llama.cpp/convert-pth-to-ggml.py

# Convert a LLaMA model checkpoint to a ggml compatible file
#
# Load the model using Torch
# Iterate over all variables and write them to a binary file.
#
# For each variable, write the following:
#   - Number of dimensions (int)
#   - Name length (int)
#   - Dimensions (int[n_dims])
#   - Name (char[name_length])
#   - Data (float[n_dims])
#
# By default, the bigger matrices are converted to 16-bit floats.
# This can be disabled by adding the "use-f32" CLI argument.
#
# At the start of the ggml file we write the model parameters
# and vocabulary.
#

import sys
import json
import struct
import numpy as np
import torch

from sentencepiece import SentencePieceProcessor

if len(sys.argv) < 3:
    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
dir_model = sys.argv[1]
fname_out = sys.argv[1] + "/ggml-model.bin"

fname_hparams   = sys.argv[1] + "/params.json"
fname_model     = sys.argv[1] + "/consolidated.00.pth"
fname_tokenizer = sys.argv[1] + "/../tokenizer.model"

# possible data types
#   ftype == 0 -> float32
#   ftype == 1 -> float16
#
# map from ftype to string
ftype_str = ["f32", "f16"]

ftype = 1
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"

with open(fname_hparams, "r") as f:
    hparams = json.load(f)

tokenizer = SentencePieceProcessor(fname_tokenizer)

hparams.update({"vocab_size": tokenizer.vocab_size()})

print(hparams)

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", 64)) # rot
fout.write(struct.pack("i", ftype))

# Is this correct??
for i in range(32000):
    # TODO: this is probably wrong - not sure how this tokenizer works
    text = tokenizer.decode([29889, i]).encode('utf-8')
    # remove the first byte (it's always '.')
    text = text[1:]
    fout.write(struct.pack("i", len(text)))
    fout.write(text)

for k, v in model.items():
    name = k
    shape = v.shape

    # skip layers.X.attention.inner_attention.rope.freqs
    if name[-5:] == "freqs":
        continue

    print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype)

    #data = tf.train.load_variable(dir_model, name).squeeze()
    data = v.numpy().squeeze()
    n_dims = len(data.shape);

    # for efficiency - transpose some matrices
    # "model/h.*/attn/c_attn/w"
    # "model/h.*/attn/c_proj/w"
    # "model/h.*/mlp/c_fc/w"
    # "model/h.*/mlp/c_proj/w"
    #if name[-14:] == "/attn/c_attn/w" or \
    #   name[-14:] == "/attn/c_proj/w" or \
    #   name[-11:] == "/mlp/c_fc/w" or \
    #   name[-13:] == "/mlp/c_proj/w":
    #    print("  Transposing")
    #    data = data.transpose()

    dshape = data.shape

    # default type is fp16
    ftype_cur = 1
    if ftype == 0 or n_dims == 1:
        print("  Converting to float32")
        data = data.astype(np.float32)
        ftype_cur = 0

    # header
    str = name.encode('utf-8')
    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
    for i in range(n_dims):
        fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
    fout.write(str);

    # data
    data.tofile(fout)

fout.close()

print("Done. Output file: " + fname_out)
print("")