Fix dtype_byte_size for FP8 fnuz / e8m0fnu dtypes

[lollinng]

Problem

dtype_byte_size only special-cases torch.float8_e4m3fn and torch.float8_e5m2. The other 1-byte FP8 dtypes — float8_e4m3fnuz, float8_e5m2fnuz, and float8_e8m0fnu — fall through to:

bit_search = re.search(r"[^\d](\d+)$", str(dtype))

Their string forms end in ...fnuz / ...fnu (no trailing digit), so bit_search is None and the function raises ValueError: dtype is not a valid dtype. This breaks anything that sizes such tensors — compute_module_sizes, infer_auto_device_map, estimate-memory, etc. — for models stored in those dtypes.

Fix

Handle all available 1-byte FP8 dtypes, guarded with hasattr so it stays correct on older torch where some variants don't exist.

Testing done (CPU)

Added tests/test_modeling_utils.py::ModelingUtilsTester::test_dtype_byte_size. With torch 2.12:

float8_e4m3fn      after-fix=1   before(regex)=RAISED
float8_e5m2        after-fix=1
float8_e4m3fnuz    after-fix=1   before(regex)=RAISED
float8_e5m2fnuz    after-fix=1   before(regex)=RAISED
float8_e8m0fnu     after-fix=1   before(regex)=RAISED

Standard dtypes unchanged (float32=4, float16/bfloat16=2, int8=1, float64=8, bool=1/8). ruff check / ruff format --check clean.

tests/test_modeling_utils.py::ModelingUtilsTester::test_dtype_byte_size PASSED

[SunMarc]

LGTM !

[HuggingFaceDocBuilderDev]

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.