Export Times¶
Custom Exporter¶
With a very simple model:
<<<
import time
from experimental_experiment.checks import print_import_time
print_import_time()
import torch
import experimental_experiment.torch_interpreter
class Neuron(torch.nn.Module):
def __init__(self, n_dims: int, n_targets: int):
super(Neuron, self).__init__()
self.linear = torch.nn.Linear(n_dims, n_targets)
def forward(self, x):
return torch.sigmoid(self.linear(x))
model = Neuron(3, 1)
x = torch.rand(5, 3)
begin = time.perf_counter()
onx = experimental_experiment.torch_interpreter.to_onnx(model, (x,))
print(f"time to export 1x --- {time.perf_counter() - begin}")
begin = time.perf_counter()
onx = experimental_experiment.torch_interpreter.to_onnx(model, (x,))
print(f"time to export 2x --- {time.perf_counter() - begin}")
>>>
time to import onnx --- 0.5991993869974976
time to import onnx_array_api --- 0.00012086800052202307
time to import torch --- 1.9300789539993275
'torch.export' already imported
time to import torch.export --- 2.230000973213464e-06
time to import onnxscript --- 0.13854226599869435
time to import onnxruntime --- 0.03265998099959688
time to import torch.onnx --- 0.03169962599713472
time to import torch._dynamo --- 1.41928641299819
time to import experimental_experiment.torch_interpreter --- 2.7358014860001276
time to import experimental_experiment.torch_interpreter.aten_functions --- 0.007763280998915434
time to export 1x --- 3.4872656509978697
time to export 2x --- 0.02234246900115977
[runpythonerror]
use_kernel_func_from_hub is not available in the installed kernels version. Please upgrade kernels to use this feature.
With a bigger model:
<<<
import time
import warnings
import numpy as np
from transformers import LlamaConfig
from transformers.models.llama.modeling_llama import LlamaModel
import onnx
import onnxruntime
import torch
import torch._dynamo
import torch.export
import onnxscript
import torch.onnx
import experimental_experiment
import experimental_experiment.torch_interpreter
import experimental_experiment.torch_interpreter.aten_functions
from experimental_experiment.ext_test_case import get_llama_model
model, example_args_collection = get_llama_model(
input_dims=[(2, 1024)],
hidden_size=4096,
num_hidden_layers=1,
vocab_size=32000,
intermediate_size=11008,
max_position_embeddings=2048,
num_attention_heads=32,
_attn_implementation="eager",
)
begin = time.perf_counter()
onx = experimental_experiment.torch_interpreter.to_onnx(
model, example_args_collection[0]
)
print(f"time to export 1x --- {time.perf_counter() - begin}")
begin = time.perf_counter()
onx = experimental_experiment.torch_interpreter.to_onnx(
model, example_args_collection[0]
)
print(f"time to export 2x --- {time.perf_counter() - begin}")
>>>
[runpythonerror]
use_kernel_func_from_hub is not available in the installed kernels version. Please upgrade kernels to use this feature.
Traceback (most recent call last):
File "<stdin>", line 26, in <module>
TypeError: get_llama_model() got an unexpected keyword argument 'input_dims'
Dynamo Exporter¶
<<<
import time
import warnings
from experimental_experiment.checks import print_import_time
print_import_time()
import torch
import experimental_experiment.torch_interpreter
class Neuron(torch.nn.Module):
def __init__(self, n_dims: int, n_targets: int):
super(Neuron, self).__init__()
self.linear = torch.nn.Linear(n_dims, n_targets)
def forward(self, x):
return torch.sigmoid(self.linear(x))
model = Neuron(3, 1)
x = torch.rand(5, 3)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
begin = time.perf_counter()
onx = torch.onnx.export(model, x, dynamo=True)
print(f"time to export 1x --- {time.perf_counter() - begin}")
begin = time.perf_counter()
onx = torch.onnx.export(model, x, dynamo=True)
print(f"time to export 2x --- {time.perf_counter() - begin}")
>>>
time to import onnx --- 0.5539983979979297
time to import onnx_array_api --- 0.0003355670014570933
time to import torch --- 1.8434638290018484
'torch.export' already imported
time to import torch.export --- 2.3649990907870233e-06
time to import onnxscript --- 0.12674686499667587
time to import onnxruntime --- 0.025429938999877777
time to import torch.onnx --- 0.03100906799954828
time to import torch._dynamo --- 1.1340352190018166
time to import experimental_experiment.torch_interpreter --- 3.0285749489994487
time to import experimental_experiment.torch_interpreter.aten_functions --- 0.0069390030002978165
[torch.onnx] Obtain model graph for `Neuron([...]` with `torch.export.export(..., strict=False)`...
[torch.onnx] Obtain model graph for `Neuron([...]` with `torch.export.export(..., strict=False)`... ✅
[torch.onnx] Run decompositions...
[torch.onnx] Run decompositions... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
[torch.onnx] Optimize the ONNX graph...
[torch.onnx] Optimize the ONNX graph... ✅
time to export 1x --- 1.9132904530015367
[torch.onnx] Obtain model graph for `Neuron([...]` with `torch.export.export(..., strict=False)`...
[torch.onnx] Obtain model graph for `Neuron([...]` with `torch.export.export(..., strict=False)`... ✅
[torch.onnx] Run decompositions...
[torch.onnx] Run decompositions... ✅
[torch.onnx] Translate the graph into ONNX...
[torch.onnx] Translate the graph into ONNX... ✅
[torch.onnx] Optimize the ONNX graph...
[torch.onnx] Optimize the ONNX graph... ✅
time to export 2x --- 0.8022989689998212
With a bigger model:
<<<
import time
import warnings
import numpy as np
from transformers import LlamaConfig
from transformers.models.llama.modeling_llama import LlamaModel
import onnx
import onnxruntime
import torch
import torch._dynamo
import torch.export
import onnxscript
import torch.onnx
import experimental_experiment
import experimental_experiment.torch_interpreter
import experimental_experiment.torch_interpreter.aten_functions
from experimental_experiment.ext_test_case import get_llama_model
model, example_args_collection = get_llama_model(
input_dims=[(2, 1024)],
hidden_size=4096,
num_hidden_layers=1,
vocab_size=32000,
intermediate_size=11008,
max_position_embeddings=2048,
num_attention_heads=32,
_attn_implementation="eager",
)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
begin = time.perf_counter()
onx = torch.onnx.export(model, *example_args_collection[0], dynamo=True)
print(f"time to export 1x --- {time.perf_counter() - begin}")
begin = time.perf_counter()
onx = torch.onnx.export(model, *example_args_collection[0], dynamo=True)
print(f"time to export 2x --- {time.perf_counter() - begin}")
>>>
[runpythonerror]
use_kernel_func_from_hub is not available in the installed kernels version. Please upgrade kernels to use this feature.
Traceback (most recent call last):
File "<stdin>", line 26, in <module>
TypeError: get_llama_model() got an unexpected keyword argument 'input_dims'