[Bug] Incorrect metadata.total_size in model.safetensors.index.json (reports ~3.87 TB instead of 15.79 GB)

#18
by mzubairtahir000714 - opened

Hi @inclusionAI team,

Thank you for releasing Ling-3.0-tiny!

I noticed that model.safetensors.index.json currently reports an inflated total_size in its metadata:

"metadata": {
  "total_size": 3869925683584
}

This causes Hugging Face's web viewer, automated download managers, and VRAM/storage calculators to display the model size as ~3.87 TB (3,604 GiB) instead of its real size.

Actual File Size
Summing the actual Content-Length across all 32 shards (model-00000-of-00032.safetensors to model-00031-of-00032.safetensors):

Actual Total Download Size: 15,787,992,416 bytes (~14.70 GiB / 15.79 GB)
This matches the expected ~7.89B parameter BF16 footprint.

Root Cause
The number 3,869,925,683,584 is an exact integer multiple of the 128 experts: $$\frac{3,869,925,683,584}{128} = 30,233,794,403$$

When slicing the packed 3D MoE expert weights (num_local_experts: 128) into individual keys (model.layers.i.mlp.experts.0 ... experts.127), the un-cloned slices shared the underlying PyTorch storage. The shard indexing script counted the full 128-expert storage buffer for each slice, inflating total_size by a factor of 128.

Suggested Fix
Update the metadata block in model.safetensors.index.json to:

  "metadata": {
    "total_size": 15787992416
  }

Thank you!

9:59 AM

Sign up or log in to comment