_core": False,
            "loss_mode": "checkpointed", "compile_mode": "default",
            "bf16_residuals": False,
            "activation_mode": "baseline", "gradient_mode": "eager",
            "overlap_swap_collectives": False, "overlap_runtime_estimates": False,
            "overlap_memory_budget": 0.0,
            "overlap_producer_schedule": False,
            "cache_core_weight_casts": False, "matmul_execution": None,
            "backward_layout_execution": None,
            "ddp_bucket_cap_mb": 128, "ddp_batched_grad_copy": False,
        },
        "data": {"complete": True, "signature": {"recipe": 1, "source": "fixture"}},
    }


@pytest.mark.parametrize("missing", [
    ("world_size", "compile_core", "loss_mode", "compile_mode", "bf16_residuals"),
    ("loss_mode", "compile_mode"),
    ("bf16_residuals",),
    ("matmul_execution",),
    ("backward_layout_execution",),
    ("overlap_swap_collectives", "overlap_runtime_estimates", "overlap_memory_budget"),
    ("overlap_producer_schedule",),
    ("activation_mode", "gradient_mode", "ddp_bucket_cap_mb", "ddp_batched_grad_copy", "cache_core_weight_casts"),
])
def test_legacy_execution_defaults_do_not_require_reconfiguration(missing):
    old = signature()
    for key in missing:
        del old["settings"][key]
    original = deepcopy(old)
    assert pipeline.check_resume_signature(old, signature()) is False
    assert old == original


@pytest.mark.parametrize("execution", [
    {"micro_batch": 4, "grad_accum": 1},
    {"micro_batch": 1, "grad_accum": 4},
    {"world_size": 2, "grad_accum": 1},
    {"activation_checkpointing": False},
    {"loss_chunk_size": 16},
    {"compile": True},
    {"compile_core": True},
    {"loss_mode": "full"},
    {"loss_mode": "compiled-full"},
    {"bf16_residuals": True},
    {"activation_mode": "swiglu"},
    {"activation_mode": "rmsnorm"},
    {"activation_mode": "all"},
    {"gradient_mode": "compiled-autograd"},
    {"gradient_mode": "compiled-autograd-overlap"},
    {"gradient_mode": "compiled-autograd-overlap", "overlap_swap_collectives": True},
    {"gradient_mode": "compiled-autograd-overlap", "overlap_runtime_estimates": True},
    {"gradient_mode": "compiled-autograd-overlap", "overlap_memory_budget": 0.02},
    {"gradient_mode": "compiled-autograd-overlap", "overlap_producer_schedule": True},
    {"cache_core_weight_casts": True},
    {"matmul_execution": {"sha256": "fixture", "entries": []}},
    {"backward_layout_execution": {"mode": "paired", "identity": "fixture"}},
    {"ddp_bucket_cap_mb": 64},
    {"ddp_batched_grad_copy": True},
    {"compile_core": True, "compile_mode": "reduce-overhead"},
    {"compile_core": True, "compile_mode": "max-autotune-no-cudagraphs"},
])
def test_execution_changes_require_explicit_reconfiguration(execution):
    old = signature()
    new = deepcopy(old)
    new["settings"].update(execution)
    with pytest.raises(ValueError):
        pipeline.check_resume_signature(old, new)
    assert pipeline.check_resume_signature(old, new, reconfigure=True) is True


@pytest.mark.parametrize("change", [
    {"seq_len": 16, "micro_batch": 1},  # Same effective batch, changed context.
    {"train_tokens": 105},
    {"lr": 0.003},
    {"min_lr": 0.0003},
    {"warmup_steps": 3},
    {"weight_decay": 0.2},
    {"seed": 1338},
    {"world_size": 2},  # Effective batch doubles.
    {"grad_accum": 3},
])
def test_reconfiguration_rejects_objective_or_global_batch_changes(change):
    old = signature()
    new = deepcopy(old)
    new["settings"]["loss_mode"] = "full"
    new["settings"].update(change)
    with pytest.raises(ValueError):
        pipeline.check_resume_signature(old, new, reconfigure=True)


def test_reconfiguration_rejects_changed_data():
    old = signature()
    new = deepcopy(old)
    new["data"]["signature"]["source"] = "another-corpus"
    with pytest.raises(ValueError):
        pipeline.check_resume_signature(old, new, reconfigure=True)


def test_train_signature_records_selected_loss_compile_and_residual_modes():
    expected = signature()
    expected["settings"].update(loss_mode="compiled-full", compile_core=True,
                                 compile_mode="max-autotune-no-cudagraphs", bf16_residuals=True)
    args = types.SimpleNamespace(**expected["settings"])
    assert pipeline.train_signature(args, expected["data"]) == expected


def test_train_signature_records_matmul_recipe_and_implementation_identity():
    expected = signature()
    metadata = {"format": "warpstate-matmul-plan-v1", "sha256": "recipe-sha",
                "source_sha256": {"ops/matmul_kernels.py": "implementation-sha"}, "entries": []}
    args = types.SimpleNamespace(**expected["settings"], matmul_metadata=metadata)
    expected["settings"]["matmul_execution"] = metadata
    assert pipeline.train_signature(args, expected["data"]) == expected
    changed = deepcopy(expected)
    changed["settings"]["matmul_execution"]["source_sha256"]["ops/matmul_kernels.py"] = "changed"
    with pytest.raises(ValueError):
        pipeline.check_resume_signature(expected, changed)
    assert pipeline.check_resume_signature(expected, changed, reconfigure=True)


def test_train_signature_defaults_legacy_arguments_to_fp32_residuals():
    expected = signature()
    legacy = dict(expected["settings"])
    del legacy["bf16_residuals"]
    assert pipeline.train_signature(types.SimpleNamespace(**legacy), expected["data"]) == expected


def test_legacy_checkpoint_requires_reconfiguration_to_enable_bf16_residuals():
    old = signature()
    del old["settings"]["bf16_residuals"]
    original = deepcopy(old)
    new = signature()
    new["settings"]["bf16_residuals"] = True
    with pytest.raises(ValueError):
        pipeline.check_resume_signature(old, new)
    assert pipeline.check_resume_signature(old, new, reconfigure=True) is True
    assert old == original


@pytest.mark.parametrize(("flags", "enabled"), [
    ([], False),
    (["--bf16-residuals"], True),
    (["--bf16-residuals", "--no-bf16-residuals"], False),
])
def test_bf16_residual_cli_is_explicit_and_reversible(monkeypatch, flags, enabled):
    monkeypatch.setattr(sys, "argv", ["run_40b.py", *flags])
    assert pipeline.arguments().bf16_residuals is enabled


def test_nondefault_compile_mode_requires_compilation(monkeypatch):
    monkeypatch.setattr(sys, "argv", ["run_40b.py", "--compile-mode", "reduce-overhead"])
    with pytest.raises(SystemExit) as error:
        pipeline.arguments()
    assert error.value.code == 2


@pytest.mark.parametrize("flags", [
    ["--gradient-mode", "compiled-autograd", "--ddp-batched-grad-copy"],
    ["--gradient-mode", "compiled-autograd-overlap", "--ddp-batched-grad-copy"],
    ["--ddp-bucket-cap-mb", "0"],
    ["--ddp-bucket-cap-mb", "nan"],
    ["--ddp-bucket-cap-mb", "inf"],
    ["--overlap-swap-collectives"],
    ["--gradient-mode", "compiled-autograd", "--overlap-runtime-estimates"],
    ["--overlap-memory-budget", "0.02"],
    ["--gradient-mode", "compiled-autograd-overlap", "--overlap-memory-budget", "nan"],
    ["--gradient-mode", "compiled-autograd-overlap", "--overlap-memory-budget", "inf"],
    ["--gradient-mode", "compiled-autograd-overlap", "--overlap-memory-budget", "-0.01"],
    ["--gradient-mode", "compiled-autograd-overlap", "--overlap-memory-budget", "0.26"],
    ["--activation-mode", "invalid"],
    ["--activation-mode", "swiglu", "--gradient-mode", "compiled-autograd"],
    ["--activation-mode", "rmsnorm", "--gradient-mode", "compiled-autograd-overlap"],
    ["--activation-mode", "all", "--gradient-mode", "compiled-autograd-overlap"],
    ["--matmul-plan", "configs/example.json", "--cache-core-weight-casts", "--no-activation-checkpointing"],
    ["--matmul-plan", "configs/example.json", "--compile-core", "--compile-mode", "reduce-overhead"],
    ["--matmul-plan", "configs/example.json", "--gradient-mode", "compiled-autograd-overlap", "--activation-checkpointing"],
    ["--cache-core-weight-casts"],
    ["--cache-core-weight-casts", "--no-activation-checkpointing", "--gradient-mode", "compiled-autograd"],
    ["--cache-core-weight-casts", "--no-activation-checkpointing", "--compile-core", "--compile-mode", "reduce-overhead"],
    ["--gradient-mode", "compiled-autograd", "--compile-core", "--compile-mode", "reduce-overhead"],
])
def test_optimization_cli_rejects_invalid_or_incompatible_settings(monkeypatch, flags):
    monkeypatch.setattr(sys, "argv", ["run_40b.py", *flags])
    with pytest.raises(SystemExit) as error:
        pipeline.arguments()
    assert error.value.code == 2


def test_optimization_cli_defaults_and_selected_modes(monkeypatch):
    monkeypatch.setattr(sys, "argv", ["run_40b.py"])
    defaults = pipeline.arguments()
    assert (defaults.activation_mode, defaults.gradient_mode) == ("baseline", "eager")
    assert defaults.ddp_bucket_cap_mb == 128 and not defaults.ddp_batched_grad_copy
    assert not defaults.cache_core_weight_casts
    assert defaults.matmul_plan is None
    monkeypatch.setattr(sys, "argv", ["run_40b.py", "--activation-mode", "all",
                                    "--ddp-bucket-cap-mb", "64", "--ddp-batched-grad-copy",
                                    "--cache-core-weight-casts", "--no-activation-checkpointing"])
    selected = pipeline.arguments()
    assert selected.activation_mode == "all"
    assert selected.cache_core_weight_casts
    assert selected.ddp_bucket_cap_mb == 64 and selected.ddp_batched_grad_copy


@pytest.mark.parametrize("compile_flag", ["--compile-core", "--compile"])
def test_loss_and_compile_modes_can_be_selected_together(monkeypatch, compile_flag):
    monkeypatch.setattr(sys, "argv", ["run_40b.py", compile_flag,
                                     "--compile-mode", "max-autotune-no-cudagraphs",
                                     "--loss-mode", "compiled-full"])
    args = pipeline.arguments()
    assert args.compile_mode == "max-autotune-no-cudagraphs"
    assert args.loss_mode == "compiled-full"


@pytest.mark.parametrize("world_size", [1, 2, 3])
@pytest.mark.parametrize("target", [1, 7, 8, 9, 15, 16, 17, 23, 24, 25, 63, 64, 65])
def test_update_partition_covers_tokens_once_and_keeps_ranks_in_step(world_size, target):
    micro_tokens = 8
    covered = []
    for rank in range(world_size):
        slices = list(pipeline.update_slices(target, micro_tokens, world_size, rank))
        assert len(slices) == math.ceil(target / (micro_tokens * world_size))
        for iteration, (offset, valid) in enumerate(slices):
            assert offset == (iteration * world_size + rank) * micro_tokens
            assert valid == max(0, min(micro_tokens, target - offset))
            covered.extend(range(offset, offset + valid))
    assert sorted(covered) == list(range(target))


def test_all_masked_rank_loss_is_finite_and_contributes_zero_gradients():
    cfg = WarpStateConfig(vocab_size=32, d_model=16, n_heads=2, n_cores=1,
                          logical_depth=1, ffn_hidden=32, chunk_size=8,
                          parallel_memory=True, loss_chunk_size=5)
    model = WarpStateLM(cfg)
    inputs = torch.arange(8).reshape(1, 8)
    labels = torch.full_like(inputs, -100)
    _, loss = model(inputs, labels, return_logits=False)
    assert torch.isfinite(loss) and loss.item() == 0.0
    loss.backward()
    for parameter in model.parameters():
        assert parameter.grad is not None
        assert torch.isfinite(parameter.grad).all()
        assert torch.count_nonzero(parameter.grad).item() == 0


def fixture_runs(tmp_path, names):
    cfg = WarpStateConfig(vocab_size=32768, d_model=16, n_heads=2, n_cores=1,
                          logical_depth=1, ffn_hidden=32, chunk_size=8,
                          parallel_memory=True, loss_chunk_size=5)
    config = tmp_path / "config.json"
    cfg.to_json(config)
    torch.manual_seed(133)
    weights = tmp_path / "initial.pt"
    torch.save({"config": vars(cfg), "model": WarpStateLM(cfg).state_dict(),
                "tokens_seen": 100}, weights)
    directories = []
    for name in names:
        directory = tmp_path / name
        data = directory / "data"
        data.mkdir(parents=True)
        np.arange(73, dtype="<u2").tofile(data / "train.bin")
        np.arange(256, dtype="<u2").tofile(data / "val.bin")
        args = types.SimpleNamespace(
            tokenizer=ROOT / "tokenizer.model", dataset="HuggingFaceFW/fineweb-edu",
            dataset_config="sample-100BT", revision="main", train_tokens=73,
            val_tokens=256,
        )
        (data / "manifest.json").write_text(json.dumps({
            "complete": True, "signature": pipeline.data_signature(args),
        }))
        directories.append(directory)
    common = [
        str(ROOT / "run_40b.py"), "--no-bootstrap", "--test-small-gpu",
        "--stage", "train", "--config", str(config), "--weights", str(weights),
        "--train-tokens", "73", "--val-tokens", "256", "--seq-len", "8",
        "--micro-batch", "2", "--loss-chunk-size", "5", "--workers", "2",
        "--eval-batches", "2", "--eval-micro-batch", "1", "--eval-every", "1",
        "--lr", "0.002", "--min-lr", "0.0002", "--warmup-steps", "2",
        "--no-autocast-cache", "--vram-limit-gib", "1", "--vram-headroom-gib", "1",
    ]
    return directories, common


def run_training(directory, common, world_size, extra=(), *, timeout=180, expect_async=False,
                 expect_validation_display=False):
    command = [sys.executable]
    if world_size > 1:
        command += ["-m", "torch.distributed.run", "--standalone", "--nnodes", "1",
                    "--nproc-per-node", str(world_size)]
    command += common + ["--work-dir", str(directory), "--grad-accum",
                         str(2 // world_size), *extra]
    environment = os.environ.copy()
    for key in ("RANK", "LOCAL_RANK", "WORLD_SIZE", "LOCAL_WORLD_SIZE",
                "GROUP_RANK", "ROLE_RANK", "ROLE_WORLD_SIZE", "MASTER_ADDR", "MASTER_PORT"):
        environment.pop(key, None)
    result = subprocess.run(command, capture_output=True, text=True, env=environment, timeout=timeout)
    assert result.returncode == 0, result.stdout + "\n" + result.stderr
    if expect_validation_display:
        records = {record["step"]: record for record in map(
            json.loads, (directory / "run/metrics.jsonl").read_text().splitlines())}
        step_lines = [line for line in result.stdout.splitlines() if line.startswith("step ")]
        assert step_lines, result.stdout
        for line in step_lines:
            step = int(line.split()[1].split("/")[0])
            record = records[step]
            if record["val_step"] == step:
                assert f"val {record['val_loss']:.5f} (step {record['val_step']})" in line, line
            else:
                assert " | val " not in line, line
    if expect_async:
        assert "Staging full CPU checkpoint" in result.stdout, result.stdout
        assert "for asynchronous checkpoint" in result.stdout, result.stdout
        assert "background write" in result.stdout, result.stdout
        runtime = json.loads((directory / "run/settings.json").read_text())["runtime"]
        assert runtime["async_checkpoint"] is True
    checkpoint = torch.load(directory / "run/latest.pt", map_location="cpu", weights_only=False)
    if expect_async:
        # The final full-state checkpoint must already be published when the
        # trainer exits; a queued write or leftover temporary file is insufficient.
        final_progress = f'step {checkpoint["step"]}, tokens={checkpoint["tokens_seen"]:,}; '
        assert final_progress + "background write" in result.stdout, result.stdout
        assert not (directory / "run/latest.pt.tmp").exists()
    return checkpoint


def assert_tensors_equal(first, second):
    if isinstance(first, torch.Tensor):
        torch.testing.assert_close(first, second, rtol=0, atol=0)
    elif isinstance(first, np.ndarray):
        np.testing.assert_array_equal(first, second)
    elif isinstance(first, dict):
        assert first.keys() == second.keys()
        for key in first:
            assert_tensors_equal(first[key], second[key])
    elif isinstance(first, (tuple, list)):
        assert len(first) == len(second)
        for value, other in zip(first, second, strict=True):
            assert_tensors_equal(value, other)
    else:
        assert first == second


TWO_GPUS = pytest.mark.skipif(torch.cuda.device_count() < 2, reason="Two visible CUDA GPUs required")


@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
@pytest.mark.parametrize(("compile_mode", "async_checkpoint"), [
    pytest.param("default", False, id="default"),
    pytest.param("reduce-overhead", False, id="reduce-overhead"),
    pytest.param("max-autotune-no-cudagraphs", False, id="max-autotune-no-cudagraphs"),
    pytest.param("max-autotune-no-cudagraphs", True, id="max-autotune-no-cudagraphs-async"),
])
def test_single_gpu_compiled_full_state_resume(tmp_path, compile_mode, async_checkpoint):
    (continuous, resumed), common = fixture_runs(tmp_path, ["continuous", "resumed"])
    common += ["--loss-mode", "compiled-full", "--compile-core",
               "--compile-mode", compile_mode, "--no-activation-checkpointing", "--autocast-cache"]
    if async_checkpoint:
        # Exercise overlapping periodic saves, then flush at max-steps and final exit.
        common += ["--save-every", "1"]
    initial = torch.load(tmp_path / "initial.pt", map_location="cpu", weights_only=False)
    # Allow cold Inductor/autotuning compilation while keeping each subprocess bounded.
    # The continuous reference always saves synchronously, including the async case.
    expected = run_training(continuous, common, 1, timeout=300)
    resumed_common = common + (["--async-checkpoint"] if async_checkpoint else [])
    partial = run_training(resumed, resumed_common, 1, ["--max-steps", "1"],
                           timeout=300, expect_async=async_checkpoint)
    assert partial["tokens_seen"] == 32 and partial["step"] == 1
    actual = run_training(resumed, resumed_common, 1, timeout=300, expect_async=async_checkpoint)

    assert actual["format"] == expected["format"] == "warpstate-server-v1"
    assert actual["tokens_seen"] == expected["tokens_seen"] == 73
    assert actual["step"] == expected["step"] == 3
    assert actual["origin"] == partial["origin"] == expected["origin"]
    assert actual["signature"] == partial["signature"] == expected["signature"]
    settings = actual["signature"]["settings"]
    assert settings["world_size"] == 1
    assert settings["loss_mode"] == "compiled-full"
    assert settings["compile_core"] is True
    assert settings["compile_mode"] == compile_mode
    assert settings["activation_checkpointing"] is False
    assert settings["seq_len"] * settings["micro_batch"] * settings["grad_accum"] == 32
    assert len(actual["rng_by_rank"]) == 1
    for key in ("model", "optimizer", "rng_by_rank"):
        assert_tensors_equal(actual[key], expected[key])
    assert actual["model"].keys() == initial["model"].keys()
    assert all(not key.startswith(("module.", "_orig_mod.")) for key in actual["model"])
    assert all(value.dtype == torch.float32 and torch.isfinite(value).all()
               for value in actual["model"].values())
    assert actual["optimizer"]["state"]
    for state in actual["optimizer"]["state"].values():
        assert state["step"].item() == 3
        for key in ("exp_avg", "exp_avg_sq"):
            assert state[key].dtype == torch.float32 and torch.isfinite(state[key]).all()
    assert math.isfinite(actual["val_loss"])
    assert actual["val_loss"] == expected["val_loss"]

    records = [json.loads(line) for line in (resumed / "run/metrics.jsonl").read_text().splitlines()]
    assert [record["step"] for record in records] == [1, 2, 3]
    assert [record["tokens"] for record in records] == [32, 64, 73]
    assert [record["lr"] for record in records] == [0.001, 0.002, 0.002]
    assert all(math.isfinite(record["loss"]) and math.isfinite(record["grad_norm"])
               for record in records)
    validations = [json.loads(line) for line in (resumed / "run/validation.jsonl").read_text().splitlines()]
    assert [record["tokens"] for record in validations] == [32, 64, 73]
    assert all(math.isfinite(record["loss"]) for record in validations)


@TWO_GPUS
@pytest.mark.parametrize("loss_mode", ["checkpointed", "compiled-full"])
def test_two_gpu_full_state_resume_and_empty_final_rank(tmp_path, loss_mode):
    (continuous, resumed), common = fixture_runs(tmp_path, ["continuous", "resumed"])
    common += ["--loss-mode", loss_mode]
    if loss_mode == "compiled-full":
        common += ["--compile-core", "--compile-mode", "default", "--no-activation-checkpointing"]
    expected = run_training(continuous, common, 2)
    partial = run_training(resumed, common, 2, ["--max-steps", "1"])
    assert partial["tokens_seen"] == 32 and partial["step"] == 1
    actual = run_training(resumed, common, 2)
    assert actual["tokens_seen"] == expected["tokens_seen"] == 73
    assert actual["step"] == expected["step"] == 3
    assert actual["origin"] == expected["origin"]
    assert actual["signature"]["settings"]["world_size"] == 2
    assert actual["signature"]["settings"]["loss_mode"] == loss_mode
    if loss_mode == "compiled-full":
        assert actual["signature"]["settings"]["compile_core"] is True
        assert actual["signature"]["settings"]["activation_checkpointing"] is False
    assert len(actual["rng_by_rank"]) == 2
    for key in ("model", "optimizer", "rng_by_rank"):
        assert_tensors_equal(actual[key], expected[key])
    assert all(not key.startswith(("module.", "_orig_mod.")) for key in actual["model"])
    records = [json.loads(line) for line in (resumed / "run/metrics.jsonl").read_text().splitlines()]
    assert [record["tokens"] for record in records] == [32, 64, 73]
    assert all(math.isfinite(record["loss"]) for record in records)


@TWO_GPUS
def test_two_gpu_bf16_residual_async_full_state_resume_and_empty_final_rank(tmp_path):
    (continuous, resumed), common = fixture_runs(tmp_path, ["continuous", "resumed"])
    common += ["--loss-mode", "compiled-full", "--compile-core", "--compile-mode", "default",
               "--no-activation-checkpointing", "--autocast-cache", "--bf16-residuals",
               "--eval-every", "2", "--save-every", "2"]
    initial = torch.load(tmp_path / "initial.pt", map_location="cpu", weights_only=False)
    # A synchronous continuous reference catches snapshot aliasing while updates
    # run during periodic writes. The split run also tests the final flush on
    # max-steps and the masked final rank when nine tokens remain across two GPUs.
    expected = run_training(continuous, common, 2, timeout=300, expect_validation_display=True)
    async_common = common + ["--async-checkpoint"]
    partial = run_training(resumed, async_common, 2, ["--max-steps", "1"],
                           timeout=300, expect_async=True, expect_validation_display=True)
    assert partial["tokens_seen"] == 32 and partial["step"] == 1
    assert partial["val_step"] == 1 and partial["val_tokens"] == 32
    actual = run_training(resumed, async_common, 2, timeout=300, expect_async=True,
                          expect_validation_display=True)

    assert actual["format"] == expected["format"] == "warpstate-server-v1"
    assert actual["tokens_seen"] == expected["tokens_seen"] == 73
    assert actual["step"] == expected["step"] == 3
    assert actual["origin"] == partial["origin"] == expected["origin"]
    assert actual["signature"] == partial["signature"] == expected["signature"]
    settings = actual["signature"]["settings"]
    assert settings["world_size"] == 2
    assert settings["loss_mode"] == "compiled-full"
    assert settings["compile_core"] is True
    assert settings["compile_mode"] == "default"
    assert settings["activation_checkpointing"] is False
    assert settings["bf16_residuals"] is True
    assert settings["seq_len"] * settings["micro_batch"] * settings["grad_accum"] * settings["world_size"] == 32
    assert len(actual["rng_by_rank"]) == 2
    for key in ("model", "optimizer", "rng_by_rank"):
        assert_tensors_equal(actual[key], expected[key])
    assert actual["model"].keys() == initial["model"].keys()
    assert all(not key.startswith(("module.", "_orig_mod.")) for key in actual["model"])
    assert all(value.dtype == torch.float32 and torch.isfinite(value).all()
               for value in actual["model"].values())
    assert actual["optimizer"]["state"]
    for state in actual["optimizer"]["state"].values():
        assert state["step"].item() == 3
        for key in ("exp_avg", "exp_avg_sq"):
            assert state[key].dtype == torch.float32 and torch.isfinite(state[key]).all()
    assert math.isfinite(actual["val_loss"])
    assert actual["val_loss"] == expected["val_loss"]
    assert actual["val_step"] == expected["val_step"] == 3
    assert actual["val_tokens"] == expected["val_tokens"] == 73

    continuous_records = [json.loads(line) for line in
                          (continuous / "run/metrics.jsonl").read_text().splitlines()]
    assert [record["val_step"] for record in continuous_records] == [0, 2, 3]
    assert [record["val_tokens"] for record in continuous_records] == [0, 64, 73]
    continuous_validations = [json.loads(line) for line in
                              (continuous / "run/validation.jsonl").read_text().splitlines()]
    assert [record["step"] for record in continuous_validations] == [2, 3]
    assert all(record["val_loss"] == validation["loss"] for record, validation in
               zip(continuous_records[1:], continuous_validations, strict=True))
    assert math.isfinite(continuous_records[0]["val_loss"])
    for directory in (continuous, resumed):
        runtime = json.loads((directory / "run/settings.json").read_text())["runtime"]
        assert runtime["eval_every"] == runtime["save_every"] == 2

    records = [json.loads(line) for line in (resumed / "run/metrics.jsonl").read_text().splitlines()]
    assert [record["step"] for record in records] == [1, 2, 3]
    assert [record["tokens"] for record in records] == [32, 64, 73]
    assert [record["lr"] for record in records] == [0.001, 0.002, 0.002]
    assert all(math.isfinite(record["loss"]) and math.isfinite(record["grad_norm"])
               for record in records)
    validations = [json.loads(line) for line in (resumed / "run/validation.jsonl").read_text().splitlines()]
    assert [record["tokens"] for record in validations] == [32, 64, 73]
    assert all(math.isfinite(record["loss"]) for record in validations)
    assert [record["val_step"] for record in records] == [1, 2, 3]
    assert [record["val_tokens"] for record in records] == [32, 64, 73]
    assert all(record["val_loss"] == validation["loss"] for record, validation in
               zip(records, validations, strict=True))


@TWO_GPUS
def test_one_to_two_gpu_reconfiguration_preserves_training_state(tmp_path):
    (serial, distributed), common = fixture_runs(tmp_path, ["serial", "distributed"])
    expected = run_training(serial, common, 1)
    partial = run_training(distributed, common, 1, ["--max-steps", "1"])
    assert partial["tokens_seen"] == 32 and partial["step"] == 1
    actual = run_training(distributed, common, 2, ["--reconfigure"])
    assert actual["tokens_seen"] == expected["tokens_seen"] == 73
    assert actual["step"] == expected["step"] == 3
    assert actual["origin"] == partial["origin"] == expected["origin"]
    settings = actual["signature"]["settings"]
    assert settings["seq_len"] * settings["micro_batch"] * settings["grad_accum"] * settings["world_size"] == 32
    assert settings["world_size"] == 2
    assert len(actual["rng_by_rank"]) == 2
    assert actual["optimizer"]["param_groups"] == expected["optimizer"]["param_groups"]
    for key, value in actual["model"].items():
        torch.testing.assert_close(value, expected["model"][key], rtol=0.01, atol=2e-5)
    for parameter, state in actual["optimizer"]["state"].items():
        reference = expected["optimizer"]["state"][parameter]
        assert state["step"].item() == reference["step"].item() == 3
        for key in ("exp_avg", "exp_avg_sq"):
            torch.testing.assert_close(state[key], reference[key], rtol=0.03, atol=2e-7)
    records = [json.loads(line) for line in (distributed / "run/metrics.jsonl").read_text().splitlines()]
    assert [record["tokens"] for record in records] == [32, 64, 73]
    assert [record["lr"] for record in records] == [0.001, 0.002, 0.002]