Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 22:08:46 +01:00
Code Issues Packages Projects Releases Wiki Activity

Refactoring convert-pth-to-ggml.py: more concise and readable (#109)

Browse Source 
* Refactor get_n_parts function to simplify code and improve readability

* Use f-strings instead of concatenation

* Refactoring: more concise and readable

* modularize

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
qunash 2023-03-19 20:17:39 +03:00 committed by GitHub
parent 70f01cb863
commit 467b149761
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 84 additions and 108 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

168
convert-pth-to-ggml.py
View File

@ -16,7 +16,7 @@
# At the start of the ggml file we write the model parameters # At the start of the ggml file we write the model parameters
# and vocabulary. # and vocabulary.
# #
import os import argparse
import sys import sys
import json import json
import struct import struct
@ -24,137 +24,91 @@ import numpy as np
import torch import torch
from sentencepiece import SentencePieceProcessor from sentencepiece import SentencePieceProcessor
if len(sys.argv) < 3: def parse_args():
print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
print(" ftype == 0 -> float32")
print(" ftype == 1 -> float16")
sys.exit(1)
# output in the same directory as the model parser = argparse.ArgumentParser(description='Convert a LLaMA model checkpoint to a ggml compatible file')
dir_model = sys.argv[1] parser.add_argument('dir_model', help='directory containing the model checkpoint')
parser.add_argument('ftype', type=int, choices=[0, 1], default=1, help='file type (0: float32, 1: float16)')
fname_hparams = sys.argv[1] + "/params.json" return parser.parse_args()
fname_tokenizer = sys.argv[1] + "/../tokenizer.model"
def get_n_parts(dim): def get_n_parts(dim):
if dim == 4096:
return 1 mappings = {4096: 1, 5120: 2, 6656: 4, 8192: 8}
elif dim == 5120: n_parts = mappings.get(dim)
return 2 if n_parts is None:
elif dim == 6656: print(f"Invalid dim: {dim}")
return 4
elif dim == 8192:
return 8
else:
print("Invalid dim: " + str(dim))
sys.exit(1) sys.exit(1)
# possible data types print(f"n_parts = {n_parts}\n")
# ftype == 0 -> float32 return n_parts
# ftype == 1 -> float16
#
# map from ftype to string
ftype_str = ["f32", "f16"]
ftype = 1 def load_hparams_and_tokenizer(dir_model):
if len(sys.argv) > 2:
ftype = int(sys.argv[2])
if ftype < 0 or ftype > 1:
print("Invalid ftype: " + str(ftype))
sys.exit(1)
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
if os.path.exists(fname_out): fname_hparams = f"{dir_model}/params.json"
print(f"Skip conversion, it already exists: {fname_out}") fname_tokenizer = f"{dir_model}/../tokenizer.model"
sys.exit(0)
with open(fname_hparams, "r") as f: with open(fname_hparams, "r") as f:
hparams = json.load(f) hparams = json.load(f)
print(hparams)
tokenizer = SentencePieceProcessor(fname_tokenizer) tokenizer = SentencePieceProcessor(fname_tokenizer)
hparams.update({"vocab_size": tokenizer.vocab_size()}) hparams.update({"vocab_size": tokenizer.vocab_size()})
n_parts = get_n_parts(hparams["dim"]) return hparams, tokenizer
print(hparams) def write_header(fout, hparams, ftype):
print('n_parts = ', n_parts)
for p in range(n_parts): keys = ["vocab_size", "dim", "multiple_of", "n_heads", "n_layers"]
print('Processing part ', p) values = [
0x67676d6c, # magic: ggml in hex
*[hparams[key] for key in keys],
hparams["dim"] // hparams["n_heads"], # rot (obsolete)
ftype
]
fout.write(struct.pack("i" * len(values), *values))
#fname_model = sys.argv[1] + "/consolidated.00.pth" def write_tokens(fout, tokenizer):
fname_model = sys.argv[1] + "/consolidated.0" + str(p) + ".pth"
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
if (p > 0):
fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin" + "." + str(p)
model = torch.load(fname_model, map_location="cpu")
fout = open(fname_out, "wb")
fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
fout.write(struct.pack("i", hparams["vocab_size"]))
fout.write(struct.pack("i", hparams["dim"]))
fout.write(struct.pack("i", hparams["multiple_of"]))
fout.write(struct.pack("i", hparams["n_heads"]))
fout.write(struct.pack("i", hparams["n_layers"]))
fout.write(struct.pack("i", hparams["dim"] // hparams["n_heads"])) # rot (obsolete)
fout.write(struct.pack("i", ftype))
# Is this correct??
for i in range(tokenizer.vocab_size()): for i in range(tokenizer.vocab_size()):
if tokenizer.is_unknown(i): if tokenizer.is_unknown(i):
# "<unk>" token (translated as ??)
text = " \u2047 ".encode("utf-8") text = " \u2047 ".encode("utf-8")
fout.write(struct.pack("i", len(text)))
fout.write(text)
elif tokenizer.is_control(i): elif tokenizer.is_control(i):
# "<s>"/"</s>" tokens text = b""
fout.write(struct.pack("i", 0))
elif tokenizer.is_byte(i): elif tokenizer.is_byte(i):
# "<U+XX>" tokens (which may be invalid UTF-8)
piece = tokenizer.id_to_piece(i) piece = tokenizer.id_to_piece(i)
if len(piece) != 6: if len(piece) != 6:
print("Invalid token: " + piece) print(f"Invalid token: {piece}")
sys.exit(1) sys.exit(1)
byte_value = int(piece[3:-1], 16) byte_value = int(piece[3:-1], 16)
fout.write(struct.pack("i", 1)) text = struct.pack("B", byte_value)
fout.write(struct.pack("B", byte_value))
else: else:
# normal token. Uses U+2581 (LOWER ONE EIGHTH BLOCK) to represent spaces.
text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8") text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
fout.write(struct.pack("i", len(text))) fout.write(struct.pack("i", len(text)))
fout.write(text) fout.write(text)
for k, v in model.items(): def process_and_write_variables(fout, model, ftype):
name = k
shape = v.shape
# skip layers.X.attention.inner_attention.rope.freqs for name, data in model.items():
if name[-5:] == "freqs":
if name.endswith("freqs"):
continue continue
print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype) shape = data.shape
#data = tf.train.load_variable(dir_model, name).squeeze() print(f"Processing variable: {name} with shape: {shape} and type: {data.dtype}\n")
data = v.numpy().squeeze()
n_dims = len(data.shape); data = np.squeeze(data)
n_dims = len(shape)
# for efficiency - transpose some matrices # for efficiency - transpose some matrices
# "model/h.*/attn/c_attn/w" # "model/h.*/attn/c_attn/w"
# "model/h.*/attn/c_proj/w" # "model/h.*/attn/c_proj/w"
# "model/h.*/mlp/c_fc/w" # "model/h.*/mlp/c_fc/w"
# "model/h.*/mlp/c_proj/w" # "model/h.*/mlp/c_proj/w"
#if name[-14:] == "/attn/c_attn/w" or \ #if name.endswith(("/attn/c_attn/w", "/attn/c_proj/w", "/mlp/c_fc/w", "/mlp/c_proj/w")):
# name[-14:] == "/attn/c_proj/w" or \
# name[-11:] == "/mlp/c_fc/w" or \
# name[-13:] == "/mlp/c_proj/w":
# print("Transposing") # print("Transposing")
# data = data.transpose() # data = data.transpose()
dshape = data.shape
# default type is fp16 # default type is fp16
ftype_cur = 1 ftype_cur = 1
if ftype == 0 or n_dims == 1: if ftype == 0 or n_dims == 1:
@ -164,18 +118,40 @@ for p in range(n_parts):
# header # header
sname = name.encode('utf-8') sname = name.encode('utf-8')
fout.write(struct.pack("iii", n_dims, len(sname), ftype_cur)) fout.write(struct.pack("iii", len(data.shape), len(sname), ftype_cur))
for i in range(n_dims): for dim in reversed(data.shape):
fout.write(struct.pack("i", dshape[n_dims - 1 - i])) fout.write(struct.pack("i", dim))
fout.write(sname); fout.write(sname)
# data # data
data.tofile(fout) data.tofile(fout)
# I hope this deallocates the memory .. def main():
model = None
fout.close() args = parse_args()
dir_model = args.dir_model
ftype = args.ftype
ftype_str = ["f32", "f16"]
print("Done. Output file: " + fname_out + ", (part ", p, ")") hparams, tokenizer = load_hparams_and_tokenizer(dir_model)
print("") n_parts = get_n_parts(hparams["dim"])
for p in range(n_parts):
print(f"Processing part {p}\n")
fname_model = f"{dir_model}/consolidated.0{p}.pth"
fname_out = f"{dir_model}/ggml-model-{ftype_str[ftype]}.bin{'' if p == 0 else '.' + str(p)}"
model = torch.load(fname_model, map_location="cpu")
with open(fname_out, "wb") as fout:
write_header(fout, hparams, ftype)
write_tokens(fout, tokenizer)
process_and_write_variables(fout, model, ftype)
del model
print(f"Done. Output file: {fname_out}, (part {p})\n")
if __name__ == "__main__":
main()