Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-03 17:51:09 +01:00
Code Issues Packages Projects Releases Wiki Activity

gguf : refactor pth to gguf conversion script

Browse Source
This commit is contained in:
M. Yusuf Sarıgöz 2023-08-17 19:58:27 +03:00
parent f71704177f
commit 1d93d04ce2
1 changed files with 39 additions and 81 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

120
convert-llama-7b-pth-to-gguf.py
View File

@ -18,6 +18,7 @@ from sentencepiece import SentencePieceProcessor
# compatible with python < 3.9 # compatible with python < 3.9
NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]' NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
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):
@ -28,10 +29,12 @@ 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")
print(" ftype == 1 -> float16") print(" ftype == 1 -> float16")
sys.exit(1) sys.exit(1)
@ -43,7 +46,7 @@ last_dir = os.path.basename(os.path.normpath(dir_model))
# possible tensor data types # possible tensor data types
# ftype == 0 -> float32 # ftype == 0 -> float32
# ftype == 1 -> float16 # ftype == 1 -> float16
#
# map from ftype to string # map from ftype to string
ftype_str = ["f32", "f16"] ftype_str = ["f32", "f16"]
@ -52,6 +55,7 @@ if len(sys.argv) > 2:
ftype = int(sys.argv[2]) ftype = int(sys.argv[2])
if ftype < 0 or ftype > 1: if ftype < 0 or ftype > 1:
print("Invalid ftype: " + str(ftype)) print("Invalid ftype: " + str(ftype))
sys.exit(1) sys.exit(1)
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf" fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".gguf"
@ -70,14 +74,14 @@ num_parts = count_model_parts(dir_model)
if num_parts > 1: 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()
gguf_writer = gguf.GGUFWriter.open(fname_out) sys.exit()
llm_arch = "llama"
gguf_writer = gguf.GGUFWriter(fname_out, arch=llm_arch)
print("gguf: get model metadata") print("gguf: get model metadata")
llm_arch = "llama"
block_count = hparams["num_hidden_layers"] block_count = hparams["num_hidden_layers"]
head_count = hparams["num_attention_heads"] head_count = hparams["num_attention_heads"]
@ -89,21 +93,20 @@ else:
if "_name_or_path" in hparams: if "_name_or_path" in hparams:
hf_repo = hparams["_name_or_path"] hf_repo = hparams["_name_or_path"]
else: else:
hf_repo="" hf_repo = ""
gguf_writer.add_architecture(llm_arch) gguf_writer.add_architecture()
gguf_writer.add_name(last_dir) gguf_writer.add_name(last_dir)
gguf_writer.add_file_type( "All tensors F32" if ftype == 0 else "Most tensors F16, some F32")
gguf_writer.add_source_hf_repo(hf_repo) gguf_writer.add_source_hf_repo(hf_repo)
gguf_writer.add_tensor_data_layout(llm_arch, "Meta AI original pth") gguf_writer.add_tensor_data_layout("Meta AI original pth")
gguf_writer.add_context_length(llm_arch, hparams["max_position_embeddings"]) gguf_writer.add_context_length(hparams["max_position_embeddings"])
gguf_writer.add_embedding_length(llm_arch, hparams["hidden_size"]) gguf_writer.add_embedding_length(hparams["hidden_size"])
gguf_writer.add_block_count(llm_arch, block_count) gguf_writer.add_block_count(block_count)
gguf_writer.add_feed_forward_length(llm_arch, hparams["intermediate_size"]) gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
gguf_writer.add_rope_dimension_count(llm_arch, hparams["hidden_size"] // hparams["num_attention_heads"]) gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
gguf_writer.add_head_count(llm_arch, head_count) gguf_writer.add_head_count(head_count)
gguf_writer.add_head_count_kv(llm_arch, head_count_kv) gguf_writer.add_head_count_kv(head_count_kv)
gguf_writer.add_layer_norm_rms_eps(llm_arch, hparams["rms_norm_eps"]) gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
# TOKENIZATION # TOKENIZATION
@ -125,19 +128,23 @@ if Path(dir_model + "/tokenizer.model").is_file():
score: float score: float
piece = tokenizer.id_to_piece(i) piece = tokenizer.id_to_piece(i)
text = piece.encode("utf-8") text = piece.encode("utf-8")
score = tokenizer.get_score(i) score = tokenizer.get_score(i)
toktype = 1 # defualt to normal token type toktype = 1 # defualt to normal token type
if tokenizer.is_unknown(i): toktype = 2 if tokenizer.is_unknown(i):
if tokenizer.is_control(i): toktype = 3 toktype = 2
if tokenizer.is_control(i):
toktype = 3
# TODO: How to determinate if a token is user defined? # TODO: How to determinate if a token is user defined?
# ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
# if tokenizer.is_user_defined(i): toktype = 4 # if tokenizer.is_user_defined(i): toktype = 4
if tokenizer.is_unused(i): toktype = 5 if tokenizer.is_unused(i):
if tokenizer.is_byte(i): toktype = 6 toktype = 5
if tokenizer.is_byte(i):
toktype = 6
tokens.append(text) tokens.append(text)
scores.append(score) scores.append(score)
@ -193,10 +200,10 @@ tensor_map = gguf.get_tensor_name_map(block_count)
# tensor info # tensor info
print("gguf: get tensor metadata") print("gguf: get tensor metadata")
part_names = ( f"consolidated.{n:02}.pth" for n in range(0, num_parts) ) part_names = (f"consolidated.{n:02}.pth" for n in range(0, num_parts))
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():
@ -218,11 +225,12 @@ 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
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:
@ -236,69 +244,19 @@ for part_name in part_names:
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_dtype = np.float16
data_nbytes = data.size * 2 if data_dtype == np.float16 else data.size * 4 print(name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data_dtype))
gguf_writer.add_tensor_info(name, data.shape, data_dtype, data_nbytes) data = data.astype(data_dtype)
gguf_writer.add_tensor(name, data)
print("gguf: write header") print("gguf: write header")
gguf_writer.write_header_to_file() gguf_writer.write_header_to_file()
print("gguf: write metadata") print("gguf: write metadata")
gguf_writer.write_kv_data_to_file() gguf_writer.write_kv_data_to_file()
print("gguf: write tensor metadata") print("gguf: write tensors")
gguf_writer.write_ti_data_to_file() gguf_writer.write_tensors_to_file()
# tensor data
print("gguf: convert and write tensor data")
part_names = ( f"consolidated.{n:02}.pth" for n in range(0, num_parts) )
for part_name in part_names:
print("gguf: loading model part '"+ part_name + "'")
model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
for name in model_part.keys():
data = model_part[name]
old_dtype = data.dtype
# we don't need these
if name == "rope.freqs":
continue
# convert any unsupported data types to float32
if data.dtype != torch.float16 and data.dtype != torch.float32:
data = data.to(torch.float32)
data = data.squeeze().numpy()
# map tensor names
if name.endswith(".weight") and name[:-7] in tensor_map:
name = tensor_map[name[:-7]] + ".weight"
elif name.endswith(".bias") and name[:-5] in tensor_map:
name = tensor_map[name[:-5]] + ".bias"
else:
print( "Can not map tensor '" + name + "'" )
sys.exit()
n_dims = len(data.shape)
data_dtype = data.dtype
# if f32 desired, convert any float16 to float32
if ftype == 0 and data.dtype == np.float16:
data = data.astype(np.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:
data = data.astype(np.float32)
# 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:
data = data.astype(np.float16)
print( name + ", shape " + str(len(data.shape)) + ", " + str(old_dtype) + " --> " + str(data.dtype))
gguf_writer.write_tensor_data(data)
gguf_writer.close() gguf_writer.close()