ctopd / tests /test_single_b200.py
codemaivanngu's picture
fix(runtime): automatically select and ensure compatible node runtime
e1515b2 unverified
Raw History Blame Contribute Delete
13 kB
"""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()