Skip to content

Commit ada4e93

Browse files
Ronald Tseronaldtse
authored andcommitted
fix: rebuild_int8_head32 shares the resumable reference decode
The sweep built four head32 zips then died mid-reference-decode (same traceback-less container kill; no margins written). Same resume file per model dir — parity and rebuild progress compose.
1 parent 48a4e32 commit ada4e93

1 file changed

Lines changed: 6 additions & 1 deletion

File tree

src/gpu/modal_export.py

Lines changed: 6 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -478,7 +478,12 @@ def rebuild_int8_head32(model_id: str, limit: int = 0) -> dict:
478478
else:
479479
dst.writestr(name, src.read(name))
480480

481-
reference = reference_decode(model, [s for s, _ in pairs], max_len=128)
481+
reference = reference_decode(
482+
model,
483+
[s for s, _ in pairs],
484+
max_len=128,
485+
resume_path=Path("/outputs/imf") / model_id / "reference_progress.jsonl",
486+
)
482487
report = run_parity(model, new_zip, pairs, max_len=128, reference=reference)
483488
if not report.passed:
484489
raise RuntimeError(f"parity gate FAILED for {new_zip.name}: {report}")

0 commit comments

Comments
 (0)