serialize & deserialize does not work as the documentation specify.

[csaybar]

System Info

• transformers version: 4.42.3
• Platform: Linux-6.8.0-51-generic-x86_64-with-glibc2.35
• Python version: 3.10.12
• Huggingface_hub version: 0.25.2
• Safetensors version: 0.5.2
• Accelerate version: 0.27.0
• Accelerate config: not found
• PyTorch version (GPU?): 2.3.1+cu121 (True)
• Tensorflow version (GPU?): 2.15.0 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?:
• Using GPU in script?:
• GPU type: NVIDIA GeForce RTX 3050 Laptop GPU

Information

• The official example scripts
• My own modified scripts

Reproduction

Hi,

I’m unsure if this is expected behavior or a bug since it does not align with what the documentation for these functions describes. Below is the code to reproduce the issue:

Expected behavior

import numpy as np
import safetensors
from safetensors.numpy  import save_file, load

# Save as a safetensors file
data_ran_uint16 = np.random.randint(0, 255, (2, 2, 2)).astype(np.uint16)
save_file({"toy": data_ran_uint16}, "toy.safetensors")

# Deserialize the file
with open("toy.safetensors", "rb") as f:
    fbytes = safetensors.deserialize(f.read())

# Expected to work
serialized = safetensors.serialize({"toy": fbytes[0][1]})

# Workaround
fbytes[0][1]["data"] = bytes(fbytes[0][1]["data"]) # I had to convert the bytearray to bytes
fbytes[0][1]["dtype"] = "uint16" # I had to change the dtype to uint16
fbytes[0][1]["shape"]
serialized = safetensors.serialize({"toy": fbytes[0][1]})
load(serialized)