Alex Sadleir
commited on
Commit
·
a1edf95
1
Parent(s):
d5fb6c3
add int4/int8
Browse files- .gitignore +1 -0
- README.md +1 -1
- __pycache__/onnx_gemma3_pipeline.cpython-312.pyc +0 -0
- download_missing_hf_files.py +33 -0
- embeddinggemma-300m/model.onnx +2 -2
- embeddinggemma-300m/onnx/model_int4.onnx +3 -0
- embeddinggemma-300m/onnx/model_int8.onnx +3 -0
- float16.py +878 -0
- quantize_extended.py +189 -0
- utils.py +67 -0
.gitignore
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
__pycache__/
|
README.md
CHANGED
|
@@ -33,7 +33,7 @@ This demonstrates how ONNX conversion can offload more computation for faster, h
|
|
| 33 |
```
|
| 34 |
2. Export the ONNX model:
|
| 35 |
```sh
|
| 36 |
-
optimum-cli export onnx --model google/embeddinggemma-300m-qat-q4_0-unquantized embeddinggemma-300m-onnx
|
| 37 |
python download_missing_hf_files.py
|
| 38 |
```
|
| 39 |
|
|
|
|
| 33 |
```
|
| 34 |
2. Export the ONNX model:
|
| 35 |
```sh
|
| 36 |
+
optimum-cli export onnx --model google/embeddinggemma-300m-qat-q4_0-unquantized --optimize O3 --slim embeddinggemma-300m-onnx
|
| 37 |
python download_missing_hf_files.py
|
| 38 |
```
|
| 39 |
|
__pycache__/onnx_gemma3_pipeline.cpython-312.pyc
DELETED
|
Binary file (6.63 kB)
|
|
|
download_missing_hf_files.py
CHANGED
|
@@ -67,3 +67,36 @@ torch.onnx.export(
|
|
| 67 |
opset_version=14
|
| 68 |
)
|
| 69 |
print("Exported dense2.onnx")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
opset_version=14
|
| 68 |
)
|
| 69 |
print("Exported dense2.onnx")
|
| 70 |
+
|
| 71 |
+
# # Quantize dense1.onnx and dense2.onnx to int4 using ONNX Runtime matmul_4bits_quantizer
|
| 72 |
+
# from onnxruntime.quantization import (
|
| 73 |
+
# matmul_nbits_quantizer,
|
| 74 |
+
# quant_utils
|
| 75 |
+
# )
|
| 76 |
+
# from pathlib import Path
|
| 77 |
+
|
| 78 |
+
# onnx_dir = Path(onnx_dir)
|
| 79 |
+
# for dense_name in ["dense1.onnx", "dense2.onnx"]:
|
| 80 |
+
# model_fp32_path = onnx_dir / dense_name
|
| 81 |
+
# model_int4_path = model_fp32_path # Overwrite original file
|
| 82 |
+
# quant_config = matmul_nbits_quantizer.DefaultWeightOnlyQuantConfig(
|
| 83 |
+
# block_size=128,
|
| 84 |
+
# is_symmetric=True,
|
| 85 |
+
# accuracy_level=4,
|
| 86 |
+
# quant_format=quant_utils.QuantFormat.QOperator,
|
| 87 |
+
# op_types_to_quantize=("MatMul", "Gather"),
|
| 88 |
+
# quant_axes=( ("MatMul", 0), ("Gather", 1) )
|
| 89 |
+
# )
|
| 90 |
+
# model = quant_utils.load_model_with_shape_infer(model_fp32_path)
|
| 91 |
+
# quant = matmul_nbits_quantizer.MatMulNBitsQuantizer(
|
| 92 |
+
# model,
|
| 93 |
+
# nodes_to_exclude=None,
|
| 94 |
+
# nodes_to_include=None,
|
| 95 |
+
# algo_config=quant_config,
|
| 96 |
+
# )
|
| 97 |
+
# quant.process()
|
| 98 |
+
# quant.model.save_model_to_file(
|
| 99 |
+
# str(model_int4_path),
|
| 100 |
+
# True
|
| 101 |
+
# )
|
| 102 |
+
# print(f"Quantized {dense_name} to int4 and overwrote original file.")
|
embeddinggemma-300m/model.onnx
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:511afd7b7ed2b58a61876a6aef4c1113a93b2afad17cf363753eef48f4669a41
|
| 3 |
+
size 1212258625
|
embeddinggemma-300m/onnx/model_int4.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f79136a20219163cdf5701a85422ccde126db44d2d51f6bcfc07b63edc0efab9
|
| 3 |
+
size 869894596
|
embeddinggemma-300m/onnx/model_int8.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7c74d91f9c44be1b3a7cbe5875d61597b493ae2789e7a62f540b351eaf0cdc57
|
| 3 |
+
size 265916124
|
float16.py
ADDED
|
@@ -0,0 +1,878 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# MIT License
|
| 2 |
+
#
|
| 3 |
+
# Copyright (c) Microsoft Corporation, Hugging Face. All rights reserved.
|
| 4 |
+
#
|
| 5 |
+
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 6 |
+
# of this software and associated documentation files (the "Software"), to deal
|
| 7 |
+
# in the Software without restriction, including without limitation the rights
|
| 8 |
+
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 9 |
+
# copies of the Software, and to permit persons to whom the Software is
|
| 10 |
+
# furnished to do so, subject to the following conditions:
|
| 11 |
+
#
|
| 12 |
+
# The above copyright notice and this permission notice shall be included in all
|
| 13 |
+
# copies or substantial portions of the Software.
|
| 14 |
+
#
|
| 15 |
+
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 16 |
+
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 17 |
+
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 18 |
+
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 19 |
+
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 20 |
+
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 21 |
+
# SOFTWARE.
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
from typing import Optional
|
| 25 |
+
import itertools
|
| 26 |
+
import numpy as np
|
| 27 |
+
import onnx
|
| 28 |
+
import packaging.version as pv
|
| 29 |
+
import warnings
|
| 30 |
+
from onnx import helper, numpy_helper
|
| 31 |
+
from onnx import onnx_pb as onnx_proto
|
| 32 |
+
import onnxslim.third_party.onnx_graphsurgeon as gs
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
FLOAT32 = 1
|
| 36 |
+
FLOAT16 = 10
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
def _npfloat16_to_int(np_list):
|
| 40 |
+
"""
|
| 41 |
+
Convert numpy float16 to python int.
|
| 42 |
+
|
| 43 |
+
:param np_list: numpy float16 list
|
| 44 |
+
:return int_list: python int list
|
| 45 |
+
"""
|
| 46 |
+
return [int(bin(_.view("H"))[2:].zfill(16), 2) for _ in np_list]
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
def convert_np_to_float16(np_array, min_positive_val=1e-7, max_finite_val=1e4):
|
| 50 |
+
"""
|
| 51 |
+
Convert float32 numpy array to float16 without changing sign or finiteness.
|
| 52 |
+
Positive values less than min_positive_val are mapped to min_positive_val.
|
| 53 |
+
Positive finite values greater than max_finite_val are mapped to max_finite_val.
|
| 54 |
+
Similar for negative values. NaN, 0, inf, and -inf are unchanged.
|
| 55 |
+
"""
|
| 56 |
+
|
| 57 |
+
def between(a, b, c):
|
| 58 |
+
return np.logical_and(a < b, b < c)
|
| 59 |
+
|
| 60 |
+
positive_values = np_array[np.where(np_array > 0)]
|
| 61 |
+
if positive_values.shape[0] > 0:
|
| 62 |
+
pos_max = positive_values.max()
|
| 63 |
+
pos_min = positive_values.min()
|
| 64 |
+
|
| 65 |
+
if pos_max >= max_finite_val:
|
| 66 |
+
warnings.warn(
|
| 67 |
+
"the float32 number {} will be truncated to {}".format(
|
| 68 |
+
pos_max, max_finite_val
|
| 69 |
+
)
|
| 70 |
+
)
|
| 71 |
+
|
| 72 |
+
if pos_min <= min_positive_val:
|
| 73 |
+
warnings.warn(
|
| 74 |
+
"the float32 number {} will be truncated to {}".format(
|
| 75 |
+
pos_min, min_positive_val
|
| 76 |
+
)
|
| 77 |
+
)
|
| 78 |
+
|
| 79 |
+
negative_values = np_array[np.where(np_array < 0)]
|
| 80 |
+
if negative_values.shape[0] > 0:
|
| 81 |
+
neg_max = negative_values.max()
|
| 82 |
+
neg_min = negative_values.min()
|
| 83 |
+
|
| 84 |
+
if neg_min <= -max_finite_val:
|
| 85 |
+
warnings.warn(
|
| 86 |
+
"the float32 number {} will be truncated to {}".format(
|
| 87 |
+
neg_min, -max_finite_val
|
| 88 |
+
)
|
| 89 |
+
)
|
| 90 |
+
|
| 91 |
+
if neg_max >= -min_positive_val:
|
| 92 |
+
warnings.warn(
|
| 93 |
+
"the float32 number {} will be truncated to {}".format(
|
| 94 |
+
neg_max, -min_positive_val
|
| 95 |
+
)
|
| 96 |
+
)
|
| 97 |
+
|
| 98 |
+
np_array = np.where(
|
| 99 |
+
between(0, np_array, min_positive_val), min_positive_val, np_array
|
| 100 |
+
)
|
| 101 |
+
np_array = np.where(
|
| 102 |
+
between(-min_positive_val, np_array, 0), -min_positive_val, np_array
|
| 103 |
+
)
|
| 104 |
+
np_array = np.where(
|
| 105 |
+
between(max_finite_val, np_array, float("inf")), max_finite_val, np_array
|
| 106 |
+
)
|
| 107 |
+
np_array = np.where(
|
| 108 |
+
between(float("-inf"), np_array, -max_finite_val), -max_finite_val, np_array
|
| 109 |
+
)
|
| 110 |
+
return np.float16(np_array)
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
def convert_tensor_float_to_float16(tensor, min_positive_val=1e-7, max_finite_val=1e4):
|
| 114 |
+
"""
|
| 115 |
+
Convert tensor float to float16.
|
| 116 |
+
|
| 117 |
+
:param tensor: TensorProto object
|
| 118 |
+
:return tensor_float16: converted TensorProto object
|
| 119 |
+
"""
|
| 120 |
+
if not isinstance(tensor, onnx_proto.TensorProto):
|
| 121 |
+
raise ValueError(
|
| 122 |
+
"Expected input type is an ONNX TensorProto but got %s" % type(tensor)
|
| 123 |
+
)
|
| 124 |
+
|
| 125 |
+
if tensor.data_type == onnx_proto.TensorProto.FLOAT:
|
| 126 |
+
tensor.data_type = onnx_proto.TensorProto.FLOAT16
|
| 127 |
+
# convert float_data (float type) to float16 and write to int32_data
|
| 128 |
+
if tensor.float_data:
|
| 129 |
+
float16_data = convert_np_to_float16(
|
| 130 |
+
np.array(tensor.float_data), min_positive_val, max_finite_val
|
| 131 |
+
)
|
| 132 |
+
int_list = _npfloat16_to_int(float16_data)
|
| 133 |
+
tensor.int32_data[:] = int_list
|
| 134 |
+
tensor.float_data[:] = []
|
| 135 |
+
# convert raw_data (bytes type)
|
| 136 |
+
if tensor.raw_data:
|
| 137 |
+
# convert n.raw_data to float
|
| 138 |
+
float32_list = np.fromstring(tensor.raw_data, dtype="float32")
|
| 139 |
+
# convert float to float16
|
| 140 |
+
float16_list = convert_np_to_float16(
|
| 141 |
+
float32_list, min_positive_val, max_finite_val
|
| 142 |
+
)
|
| 143 |
+
# convert float16 to bytes and write back to raw_data
|
| 144 |
+
tensor.raw_data = float16_list.tostring()
|
| 145 |
+
return tensor
|
| 146 |
+
|
| 147 |
+
|
| 148 |
+
def make_value_info_from_tensor(tensor):
|
| 149 |
+
shape = numpy_helper.to_array(tensor).shape
|
| 150 |
+
return helper.make_tensor_value_info(tensor.name, tensor.data_type, shape)
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
DEFAULT_OP_BLOCK_LIST = [
|
| 154 |
+
"ArrayFeatureExtractor",
|
| 155 |
+
"Binarizer",
|
| 156 |
+
"CastMap",
|
| 157 |
+
"CategoryMapper",
|
| 158 |
+
"DictVectorizer",
|
| 159 |
+
"FeatureVectorizer",
|
| 160 |
+
"Imputer",
|
| 161 |
+
"LabelEncoder",
|
| 162 |
+
"LinearClassifier",
|
| 163 |
+
"LinearRegressor",
|
| 164 |
+
"Normalizer",
|
| 165 |
+
"OneHotEncoder",
|
| 166 |
+
"RandomUniformLike",
|
| 167 |
+
"SVMClassifier",
|
| 168 |
+
"SVMRegressor",
|
| 169 |
+
"Scaler",
|
| 170 |
+
"TreeEnsembleClassifier",
|
| 171 |
+
"TreeEnsembleRegressor",
|
| 172 |
+
"ZipMap",
|
| 173 |
+
"NonMaxSuppression",
|
| 174 |
+
"TopK",
|
| 175 |
+
"RoiAlign",
|
| 176 |
+
"Resize",
|
| 177 |
+
# 'Range',
|
| 178 |
+
"CumSum",
|
| 179 |
+
"Min",
|
| 180 |
+
"Max",
|
| 181 |
+
"Upsample",
|
| 182 |
+
# NEW:
|
| 183 |
+
"RandomNormalLike",
|
| 184 |
+
# TODO: Ideally, "Cast" nodes should not be here, for the following reasons:
|
| 185 |
+
# - It breaks the semantics that the default list contains "ops that are not supported for float16 in ONNX Runtime".
|
| 186 |
+
# - When fp32 casts already exist in the model (e.g., for rotary embeddings), this script will insert redundant casts around it.
|
| 187 |
+
# However, without it, the graphs produced are invalid. Eventually, we will resolve this.
|
| 188 |
+
"Cast",
|
| 189 |
+
]
|
| 190 |
+
|
| 191 |
+
|
| 192 |
+
def initial_checking(model, disable_shape_infer):
|
| 193 |
+
func_infer_shape = None
|
| 194 |
+
if not disable_shape_infer and pv.Version(onnx.__version__) >= pv.Version("1.2"):
|
| 195 |
+
try:
|
| 196 |
+
from onnx.shape_inference import infer_shapes
|
| 197 |
+
|
| 198 |
+
func_infer_shape = infer_shapes
|
| 199 |
+
finally:
|
| 200 |
+
pass
|
| 201 |
+
|
| 202 |
+
if not isinstance(model, onnx_proto.ModelProto):
|
| 203 |
+
raise ValueError(
|
| 204 |
+
"Expected model type is an ONNX ModelProto but got %s" % type(model)
|
| 205 |
+
)
|
| 206 |
+
|
| 207 |
+
if func_infer_shape is not None:
|
| 208 |
+
model = func_infer_shape(model)
|
| 209 |
+
|
| 210 |
+
is_fp16_ready_flag = check_if_fp16_ready(model.graph)
|
| 211 |
+
|
| 212 |
+
return model, func_infer_shape, is_fp16_ready_flag
|
| 213 |
+
|
| 214 |
+
|
| 215 |
+
def convert_float_to_float16(
|
| 216 |
+
model,
|
| 217 |
+
min_positive_val=1e-7,
|
| 218 |
+
max_finite_val=1e4,
|
| 219 |
+
keep_io_types=False,
|
| 220 |
+
disable_shape_infer=False,
|
| 221 |
+
op_block_list=None,
|
| 222 |
+
node_block_list=None,
|
| 223 |
+
check_fp16_ready=True,
|
| 224 |
+
):
|
| 225 |
+
|
| 226 |
+
# create blocklists
|
| 227 |
+
if op_block_list is None:
|
| 228 |
+
op_block_list = DEFAULT_OP_BLOCK_LIST
|
| 229 |
+
if node_block_list is None:
|
| 230 |
+
node_block_list = []
|
| 231 |
+
op_block_list = set(op_block_list)
|
| 232 |
+
node_block_list = set(node_block_list)
|
| 233 |
+
|
| 234 |
+
global_input_name_dict = (
|
| 235 |
+
{}
|
| 236 |
+
) # key: input name, value: new output name after Cast node
|
| 237 |
+
# basic checking, including shape inference
|
| 238 |
+
model, func_infer_shape, is_fp16_ready_flag = initial_checking(
|
| 239 |
+
model, disable_shape_infer
|
| 240 |
+
)
|
| 241 |
+
if is_fp16_ready_flag and check_fp16_ready:
|
| 242 |
+
raise ValueError(
|
| 243 |
+
"The model is already converted to float16, if convert again, the model might be wrong. \n If you are sure to convert again, please set check_fp16_ready=False."
|
| 244 |
+
)
|
| 245 |
+
|
| 246 |
+
graph_stack = [model.graph]
|
| 247 |
+
|
| 248 |
+
is_top_level = True
|
| 249 |
+
while graph_stack:
|
| 250 |
+
next_level = []
|
| 251 |
+
for curr_graph in graph_stack:
|
| 252 |
+
process_graph_input(
|
| 253 |
+
curr_graph, is_top_level, keep_io_types, global_input_name_dict
|
| 254 |
+
)
|
| 255 |
+
value_info_block_list = process_tensor_in_node(
|
| 256 |
+
curr_graph,
|
| 257 |
+
op_block_list,
|
| 258 |
+
node_block_list,
|
| 259 |
+
min_positive_val,
|
| 260 |
+
max_finite_val,
|
| 261 |
+
)
|
| 262 |
+
process_value_info(curr_graph, value_info_block_list)
|
| 263 |
+
process_node_in_block_list(
|
| 264 |
+
curr_graph, global_input_name_dict, op_block_list, node_block_list
|
| 265 |
+
)
|
| 266 |
+
process_initializers(
|
| 267 |
+
curr_graph,
|
| 268 |
+
op_block_list,
|
| 269 |
+
node_block_list,
|
| 270 |
+
min_positive_val,
|
| 271 |
+
max_finite_val,
|
| 272 |
+
)
|
| 273 |
+
process_graph_output(curr_graph, is_top_level, keep_io_types)
|
| 274 |
+
sub_graph_list = get_next_level_graph(
|
| 275 |
+
curr_graph, op_block_list, node_block_list
|
| 276 |
+
)
|
| 277 |
+
if len(sub_graph_list) > 0:
|
| 278 |
+
next_level.extend(sub_graph_list)
|
| 279 |
+
|
| 280 |
+
if not is_top_level:
|
| 281 |
+
process_node_input_output(curr_graph, global_input_name_dict)
|
| 282 |
+
is_top_level = False # Going to process sub-graph
|
| 283 |
+
graph_stack = next_level
|
| 284 |
+
|
| 285 |
+
remove_unnecessary_cast_node(model.graph)
|
| 286 |
+
|
| 287 |
+
# Topologically sort the graph
|
| 288 |
+
# NOTE: We do not perform another round of optimization as the model is already optimized
|
| 289 |
+
graph = gs.import_onnx(model)
|
| 290 |
+
graph.toposort()
|
| 291 |
+
model = gs.export_onnx(graph)
|
| 292 |
+
|
| 293 |
+
return model
|
| 294 |
+
|
| 295 |
+
|
| 296 |
+
# Change the input/output of the node to the new output name after Cast node for sub-graph
|
| 297 |
+
# Because there have NO value_info start from
|
| 298 |
+
def process_node_input_output(
|
| 299 |
+
graph: onnx_proto.GraphProto, global_input_name_dict: dict
|
| 300 |
+
):
|
| 301 |
+
for node in graph.node:
|
| 302 |
+
for i, input_name in enumerate(node.input):
|
| 303 |
+
if input_name in global_input_name_dict:
|
| 304 |
+
node.input[i] = global_input_name_dict[input_name]
|
| 305 |
+
for i, output_name in enumerate(node.output):
|
| 306 |
+
if output_name in global_input_name_dict:
|
| 307 |
+
node.output[i] = global_input_name_dict[output_name]
|
| 308 |
+
|
| 309 |
+
|
| 310 |
+
def process_graph_input(
|
| 311 |
+
graph: onnx_proto.GraphProto,
|
| 312 |
+
is_top_level: bool,
|
| 313 |
+
is_io_fp32: bool,
|
| 314 |
+
global_input_name_dict: dict,
|
| 315 |
+
):
|
| 316 |
+
# The input dtype is float32, need to cast to fp16
|
| 317 |
+
if is_top_level and is_io_fp32:
|
| 318 |
+
for graph_input in graph.input: # n_input is ValueInfoProto
|
| 319 |
+
if graph_input.type.tensor_type.elem_type == onnx_proto.TensorProto.FLOAT:
|
| 320 |
+
downstream_nodes = find_downstream_node_by_input_name(
|
| 321 |
+
graph, graph_input.name
|
| 322 |
+
)
|
| 323 |
+
for d_node in downstream_nodes:
|
| 324 |
+
# More than one node may consume the model input, so we only create
|
| 325 |
+
# a single cast node, and then reuse this node when needed.
|
| 326 |
+
cast_exists = graph_input.name in global_input_name_dict
|
| 327 |
+
if cast_exists:
|
| 328 |
+
cast_node_output_name = global_input_name_dict[graph_input.name]
|
| 329 |
+
else:
|
| 330 |
+
cast_node_output_name = graph_input.name + "_fp16"
|
| 331 |
+
add_cast_node(
|
| 332 |
+
graph,
|
| 333 |
+
[graph_input.name],
|
| 334 |
+
[cast_node_output_name],
|
| 335 |
+
cast_node_output_name, # Set node name same as output name
|
| 336 |
+
FLOAT16,
|
| 337 |
+
)
|
| 338 |
+
add_new_value_info(
|
| 339 |
+
graph,
|
| 340 |
+
graph_input,
|
| 341 |
+
cast_node_output_name,
|
| 342 |
+
onnx_proto.TensorProto.FLOAT16,
|
| 343 |
+
)
|
| 344 |
+
for i, input_name in enumerate(d_node.input):
|
| 345 |
+
if input_name == graph_input.name:
|
| 346 |
+
d_node.input[i] = (
|
| 347 |
+
cast_node_output_name # Change the input of the second node
|
| 348 |
+
)
|
| 349 |
+
global_input_name_dict[graph_input.name] = (
|
| 350 |
+
cast_node_output_name
|
| 351 |
+
)
|
| 352 |
+
|
| 353 |
+
# For the sub-graph, don't do cast
|
| 354 |
+
else: # Change the input dtype to fp16 without any cast
|
| 355 |
+
for graph_input in graph.input:
|
| 356 |
+
if graph_input.type.tensor_type.elem_type == onnx_proto.TensorProto.FLOAT:
|
| 357 |
+
graph_input.type.tensor_type.elem_type = onnx_proto.TensorProto.FLOAT16
|
| 358 |
+
|
| 359 |
+
|
| 360 |
+
def process_graph_output(
|
| 361 |
+
graph: onnx_proto.GraphProto, is_top_level: bool, is_io_fp32: bool
|
| 362 |
+
):
|
| 363 |
+
if is_top_level and is_io_fp32: # the output dtype is float32, need to cast to fp16
|
| 364 |
+
for i, graph_output in enumerate(graph.output):
|
| 365 |
+
if graph_output.type.tensor_type.elem_type == onnx_proto.TensorProto.FLOAT:
|
| 366 |
+
new_producer_name = graph_output.name + "_fp16"
|
| 367 |
+
original_name = graph_output.name # The correct output name
|
| 368 |
+
|
| 369 |
+
# Get the node(s) that produce the model output
|
| 370 |
+
# These will most likely be fp16, but could be fp32 if the previous node is in block_list
|
| 371 |
+
upstream_nodes = find_upstream_node_by_output_name(graph, original_name)
|
| 372 |
+
assert len(upstream_nodes) == 1 # Should be only one node
|
| 373 |
+
|
| 374 |
+
producer_node = upstream_nodes[0]
|
| 375 |
+
|
| 376 |
+
for i, output_name in enumerate(producer_node.output):
|
| 377 |
+
if output_name == original_name:
|
| 378 |
+
producer_node.output[i] = new_producer_name
|
| 379 |
+
|
| 380 |
+
cast_node_name = new_producer_name + "_input_cast" + str(i)
|
| 381 |
+
add_cast_node(
|
| 382 |
+
graph,
|
| 383 |
+
[new_producer_name],
|
| 384 |
+
[original_name],
|
| 385 |
+
cast_node_name,
|
| 386 |
+
onnx_proto.TensorProto.FLOAT,
|
| 387 |
+
)
|
| 388 |
+
for value_info in graph.value_info:
|
| 389 |
+
if original_name == value_info.name:
|
| 390 |
+
value_info.type.tensor_type.elem_type = (
|
| 391 |
+
onnx_proto.TensorProto.FLOAT
|
| 392 |
+
)
|
| 393 |
+
|
| 394 |
+
# Get the node(s) that consume the model output
|
| 395 |
+
downstream_nodes = find_downstream_node_by_input_name(
|
| 396 |
+
graph,
|
| 397 |
+
original_name,
|
| 398 |
+
include_subgraphs=False,
|
| 399 |
+
)
|
| 400 |
+
|
| 401 |
+
# It is possible that the producer node is also input to downstream nodes
|
| 402 |
+
# So, we update the inputs of these downstream nodes
|
| 403 |
+
for d_node in downstream_nodes:
|
| 404 |
+
for i, input_name in enumerate(d_node.input):
|
| 405 |
+
if input_name == original_name:
|
| 406 |
+
d_node.input[i] = new_producer_name
|
| 407 |
+
|
| 408 |
+
else: # change the output dtype to fp16 in tensor
|
| 409 |
+
for graph_output in graph.output:
|
| 410 |
+
if graph_output.type.tensor_type.elem_type == onnx_proto.TensorProto.FLOAT:
|
| 411 |
+
graph_output.type.tensor_type.elem_type = onnx_proto.TensorProto.FLOAT16
|
| 412 |
+
|
| 413 |
+
|
| 414 |
+
def process_node_in_block_list(
|
| 415 |
+
graph: onnx_proto.GraphProto,
|
| 416 |
+
global_input_name_dict: dict,
|
| 417 |
+
op_block_list: list,
|
| 418 |
+
node_block_list: list,
|
| 419 |
+
):
|
| 420 |
+
# NB: Important to create a copy of the nodes in the graph to avoid modifying
|
| 421 |
+
# the graph in-place while iterating (causing an infinite loop)
|
| 422 |
+
for node in list(graph.node):
|
| 423 |
+
if (node.op_type in op_block_list) or (node.name in node_block_list):
|
| 424 |
+
insert_cast32_before_node(graph, node, global_input_name_dict)
|
| 425 |
+
insert_cast16_after_node(graph, node, global_input_name_dict)
|
| 426 |
+
|
| 427 |
+
|
| 428 |
+
# Todo: global_input_name_dict still not fill value
|
| 429 |
+
def insert_cast32_before_node(
|
| 430 |
+
graph: onnx_proto.GraphProto, node: onnx_proto.NodeProto, global_input_name_dict
|
| 431 |
+
):
|
| 432 |
+
for i, input_name in enumerate(node.input):
|
| 433 |
+
for value_info in itertools.chain(graph.value_info, graph.input):
|
| 434 |
+
if input_name == value_info.name:
|
| 435 |
+
if (
|
| 436 |
+
value_info.type.tensor_type.elem_type
|
| 437 |
+
!= onnx_proto.TensorProto.FLOAT16
|
| 438 |
+
):
|
| 439 |
+
break
|
| 440 |
+
cast_output_name = node.name + "_input_cast_" + str(i)
|
| 441 |
+
add_new_value_info(
|
| 442 |
+
graph, value_info, cast_output_name, onnx_proto.TensorProto.FLOAT
|
| 443 |
+
)
|
| 444 |
+
cast_node_name = node.name + "_input_cast" + str(i)
|
| 445 |
+
add_cast_node(
|
| 446 |
+
graph,
|
| 447 |
+
[input_name],
|
| 448 |
+
[cast_output_name],
|
| 449 |
+
cast_node_name,
|
| 450 |
+
onnx_proto.TensorProto.FLOAT,
|
| 451 |
+
)
|
| 452 |
+
node.input[i] = cast_output_name
|
| 453 |
+
break
|
| 454 |
+
|
| 455 |
+
|
| 456 |
+
# Todo: global_input_name_dict still not fill value
|
| 457 |
+
def insert_cast16_after_node(
|
| 458 |
+
graph: onnx_proto.GraphProto, node: onnx_proto.NodeProto, global_input_name_dict
|
| 459 |
+
):
|
| 460 |
+
for i, output_name in enumerate(node.output):
|
| 461 |
+
for value_info in itertools.chain(graph.value_info, graph.output):
|
| 462 |
+
if output_name == value_info.name:
|
| 463 |
+
if (
|
| 464 |
+
value_info.type.tensor_type.elem_type
|
| 465 |
+
!= onnx_proto.TensorProto.FLOAT
|
| 466 |
+
):
|
| 467 |
+
break
|
| 468 |
+
cast_input_name = node.name + "_output_cast_" + str(i)
|
| 469 |
+
add_new_value_info(
|
| 470 |
+
graph, value_info, cast_input_name, onnx_proto.TensorProto.FLOAT
|
| 471 |
+
)
|
| 472 |
+
value_info.type.tensor_type.elem_type = onnx_proto.TensorProto.FLOAT16
|
| 473 |
+
cast_node_name = node.name + "_output_cast" + str(i)
|
| 474 |
+
add_cast_node(
|
| 475 |
+
graph,
|
| 476 |
+
[cast_input_name],
|
| 477 |
+
[output_name],
|
| 478 |
+
cast_node_name,
|
| 479 |
+
onnx_proto.TensorProto.FLOAT16,
|
| 480 |
+
)
|
| 481 |
+
node.output[i] = cast_input_name
|
| 482 |
+
break
|
| 483 |
+
|
| 484 |
+
|
| 485 |
+
# Process tensor data in attribute of the node
|
| 486 |
+
def process_tensor_in_node(
|
| 487 |
+
graph: onnx_proto.GraphProto,
|
| 488 |
+
op_block_list: list,
|
| 489 |
+
node_block_list: list,
|
| 490 |
+
min_positive_val,
|
| 491 |
+
max_finite_val,
|
| 492 |
+
):
|
| 493 |
+
value_info_block_list = set() # This is for later use, not in this step
|
| 494 |
+
for node in graph.node:
|
| 495 |
+
# NOTE: "Cast" operation cannot change its output type because it is strongly typed.
|
| 496 |
+
if (
|
| 497 |
+
(node.op_type in op_block_list)
|
| 498 |
+
or (node.name in node_block_list)
|
| 499 |
+
or (node.op_type == "Cast")
|
| 500 |
+
):
|
| 501 |
+
# if (node.op_type in op_block_list) or (node.name in node_block_list):
|
| 502 |
+
# Only need to block the output value_info changing
|
| 503 |
+
for output_name in node.output:
|
| 504 |
+
value_info_block_list.add(output_name)
|
| 505 |
+
else:
|
| 506 |
+
for attr in node.attribute:
|
| 507 |
+
# one tensor
|
| 508 |
+
if attr.t.data_type == onnx_proto.TensorProto.FLOAT:
|
| 509 |
+
attr.t.CopyFrom(
|
| 510 |
+
convert_tensor_float_to_float16(
|
| 511 |
+
attr.t, min_positive_val, max_finite_val
|
| 512 |
+
)
|
| 513 |
+
)
|
| 514 |
+
# list of tensor
|
| 515 |
+
for t in attr.tensors:
|
| 516 |
+
if t.data_type == onnx_proto.TensorProto.FLOAT:
|
| 517 |
+
t.CopyFrom(
|
| 518 |
+
convert_tensor_float_to_float16(
|
| 519 |
+
t, min_positive_val, max_finite_val
|
| 520 |
+
)
|
| 521 |
+
)
|
| 522 |
+
return value_info_block_list
|
| 523 |
+
|
| 524 |
+
|
| 525 |
+
# Change all the value info type from float32 to float16 if not in block list
|
| 526 |
+
def process_value_info(graph: onnx_proto.GraphProto, value_info_block_list: list):
|
| 527 |
+
for value_info in graph.value_info:
|
| 528 |
+
if value_info.name in value_info_block_list:
|
| 529 |
+
continue
|
| 530 |
+
else:
|
| 531 |
+
if value_info.type.tensor_type.elem_type == onnx_proto.TensorProto.FLOAT:
|
| 532 |
+
value_info.type.tensor_type.elem_type = onnx_proto.TensorProto.FLOAT16
|
| 533 |
+
|
| 534 |
+
|
| 535 |
+
# Initializer is 'edge' type, so doesn't have value_info
|
| 536 |
+
def process_initializers(
|
| 537 |
+
graph: onnx_proto.GraphProto,
|
| 538 |
+
op_block_list,
|
| 539 |
+
node_block_list,
|
| 540 |
+
min_positive_val,
|
| 541 |
+
max_finite_val,
|
| 542 |
+
):
|
| 543 |
+
# Find the input of the block node, don't need to change this kind of initializer
|
| 544 |
+
initializer_block_list = set()
|
| 545 |
+
for node in graph.node:
|
| 546 |
+
if (node.op_type in op_block_list) or (node.name in node_block_list):
|
| 547 |
+
for (
|
| 548 |
+
input_name
|
| 549 |
+
) in (
|
| 550 |
+
node.input
|
| 551 |
+
): # some is initializer, some is value_info, can't distinguish but doesn't matter
|
| 552 |
+
initializer_block_list.add(input_name)
|
| 553 |
+
# Process initializers
|
| 554 |
+
for initializer in graph.initializer:
|
| 555 |
+
if initializer.name not in initializer_block_list:
|
| 556 |
+
if initializer.data_type == onnx_proto.TensorProto.FLOAT:
|
| 557 |
+
convert_tensor_float_to_float16(
|
| 558 |
+
initializer, min_positive_val, max_finite_val
|
| 559 |
+
)
|
| 560 |
+
|
| 561 |
+
|
| 562 |
+
def get_next_level_graph(
|
| 563 |
+
graph: onnx_proto.GraphProto, op_block_list: list, node_block_list: list
|
| 564 |
+
):
|
| 565 |
+
sub_graph_list = []
|
| 566 |
+
for node in graph.node:
|
| 567 |
+
if node.op_type in op_block_list or node.name in node_block_list:
|
| 568 |
+
continue
|
| 569 |
+
for attr in node.attribute:
|
| 570 |
+
# Check if sub-graph exist
|
| 571 |
+
if len(attr.g.node) > 0: # single sub-graph
|
| 572 |
+
sub_graph_list.append(attr.g)
|
| 573 |
+
for g in attr.graphs:
|
| 574 |
+
if len(g.node) > 0: # multiple sub-graphs
|
| 575 |
+
sub_graph_list.append(g)
|
| 576 |
+
return sub_graph_list
|
| 577 |
+
|
| 578 |
+
|
| 579 |
+
def add_cast_node(
|
| 580 |
+
graph: onnx_proto.GraphProto,
|
| 581 |
+
inputs: list,
|
| 582 |
+
outputs: list,
|
| 583 |
+
node_name: str,
|
| 584 |
+
to_type: int,
|
| 585 |
+
):
|
| 586 |
+
new_node = [helper.make_node("Cast", inputs, outputs, to=to_type, name=node_name)]
|
| 587 |
+
graph.node.extend(new_node)
|
| 588 |
+
|
| 589 |
+
|
| 590 |
+
def add_new_value_info(
|
| 591 |
+
graph: onnx_proto.GraphProto,
|
| 592 |
+
exist_value_info: onnx_proto.ValueInfoProto,
|
| 593 |
+
name: str,
|
| 594 |
+
dtype: int,
|
| 595 |
+
):
|
| 596 |
+
new_value_info = graph.value_info.add()
|
| 597 |
+
new_value_info.CopyFrom(exist_value_info)
|
| 598 |
+
new_value_info.name = name
|
| 599 |
+
new_value_info.type.tensor_type.elem_type = dtype
|
| 600 |
+
|
| 601 |
+
|
| 602 |
+
# Find the node that has the specified output name
|
| 603 |
+
def find_upstream_node_by_output_name(graph: onnx_proto.GraphProto, output_name: str):
|
| 604 |
+
nodes = []
|
| 605 |
+
for node in graph.node:
|
| 606 |
+
if output_name in node.output:
|
| 607 |
+
nodes.append(node)
|
| 608 |
+
assert len(nodes) <= 1 # Suppose there is less than one node found
|
| 609 |
+
return nodes
|
| 610 |
+
|
| 611 |
+
|
| 612 |
+
# Find the node that has the specified input name, including in subgraphs
|
| 613 |
+
def find_downstream_node_by_input_name(
|
| 614 |
+
graph: onnx_proto.GraphProto, input_name: str, include_subgraphs=True
|
| 615 |
+
):
|
| 616 |
+
nodes = []
|
| 617 |
+
|
| 618 |
+
# Check nodes in current graph
|
| 619 |
+
for node in graph.node:
|
| 620 |
+
if input_name in node.input:
|
| 621 |
+
nodes.append(node)
|
| 622 |
+
|
| 623 |
+
if not include_subgraphs:
|
| 624 |
+
continue
|
| 625 |
+
|
| 626 |
+
# Recursively check subgraphs in node attributes
|
| 627 |
+
for attr in node.attribute:
|
| 628 |
+
if attr.type == onnx_proto.AttributeProto.GRAPH:
|
| 629 |
+
# Single subgraph
|
| 630 |
+
if len(attr.g.node) > 0:
|
| 631 |
+
nodes.extend(find_downstream_node_by_input_name(attr.g, input_name))
|
| 632 |
+
|
| 633 |
+
# Multiple subgraphs
|
| 634 |
+
if attr.type == onnx_proto.AttributeProto.GRAPHS:
|
| 635 |
+
for g in attr.graphs:
|
| 636 |
+
if len(g.node) > 0:
|
| 637 |
+
nodes.extend(find_downstream_node_by_input_name(g, input_name))
|
| 638 |
+
|
| 639 |
+
return nodes
|
| 640 |
+
|
| 641 |
+
|
| 642 |
+
# Remove identity node
|
| 643 |
+
def remove_identity_node_from_model(model: onnx_proto.ModelProto):
|
| 644 |
+
remove_identity_node_from_graph(model.graph)
|
| 645 |
+
try:
|
| 646 |
+
from onnx.shape_inference import infer_shapes
|
| 647 |
+
|
| 648 |
+
func_infer_shape = infer_shapes
|
| 649 |
+
model = func_infer_shape(model)
|
| 650 |
+
return model
|
| 651 |
+
finally:
|
| 652 |
+
pass
|
| 653 |
+
|
| 654 |
+
|
| 655 |
+
# Remove identity node
|
| 656 |
+
def remove_identity_node_from_graph(graph: onnx_proto.GraphProto):
|
| 657 |
+
for curr_node in graph.node:
|
| 658 |
+
if curr_node.op_type == "Identity":
|
| 659 |
+
for input_name in curr_node.input:
|
| 660 |
+
upstream_nodes = find_upstream_node_by_output_name(graph, input_name)
|
| 661 |
+
for u_node in upstream_nodes:
|
| 662 |
+
if u_node is not None:
|
| 663 |
+
u_node.output[0] = curr_node.output[0]
|
| 664 |
+
graph.node.remove(curr_node)
|
| 665 |
+
|
| 666 |
+
|
| 667 |
+
def convert_float_to_float16_model_path(
|
| 668 |
+
model_path, min_positive_val=1e-7, max_finite_val=1e4, keep_io_types=False
|
| 669 |
+
):
|
| 670 |
+
"""
|
| 671 |
+
Convert tensor float type in the ONNX Model to tensor float16.
|
| 672 |
+
*It is to fix an issue that infer_shapes func cannot be used to infer >2GB models.
|
| 673 |
+
*But this function can be applied to all model sizes.
|
| 674 |
+
:param model_path: ONNX Model path
|
| 675 |
+
:return: converted ONNX ModelProto object
|
| 676 |
+
Examples
|
| 677 |
+
::
|
| 678 |
+
#Convert to ONNX ModelProto object and save model binary file:
|
| 679 |
+
from onnxmltools.utils.float16_converter import convert_float_to_float16_model_path
|
| 680 |
+
new_onnx_model = convert_float_to_float16_model_path('model.onnx')
|
| 681 |
+
onnx.save(new_onnx_model, 'new_model.onnx')
|
| 682 |
+
"""
|
| 683 |
+
|
| 684 |
+
disable_shape_infer = False
|
| 685 |
+
if pv.Version(onnx.__version__) >= pv.Version("1.8"):
|
| 686 |
+
try:
|
| 687 |
+
# infer_shapes_path can be applied to all model sizes
|
| 688 |
+
from onnx.shape_inference import infer_shapes_path
|
| 689 |
+
import tempfile
|
| 690 |
+
import os
|
| 691 |
+
|
| 692 |
+
# shape_infer_model_path should be in the same folder of model_path
|
| 693 |
+
with tempfile.NamedTemporaryFile(
|
| 694 |
+
dir=os.path.dirname(model_path)
|
| 695 |
+
) as tmpfile:
|
| 696 |
+
shape_infer_model_path = tmpfile.name
|
| 697 |
+
infer_shapes_path(model_path, shape_infer_model_path)
|
| 698 |
+
model = onnx.load(shape_infer_model_path)
|
| 699 |
+
disable_shape_infer = True
|
| 700 |
+
finally:
|
| 701 |
+
pass
|
| 702 |
+
if not disable_shape_infer:
|
| 703 |
+
model = onnx.load(model_path)
|
| 704 |
+
return convert_float_to_float16(
|
| 705 |
+
model, min_positive_val, max_finite_val, keep_io_types, disable_shape_infer
|
| 706 |
+
)
|
| 707 |
+
|
| 708 |
+
|
| 709 |
+
def remove_unnecessary_cast_node(graph_proto: onnx_proto.GraphProto):
|
| 710 |
+
# 1. find all cast nodes in the graph
|
| 711 |
+
cast_node_list = []
|
| 712 |
+
input_name_to_cast_node_dict = {}
|
| 713 |
+
output_name_to_cast_node_dict = {}
|
| 714 |
+
# using name as key to point to a node. because node object cannot be key
|
| 715 |
+
name_to_node_dict = {}
|
| 716 |
+
for node in graph_proto.node:
|
| 717 |
+
if node.op_type == "Cast":
|
| 718 |
+
# if node.name not in ["graph_input_cast0", "graph_output_cast0"]:
|
| 719 |
+
cast_node_list.append(node)
|
| 720 |
+
|
| 721 |
+
name_to_node_dict[node.name] = node
|
| 722 |
+
for input_name in node.input:
|
| 723 |
+
input_name_to_cast_node_dict[input_name] = node
|
| 724 |
+
for output_name in node.output:
|
| 725 |
+
output_name_to_cast_node_dict[output_name] = node
|
| 726 |
+
|
| 727 |
+
# 2. find upstream and downstream node of the cast node
|
| 728 |
+
cast_node_upstream_dict = {} # mapping cast node(name) to its upstream node
|
| 729 |
+
cast_node_downstream_dict = {} # mapping cast node(name) to its downstream node
|
| 730 |
+
for current_node in graph_proto.node:
|
| 731 |
+
# find the downstream node(s)
|
| 732 |
+
for input_name in current_node.input:
|
| 733 |
+
if input_name in output_name_to_cast_node_dict:
|
| 734 |
+
# found the downstream node of the cast node, might be multiple
|
| 735 |
+
cast_node = output_name_to_cast_node_dict[input_name]
|
| 736 |
+
if cast_node.name not in cast_node_downstream_dict:
|
| 737 |
+
cast_node_downstream_dict[cast_node.name] = current_node
|
| 738 |
+
else: # already exists one downstream node, make it a list
|
| 739 |
+
existing_downstream_nodes = cast_node_downstream_dict[
|
| 740 |
+
cast_node.name
|
| 741 |
+
]
|
| 742 |
+
if isinstance(existing_downstream_nodes, list):
|
| 743 |
+
existing_downstream_nodes.append(current_node)
|
| 744 |
+
else: # make a list
|
| 745 |
+
existing_downstream_nodes = [
|
| 746 |
+
existing_downstream_nodes,
|
| 747 |
+
current_node,
|
| 748 |
+
]
|
| 749 |
+
cast_node_downstream_dict[cast_node.name] = (
|
| 750 |
+
existing_downstream_nodes
|
| 751 |
+
)
|
| 752 |
+
# find the upstream node
|
| 753 |
+
for output_name in current_node.output:
|
| 754 |
+
if output_name in input_name_to_cast_node_dict:
|
| 755 |
+
# found the upstream node of the cast node, should be unique
|
| 756 |
+
cast_node = input_name_to_cast_node_dict[output_name]
|
| 757 |
+
cast_node_upstream_dict[cast_node.name] = current_node
|
| 758 |
+
|
| 759 |
+
# 3. remove the cast node which upstream is 'Constant'
|
| 760 |
+
for cast_node_name, upstream_node in cast_node_upstream_dict.items():
|
| 761 |
+
cast_node = name_to_node_dict[cast_node_name]
|
| 762 |
+
if upstream_node.op_type == "Constant":
|
| 763 |
+
cast_node_list.remove(cast_node)
|
| 764 |
+
|
| 765 |
+
# 4. find (cast_to_fp16, cast_to_fp32) pairs where --fp32--> cast_to_fp16 --fp16--> cast_to_fp32.
|
| 766 |
+
remove_candidate = []
|
| 767 |
+
|
| 768 |
+
name_to_value_info = {
|
| 769 |
+
value_info.name: value_info
|
| 770 |
+
for value_info in itertools.chain(graph_proto.value_info, graph_proto.input)
|
| 771 |
+
}
|
| 772 |
+
|
| 773 |
+
def get_type(name: str) -> Optional[int]:
|
| 774 |
+
if name in name_to_value_info:
|
| 775 |
+
return name_to_value_info[name].type
|
| 776 |
+
else:
|
| 777 |
+
# `name` has no value info.
|
| 778 |
+
return None
|
| 779 |
+
|
| 780 |
+
for cast_node_name, downstream_node in cast_node_downstream_dict.items():
|
| 781 |
+
cast_node = name_to_node_dict[cast_node_name]
|
| 782 |
+
if len(cast_node.input) != 1:
|
| 783 |
+
raise RuntimeError(
|
| 784 |
+
f"Cast node {cast_node_name} should have only one input, but has {len(cast_node.input)}."
|
| 785 |
+
)
|
| 786 |
+
|
| 787 |
+
input_type = get_type(cast_node.input[0])
|
| 788 |
+
if input_type != onnx_proto.TensorProto.FLOAT:
|
| 789 |
+
continue
|
| 790 |
+
if isinstance(downstream_node, list):
|
| 791 |
+
for dn in downstream_node:
|
| 792 |
+
if (
|
| 793 |
+
dn.op_type == "Cast"
|
| 794 |
+
and dn.attribute[0].i == 32
|
| 795 |
+
and cast_node.attribute[0].i == 16
|
| 796 |
+
and dn in cast_node_list
|
| 797 |
+
and cast_node in cast_node_list
|
| 798 |
+
):
|
| 799 |
+
remove_candidate.append((cast_node, dn))
|
| 800 |
+
else:
|
| 801 |
+
if (
|
| 802 |
+
downstream_node.op_type == "Cast"
|
| 803 |
+
and cast_node.attribute[0].i == FLOAT16
|
| 804 |
+
and downstream_node.attribute[0].i == FLOAT32
|
| 805 |
+
and downstream_node in cast_node_list
|
| 806 |
+
and cast_node in cast_node_list
|
| 807 |
+
):
|
| 808 |
+
remove_candidate.append((cast_node, downstream_node))
|
| 809 |
+
|
| 810 |
+
# 5. change "upstream --fp32--> cast_to_fp16 --fp16--> cast_to_fp32 --fp32--> downstream" to
|
| 811 |
+
# "upstream --fp32--> downstream".
|
| 812 |
+
for cast_node_pair in remove_candidate:
|
| 813 |
+
first_cast_node = cast_node_pair[0]
|
| 814 |
+
second_cast_node = cast_node_pair[1]
|
| 815 |
+
upstream_node = cast_node_upstream_dict.get(first_cast_node.name)
|
| 816 |
+
downstream_node = cast_node_downstream_dict.get(second_cast_node.name)
|
| 817 |
+
if upstream_node is None and downstream_node is not None:
|
| 818 |
+
# The upstream_node should be graph input
|
| 819 |
+
out = first_cast_node.input[0]
|
| 820 |
+
for i, input_name in enumerate(downstream_node.input):
|
| 821 |
+
for output_name in second_cast_node.output:
|
| 822 |
+
if input_name == output_name:
|
| 823 |
+
# change the input as the upstream node's output
|
| 824 |
+
downstream_node.input[i] = out
|
| 825 |
+
elif upstream_node is not None and downstream_node is None:
|
| 826 |
+
raise ValueError(
|
| 827 |
+
"The downstream node of the second cast node should be graph output"
|
| 828 |
+
)
|
| 829 |
+
else:
|
| 830 |
+
# find the upstream node's output to first_cast_node
|
| 831 |
+
out = None
|
| 832 |
+
for output_name in upstream_node.output:
|
| 833 |
+
if output_name == first_cast_node.input[0]:
|
| 834 |
+
out = output_name
|
| 835 |
+
break
|
| 836 |
+
# find the downstream node's input as second_cast_node's output
|
| 837 |
+
for i, input_name in enumerate(downstream_node.input):
|
| 838 |
+
for output_name in second_cast_node.output:
|
| 839 |
+
if input_name == output_name:
|
| 840 |
+
# change the input as the upstream node's output
|
| 841 |
+
downstream_node.input[i] = out
|
| 842 |
+
|
| 843 |
+
# 6. remove the cast node pair
|
| 844 |
+
for cast_node_pair in remove_candidate:
|
| 845 |
+
graph_proto.node.remove(cast_node_pair[0])
|
| 846 |
+
graph_proto.node.remove(cast_node_pair[1])
|
| 847 |
+
|
| 848 |
+
|
| 849 |
+
# Check if the model is already converted to float16
|
| 850 |
+
def check_if_fp16_ready(graph_proto):
|
| 851 |
+
# Check graph input and ouput
|
| 852 |
+
is_value_info_fp16 = False
|
| 853 |
+
for value_info in itertools.chain(
|
| 854 |
+
graph_proto.output, graph_proto.input, graph_proto.value_info
|
| 855 |
+
):
|
| 856 |
+
if value_info.type.tensor_type.elem_type == onnx_proto.TensorProto.FLOAT16:
|
| 857 |
+
is_value_info_fp16 = True
|
| 858 |
+
break
|
| 859 |
+
|
| 860 |
+
# Check initializer
|
| 861 |
+
is_initializer_fp16 = False
|
| 862 |
+
for initializer in graph_proto.initializer:
|
| 863 |
+
if initializer.data_type == onnx_proto.TensorProto.FLOAT16:
|
| 864 |
+
is_initializer_fp16 = True
|
| 865 |
+
break
|
| 866 |
+
|
| 867 |
+
# Check cast node
|
| 868 |
+
has_cast_node_fp16 = False
|
| 869 |
+
for node in graph_proto.node:
|
| 870 |
+
if node.op_type == "Cast" and node.attribute[0].i == FLOAT16:
|
| 871 |
+
has_cast_node_fp16 = True
|
| 872 |
+
break
|
| 873 |
+
|
| 874 |
+
# Any of above flags is True, return True
|
| 875 |
+
if is_value_info_fp16 or is_initializer_fp16 or has_cast_node_fp16:
|
| 876 |
+
return True # already converted to float16
|
| 877 |
+
else:
|
| 878 |
+
return False # not converted to float16 yet
|
quantize_extended.py
ADDED
|
@@ -0,0 +1,189 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Script to quantize ONNX models to additional formats: int4, int8, etc.
|
| 3 |
+
Based on transformers.js/scripts/quantize.py, extended for more quantization options.
|
| 4 |
+
"""
|
| 5 |
+
|
| 6 |
+
from enum import Enum
|
| 7 |
+
from tqdm import tqdm
|
| 8 |
+
from typing import Set, List, Optional
|
| 9 |
+
import onnx
|
| 10 |
+
import os
|
| 11 |
+
from dataclasses import dataclass, field
|
| 12 |
+
from transformers import HfArgumentParser
|
| 13 |
+
from onnxruntime.quantization import QuantType, QuantizationMode
|
| 14 |
+
from onnxruntime.quantization.onnx_quantizer import ONNXQuantizer
|
| 15 |
+
from onnxruntime.quantization.registry import IntegerOpsRegistry
|
| 16 |
+
from onnxruntime.quantization.matmul_nbits_quantizer import MatMulNBitsQuantizer
|
| 17 |
+
from onnxruntime.quantization.matmul_bnb4_quantizer import MatMulBnb4Quantizer
|
| 18 |
+
import float16
|
| 19 |
+
import utils
|
| 20 |
+
|
| 21 |
+
class QuantMode(Enum):
|
| 22 |
+
FP16 = "fp16"
|
| 23 |
+
Q8 = "q8"
|
| 24 |
+
QI8 = "int8"
|
| 25 |
+
QU8 = "uint8"
|
| 26 |
+
Q4 = "q4"
|
| 27 |
+
Q4F16 = "q4f16"
|
| 28 |
+
BNB4 = "bnb4"
|
| 29 |
+
INT4 = "int4"
|
| 30 |
+
INT8 = "int8"
|
| 31 |
+
|
| 32 |
+
QUANTIZE_SUFFIX_MAPPING = {
|
| 33 |
+
QuantMode.Q8: "quantized",
|
| 34 |
+
QuantMode.INT4: "int4",
|
| 35 |
+
QuantMode.INT8: "int8",
|
| 36 |
+
}
|
| 37 |
+
|
| 38 |
+
QUANTIZE_OPTIONS = tuple(x.value for x in QuantMode)
|
| 39 |
+
QUINT8_OPS = (
|
| 40 |
+
"Conv",
|
| 41 |
+
"GroupQueryAttention",
|
| 42 |
+
"MultiHeadAttention",
|
| 43 |
+
)
|
| 44 |
+
|
| 45 |
+
@dataclass
|
| 46 |
+
class IOArguments:
|
| 47 |
+
input_folder: str = field(metadata={"help": "Path of the input folder containing the .onnx models to quantize"})
|
| 48 |
+
output_folder: str = field(metadata={"help": "Path of the output folder where the quantized .onnx models will be saved"})
|
| 49 |
+
|
| 50 |
+
@dataclass
|
| 51 |
+
class QuantizationArguments:
|
| 52 |
+
modes: QuantMode = field(default=QUANTIZE_OPTIONS, metadata={"help": "Quantization mode to use.", "choices": QUANTIZE_OPTIONS, "nargs": "+",})
|
| 53 |
+
per_channel: bool = field(default=None, metadata={"help": "Whether to quantize weights per channel"})
|
| 54 |
+
reduce_range: bool = field(default=None, metadata={"help": "Whether to quantize weights with 7-bits."})
|
| 55 |
+
block_size: int = field(default=None, metadata={"help": "Block size for blockwise quantization."})
|
| 56 |
+
is_symmetric: bool = field(default=True, metadata={"help": "Indicate whether to quantize the model symmetrically"})
|
| 57 |
+
accuracy_level: int = field(default=None, metadata={"help": "Accuracy level of the 4-bit quantized MatMul computation."})
|
| 58 |
+
quant_type: int = field(default=MatMulBnb4Quantizer.NF4, metadata={"help": "Quantization data type. 0: FP4, 1: NF4", "choices": [MatMulBnb4Quantizer.FP4, MatMulBnb4Quantizer.NF4],})
|
| 59 |
+
op_block_list: List[str] = field(default=None, metadata={"help": "List of operators to exclude from quantization.", "nargs": "+",})
|
| 60 |
+
|
| 61 |
+
def quantize_int4(
|
| 62 |
+
model: onnx.ModelProto,
|
| 63 |
+
save_path: str,
|
| 64 |
+
block_size: int = 32,
|
| 65 |
+
is_symmetric: bool = True,
|
| 66 |
+
accuracy_level: int = 4,
|
| 67 |
+
):
|
| 68 |
+
"""
|
| 69 |
+
Quantize the weights of the model from float32 to 4-bit int using MatMulNBitsQuantizer
|
| 70 |
+
"""
|
| 71 |
+
quantizer = MatMulNBitsQuantizer(
|
| 72 |
+
model=model,
|
| 73 |
+
block_size=block_size,
|
| 74 |
+
is_symmetric=is_symmetric,
|
| 75 |
+
accuracy_level=accuracy_level,
|
| 76 |
+
)
|
| 77 |
+
quantizer.process()
|
| 78 |
+
utils.check_and_save_model(quantizer.model.model, save_path)
|
| 79 |
+
return quantizer.model.model
|
| 80 |
+
|
| 81 |
+
def quantize_int8(
|
| 82 |
+
model: onnx.ModelProto,
|
| 83 |
+
save_path: str,
|
| 84 |
+
per_channel: bool = False,
|
| 85 |
+
reduce_range: bool = False,
|
| 86 |
+
weight_type: QuantType = QuantType.QInt8,
|
| 87 |
+
op_block_list: Optional[List[str]] = None,
|
| 88 |
+
):
|
| 89 |
+
"""
|
| 90 |
+
Quantize the weights of the model from float32 to int8
|
| 91 |
+
"""
|
| 92 |
+
op_types_to_quantize = set(IntegerOpsRegistry.keys())
|
| 93 |
+
if op_block_list is not None:
|
| 94 |
+
op_types_to_quantize.difference_update(op_block_list)
|
| 95 |
+
|
| 96 |
+
quantizer = ONNXQuantizer(
|
| 97 |
+
model,
|
| 98 |
+
per_channel,
|
| 99 |
+
reduce_range,
|
| 100 |
+
mode=QuantizationMode.IntegerOps,
|
| 101 |
+
static=False,
|
| 102 |
+
weight_qType=weight_type,
|
| 103 |
+
activation_qType=QuantType.QUInt8,
|
| 104 |
+
tensors_range=None,
|
| 105 |
+
nodes_to_quantize=[],
|
| 106 |
+
nodes_to_exclude=[],
|
| 107 |
+
op_types_to_quantize=op_types_to_quantize,
|
| 108 |
+
extra_options=dict(EnableSubgraph=True, MatMulConstBOnly=True),
|
| 109 |
+
)
|
| 110 |
+
quantizer.quantize_model()
|
| 111 |
+
utils.check_and_save_model(quantizer.model.model, save_path)
|
| 112 |
+
return quantizer.model.model
|
| 113 |
+
|
| 114 |
+
def main():
|
| 115 |
+
parser = HfArgumentParser((IOArguments, QuantizationArguments))
|
| 116 |
+
io_args, quantization_args = parser.parse_args_into_dataclasses()
|
| 117 |
+
input_folder = io_args.input_folder
|
| 118 |
+
output_folder = io_args.output_folder
|
| 119 |
+
if not quantization_args.modes:
|
| 120 |
+
raise ValueError("At least one quantization mode must be specified")
|
| 121 |
+
|
| 122 |
+
if not os.path.exists(input_folder):
|
| 123 |
+
raise ValueError(f"Input folder {input_folder} does not exist")
|
| 124 |
+
|
| 125 |
+
model_names_or_paths = [
|
| 126 |
+
os.path.join(input_folder, file)
|
| 127 |
+
for file in os.listdir(input_folder)
|
| 128 |
+
if file.endswith(".onnx")
|
| 129 |
+
]
|
| 130 |
+
if not model_names_or_paths:
|
| 131 |
+
raise ValueError(f"No .onnx models found in {input_folder}")
|
| 132 |
+
|
| 133 |
+
os.makedirs(output_folder, exist_ok=True)
|
| 134 |
+
|
| 135 |
+
for model_path in tqdm(model_names_or_paths, desc="Models"):
|
| 136 |
+
file_name_without_extension = os.path.splitext(os.path.basename(model_path))[0]
|
| 137 |
+
model = onnx.load_model(model_path)
|
| 138 |
+
for mode in tqdm(quantization_args.modes, desc="Modes"):
|
| 139 |
+
try:
|
| 140 |
+
suffix = QUANTIZE_SUFFIX_MAPPING.get(QuantMode(mode), mode)
|
| 141 |
+
except Exception:
|
| 142 |
+
suffix = mode
|
| 143 |
+
save_path = os.path.join(output_folder, f"{file_name_without_extension}_{suffix}.onnx")
|
| 144 |
+
mode_enum = QuantMode(mode)
|
| 145 |
+
try:
|
| 146 |
+
if mode_enum == QuantMode.FP16:
|
| 147 |
+
float16.convert_float_to_float16(
|
| 148 |
+
model,
|
| 149 |
+
keep_io_types=True,
|
| 150 |
+
disable_shape_infer=False,
|
| 151 |
+
op_block_list=quantization_args.op_block_list or []
|
| 152 |
+
)
|
| 153 |
+
|
| 154 |
+
elif mode_enum == QuantMode.INT4 or mode_enum == QuantMode.Q4:
|
| 155 |
+
quantize_int4(
|
| 156 |
+
model,
|
| 157 |
+
save_path,
|
| 158 |
+
block_size=quantization_args.block_size or 32,
|
| 159 |
+
is_symmetric=quantization_args.is_symmetric,
|
| 160 |
+
accuracy_level=quantization_args.accuracy_level or 0,
|
| 161 |
+
)
|
| 162 |
+
|
| 163 |
+
elif mode_enum == QuantMode.INT8 or mode_enum == QuantMode.QI8:
|
| 164 |
+
quantize_int8(
|
| 165 |
+
model,
|
| 166 |
+
save_path,
|
| 167 |
+
per_channel=quantization_args.per_channel or False,
|
| 168 |
+
reduce_range=quantization_args.reduce_range or False,
|
| 169 |
+
weight_type=QuantType.QInt8,
|
| 170 |
+
op_block_list=quantization_args.op_block_list,
|
| 171 |
+
)
|
| 172 |
+
|
| 173 |
+
elif mode_enum == QuantMode.Q8:
|
| 174 |
+
quantize_int8(
|
| 175 |
+
model,
|
| 176 |
+
save_path,
|
| 177 |
+
per_channel=quantization_args.per_channel or False,
|
| 178 |
+
reduce_range=quantization_args.reduce_range or False,
|
| 179 |
+
weight_type=QuantType.QUInt8,
|
| 180 |
+
op_block_list=quantization_args.op_block_list,
|
| 181 |
+
)
|
| 182 |
+
|
| 183 |
+
# Add other modes as needed (Q4F16, BNB4, QU8, etc.)
|
| 184 |
+
except Exception as e:
|
| 185 |
+
print(f"[WARN] Quantization mode '{mode}' failed for model '{model_path}': {e}")
|
| 186 |
+
continue
|
| 187 |
+
|
| 188 |
+
if __name__ == "__main__":
|
| 189 |
+
main()
|
utils.py
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import onnx
|
| 2 |
+
from typing import Optional, Union
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
import os
|
| 5 |
+
import logging
|
| 6 |
+
|
| 7 |
+
logger = logging.getLogger(__name__)
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
# https://github.com/onnx/onnx/pull/6556
|
| 11 |
+
MAXIMUM_PROTOBUF = 2147483648 # 2GiB
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def strict_check_model(model_or_path: Union[onnx.ModelProto, str, Path]):
|
| 15 |
+
try:
|
| 16 |
+
onnx.checker.check_model(model_or_path, full_check=True)
|
| 17 |
+
except Exception as e:
|
| 18 |
+
if "No Op registered for" in str(e):
|
| 19 |
+
pass
|
| 20 |
+
else:
|
| 21 |
+
raise e
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def check_and_save_model(model: onnx.ModelProto, save_path: Optional[Union[str, Path]]):
|
| 25 |
+
if model.ByteSize() < MAXIMUM_PROTOBUF:
|
| 26 |
+
strict_check_model(model)
|
| 27 |
+
if save_path:
|
| 28 |
+
save_path = Path(save_path).as_posix()
|
| 29 |
+
external_file_name = os.path.basename(save_path) + "_data"
|
| 30 |
+
external_path = os.path.join(os.path.dirname(save_path), external_file_name)
|
| 31 |
+
|
| 32 |
+
if save_path.endswith(".onnx") and os.path.isfile(save_path):
|
| 33 |
+
os.remove(save_path)
|
| 34 |
+
if os.path.isfile(external_path):
|
| 35 |
+
os.remove(external_path)
|
| 36 |
+
|
| 37 |
+
onnx.save(
|
| 38 |
+
model,
|
| 39 |
+
save_path,
|
| 40 |
+
convert_attribute=True,
|
| 41 |
+
)
|
| 42 |
+
elif save_path is not None:
|
| 43 |
+
# path/to/model.onnx
|
| 44 |
+
save_path = Path(save_path).as_posix()
|
| 45 |
+
|
| 46 |
+
external_file_name = os.path.basename(save_path) + "_data"
|
| 47 |
+
# path/to/model.onnx_data
|
| 48 |
+
external_path = os.path.join(os.path.dirname(save_path), external_file_name)
|
| 49 |
+
|
| 50 |
+
if save_path.endswith(".onnx") and os.path.isfile(save_path):
|
| 51 |
+
os.remove(save_path)
|
| 52 |
+
if os.path.isfile(external_path):
|
| 53 |
+
os.remove(external_path)
|
| 54 |
+
|
| 55 |
+
onnx.save(
|
| 56 |
+
model,
|
| 57 |
+
save_path,
|
| 58 |
+
save_as_external_data=True,
|
| 59 |
+
all_tensors_to_one_file=True,
|
| 60 |
+
location=external_file_name,
|
| 61 |
+
convert_attribute=True,
|
| 62 |
+
)
|
| 63 |
+
|
| 64 |
+
else:
|
| 65 |
+
logger.info(
|
| 66 |
+
"Merged ONNX model exceeds 2GB, the model will not be checked without `save_path` given."
|
| 67 |
+
)
|