vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

truncate intermediate fp32 if converting bf16 to bf16

Browse source
This commit is contained in:
Sigbjørn Skjæret 2024-06-10 04:30:47 +02:00 • committed by GitHub
parent 6a52bfe332
commit 46054d1aab
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
convert-hf-to-gguf.py
View file

@ -295,8 +295,8 @@ class Model:
if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
if self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
data = gguf.quantize_bf16(data)
assert data.dtype == np.int16
data = gguf.truncate_bf16(data) if old_dtype == torch.bfloat16 else gguf.quantize_bf16(data)
assert data.dtype in (np.int16, np.uint16)
data_qtype = gguf.GGMLQuantizationType.BF16
elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0 and gguf.can_quantize_to_q8_0(data):