Knowledge note
HiFloat4 (HiF4)
HiFloat4 (HiF4) is a 4.5-bit-per-value block floating-point format that packs 64 S1P2 payloads with a three-level shared scaling hierarchy.
HiFloat4 (HiF4)
HiFloat4 (HiF4) is a 4.5-bit-per-value block floating-point format that stores 64 four-bit S1P2 values with 32 bits of shared E6M2, E1_8, and E1_16 scaling metadata.
Why It Exists
Four-bit scalar payloads have too little precision and range for many LLM tensors. HiF4 amortizes exponent information across a 64-value unit so the payload can keep a three-bit significand while local one-bit flags follow smaller range changes.
How It Works
One E6M2 scale supplies the shared base magnitude. Eight one-bit E1_8 flags cover groups of eight values, and sixteen one-bit E1_16 flags cover groups of four. The payload uses sign-magnitude S1P2, equivalent to E1M2, so local flags can be absorbed as shifts before the dot-product reduction.
Tradeoffs
HiF4’s benefits depend on support for its conversion path and 64-wide integer-heavy dot product. The source reports inference results and hardware estimates, but leaves training evaluation and a public native implementation open.
Where It Appears
- HiFloat4 (HiF4): 4-Bit Block Floating Point for LLM Inference — Format layout, BF16 conversion, dot-product hardware flow, and LLM inference comparisons.