saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-06-12 00:59:23 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Disable some ORT graph optimizers in offline transformers optimization tool (#8923)

Browse source 
walkaround "Unsupported operator FusedMatMul" during symbolic shape inference
This commit is contained in:
Tianlei Wu 2021-09-01 15:47:57 -07:00 committed by GitHub
parent 225439193e
commit 9467f511ac
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
1 changed files with 21 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

26
onnxruntime/python/tools/transformers/optimizer.py
View file

@ -50,7 +50,8 @@ MODEL_TYPES = {
def optimize_by_onnxruntime(onnx_model_path: str,
use_gpu: bool = False,
optimized_model_path: str = None,
opt_level: int = 99) -> str:
opt_level: int = 99,
disabled_optimizers=[]) -> str:
"""
Use onnxruntime to optimize model.
@ -59,7 +60,7 @@ def optimize_by_onnxruntime(onnx_model_path: str,
use_gpu (bool): whether the optimized model is targeted to run in GPU.
optimized_model_path (str or None): the path of optimized model.
opt_level (int): graph optimization level.
disabled_optimizers (List[str]): a list of names of disabled optimizers
Returns:
optimized_model_path (str): the path of optimized model
"""
@ -84,10 +85,17 @@ def optimize_by_onnxruntime(onnx_model_path: str,
sess_options.optimized_model_filepath = optimized_model_path
kwargs = {}
if disabled_optimizers:
kwargs["disabled_optimizers"] = disabled_optimizers
if not use_gpu:
session = onnxruntime.InferenceSession(onnx_model_path, sess_options, providers=['CPUExecutionProvider'])
session = onnxruntime.InferenceSession(onnx_model_path,
sess_options,
providers=['CPUExecutionProvider'],
**kwargs)
else:
session = onnxruntime.InferenceSession(onnx_model_path, sess_options)
session = onnxruntime.InferenceSession(onnx_model_path, sess_options, **kwargs)
assert 'CUDAExecutionProvider' in session.get_providers() # Make sure there is GPU
assert os.path.exists(optimized_model_path) and os.path.isfile(optimized_model_path)
@ -254,7 +262,15 @@ def optimize_model(input,
temp_model_path = None
if opt_level > 1:
temp_model_path = optimize_by_onnxruntime(input, use_gpu=use_gpu, opt_level=opt_level)
# Disable some optimizers that might cause failure in symbolic shape inference or attention fusion.
disabled_optimizers = [] if only_onnxruntime else [
'MatMulScaleFusion', 'MatMulAddFusion'
'SimplifiedLayerNormFusion', 'GemmActivationFusion', 'BiasSoftmaxFusion'
]
temp_model_path = optimize_by_onnxruntime(input,
use_gpu=use_gpu,
opt_level=opt_level,
disabled_optimizers=disabled_optimizers)
elif opt_level == 1:
# basic optimizations (like constant folding and cast elimation) are not specified to exection provider.
# CPU provider is used here so that there is no extra node for GPU memory copy.