Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-07 11:23:56 +01:00
Code Issues Packages Projects Releases Wiki Activity

fix conflicts

Browse Source
This commit is contained in:
M. Yusuf Sarıgöz 2023-08-13 16:55:42 +03:00
parent 6beebf3fd9
commit 24f48833ab
1 changed files with 38 additions and 44 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

82
convert-llama-h5-to-gguf.py
View File

@ -23,7 +23,6 @@ def permute(weights: NDArray, n_head: int) -> NDArray:
.swapaxes(1, 2) .swapaxes(1, 2)
.reshape(weights.shape)) .reshape(weights.shape))
def count_model_parts(dir_model: str) -> int: def count_model_parts(dir_model: str) -> int:
num_parts = 0 num_parts = 0
for filename in os.listdir(dir_model): for filename in os.listdir(dir_model):
@ -34,7 +33,6 @@ def count_model_parts(dir_model: str) -> int:
print("gguf: found " + str(num_parts) + " model parts") print("gguf: found " + str(num_parts) + " model parts")
return num_parts return num_parts
if len(sys.argv) < 3: if len(sys.argv) < 3:
print("Usage: convert-h5-to-ggml.py dir-model ftype\n") print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
print(" ftype == 0 -> float32") print(" ftype == 0 -> float32")
@ -188,7 +186,7 @@ else:
) )
for part_name in part_names: for part_name in part_names:
print("gguf: loading model part '" + part_name + "'") print("gguf: loading model part '"+ part_name + "'")
model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
for name in model_part.keys(): for name in model_part.keys():
@ -206,7 +204,7 @@ for part_name in part_names:
# permute these # permute these
if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"): if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"):
data = permute(data, head_count) data = permute(data,head_count)
# map tensor names # map tensor names
if name.endswith(".weight") and name[:-7] in tensor_map: if name.endswith(".weight") and name[:-7] in tensor_map:
@ -214,11 +212,11 @@ for part_name in part_names:
elif name.endswith(".bias") and name[:-5] in tensor_map: elif name.endswith(".bias") and name[:-5] in tensor_map:
name = tensor_map[name[:-5]] + ".bias" name = tensor_map[name[:-5]] + ".bias"
else: else:
print("Can not map tensor '" + name + "'") print( "Can not map tensor '" + name + "'" )
sys.exit() sys.exit()
n_dims = len(data.shape) n_dims = len(data.shape)
data_dtype = data.dtype data_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:
@ -255,60 +253,56 @@ else:
) )
for part_name in part_names: for part_name in part_names:
print("gguf: loading model part '" + part_name + "'") print("gguf: loading model part '"+ part_name + "'")
model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
for name in model_part.keys(): for name in model_part.keys():
data = model_part[name] data = model_part[name]
<< << << < HEAD
n_dims = len(data.shape) old_dtype = data.dtype
data_dtype = data.dtype
== == == =
old_dtype = data.dtype
# we don't need these # we don't need these
if name.endswith(".rotary_emb.inv_freq"): if name.endswith(".rotary_emb.inv_freq"):
continue continue
>>>>>> > 17800cd80fec468411481dc34a51d42a936442f1
# 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)
data = data.squeeze().numpy() data = data.squeeze().numpy()
# permute these # permute these
if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"): if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"):
data = permute(data, head_count) data = permute(data, head_count)
# map tensor names # map tensor names
if name.endswith(".weight") and name[:-7] in tensor_map: if name.endswith(".weight") and name[:-7] in tensor_map:
name = tensor_map[name[:-7]] + ".weight" name = tensor_map[name[:-7]] + ".weight"
elif name.endswith(".bias") and name[:-5] in tensor_map: elif name.endswith(".bias") and name[:-5] in tensor_map:
name = tensor_map[name[:-5]] + ".bias" name = tensor_map[name[:-5]] + ".bias"
else: else:
print("Can not map tensor '" + name + "'" ) print( "Can not map tensor '" + name + "'" )
sys.exit() sys.exit()
n_dims = len(data.shape) n_dims = len(data.shape)
data_dtype = data.dtype data_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 = data.astype(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 = data.astype(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 = data.astype(np.float16) data = data.astype(np.float16)
print(name + ", shape " + str(len(data.shape)) + ", " + str(old_dtype) + " --> " + str(data.dtype)) print( name + ", shape " + str(len(data.shape)) + ", " + str(old_dtype) + " --> " + str(data.dtype))
gguf_writer.write_tensor_to_file(data) gguf_writer.write_tensor_to_file(data)
gguf_writer.close() gguf_writer.close()