root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-28 12:24:35 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

convert-llama-7b-pth-to-gguf.py : fixes

Browse Source
This commit is contained in:
klosax 2023-08-17 21:44:34 +02:00 committed by GitHub
parent 9e2d4dd48e
commit 3c1b7217a9
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 13 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
convert-llama-7b-pth-to-gguf.py
View File

@ -76,8 +76,9 @@ if num_parts > 1:
print("gguf: Only models with a single datafile are supported.") print("gguf: Only models with a single datafile are supported.")
sys.exit() sys.exit()
llm_arch = "llama"
gguf_writer = gguf.GGUFWriter(fname_out, arch=llm_arch) ARCH=gguf.MODEL_ARCH.LLAMA
gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
print("gguf: get model metadata") print("gguf: get model metadata")
@ -195,7 +196,7 @@ if Path(dir_model + "/tokenizer.json").is_file():
# TENSORS # TENSORS
tensor_map = gguf.get_tensor_name_map(block_count) tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
# tensor info # tensor info
print("gguf: get tensor metadata") print("gguf: get tensor metadata")
@ -213,6 +214,8 @@ for part_name in part_names:
if name == "rope.freqs": if name == "rope.freqs":
continue continue
old_dtype = data.dtype
# convert any unsupported data types to float32 # convert any unsupported data types to float32
if data.dtype != torch.float16 and data.dtype != torch.float32: if data.dtype != torch.float16 and data.dtype != torch.float32:
data = data.to(torch.float32) data = data.to(torch.float32)
@ -230,23 +233,20 @@ for part_name in part_names:
n_dims = len(data.shape) n_dims = len(data.shape)
data_dtype = data.dtype data_dtype = data.dtype
old_dtype = data_dtype
# if f32 desired, convert any float16 to float32 # if f32 desired, convert any float16 to float32
if ftype == 0 and data.dtype == np.float16: if ftype == 0 and data_dtype == np.float16:
data_dtype = np.float32 data = data.astype(np.float32)
# TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
if ftype == 1 and data_dtype == np.float16 and n_dims == 1: if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
data_dtype = np.float32 data = data.astype(np.float32)
# if f16 desired, convert any float32 2-dim weight tensors to float16 # if f16 desired, convert any float32 2-dim weight tensors to float16
if ftype == 1 and data.dtype == np.float32 and name.endswith(".weight") and n_dims == 2: if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
data_dtype = np.float16 data = data.astype(np.float16)
print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data_dtype)) print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
data = data.astype(data_dtype)
gguf_writer.add_tensor(name, data) gguf_writer.add_tensor(name, data)