File size: 12,995 Bytes
0ce2d91
 
e1515b2
 
0ce2d91
 
 
 
 
 
 
 
e1515b2
0ce2d91
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
43eefa6
 
 
 
 
 
 
 
b3af95b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
06ba1d4
b3af95b
 
06ba1d4
 
 
 
 
 
 
 
 
 
 
619cff6
 
 
 
 
 
 
 
 
 
 
 
 
 
0ce2d91
 
 
 
 
 
 
 
e1515b2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0ce2d91
 
 
 
47d5468
0ce2d91
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Offline checks for the real-workload launcher; no ML imports or GPU required."""
import ast
import contextlib
import io
import importlib.util
import json
import pathlib
import subprocess
import sys
import tempfile
import types
import unittest
from unittest.mock import patch

ROOT = pathlib.Path(__file__).resolve().parents[1]
SPEC = importlib.util.spec_from_file_location('launcher', ROOT / 'tools/launch_single_b200.py')
launcher = importlib.util.module_from_spec(SPEC)
SPEC.loader.exec_module(launcher)


def method_without_ml_imports(relative, name):
    tree = ast.parse((ROOT / relative).read_text())
    method = next(node for node in ast.walk(tree) if isinstance(node, ast.FunctionDef) and node.name == name)
    module = ast.Module(body=[ast.ImportFrom(module='__future__', names=[ast.alias(name='annotations')], level=0), method], type_ignores=[])
    scope = {}
    exec(compile(ast.fix_missing_locations(module), relative, 'exec'), scope)
    return scope[name]


class LauncherTests(unittest.TestCase):
    def plan(self, *arguments):
        return launcher.build_plan(launcher.parser().parse_args(['--run-id', 'test-real', *arguments]))

    def test_real_workload_and_one_gpu_topology(self):
        plan = self.plan()
        flags = plan['flags']
        for key, value in {'train_batch_size': 512, 'rollout_batch_size': 512, 'max_steps': 100,
                           'generate_max_len': 8192, 'prompt_max_len': 2048, 'max_len': 10240,
                           'micro_train_batch_size': 4, 'ctkd_topk': 16, 'eval_math_n_samples': 32,
                           'eval_math_max_tokens': 8192, 'eval_steps': 10, 'save_steps': 10}.items():
            self.assertEqual(flags[key], value, key)
        for key in ['num_nodes', 'num_gpus_per_node', 'teacher_dp_size', 'teacher_tp_size',
                    'teacher_pp_size', 'rollout_num_engines', 'rollout_tp_size']:
            self.assertEqual(flags[key], 1, key)
        self.assertEqual(plan['gradient_accumulation_microsteps'], 128)

    def test_flags_exist_in_actual_argument_dataclasses(self):
        fields = set()
        for path in (ROOT / 'kdflow/arguments').glob('*_args.py'):
            for node in ast.walk(ast.parse(path.read_text())):
                if isinstance(node, ast.AnnAssign) and isinstance(node.target, ast.Name):
                    fields.add(node.target.id)
        self.assertFalse(set(self.plan()['flags']) - fields)

    def test_micro_batch_changes_preserve_global_workload_and_baseline(self):
        plan = self.plan('--micro-batch', '1', '--algorithm', 'strict-full')
        self.assertEqual(plan['gradient_accumulation_microsteps'], 512)
        self.assertEqual(plan['flags']['train_batch_size'], 512)
        self.assertEqual(plan['flags']['kd_algorithm'], 'byte_align_ctkd')

    def test_simct_preset_changes_only_algorithm_and_adds_disabled_guard(self):
        baseline = self.plan('--algorithm', 'strict-full')['flags']
        flags = self.plan('--algorithm', 'simct')['flags']
        self.assertEqual(flags['kd_algorithm'], 'simct')
        self.assertEqual(flags['simct_gh_mask_threshold'], 0)
        self.assertEqual({k: v for k, v in flags.items() if k not in {'kd_algorithm', 'simct_gh_mask_threshold'}},
                         {k: v for k, v in baseline.items() if k != 'kd_algorithm'})

    def test_mp_mode_presets_preserve_full_workload(self):
        baseline = self.plan()['flags']
        variants = ['mp-atomic', *['mp-fixed' + str(n) for n in range(1, 6)],
                    'mp-random1-2', 'mp-random2-5', 'mp-grass-chunk', 'mp-align']
        for name in variants:
            plan = self.plan('--algorithm', name)
            self.assertTrue(plan['flags']['kd_algorithm'].startswith('mp_opd_'))
            for key, value in baseline.items():
                if key not in {'kd_algorithm'}:
                    self.assertEqual(plan['flags'][key], value, (name, key))
            self.assertIn('no B200/full-training qualification', plan['qualification'])
        for n in range(1, 6):
            flags = self.plan('--algorithm', 'mp-fixed' + str(n))['flags']
            self.assertEqual(flags['kd_algorithm'], 'mp_opd_fixed')
            self.assertEqual(flags['mp_opd_fixed_span_length'], n)
        for name, interval in [('mp-random1-2', (1, 2)), ('mp-random2-5', (2, 5))]:
            flags = self.plan('--algorithm', name)['flags']
            self.assertEqual((flags['mp_opd_random_min_length'], flags['mp_opd_random_max_length']), interval)

    def test_mp_flags_exist_in_argument_dataclasses(self):
        fields = set()
        for path in (ROOT / 'kdflow/arguments').glob('*_args.py'):
            fields |= {node.target.id for node in ast.walk(ast.parse(path.read_text()))
                       if isinstance(node, ast.AnnAssign) and isinstance(node.target, ast.Name)}
        for name in ['mp-atomic', 'mp-fixed5', 'mp-random2-5', 'mp-grass-chunk', 'mp-align', 'mp-align-new']:
            self.assertFalse(set(self.plan('--algorithm', name)['flags']) - fields)

    def test_align_new_preset_and_exact_head_guard(self):
        flags = self.plan('--algorithm', 'mp-align-new')['flags']
        self.assertEqual(flags['kd_algorithm'], 'mp_opd_align_new')
        self.assertEqual(flags['mp_opd_align_new_min_descent_fraction'], .1)
        self.assertEqual(flags['mp_opd_align_new_eps_g'], 1e-12)
        self.assertEqual(flags['mp_opd_chunk_source'], 'xtoken')
        for key in ('train_batch_size','rollout_batch_size','max_steps','generate_max_len','eval_steps','eval_math_n_samples'):
            self.assertEqual(flags[key], self.plan()['flags'][key])
        with self.assertRaisesRegex(ValueError, 'exact_head'):
            self.plan('--algorithm','mp-align-new','--mp-geometry','diag')

    def test_trust_and_simct_paper_presets_keep_common_workload(self):
        baseline=self.plan('--algorithm','strict-full')['flags']
        for name,mode in [('simct_paper','simct_paper'),('mp-trust-r','mp_opd_trust_r'),
                          ('mp-trust-b','mp_opd_trust_b'),('mp-trust-p','mp_opd_trust_p')]:
            flags=self.plan('--algorithm',name)['flags']
            self.assertEqual(flags['kd_algorithm'],mode)
            for key in ('train_batch_size','rollout_batch_size','max_steps','generate_max_len',
                        'learning_rate','eval_steps','eval_math_n_samples','student_name_or_path','train_dataset_path'):
                self.assertEqual(flags[key],baseline[key])
            if name.startswith('mp-trust'):
                self.assertEqual(flags['mp_opd_grass_geometry'],'exact_head')
                with self.assertRaisesRegex(ValueError,'exact_head'):
                    self.plan('--algorithm',name,'--mp-geometry','diag')

    def test_dry_run_creates_no_files(self):
        with tempfile.TemporaryDirectory() as folder:
            destination = pathlib.Path(folder) / 'absent'
            raw = subprocess.check_output([sys.executable, str(ROOT / 'tools/launch_single_b200.py'),
                '--root', str(destination), '--dry-run'], text=True)
            self.assertEqual(json.loads(raw)['flags']['max_steps'], 100)
            self.assertFalse(destination.exists())

    def test_runtime_selector_respects_overrides_and_reuses_compatible_sibling(self):
        from tools.runtime_paths import resolve_runtime, EXPECTED_SHA
        with tempfile.TemporaryDirectory() as folder:
            base=pathlib.Path(folder)/'runtime'; sibling=base.with_name('runtime-u22')
            self.assertEqual(resolve_runtime(environ={},host_glibc=(2,35),base=base),base)
            base.mkdir()
            (base/'.ctopd-runtime.json').write_text(json.dumps({'archive_sha256':EXPECTED_SHA}))
            self.assertEqual(resolve_runtime(environ={},host_glibc=(2,35),base=base),sibling)
            self.assertEqual(resolve_runtime(environ={},host_glibc=(2,39),base=base),base)
            self.assertEqual(resolve_runtime(base,environ={},host_glibc=(2,35)),base)
            self.assertEqual(resolve_runtime(environ={'CTOPD_RUNTIME_PREFIX':str(base)},host_glibc=(2,35)),base)
            (base/'.ctopd-runtime.json').unlink(); base.rmdir(); sibling.mkdir()
            self.assertEqual(resolve_runtime(environ={},host_glibc=(2,39),base=base),sibling)

    def test_runtime_selector_preserves_unknown_installs_and_rejects_symlinks(self):
        from tools.runtime_paths import resolve_runtime
        with tempfile.TemporaryDirectory() as folder:
            base=pathlib.Path(folder)/'runtime'; base.mkdir()
            (base/'keep').write_text('unmanaged')
            self.assertEqual(resolve_runtime(environ={},host_glibc=(2,35),base=base),base)
            self.assertEqual((base/'keep').read_text(),'unmanaged')
            link=pathlib.Path(folder)/'link'
            try: link.symlink_to(base,target_is_directory=True)
            except OSError: self.skipTest('symlink unavailable on this host')
            with self.assertRaisesRegex(ValueError,'non-symlink'): resolve_runtime(link,environ={})

    def test_launcher_runtime_ensure_inherits_gpu_and_runs_only_for_real_launch(self):
        env={'CUDA_VISIBLE_DEVICES':'GPU-fixture-assigned'}
        plan=self.plan('--runtime','/opt/custom-runtime')
        with patch.object(launcher.subprocess,'run') as run:
            launcher.ensure_plan_runtime(plan,env)
        call=run.call_args
        self.assertIn(str(ROOT/'ensure_runtime.sh'),call.args[0])
        self.assertEqual(call.kwargs['env']['CUDA_VISIBLE_DEVICES'],'GPU-fixture-assigned')
        self.assertEqual(env['CTOPD_RUNTIME_PREFIX'],plan['runtime'])
        self.assertEqual(env['CTOPD_ROOT'],plan['root'])
        with patch.object(launcher,'ensure_plan_runtime') as ensure, contextlib.redirect_stdout(io.StringIO()):
            launcher.main(['--dry-run','--run-id','preview-auto-runtime'])
        ensure.assert_not_called()

    def test_existing_output_and_path_traversal_are_refused(self):
        with self.assertRaises(ValueError):
            self.plan('--run-id', '../other-run')
        with tempfile.TemporaryDirectory() as folder:
            plan = self.plan('--root', folder, '--runtime', str(pathlib.Path(folder) / 'runtime'))
            for key in ['student_name_or_path', 'teacher_name_or_path']:
                model = pathlib.Path(plan['flags'][key])
                model.mkdir(parents=True)
                for name in ['config.json', 'tokenizer_config.json', 'tokenizer.json', 'model.safetensors']:
                    (model / name).write_text('{}')
            for path in [pathlib.Path(plan['runtime']) / 'python-b200.sh',
                         pathlib.Path(plan['flags']['train_dataset_path']),
                         pathlib.Path(plan['flags']['eval_math_dataset_path'])]:
                path.parent.mkdir(parents=True, exist_ok=True)
                path.touch()
            pathlib.Path(plan['output']).mkdir(parents=True)
            with self.assertRaises(FileExistsError):
                launcher.validate_files(plan)

    def test_actual_token_normalization_across_world_and_micro_batch_sizes(self):
        prepare = method_without_ml_imports('kdflow/trainer/on_policy_kd_trainer.py', '_prepare_global_batches')
        backward = method_without_ml_imports('kdflow/backend/fsdp/fsdp_strategy.py', 'backward')
        # Unequal sequence lengths make an incorrect mean-of-means observable.
        lengths = [1 + (index * 13 % 97) for index in range(512)]
        gradients = [((index % 7) - 3) * length for index, length in enumerate(lengths)]
        reference = sum(gradients) / sum(lengths)

        class Mask:
            def __init__(self, count): self.count = count
            def sum(self): return self.count

        class Loss:
            def __init__(self, value, sink): self.value, self.sink = value, sink
            def __truediv__(self, divisor): return Loss(self.value / divisor, self.sink)
            def backward(self): self.sink.append(self.value)

        for world, micro in [(8, 4), (1, 4), (1, 1)]:
            batches = [{'stu_loss_mask': Mask(sum(lengths[i:i + micro])),
                        'gradient': sum(gradients[i:i + micro])} for i in range(0, 512, micro)]
            trainer = types.SimpleNamespace(args=types.SimpleNamespace(train=types.SimpleNamespace(use_dynamic_bsz=False)))
            batches = prepare(trainer, batches, 512 // micro)[0]
            rank_gradients = []
            for rank in range(world):
                strategy = types.SimpleNamespace(step=0, accumulated_gradient=512 // micro // world)
                sink = []
                for batch in batches[rank::world]:
                    loss = Loss(batch['gradient'] / batch['avg_micro_batch_token_num'], sink)
                    backward(strategy, loss, None, None)
                self.assertEqual(strategy.step, 0)
                rank_gradients.append(sum(sink))
            self.assertAlmostEqual(sum(rank_gradients) / world, reference, places=12)


if __name__ == '__main__':
    unittest.main()