Add F4 (float4_e2m1fn_x2) and F8_E8M0 (float8_e8m0fnu) dtype support

[uasind]

Summary

Adds support for two dtypes used by modern mixed-precision models (DeepSeek V4-Flash, DeepSeek V3, etc.) that currently cause add_filenames() to fail:

• F8_E8M0 (torch.float8_e8m0fnu, PyTorch 2.5+): unsigned 8-bit exponent-only format used for per-tile quantization scales. One byte per element. Straightforward addition — same storage size as F8_E4M3.

• F4 (torch.float4_e2m1fn_x2, PyTorch 2.10+): packed FP4 format, two 4-bit values per byte. safetensors stores the shape in FP4-element count while PyTorch float4_e2m1fn_x2 counts packed pairs (one byte each). This requires shape adjustment on load.

Without these, any safetensors file containing F4 or F8_E8M0 tensors raises:

ValueError: 'F8_E8M0' is not a valid DType
ValueError: 'F4' is not a valid DType

Changes

fastsafetensors/st_types.py

• Add F4 and F8_E8M0 to DType enum

fastsafetensors/frameworks/_torch.py

• Map DType.F8_E8M0 → torch.float8_e8m0fnu (guarded by hasattr)
• Map DType.F4 → torch.float4_e2m1fn_x2 (guarded by hasattr)
• Add U8 NCCL workaround for both (NCCL has no float8_e8m0/float4 support)
• get_dtype_size() returns 0.5 for F4 (two FP4 values per byte)
• get_storage_shape(): collapses packed sub-byte shapes to flat byte count for DLPack so the workaround-dtype view doesn't overread the buffer
• get_native_shape(): converts safetensors FP4-element shape to PyTorch packed-pair shape
• get_empty_tensor(): uses get_native_shape() for correct allocation shape
• TorchTensor.reshape(): needed by common.py after DLPack import

fastsafetensors/frameworks/__init__.py

• get_dtype_size() return type: int → float (to accommodate 0.5 for F4)
• get_storage_shape() added (default: identity — no change for existing dtypes)
• get_native_shape() added (default: identity — no change for existing dtypes)
• TensorBase.reshape() added (default: NotImplementedError)

fastsafetensors/dlpack.py

• F4 and F8_E8M0: mapped as opaque (kDLUInt, 8, 1) — consistent with their U8 workaround

fastsafetensors/common.py

• Validation: nbytes = int(nelements * get_dtype_size()) — safe for fractional sizes
• get_tensors(): uses get_storage_shape() / get_native_shape() for packed dtypes

tests/test_fastsafetensors.py

• test_float8_e8m0fnu: bit-exact round-trip via _test_type
• test_float4_e2m1fn_x2: bit-exact round-trip (uses uint8 view, since randn/cast don't support float4)

F4 shape handling detail

safetensors stores an F4 weight of logical shape [2048, 4096] (8M FP4 values) as 4 MiB (0.5 bytes per FP4 value). torch.float4_e2m1fn_x2 represents paired values: [2048, 4096] in that dtype would be 8 MiB — wrong. The correct PyTorch shape is [2048, 2048] (4M packed pairs = 4 MiB).

get_storage_shape() handles DLPack by using a flat [4194304] uint8 shape (avoids buffer overread). get_native_shape() then reshapes to [2048, 2048] after the view. get_empty_tensor() also uses get_native_shape() so broadcast/scatter allocations are correctly sized.

Validation

Tested against DeepSeek V4-Flash MP=2 safetensors shards (~82 GiB each) containing all six dtypes (BF16, F32, F4, F8_E4M3, F8_E8M0, I64). With this PR:

• add_filenames() succeeds on V4-Flash shards
• All tensors load bit-exactly vs safetensors.torch.load_file
• 41 tests pass (39 existing + 2 new), zero regressions

Notes

• get_storage_shape() and get_native_shape() default to identity on FrameworkOpBase, so non-PyTorch frameworks (e.g. Paddle) are unaffected unless they add their own F4/F8_E8M0 dtype mappings.
• F8_E8M0 requires PyTorch ≥ 2.5; F4 requires PyTorch ≥ 2.10. Both additions are guarded by hasattr() so older PyTorch versions still work for all currently-supported dtypes.

[takeshi-yoshimura]

@uasind
Thanks for your contribution! While I review and test your code, please fix the lint issue and add signed-off-by like other commits.

[takeshi-yoshimura]

@uasind
Can you please resolve lint issues, DCO, and my comments. Thanks!

[takeshi-yoshimura]

Can you check st_shape[-1] % ratio == 0 and raise Exception on malformed safetensors files?

[takeshi-yoshimura]

Our CI does not include PyTorch 2.10 currently. Can you please add a new workflow to run only the new tests by modifying .github/workflows/test-torch.yaml?