Bug fix: Update create_dynamic_map to always return a float32 tensor

[mitchellgoffpc]

bitsandbytes.functional.create_dynamic_map doesn't specify a dtype for the result tensor it creates, so it will use the torch default dtype. However, all of the cquantize_blockwise_*/cdequantize_blockwise_* functions in pythoninterface.cpp expect a float * for code. This discrepancy causes quantize_blockwise and dequantize_blockwise to crash or give incorrect outputs whenever the default torch dtype is modified before name2qmap["dynamic"] is created on the first call. (Obviously using torch.set_default_dtype isn't necessarily the best practice, but at the time of writing it's still in fairly common use, e.g. https://github.com/meta-llama/llama-models/blob/main/models/llama3/reference_impl/generation.py#L157.)

To reproduce:

import torch
from bitsandbytes.functional import quantize_blockwise, name2qmap

torch.manual_seed(42)
X = torch.randn(65536).cuda()
Q1, _ = quantize_blockwise(X)
Q2, _ = quantize_blockwise(X)
torch.testing.assert_close(Q1, Q2)

del name2qmap['dynamic']
torch.set_default_dtype(torch.float16)
Q3, _ = quantize_blockwise(X)
torch.testing.assert_close(Q1, Q3)

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[matthewdouglas]

Makes sense, thanks!