Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Support wildcard for source files in convert utilities #399

Merged
merged 2 commits into from
Jul 2, 2023
Merged

Support wildcard for source files in convert utilities #399

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a1e4bea
Support wildcard for source files
guillaume-be Jun 25, 2023
20878ee
- Support offloading at end of every file for conversion
guillaume-be Jun 25, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
156 changes: 112 additions & 44 deletions utils/convert_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
# Copyright 2019-2023 Guillaume Becquin
# Copyright 2023 https://github.com/starkat99/half-rs
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
Expand All @@ -9,15 +10,51 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# Copyright (c) 2005-2023, NumPy Developers.
# All rights reserved.
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are
# met:
#
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
#
# * Redistributions in binary form must reproduce the above
# copyright notice, this list of conditions and the following
# disclaimer in the documentation and/or other materials provided
# with the distribution.
#
# * Neither the name of the NumPy Developers nor the names of any
# contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.


import argparse
import numpy as np
import glob
import logging
import subprocess
import sys
import torch

import zipfile
from pathlib import Path
from typing import Dict

import numpy as np
import torch
from numpy.lib.format import write_array
from numpy.lib.npyio import zipfile_factory
from torch import Tensor


Expand All @@ -32,15 +69,37 @@ def get_bf16_repr(input_tensor: torch.Tensor) -> np.ndarray:
nan_mask = np.logical_and(byte_array, 0x7FFF_FFFF) > 0x7F80_0000
round_bit = 0x0000_8000
output_val = np.right_shift(byte_array, 16)
threshold_mask = (np.logical_and(byte_array, round_bit) != 0) & (np.logical_and(byte_array, (3*round_bit-1)) != 0)
output = np.where(nan_mask, nan_value, np.where(threshold_mask, output_val+1, output_val)).astype(np.uint16)
threshold_mask = (np.logical_and(byte_array, round_bit) != 0) & (
np.logical_and(byte_array, (3 * round_bit - 1)) != 0
)
output = np.where(
nan_mask, nan_value, np.where(threshold_mask, output_val + 1, output_val)
).astype(np.uint16)
return output


def append_to_zipf(
array_dict: Dict[str, np.ndarray], parent_zipfile: zipfile.ZipFile
) -> None:
"""Append a dictionary of arrays to a parent zipfile.

Inspired from https://github.com/numpy/numpy/blob/main/numpy/lib/npyio.py
shared under BSD 3-Clause license by the numpy team
"""
for key, array in array_dict.items():
internal_filename = key + ".npy"
array = np.asanyarray(array)
with parent_zipfile.open(internal_filename, "w", force_zip64=True) as f_in:
write_array(f_in, array, allow_pickle=True, pickle_kwargs=None)


if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"source_file", nargs="+", help="Absolute path to the Pytorch weights file to convert"
"source_file",
nargs="+",
help="""Absolute path (or file pattern) to the Pytorch weights file(s) to convert.
A single file, list of files, glob pattern or list of glob patterns can be provided.""",
)
parser.add_argument(
"--skip_embeddings",
Expand All @@ -67,45 +126,54 @@ def get_bf16_repr(input_tensor: torch.Tensor) -> np.ndarray:
)
args = parser.parse_args()

nps = {}
target_folder = Path(args.source_file[0]).parent
with zipfile_factory(
target_folder / "model.npz", mode="w", compression=False
) as output_zipfile:
for source_file_or_pattern in args.source_file:
source_files = glob.glob(source_file_or_pattern)
for source_file in source_files:
logging.info(f"Processing source file {source_file}...")
nps = {}
source_file = Path(source_file)
weights = torch.load(str(source_file), map_location="cpu")

for source_file in args.source_file:
source_file = Path(source_file)
weights = torch.load(str(source_file), map_location="cpu")

for k, v in weights.items():
k = k.replace("gamma", "weight").replace("beta", "bias")
if args.skip_embeddings:
if k in {
"model.encoder.embed_tokens.weight",
"encoder.embed_tokens.weight",
"model.decoder.embed_tokens.weight",
"decoder.embed_tokens.weight",
}:
continue
if args.skip_lm_head:
if k in {
"lm_head.weight",
}:
continue
if args.prefix:
k = args.prefix + k
if args.suffix:
k = k.split(".")[-1]
if isinstance(v, Tensor):
if v.dtype == torch.bfloat16:
tensor = get_bf16_repr(v)
else:
tensor = v.cpu().numpy()
if args.dtype is not None:
nps[k] = np.ascontiguousarray(tensor.astype(np.dtype(args.dtype)))
else:
nps[k] = np.ascontiguousarray(tensor)
print(f"converted {k} - {str(sys.getsizeof(nps[k]))} bytes")
else:
print(f"skipped non-tensor object: {k}")
np.savez(target_folder / "model.npz", **nps)
for k, v in weights.items():
k = k.replace("gamma", "weight").replace("beta", "bias")
if args.skip_embeddings:
if k in {
"model.encoder.embed_tokens.weight",
"encoder.embed_tokens.weight",
"model.decoder.embed_tokens.weight",
"decoder.embed_tokens.weight",
}:
continue
if args.skip_lm_head:
if k in {
"lm_head.weight",
}:
continue
if args.prefix:
k = args.prefix + k
if args.suffix:
k = k.split(".")[-1]
if isinstance(v, Tensor):
if v.dtype == torch.bfloat16:
tensor = get_bf16_repr(v)
else:
tensor = v.cpu().numpy()
if args.dtype is not None:
nps[k] = np.ascontiguousarray(
tensor.astype(np.dtype(args.dtype))
)
else:
nps[k] = np.ascontiguousarray(tensor)
logging.info(
f"converted {k} - {str(sys.getsizeof(nps[k]))} bytes"
)
else:
logging.info(f"skipped non-tensor object: {k}")
append_to_zipf(nps, output_zipfile)

source = str(target_folder / "model.npz")
target = str(target_folder / "rust_model.ot")
Expand All @@ -119,7 +187,7 @@ def get_bf16_repr(input_tensor: torch.Tensor) -> np.ndarray:
"--",
source,
target,
]
]
if args.download_libtorch:
cargo_args += ["--features", "download-libtorch"]
subprocess.run(cargo_args)
45 changes: 28 additions & 17 deletions utils/download-dependencies_distilbert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,35 +1,46 @@
from pathlib import Path
import os
import numpy as np
import torch
import subprocess
from pathlib import Path

import numpy as np
import requests
import torch

if __name__ == "__main__":

target_path = Path.home() / 'rustbert' / 'distilbert'
target_path = Path.home() / "rustbert" / "distilbert"
os.makedirs(str(target_path), exist_ok=True)

weights_url = "https://huggingface.co/sshleifer/tiny-distilbert-base-cased/resolve/main/pytorch_model.bin"
r = requests.get(weights_url, allow_redirects=True)
(target_path / 'pytorch_model.bin').open('wb').write(r.content)
(target_path / "pytorch_model.bin").open("wb").write(r.content)

weights = torch.load(target_path / 'pytorch_model.bin', map_location='cpu')
weights = torch.load(target_path / "pytorch_model.bin", map_location="cpu")
nps = {}
for k, v in weights.items():
nps[k] = np.ascontiguousarray(v.cpu().numpy())

np.savez(target_path / 'model.npz', **nps)
np.savez(target_path / "model.npz", **nps)

source = str(target_path / 'model.npz')
target = str(target_path / 'model.ot')
source = str(target_path / "model.npz")
target = str(target_path / "model.ot")

toml_location = (Path(__file__).resolve() / '..' / '..' / 'Cargo.toml').resolve()
toml_location = (Path(__file__).resolve() / ".." / ".." / "Cargo.toml").resolve()

subprocess.call(
['cargo', 'run', '--bin=convert-tensor', '--features', 'download-libtorch', '--manifest-path=%s' % toml_location, '--', source, target])

os.remove(str(target_path / 'pytorch_model.bin'))
os.remove(str(target_path / 'model.npz'))

assert (target_path / 'model.ot').exists(), "Conversion of the model failed."
[
"cargo",
"run",
"--bin=convert-tensor",
"--features",
"download-libtorch",
"--manifest-path=%s" % toml_location,
"--",
source,
target,
]
)

os.remove(str(target_path / "pytorch_model.bin"))
os.remove(str(target_path / "model.npz"))

assert (target_path / "model.ot").exists(), "Conversion of the model failed."
Loading