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 os
import sys
import json
import struct
import numpy as np
import torch
import transformers
import torch
from sentencepiece import SentencePieceProcessor

# if models/ directory does not exist, raise an error
if not os.path.exists("models"):
    raise Exception("Please create a models/ directory and add the LLaMa model files there.")

# if adapter_model.bin is not in directory, download it from https://huggingface.co/tloen/alpaca-lora-7b/resolve/main/adapter_model.bin
if not os.path.exists("adapter_model.bin"):
    import urllib.request
    url = "https://huggingface.co/tloen/alpaca-lora-7b/resolve/main/adapter_model.bin"
    urllib.request.urlretrieve(url, "adapter_model.bin")

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_hparams   = sys.argv[1] + "/params.json"
fname_tokenizer = sys.argv[1] + "/../tokenizer.model"

def get_n_parts(dim):
    assert dim == 4096, "Only LLaMa 7B has been instruct tuned."
    if dim == 4096:
        return 1
    elif dim == 5120:
        return 2
    elif dim == 6656:
        return 4
    elif dim == 8192:
        return 8
    else:
        print("Invalid dim: " + str(dim))
        sys.exit(1)

# 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()})

n_parts = get_n_parts(hparams["dim"])

print(hparams)
print('n_parts = ', n_parts)

##### Set up LoRa ####
RANK = 8
LORA_ALPHA = 16 # https://huggingface.co/tloen/alpaca-lora-7b/blob/431cdc6483f8535d2c745b2d7f8a2fd03df7b6ec/adapter_config.json#L7
LORA_SCALING = LORA_ALPHA / RANK # https://github.com/huggingface/peft/blob/64f63a7df2a02cfd144592d9aa9c871b59258c55/src/peft/tuners/lora.py#LL308C18-L308C52
lora_tensors = {}
lora_wts = torch.load("adapter_model.bin", map_location=torch.device('cpu'))
for key in lora_wts.keys():
    lora_tensors[key] = lora_wts[key]
    # if key ends with lora_A.weight, and its first dimension is 8, then concat zeros to change first dimension of the torch.tensor to 32
    if key.endswith("lora_A.weight") and lora_tensors[key].shape[0] == 8:
        shape_0 = lora_tensors[key].shape[0]
        shape_1 = lora_tensors[key].shape[1]
        zeros = torch.zeros((32 - shape_0, shape_1))
        lora_tensors[key] = torch.cat((lora_tensors[key], zeros), 0)
    # if key ends with lora_B.weight, and its second dimension is 8, then concat zeros to change second dimension of the torch.tensor to 32
    if key.endswith("lora_B.weight") and lora_tensors[key].shape[1] == 8:
        shape_0 = lora_tensors[key].shape[0]
        shape_1 = lora_tensors[key].shape[1]
        zeros = torch.zeros((shape_0, 32 - shape_1))
        lora_tensors[key] = torch.cat((lora_tensors[key], zeros), 1)
    # Scale the tensors
    lora_tensors[key] = lora_tensors[key]

p = 0
print('Processing part ', p)

#fname_model = sys.argv[1] + "/consolidated.00.pth"
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()):
    if tokenizer.is_unknown(i):
        # "<unk>" token (translated as ??)
        text = " \u2047 ".encode("utf-8")
        fout.write(struct.pack("i", len(text)))
        fout.write(text)
    elif tokenizer.is_control(i):
        # "<s>"/"</s>" tokens
        fout.write(struct.pack("i", 0))
    elif tokenizer.is_byte(i):
        # "<U+XX>" tokens (which may be invalid UTF-8)
        piece = tokenizer.id_to_piece(i)
        if len(piece) != 6:
            print("Invalid token: " + piece)
            sys.exit(1)
        byte_value = int(piece[3:-1], 16)
        fout.write(struct.pack("i", 1))
        fout.write(struct.pack("B", byte_value))
    else:
        # normal token. Uses U+2581 (LOWER ONE EIGHTH BLOCK) to represent spaces.
        text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
        fout.write(struct.pack("i", len(text)))
        fout.write(text)

for k, v in model.items():
    if '.attention.wq.weight' in k: # then the key is of the format "layers.{LAYER_NUMBER}.attention.wq.weight"
        # Get the layer number
        layer_number = int(k.split('.')[1])
        # Get corresponding lora_A tensor 'layers.{LAYER_NUMBER}.self_attn.q_proj.lora_A.weight'
        lora_A_tensor = lora_tensors['base_model.model.model.layers.'
                                     + str(layer_number) + '.self_attn.q_proj.lora_A.weight']
        # Get corresponding lora_B tensor 'layers.{LAYER_NUMBER}.self_attn.q_proj.lora_B.weight'
        lora_B_tensor = lora_tensors['base_model.model.model.layers.'
                                     + str(layer_number) + '.self_attn.q_proj.lora_B.weight']
        # Merge the two tensors https://github.com/huggingface/peft/blob/64f63a7df2a02cfd144592d9aa9c871b59258c55/src/peft/tuners/lora.py
        v += (lora_B_tensor @ lora_A_tensor).T * LORA_SCALING
    # Similarly for 'attention.wv.weight'
    if '.attention.wv.weight' in k: # then the key is of the format "layers.{LAYER_NUMBER}.attention.wv.weight"
        # Get the layer number
        layer_number = int(k.split('.')[1])
        # Get corresponding lora_A tensor 'layers.{LAYER_NUMBER}.self_attn.v_proj.lora_A.weight'
        lora_A_tensor = lora_tensors['base_model.model.model.layers.' +
                                     str(layer_number) + '.self_attn.v_proj.lora_A.weight']
        # Get corresponding lora_B tensor 'layers.{LAYER_NUMBER}.self_attn.v_proj.lora_B.weight'
        lora_B_tensor = lora_tensors['base_model.model.model.layers.' +
                                     str(layer_number) + '.self_attn.v_proj.lora_B.weight']
        # Merge the two tensors https://github.com/huggingface/peft/blob/64f63a7df2a02cfd144592d9aa9c871b59258c55/src/peft/tuners/lora.py
        v += (lora_B_tensor @ lora_A_tensor).T * LORA_SCALING

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

    # data
    data.tofile(fout)

# I hope this deallocates the memory.
model = None

fout.close()

print("Done. Output file: " + fname_out + ", (part ", p, ")")
print("")