def optimize_onnx(onnx_model_path,
optimized_model_path,
is_float16,
num_attention_heads,
hidden_size,
use_external_data_format=False,
**kwargs):
""" Optimize ONNX model with an option to convert it to use mixed precision.
"""
from optimizer import optimize_model
from fusion_options import FusionOptions
optimization_options = FusionOptions('gpt2')
#optimization_options.enable_gelu = False
#optimization_options.enable_layer_norm = False
#optimization_options.enable_attention = False
m = optimize_model(onnx_model_path,
model_type='gpt2',
num_heads=num_attention_heads,
hidden_size=hidden_size,
opt_level=0,
optimization_options=optimization_options,
use_gpu=False)
if is_float16:
op_full_list = set([node.op_type for node in m.nodes()])
op_block_list = set(kwargs["op_block_list"]
) if "op_block_list" in kwargs else set()
op_remain_list = op_full_list.difference(op_block_list)
logger.info(
f"op_block_list={op_block_list} op_remain_list={op_remain_list}"
)
m.convert_float_to_float16(use_symbolic_shape_infer=True, **kwargs)
m.save_model_to_file(optimized_model_path, use_external_data_format)
def optimize_onnx(onnx_model_path,
optimized_model_path,
is_float16,
num_attention_heads,
hidden_size,
use_external_data_format=False,
auto_mixed_precision=False,
**kwargs):
""" Optimize ONNX model with an option to convert it to use mixed precision.
"""
from optimizer import optimize_model
from fusion_options import FusionOptions
optimization_options = FusionOptions('gpt2')
#optimization_options.enable_gelu = False
#optimization_options.enable_layer_norm = False
#optimization_options.enable_attention = False
m = optimize_model(onnx_model_path,
model_type='gpt2',
num_heads=num_attention_heads,
hidden_size=hidden_size,
opt_level=0,
optimization_options=optimization_options,
use_gpu=False)
if is_float16:
if auto_mixed_precision:
Gpt2Helper.auto_mixed_precision(m)
else:
m.convert_float_to_float16(use_symbolic_shape_infer=True,
**kwargs)
m.save_model_to_file(optimized_model_path, use_external_data_format)
def optimize_onnx_model(
model_name,
onnx_model_path,
optimized_model_path,
model_type,
num_attention_heads,
hidden_size,
use_gpu,
precision,
use_raw_attention_mask,
overwrite,
model_fusion_statistics,
use_external_data_format,
optimization_options=None,
):
if overwrite or not os.path.exists(optimized_model_path):
Path(optimized_model_path).parent.mkdir(parents=True, exist_ok=True)
from fusion_options import FusionOptions
from optimizer import optimize_model
if optimization_options == None:
optimization_options = FusionOptions(model_type)
optimization_options.use_raw_attention_mask(use_raw_attention_mask)
if Precision.FLOAT16 == precision:
optimization_options.enable_gelu_approximation = True
if Precision.INT8 == precision:
optimization_options.enable_embed_layer_norm = False
# Use script to optimize model.
# Use opt_level <= 1 for models to be converted to fp16, because some fused op (like FusedGemm) has only fp32 and no fp16.
# It is better to be conservative so we use opt_level=0 here, in case MemcpyFromHost is added to the graph by OnnxRuntime.
opt_model = optimize_model(
onnx_model_path,
model_type,
num_heads=num_attention_heads,
hidden_size=hidden_size,
opt_level=0,
optimization_options=optimization_options,
use_gpu=use_gpu,
only_onnxruntime=False,
)
if model_type == "bert_keras" or model_type == "bert_tf":
opt_model.use_dynamic_axes()
model_fusion_statistics[
optimized_model_path] = opt_model.get_fused_operator_statistics()
if Precision.FLOAT16 == precision:
opt_model.convert_float_to_float16(keep_io_types=True)
opt_model.save_model_to_file(optimized_model_path,
use_external_data_format)
else:
logger.info(
f"Skip optimization since model existed: {optimized_model_path}")
def optimize_by_fusion(
model: ModelProto,
model_type: str = "bert",
num_heads: int = 0,
hidden_size: int = 0,
optimization_options: Optional[FusionOptions] = None,
):
"""Optimize Model by graph fusion logic.
Note that ONNXRuntime graph optimizations (like constant folding) will not be applied. So it is better to enable
constant folding during exporting ONNX model, or run optimize_by_onnxruntime on the model first like optimize_model.
For BERT model, num_heads and hidden_size are optional. For other model types, you need specify these parameters.
Args:
model (ModelProto): model object
model_type (str, optional): model type - like bert, bert_tf, bert_keras or gpt2. Defaults to 'bert'.
num_heads (int, optional): number of attention heads. Defaults to 0.
0 allows detect the parameter from graph automatically (for model_type "bert" only).
hidden_size (int, optional): hidden size. Defaults to 0.
0 allows detect the parameter from graph automatically (for model_type "bert" only).
optimization_options (FusionOptions, optional): optimization options that turn on/off some fusions. Defaults to None.
Returns:
object of an optimizer class.
"""
if model_type != "bert" and (num_heads == 0 or hidden_size == 0):
logger.warning(
"Please specify parameters of num_heads and hidden_size when model_type is not 'bert'"
)
(optimizer_class, producer, _) = MODEL_TYPES[model_type]
if model.producer_name and producer != model.producer_name:
logger.warning(
f'Model producer not matched: Expected "{producer}", Got "{model.producer_name}".'
"Please specify correct --model_type parameter.")
if optimization_options is None:
optimization_options = FusionOptions(model_type)
optimizer = optimizer_class(model, num_heads, hidden_size)
optimizer.optimize(optimization_options)
optimizer.topological_sort()
optimizer.model.producer_name = "onnxruntime.transformers"
from onnxruntime import __version__ as onnxruntime_version
optimizer.model.producer_version = onnxruntime_version
return optimizer
def optimize_model(input,
model_type='bert',
num_heads=0,
hidden_size=0,
optimization_options=None,
opt_level=None,
use_gpu=False,
only_onnxruntime=False):
""" Optimize Model by OnnxRuntime and/or python fusion logic.
ONNX Runtime has graph optimizations (https://onnxruntime.ai/docs/resources/graph-optimizations.html).
However, the coverage is limited. We also have graph fusions that implemented in Python to improve the coverage.
They can combined: ONNX Runtime will run first when opt_level > 0, then graph fusions in Python will be applied.
To use ONNX Runtime only and no Python fusion logic, use only_onnxruntime flag and a positive opt_level like
optimize_model(input, opt_level=1, use_gpu=False, only_onnxruntime=True)
When opt_level is None, we will choose default optimization level according to model type.
When opt_level is 0 and only_onnxruntime is False, only python fusion logic is used and onnxruntime is disabled.
When opt_level > 1, use_gpu shall set properly since the optimized graph might contain operators for GPU or CPU only.
If your model is intended for GPU inference only (especially float16 or mixed precision model), it is recommended to
set use_gpu to be True, otherwise the model is not optimized for GPU inference.
For BERT model, num_heads and hidden_size are optional. For other model types, you need specify these parameters.
Args:
input (str): input model path.
model_type (str, optional): model type - like bert, bert_tf, bert_keras or gpt2. Defaults to 'bert'.
num_heads (int, optional): number of attention heads. Defaults to 0.
0 allows detect the parameter from graph automatically (for model_type "bert" only).
hidden_size (int, optional): hidden size. Defaults to 0.
0 allows detect the parameter from graph automatically (for model_type "bert" only).
optimization_options (FusionOptions, optional): optimization options that turn on/off some fusions. Defaults to None.
opt_level (int, optional): onnxruntime graph optimization level (0, 1, 2 or 99) or None. Defaults to None.
When the value is None, default value (1 for bert and gpt2, 0 for other model types) will be used.
When the level > 0, onnxruntime will be used to optimize model first.
use_gpu (bool, optional): use gpu or not for onnxruntime. Defaults to False.
only_onnxruntime (bool, optional): only use onnxruntime to optimize model, and no python fusion. Defaults to False.
Returns:
object of an optimizer class.
"""
assert opt_level is None or opt_level in [0, 1, 2, 99]
if model_type != "bert" and (num_heads == 0 or hidden_size == 0):
logger.warning("Please specify parameters of num_heads and hidden_size when model_type is not 'bert'")
(optimizer_class, producer, default_opt_level) = MODEL_TYPES[model_type]
if opt_level is None:
opt_level = default_opt_level
temp_model_path = None
if opt_level > 1:
temp_model_path = optimize_by_onnxruntime(input, use_gpu=use_gpu, opt_level=opt_level)
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.
temp_model_path = optimize_by_onnxruntime(input, use_gpu=False, opt_level=1)
if only_onnxruntime and not temp_model_path:
logger.warning("Please specify a positive value for opt_level when only_onnxruntime is True")
model = load_model(temp_model_path or input, format=None, load_external_data=True)
if model.producer_name and producer != model.producer_name:
logger.warning(
f"Model producer not matched: Expect {producer}, Got {model.producer_name} {model.producer_version}. Please specify correct --model_type parameter."
)
if optimization_options is None:
optimization_options = FusionOptions(model_type)
optimizer = optimizer_class(model, num_heads, hidden_size)
if not only_onnxruntime:
optimizer.optimize(optimization_options)
# Remove the temporary model.
if temp_model_path:
os.remove(temp_model_path)
logger.debug("Remove tempoary model: {}".format(temp_model_path))
optimizer.model.producer_name = "onnxruntime.transformers"
from onnxruntime import __version__ as onnxruntime_version
optimizer.model.producer_version = onnxruntime_version
return optimizer
