[Bugfix][Quant] Read safetensors headers of checkpoints with other file names (#60411)

Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
stefankoncarevic
2026-10-08 01:38:41 -05:00
committed by GitHub
co-authored by Andreas Karatzas
parent ee406adf0c
commit 814301b268
4 changed files with 90 additions and 2 deletions
+39 -1
View File
@@ -1355,6 +1355,30 @@ def _read_safetensors_metadata_in_dir(local_dir: Path) -> dict[str, Any]:
}
def _read_safetensors_metadata_in_repo(
model: str,
*,
revision: str | None = None,
) -> dict[str, Any]:
api = hf_api()
try:
filenames = [
filename
for filename in api.list_repo_files(model, revision=revision)
if filename.endswith(".safetensors")
]
return {
param_name: asdict(info)
for filename in filenames
for param_name, info in api.parse_safetensors_file_metadata(
model, filename, revision=revision
).tensors.items()
}
except Exception as e:
logger.debug("Could not read safetensors headers for %s: %s", model, e)
return {}
def get_safetensors_params_metadata(
model: str,
*,
@@ -1372,6 +1396,12 @@ def get_safetensors_params_metadata(
for param_name, info in file_mt.tensors.items()
}
# `get_safetensors_metadata` only looks for `model.safetensors` and its
# index, so older checkpoints naming their weights differently land here
# even though their headers are perfectly readable.
if metadata := _read_safetensors_metadata_in_repo(model, revision=revision):
return metadata
# Hub fetch failed (e.g. 429, network unreachable). Fall back to the
# local HF cache: weights may already be cached from a prior run, and
# weight loading itself uses the same cache.
@@ -1390,7 +1420,15 @@ def get_safetensors_params_metadata(
str(e),
)
return {}
return _read_safetensors_metadata_in_dir(Path(local_dir))
metadata = _read_safetensors_metadata_in_dir(Path(local_dir))
if not metadata:
logger.warning_once(
"Could not retrieve safetensors metadata for %s "
"(Hub fetch failed and the cached snapshot holds no weights yet).",
model,
)
return metadata
@cache