Add Qwen235 state replay diagnosis
This commit is contained in:
@@ -135,6 +135,58 @@ class FidelityEnvelopeTest(unittest.TestCase):
|
||||
self.assertEqual(len(evidence), 1)
|
||||
self.assertEqual(evidence[0]["log"], "stdout.log")
|
||||
|
||||
def test_qwen235_state_replay_changes_observation_flags_only(self) -> None:
|
||||
module = load("run_qwen235_fixed_pd_state_replay.py")
|
||||
base = [
|
||||
"/usr/bin/python3",
|
||||
"-m",
|
||||
"frontier.main",
|
||||
"--metrics_config_output_dir",
|
||||
"/frozen/metrics",
|
||||
"--metrics_config_run_id",
|
||||
"frozen",
|
||||
"--no-metrics_config_store_frontier_stage_batch_ledger",
|
||||
"--attn_tensor_parallel_size",
|
||||
"8",
|
||||
"--moe_expert_parallel_size",
|
||||
"8",
|
||||
"--communication_collective_profile_path",
|
||||
"/profiles/real-allreduce.csv",
|
||||
]
|
||||
transformed = module.transform_command(
|
||||
base, metrics_root=Path("/new/metrics"), run_id="state-run"
|
||||
)
|
||||
self.assertEqual(
|
||||
module.option_value(transformed, "--attn_tensor_parallel_size"), "8"
|
||||
)
|
||||
self.assertEqual(
|
||||
module.option_value(transformed, "--communication_collective_profile_path"),
|
||||
"/profiles/real-allreduce.csv",
|
||||
)
|
||||
self.assertIn(
|
||||
"--metrics_config_store_frontier_stage_batch_ledger", transformed
|
||||
)
|
||||
self.assertIn("--metrics_config_keep_individual_batch_metrics", transformed)
|
||||
self.assertNotIn(
|
||||
"--no-metrics_config_store_frontier_stage_batch_ledger", transformed
|
||||
)
|
||||
|
||||
def test_qwen235_state_replay_rejects_implicit_ledger_base(self) -> None:
|
||||
module = load("run_qwen235_fixed_pd_state_replay.py")
|
||||
base = [
|
||||
"/usr/bin/python3",
|
||||
"-m",
|
||||
"frontier.main",
|
||||
"--metrics_config_output_dir",
|
||||
"/frozen/metrics",
|
||||
"--metrics_config_run_id",
|
||||
"frozen",
|
||||
]
|
||||
with self.assertRaisesRegex(ValueError, "explicitly disable one full Frontier ledger"):
|
||||
module.transform_command(
|
||||
base, metrics_root=Path("/new/metrics"), run_id="state-run"
|
||||
)
|
||||
|
||||
def test_materialize_qwen235_allreduce_requires_serving_contract(self) -> None:
|
||||
module = load("materialize_qwen235_v020_allreduce.py")
|
||||
with tempfile.TemporaryDirectory() as temporary:
|
||||
|
||||
Reference in New Issue
Block a user