Fix crash on scalar tensors in tensors()

[Bafoleco]

Summary

• view(*shape) in _impl.py crashes when shape is [] (scalar tensor) because view(*[]) expands to view() with no arguments, raising TypeError
• Fix: reshape(shape) instead of view(*shape) — reshape is zero-copy here since the tensor is always contiguous (freshly created from torch.from_dlpack).

Context

This occurs with FP8-quantized models (e.g. DeepSeek V3/V3.2) which store per-tensor quantization scales (input_scale, weight_scale_2) as scalar float32 tensors with shape []. These models have thousands of such tensors.

Minimal repro:

torch.zeros(4).view(torch.float32).view(*[])
# TypeError: view() received an invalid combination of arguments

torch.zeros(4).view(torch.float32).reshape([])
# Works — returns scalar tensor with shape torch.Size([])

Testing

Ran test script locally and seems things still pass

[Bafoleco]

Hey @arlo-scitix, noticed the existing implementation seems to crash for DeepSeek-V3.2, so wondering if we could make a small fix here? Super excited to use this project btw, speeds look really good!

[arlo-scitix]

Thanks.
Are you using the official model? Could you provide the model’s Hugging Face URL?

[Bafoleco]

Thanks. Are you using the official model? Could you provide the model’s Hugging Face URL?

I'm using the NVFP4 quantized version: https://huggingface.co/nvidia/DeepSeek-V3.2-NVFP4/blob/main/model-00016-of-000163.safetensors

[arlo-scitix]

Got it. Merged.