File size: 10,458 Bytes
dc72977 | 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 | """CPU-only resume/provenance and two-worker scheduling checks."""
import io
import json
from pathlib import Path
import sys
import tempfile
import threading
import unittest
from contextlib import redirect_stdout
from dataclasses import replace
from unittest.mock import patch
sys.path.insert(0, str(Path(__file__).resolve().parents[1] / "scripts"))
import run_elf_remaining as q
from elf_common import SOURCE_COMMIT, default_sampling, sha256, write_json
class QueueTests(unittest.TestCase):
def setUp(self):
self.temp = tempfile.TemporaryDirectory()
self.addCleanup(self.temp.cleanup)
self.root = Path(self.temp.name)
self.assets = {t: {"checkpoint": "model", "sha256": {"model": "locked"}}
for t in ("owt", "wmt14", "xsum")}
write_json(self.root / "artifacts/elf_lock.json", {})
write_json(self.root / "data/elf/assets.json", {
"assets": self.assets, "lock_sha256": sha256(self.root / "artifacts/elf_lock.json")})
self.asset_hash = sha256(self.root / "data/elf/assets.json")
def input(self, task, split, count):
prefix = task == "owt-prefix"
path = self.root / "data/elf" / ("owt-prefix.jsonl" if prefix else f"{task}-{split}.jsonl")
path.write_text(''.join(json.dumps({"id": i}) + '\n' for i in range(count)))
meta = {"task": task, "split": split, "count": count, "sha256": sha256(path)}
if prefix:
source = self.root / "data/conditional/owt-c64/prompts.jsonl"
source.parent.mkdir(parents=True, exist_ok=True)
source.write_text("source\n")
meta.update(protocol="c64_text_t5_v1", source_sha256=sha256(source))
write_json(path.with_suffix(".manifest.json"), meta)
return path
def request(self, cell):
return {
"task": cell.task, "steps": cell.steps, "seed": 42, "compile": False, "batch_size": 8,
"source_commit": SOURCE_COMMIT, "assets_manifest_sha256": self.asset_hash,
"checkpoint_task": "owt" if cell.task == "owt-prefix" else cell.task,
"checkpoint_selection": "ema_params1", "checkpoint_sha256": "locked",
"sampling": default_sampling(cell.task, cell.steps),
"model_canvas_t5_tokens": 1088 if cell.task == "xsum" else 128 if cell.task == "wmt14" else 1024,
"protocol": "c64_text_t5_v1" if cell.task == "owt-prefix" else "elf_native_v1",
"condition_tokenization_policy": q.EOS_POLICY if cell.task in ("wmt14", "xsum") else "none",
"input": {"sha256": cell.input_sha256} if cell.input_sha256 else None,
}
def result(self, path, cell, status="quality_valid"):
request = self.request(cell)
write_json(path / "request.json", request)
if status == "partial":
return
pairs = [(i, 0) for i in range(cell.count)]
if cell.task == "owt-prefix":
pairs += [(i, c) for c in range(1, 5) for i in range(min(256, cell.count))]
samples = path / "samples.jsonl"
samples.write_text(''.join(json.dumps({"id": j, "prompt_id": i, "completion_id": c}) + '\n'
for j, (i, c) in enumerate(pairs)))
gen = {**request, "sample_count": len(pairs), "prompt_count": cell.count,
"diversity": cell.task == "owt-prefix", "samples_sha256": sha256(samples)}
write_json(path / "generation.json", gen)
if status == "needs_evaluation":
return
metric = {"task": cell.task, "sample_count": cell.count, "samples_sha256": sha256(samples),
"generation_manifest_sha256": sha256(path / "generation.json"),
"valid": status == "quality_valid", "metrics": {"entropy_gpt2_nats": 5.1}}
if not metric["valid"]:
metric["reason"] = "At least one completion has fewer than two GPT-2 tokens; no samples were dropped"
write_json(path / "metrics.json", metric)
def reusable(self, cell):
return q.reusable_result(self.root, cell, self.asset_hash, self.assets)
def test_full_matrix_and_input_block_isolation(self):
cells, blocked = q.build_matrix(self.root, q.TASKS)
self.assertEqual(len(cells), 7)
self.assertEqual(len(blocked), 5)
for task, split, count in (("owt-prefix", "c64-text-t5", 1024),
("wmt14", "validation", 3000), ("wmt14", "test", 3003),
("xsum", "validation", 11332), ("xsum", "test", 11334)):
self.input(task, split, count)
cells, blocked = q.build_matrix(self.root, q.TASKS)
self.assertFalse(blocked)
self.assertEqual(len(cells), 32)
self.assertEqual(len({c.key for c in cells}), 32)
self.assertEqual([c.steps for c in cells if c.task == "owt"], [1, 2, 4, 8, 16, 32, 1024])
source = self.root / "data/conditional/owt-c64/prompts.jsonl"
source.write_text("changed\n")
cells, blocked = q.build_matrix(self.root, q.TASKS)
self.assertEqual(len(cells), 25)
self.assertEqual(blocked[0]["task"], "owt-prefix")
def test_reuse_scores_or_generation_but_not_sanity_or_tampered_samples(self):
cell = q.Cell("owt", "unconditional", 8, 1024)
sanity = self.root / "results/elf-sanity/quality" / cell.key
self.result(sanity, replace(cell, count=1000))
self.assertIsNone(self.reusable(cell)[0])
path = self.root / "results/elf-old/quality" / cell.key
self.result(path, cell, "needs_evaluation")
self.assertEqual(self.reusable(cell)[0]["status"], "needs_evaluation")
self.result(path, cell)
self.assertEqual(self.reusable(cell)[0]["status"], "quality_valid")
with (path / "samples.jsonl").open("a") as f:
f.write('{}\n')
found, rejected = self.reusable(cell)
self.assertIsNone(found)
self.assertTrue(rejected)
def test_sampling_batch_and_no_eos_mismatch_rejected(self):
cell = q.Cell("wmt14", "validation", 8, 3000, "data", "hash")
for key, value in (("batch_size", 16), ("condition_tokenization_policy", "old"),
("sampling", {}), ("checkpoint_sha256", "different")):
request = self.request(cell)
request[key] = value
self.assertFalse(q.matches_request(request, cell, self.asset_hash, self.assets))
def test_invalid_quality_not_retried_and_conflicts_rejected(self):
cell = q.Cell("owt", "unconditional", 1, 2)
path = self.root / "results/elf-old/quality" / cell.key
self.result(path, cell, "quality_invalid")
self.assertEqual(self.reusable(cell)[0]["status"], "quality_invalid")
self.result(self.root / "results/elf-other/quality" / cell.key, cell)
with self.assertRaisesRegex(ValueError, "Conflicting"):
self.reusable(cell)
def test_partial_runs_preserved_and_prefix_requests_diversity(self):
cell = q.Cell("owt-prefix", "c64-text-t5", 32, 1024, "inputs", "digest")
root = self.root / "results/elf-gpu01"
path = root / "quality" / cell.key
self.result(path, cell, "partial")
retry = q.fresh_path(root, cell)
self.assertIn("elf-gpu01-retry1", str(retry))
self.assertTrue((path / "request.json").exists())
cmds = q.commands(self.root, sys.executable, cell, retry, True)
self.assertEqual([stage for stage, _ in cmds], ["generate", "evaluate"])
self.assertIn("--diversity", cmds[0][1])
self.assertNotIn("timing", str(cmds))
self.assertEqual([s for s, _ in q.commands(self.root, sys.executable, cell, path, False)], ["evaluate"])
self.result(path, cell)
self.assertEqual(self.reusable(cell)[0]["split"], "c64-text-t5")
def test_plan_does_not_query_gpu_or_launch_processes(self):
with patch.object(sys, "argv", ["queue", "--root", str(self.root), "--plan"]), \
patch.object(q.subprocess, "check_output") as query, \
patch.object(q.subprocess, "Popen") as popen, redirect_stdout(io.StringIO()):
self.assertEqual(q.main(), 2) # Missing conditional inputs, OWT still planned.
query.assert_not_called()
popen.assert_not_called()
self.assertFalse((self.root / "results").exists())
def test_two_workers_isolated_and_invalid_quality_does_not_stop_queue(self):
cells = [q.Cell("owt", "unconditional", s, 2) for s in (1, 2, 4)]
barrier = threading.Barrier(2, timeout=10)
launches, lock = [], threading.Lock()
def fake_popen(command, **kwargs):
stage = "generate" if "generate" in command else "evaluate"
path = Path(command[command.index("--output" if stage == "generate" else "--run") + 1])
cell = next(c for c in cells if str(path).endswith(c.key))
with lock:
launches.append((cell.steps, stage, kwargs["env"]["CUDA_VISIBLE_DEVICES"]))
class Process:
def wait(inner):
if stage == "generate" and cell.steps in (1, 2):
barrier.wait()
status = "needs_evaluation" if stage == "generate" else "quality_invalid" if cell.steps == 1 else "quality_valid"
self.result(path, cell, status)
return int(status == "quality_invalid")
return Process()
with patch.object(sys, "argv", ["queue", "--root", str(self.root)]), \
patch.object(sys, "platform", "linux"), patch.object(q, "build_matrix", return_value=(cells, [])), \
patch.object(q.subprocess, "check_output", side_effect=lambda cmd, **kw: f"GPU-{cmd[2]}, H200 NVL"), \
patch.object(q.subprocess, "Popen", side_effect=fake_popen), redirect_stdout(io.StringIO()):
self.assertEqual(q.main(), 0)
self.assertEqual(len(launches), 6)
self.assertEqual({gpu for _, _, gpu in launches}, {"GPU-0", "GPU-1"})
for cell in cells:
self.assertEqual(len({gpu for step, _, gpu in launches if step == cell.steps}), 1)
report = json.loads(next((self.root / "results/elf-gpu01/logs").glob("*/queue.json")).read_text())
self.assertEqual([r["status"] for r in report["runs"]], ["quality_invalid", "quality_valid", "quality_valid"])
if __name__ == "__main__":
unittest.main()
|