For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).
Python function
header_only_weight_files
header_only_weight_files()
max.pipelines.weights.hf_utils.header_only_weight_files(huggingface_model_id, filenames, revision=None, max_workers=8)
Stands in header-only stubs for a model’s safetensors files.
Fetches each remote file’s header and writes a sparse stub of the full
file size, instead of downloading the weights. Compilation reads only
tensor names, shapes, dtypes and the file-level __metadata__, all of
which the header carries, so a compile-only run never needs the bytes.
-
Parameters:
-
- huggingface_model_id (str) – The Hugging Face model identifier, that is,
MiniMaxAI/MiniMax-M3-MXFP8. - filenames (list[str]) – Safetensors file paths relative to the root of the Hugging Face repo.
- revision (str | None) – The Hugging Face revision to read headers from. Defaults to the Hub’s default revision.
- max_workers (int) – The number of worker threads to concurrently fetch headers.
- huggingface_model_id (str) – The Hugging Face model identifier, that is,
-
Returns:
-
Absolute stub paths, in the order of
filenames. -
Raises:
-
ValueError – If any filename is not a safetensors file.
-
Return type:
-
list[Path]