Broken shard upload: model-00004-of-00004.safetensors invalid (mmap length > file size)
The checkpoint shard model-00004-of-00004.safetensors is invalid/truncated upstream.
Downloaded from Hub:
- file size: 926,838,784 bytes (~884 MiB)
- sha256: 966d862d3b6f54507cc2bfa2e4d7fcaef4390106f2ae0d569ba3162bb5be7e2b
But safetensors header requires data beyond EOF:
- header_requires: 1,580,230,264 bytes
- missing_bytes: 653,391,480
- worst_tensor: model.norm.weight
- OK? False
This causes conversion/load failures (e.g., llama.cpp convert_hf_to_gguf.py):
ValueError: mmap length is greater than file size
Please re-upload a correct model-00004-of-00004.safetensors (or provide a fixed revision/tag).
Hi Navyagoli,
Thanks for the detailed report and the sha256 check. You are right — model-00004-of-00004.safetensors is truncated upstream (the header expects data past EOF), which causes mmap length > file size failures during load and conversion.
We just re-upload a complete, correct shard with the full parameters.
Appreciate you flagging this.
Hi YujieZhao, I checked the 'Files and versions' tab, and the file 'model-00004-of-00004.safetensors' still appears to have been last modified 3 months ago. Could you please clarify if the update was pushed correctly?
Hi Navyagoli,
The “last modified” timestamp in the Files and versions tab can look older in our case because we synced the shard via Git LFS, which may preserve the original commit/file metadata instead of showing a fresh “upload time” the way a manual re-upload would.
Thank you for the reply.