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

convert-*.py: hash pytorch array as numpy without type conversion (except for bf16 which is typecasted upward)

Browse source
This commit is contained in:
brian khuu 2024-07-28 02:29:37 +10:00
parent f05fa2a2b1
commit 6db4f52d1c
1 changed files with 5 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

7
convert_hf_to_gguf.py
View file

@ -273,8 +273,8 @@ class Model:
for name, data_torch in self.get_tensors(): for name, data_torch in self.get_tensors():
uuidv5_data_buffer: np.ndarray = data_torch.to(torch.float64).numpy() uuidv5_data_buffer: np.ndarray = data_torch.numpy()
uuidv5_sha1.update(uuidv5_data_buffer.tobytes('C')) uuidv5_sha1.update(uuidv5_data_buffer.data.tobytes('C'))
# we don't need these # we don't need these
if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")): if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
@ -3506,6 +3506,9 @@ class LazyTorchTensor(gguf.LazyBase):
torch.float16: np.float16, torch.float16: np.float16,
torch.float32: np.float32, torch.float32: np.float32,
torch.float64: np.float64, torch.float64: np.float64,
# No direct mapping avaliable. Cast upwards to avoid loss of precision
torch.bfloat16: np.float32,
} }
# used for safetensors slices # used for safetensors slices