diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-17 01:21:41 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-17 01:21:41 +0200 |
| commit | e00becb643b3a99de16462b3a032717a93b983d9 (patch) | |
| tree | 873ad2646b0ef2f07eae5a5a74387d062c77ec0c /src/dtype.h | |
| parent | b4343d6ac1c6cc7a7d228f954220def2f6e6ed67 (diff) | |
| download | ffwd-e00becb643b3a99de16462b3a032717a93b983d9.tar.gz ffwd-e00becb643b3a99de16462b3a032717a93b983d9.zip | |
all: accept F16 safetensors weights (widen to F32 at load)
Diffstat (limited to 'src/dtype.h')
| -rw-r--r-- | src/dtype.h | 36 |
1 files changed, 36 insertions, 0 deletions
diff --git a/src/dtype.h b/src/dtype.h index 5c1ce8f..2a25cae 100644 --- a/src/dtype.h +++ b/src/dtype.h @@ -16,4 +16,40 @@ static inline float embed_bf16_to_f32(uint16_t bf16) { return f; } +/* IEEE-754 binary16 -> binary32. Unlike BF16, F16 has a 5-bit exponent and a + * 10-bit mantissa, so widening is a field re-pack, not a shift: rebias the + * exponent (15 -> 127), shift the mantissa up by 13, and special-case zeros, + * subnormals, and inf/NaN. Used by the loaders to accept F16 safetensors + * snapshots (e.g. the public multilingual-e5 weights); there is no fused F16 + * kernel, so weights are widened to F32 once at load. */ +static inline float embed_f16_to_f32(uint16_t f16) { + uint32_t sign = (uint32_t)(f16 & 0x8000u) << 16; + uint32_t exp = (f16 >> 10) & 0x1fu; + uint32_t mant = f16 & 0x3ffu; + uint32_t bits; + if (exp == 0) { + if (mant == 0) { + bits = sign; /* signed zero */ + } else { + /* Subnormal half: renormalize into a normal float. Shift the + * mantissa left until its leading 1 reaches the hidden-bit + * position, decrementing the exponent for each shift. */ + exp = 127 - 15 + 1; + while ((mant & 0x400u) == 0) { + mant <<= 1; + exp--; + } + mant &= 0x3ffu; + bits = sign | (exp << 23) | (mant << 13); + } + } else if (exp == 0x1fu) { + bits = sign | 0x7f800000u | (mant << 13); /* inf or NaN */ + } else { + bits = sign | ((exp + (127 - 15)) << 23) | (mant << 13); + } + float f; + memcpy(&f, &bits, sizeof(f)); + return f; +} + #endif /* EMBED_DTYPE_H */ |
