Note
Go to the end to download the full example code.
201: Evaluate different ways to export a torch model to ONNX¶
The example evaluates the performance of onnxruntime of a simple torch model after it was converted into ONNX through different processes:
TorchScript-based ONNX Exporter, let’s call it script
TorchDynamo-based ONNX Exporter, let’s call it dynamo
if available, the previous model but optimized, dynopt
a custom exporter cus_p0, this exporter supports a very limited set of models, as dynamo, it relies on torch.fx but the design is closer to what tensorflow-onnx does.
the same exporter but unused nodes were removed and constants were folded, cus_p2
To run the script:
python _doc/examples/plot_torch_export --help
The script takes around 12 minutes with a larger models.
Some helpers¶
from experimental_experiment.args import get_parsed_args
script_args = get_parsed_args(
"plot_torch_export",
description=__doc__,
scenarios={
"small": "small model to test",
"middle": "55Mb model",
"large": "1Gb model",
},
warmup=5,
repeat=5,
maxtime=(
2,
"maximum time to run a model to measure the computation time, "
"it is 0.1 when scenario is small",
),
expose="scenarios,repeat,warmup",
)
import contextlib
import itertools
import os
import platform
import pprint
import multiprocessing
import time
import cProfile
import pstats
import io
import warnings
import logging
from pstats import SortKey
try:
with warnings.catch_warnings():
warnings.simplefilter("ignore")
import onnxruntime
has_cuda = "CUDAExecutionProvider" in onnxruntime.get_available_providers()
except ImportError:
print("onnxruntime not available.")
import sys
sys.exit(0)
import numpy as np
import matplotlib.pyplot as plt
import pandas
import onnx
from onnx_array_api.profiling import profile2graph
import torch
from torch import nn
import torch.nn.functional as F
import experimental_experiment
from experimental_experiment.torch_interpreter import to_onnx
from experimental_experiment.xbuilder import OptimizationOptions
from experimental_experiment.plotting.memory import memory_peak_plot
from experimental_experiment.ext_test_case import measure_time, get_figure
from experimental_experiment.memory_peak import start_spying_on
from experimental_experiment.ext_test_case import unit_test_going
from experimental_experiment.helpers import pretty_onnx
from tqdm import tqdm
has_cuda = has_cuda and torch.cuda.is_available()
logging.disable(logging.ERROR)
def system_info():
obs = {}
obs["processor"] = platform.processor()
obs["cores"] = multiprocessing.cpu_count()
try:
obs["cuda"] = 1 if torch.cuda.is_available() else 0
obs["cuda_count"] = torch.cuda.device_count()
obs["cuda_name"] = torch.cuda.get_device_name()
obs["cuda_capa"] = torch.cuda.get_device_capability()
except (RuntimeError, AssertionError):
# no cuda
pass
return obs
pprint.pprint(system_info())
{'cores': 20,
'cuda': 1,
'cuda_capa': (8, 9),
'cuda_count': 1,
'cuda_name': 'NVIDIA GeForce RTX 4060 Laptop GPU',
'processor': 'x86_64'}
Scripts arguments
if script_args.scenario in (None, "small"):
script_args.maxtime = 0.1
if unit_test_going():
script_args.warmup = 1
script_args.repeat = 1
script_args.maxtime = 0.1
script_args.scenario = "small"
print(f"scenario={script_args.scenario or 'small'}")
print(f"warmup={script_args.warmup}")
print(f"repeat={script_args.repeat}")
print(f"maxtime={script_args.maxtime}")
scenario=small
warmup=5
repeat=5
maxtime=0.1
The model¶
A simple model to convert.
class MyModelClass(nn.Module):
def __init__(self, scenario=script_args.scenario):
super().__init__()
if scenario == "middle":
self.large = False
self.conv1 = nn.Conv2d(1, 128, 5)
self.conv2 = nn.Conv2d(128, 16, 5)
self.fc1 = nn.Linear(13456, 1024)
self.fcs = []
self.fc2 = nn.Linear(1024, 128)
self.fc3 = nn.Linear(128, 10)
elif scenario in (None, "small"):
self.large = False
self.conv1 = nn.Conv2d(1, 16, 5)
self.conv2 = nn.Conv2d(16, 16, 5)
self.fc1 = nn.Linear(16, 512)
self.fcs = []
self.fc2 = nn.Linear(512, 128)
self.fc3 = nn.Linear(128, 10)
elif scenario in (None, "large"):
self.large = True
self.conv1 = nn.Conv2d(1, 128, 5)
self.conv2 = nn.Conv2d(128, 16, 5)
self.fc1 = nn.Linear(13456, 4096)
# torch script does not support loops.
self.fca = nn.Linear(4096, 4096)
self.fcb = nn.Linear(4096, 4096)
self.fcc = nn.Linear(4096, 4096)
self.fcd = nn.Linear(4096, 4096)
self.fce = nn.Linear(4096, 4096)
self.fcf = nn.Linear(4096, 4096)
self.fcg = nn.Linear(4096, 4096)
self.fch = nn.Linear(4096, 4096)
self.fci = nn.Linear(4096, 4096)
self.fck = nn.Linear(4096, 4096)
self.fcl = nn.Linear(4096, 4096)
self.fcm = nn.Linear(4096, 4096)
self.fcn = nn.Linear(4096, 4096)
# end of the unfolded loop.
self.fc2 = nn.Linear(4096, 128)
self.fc3 = nn.Linear(128, 10)
else:
raise ValueError(f"Unsupported scenario={scenario!r}.")
def forward(self, x):
x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
x = F.max_pool2d(F.relu(self.conv2(x)), 2)
x = torch.flatten(x, 1)
x = F.relu(self.fc1(x))
if self.large:
# loop
x = F.relu(self.fca(x))
x = F.relu(self.fcb(x))
x = F.relu(self.fcc(x))
x = F.relu(self.fcd(x))
x = F.relu(self.fce(x))
x = F.relu(self.fcf(x))
x = F.relu(self.fcg(x))
x = F.relu(self.fch(x))
x = F.relu(self.fci(x))
x = F.relu(self.fck(x))
x = F.relu(self.fcl(x))
x = F.relu(self.fcm(x))
x = F.relu(self.fcn(x))
# end of the loop
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def create_model_and_input(scenario=script_args.scenario):
if scenario == "middle":
shape = [1, 1, 128, 128]
elif scenario in (None, "small"):
shape = [1, 1, 16, 16]
elif scenario == "large":
shape = [1, 1, 128, 128]
else:
raise ValueError(f"Unsupported scenario={scenario!r}.")
input_tensor = torch.rand(*shape).to(torch.float32)
model = MyModelClass(scenario=scenario)
assert model(input_tensor) is not None
return model, input_tensor
def torch_model_size(model):
size_model = 0
for param in model.parameters():
size = param.numel() * torch.finfo(param.data.dtype).bits / 8
size_model += size
return size_model
model, input_tensor = create_model_and_input()
model_size = torch_model_size(model)
print(f"model size={model_size / 2 ** 20} Mb")
model size=0.31467437744140625 Mb
The exporters¶
def export_script(filename, model, *args):
with contextlib.redirect_stdout(io.StringIO()):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
torch.onnx.export(model, *args, filename, input_names=["input"])
def export_dynamo(filename, model, *args):
with contextlib.redirect_stdout(io.StringIO()):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
export_output = torch.onnx.export(model, args, dynamo=True)
export_output.save(filename)
def export_dynopt(filename, model, *args):
with contextlib.redirect_stdout(io.StringIO()):
with warnings.catch_warnings():
warnings.simplefilter("ignore")
export_output = torch.onnx.export(model, args, dynamo=True)
model_onnx = export_output.model_proto
from experimental_experiment.convert.convert_helper import (
optimize_model_proto_oxs,
)
optimized_model = optimize_model_proto_oxs(model_onnx)
with open(filename, "wb") as f:
f.write(optimized_model.SerializeToString())
def export_cus_p0(filename, model, *args):
onx = to_onnx(model, tuple(args), input_names=["input"])
with open(filename, "wb") as f:
f.write(onx.SerializeToString())
def export_cus_p2(filename, model, *args):
onx = to_onnx(
model,
tuple(args),
input_names=["input"],
options=OptimizationOptions(
remove_unused=True,
constant_folding=True,
),
)
with open(filename, "wb") as f:
f.write(onx.SerializeToString())
Let’s check they are working.
export_functions = [
export_script,
export_dynamo,
export_dynopt,
export_cus_p0,
export_cus_p2,
]
exporters = {f.__name__.replace("export_", ""): f for f in export_functions}
supported_exporters = {}
for k, v in exporters.items():
print(f"run exporter {k}")
filename = f"plot_torch_export_{k}.onnx"
try:
v(filename, model, input_tensor)
except Exception as e:
print(f"skipped due to {str(e)[:1000]}")
continue
supported_exporters[k] = v
print(f"done. size={os.stat(filename).st_size / 2 ** 20:1.0f} Mb")
run exporter script
done. size=0 Mb
run exporter dynamo
done. size=0 Mb
run exporter dynopt
done. size=0 Mb
run exporter cus_p0
done. size=0 Mb
run exporter cus_p2
done. size=0 Mb
Exporter memory¶
def flatten(ps):
obs = ps["cpu"].to_dict(unit=2**20)
if "gpus" in ps:
for i, g in enumerate(ps["gpus"]):
for k, v in g.to_dict(unit=2**20).items():
obs[f"gpu{i}_{k}"] = v
return obs
data = []
for k, v in supported_exporters.items():
print(f"run exporter for memory {k}")
filename = f"plot_torch_export_{k}.onnx"
if has_cuda:
torch.cuda.set_device(0)
stat = start_spying_on(cuda=1 if has_cuda else 0)
v(filename, model, input_tensor)
obs = flatten(stat.stop())
print("done.")
onx = onnx.load(filename)
obs.update(dict(nodes=len(onx.graph.node), export=k))
data.append(obs)
stat = start_spying_on(cuda=1 if has_cuda else 0)
exported_mod = torch.export.export(model, (input_tensor,))
obs = flatten(stat.stop())
obs.update(dict(export="torch.fx"))
data.append(obs)
run exporter for memory script
done.
run exporter for memory dynamo
done.
run exporter for memory dynopt
done.
run exporter for memory cus_p0
done.
run exporter for memory cus_p2
done.
The result.
df1 = pandas.DataFrame(data)
df1.to_csv("plot_torch_export_memory.csv", index=False)
df1.to_excel("plot_torch_export_memory.xlsx", index=False)
print(df1)
ax = memory_peak_plot(
data,
bars=[model_size * i / 2**20 for i in range(1, 5)],
suptitle=f"Memory Consumption of the Export\nmodel size={model_size / 2**20:1.0f} Mb",
)
get_figure(ax).savefig("plot_torch_export_memory.png")
peak mean n begin end gpu0_peak gpu0_mean gpu0_n gpu0_begin gpu0_end nodes export
0 6305.843750 6305.839063 10 6305.843750 6305.839844 1835.617188 1835.617188 10 1835.617188 1835.617188 12.0 script
1 6306.984375 6306.121094 51 6305.839844 6306.984375 1835.617188 1835.617188 51 1835.617188 1835.617188 17.0 dynamo
2 6307.011719 6306.704132 97 6306.984375 6292.640625 1835.617188 1835.617188 97 1835.617188 1835.617188 16.0 dynopt
3 6292.808594 6292.782227 16 6292.750000 6292.808594 1835.617188 1835.617188 16 1835.617188 1835.617188 15.0 cus_p0
4 6292.863281 6292.820801 16 6292.808594 6292.863281 1835.617188 1835.617188 16 1835.617188 1835.617188 12.0 cus_p2
5 6292.875000 6292.875000 14 6292.875000 6292.875000 1835.617188 1835.617188 14 1835.617188 1835.617188 NaN torch.fx
Exporter speed¶
data = []
for k, v in supported_exporters.items():
print(f"run exporter {k}")
filename = f"plot_torch_export_{k}.onnx"
times = []
for _ in range(script_args.repeat):
begin = time.perf_counter()
v(filename, model, input_tensor)
duration = time.perf_counter() - begin
times.append(duration)
onx = onnx.load(filename)
print("done.")
data.append(
dict(
export=k,
time=np.mean(times),
min=min(times),
max=max(times),
first=times[0],
last=times[-1],
std=np.std(times),
nodes=len(onx.graph.node),
)
)
run exporter script
done.
run exporter dynamo
done.
run exporter dynopt
done.
run exporter cus_p0
done.
run exporter cus_p2
done.
The last export to measure time torch spends in export the model before any other export can begin the translation except the first one.
times = []
for _ in range(script_args.repeat):
begin = time.perf_counter()
exported_mod = torch.export.export(model, (input_tensor,))
duration = time.perf_counter() - begin
times.append(duration)
data.append(
dict(
export="torch.fx",
time=np.mean(times),
min=min(times),
max=max(times),
first=times[0],
last=times[-1],
std=np.std(times),
nodes=len(onx.graph.node),
)
)
The result.
df1 = pandas.DataFrame(data)
df1.to_csv("plot_torch_export_time.csv", index=False)
df1.to_excel("plot_torch_export_time.xlsx", index=False)
print(df1)
fig, ax = plt.subplots(1, 1)
dfi = df1[["export", "time", "std"]].set_index("export")
dfi["time"].plot.bar(ax=ax, title="Export time", yerr=dfi["std"], rot=30)
fig.tight_layout()
fig.savefig("plot_torch_export_time.png")
export time min max first last std nodes
0 script 0.071788 0.060297 0.084890 0.084890 0.060297 0.009607 12
1 dynamo 0.542353 0.425884 0.826942 0.472952 0.826942 0.149822 17
2 dynopt 0.517435 0.406325 0.736772 0.565939 0.455488 0.122987 16
3 cus_p0 0.108924 0.094842 0.117783 0.113209 0.114845 0.008427 15
4 cus_p2 0.098535 0.092205 0.111432 0.104525 0.092245 0.008014 12
5 torch.fx 0.074501 0.070695 0.084103 0.084103 0.070695 0.004938 12
Exporter Profiling¶
def clean_text(text):
pathes = [
os.path.abspath(os.path.normpath(os.path.join(os.path.dirname(torch.__file__), ".."))),
os.path.abspath(os.path.normpath(os.path.join(os.path.dirname(onnx.__file__), ".."))),
os.path.abspath(
os.path.normpath(
os.path.join(os.path.dirname(experimental_experiment.__file__), "..")
)
),
]
for p in pathes:
text = text.replace(p, "")
text = text.replace("experimental_experiment", "experimental_experiment".upper())
return text
def profile_function(name, export_function, verbose=False):
print(f"profile {name}: {export_function}")
pr = cProfile.Profile()
pr.enable()
for _ in range(script_args.repeat):
export_function("dummyc.onnx", model, input_tensor)
pr.disable()
s = io.StringIO()
sortby = SortKey.CUMULATIVE
ps = pstats.Stats(pr, stream=s).sort_stats(sortby)
ps.print_stats()
raw = s.getvalue()
text = "\n".join(raw.split("\n")[:200])
if verbose:
print(text)
with open(f"plot_torch_export_profile_{name}.txt", "w") as f:
f.write(raw)
root, nodes = profile2graph(ps, clean_text=clean_text)
text = root.to_text()
with open(f"plot_torch_export_profile_{name}_h.txt", "w") as f:
f.write(text)
print("done.")
profile_function("custom0", export_cus_p0, True)
profile_function("custom2", export_cus_p2)
profile custom0: <function export_cus_p0 at 0x7f622b5af2e0>
1097788 function calls (1066786 primitive calls) in 1.094 seconds
Ordered by: cumulative time
ncalls tottime percall cumtime percall filename:lineno(function)
60 0.001 0.000 1.145 0.019 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:878(call_function)
25 0.001 0.000 1.092 0.044 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/nn_module.py:342(call_function)
1080/690 0.002 0.000 0.233 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_stats.py:15(wrapper)
65 0.000 0.000 0.190 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:2151(wrap_fx_proxy)
65 0.000 0.000 0.189 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:2224(wrap_fx_proxy_cls)
60 0.001 0.000 0.186 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:2308(_wrap_fx_proxy)
90 0.001 0.000 0.183 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/utils.py:2072(wrap_fake_exception)
60 0.001 0.000 0.178 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/utils.py:2468(get_fake_value)
870 0.002 0.000 0.176 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1263(__torch_dispatch__)
870 0.011 0.000 0.174 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1788(dispatch)
280/53 0.002 0.000 0.168 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:935(step)
5 0.001 0.000 0.167 0.033 /home/xadupre/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:4468(to_onnx)
55/11 0.003 0.000 0.162 0.015 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:653(wrapper)
485 0.004 0.000 0.160 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1348(_cached_dispatch_impl)
55/11 0.000 0.000 0.160 0.015 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:2341(CALL)
55/11 0.000 0.000 0.160 0.015 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:2300(_call)
160/110 0.058 0.000 0.150 0.001 {method 'clone' of 'torch._C.TensorBase' objects}
50 0.001 0.000 0.143 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/parameter.py:63(__deepcopy__)
5175/2065 0.011 0.000 0.118 0.000 /usr/lib/python3.12/copy.py:118(deepcopy)
435/325 0.000 0.000 0.117 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_ops.py:721(__call__)
25 0.000 0.000 0.116 0.005 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/utils.py:2083(deepcopy_to_fake_tensor)
610/250 0.002 0.000 0.113 0.000 /usr/lib/python3.12/copy.py:247(_reconstruct)
55/10 0.001 0.000 0.113 0.011 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/symbolic_shapes.py:6789(run_node)
1215 0.003 0.000 0.112 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1231(__torch_function__)
255/120 0.002 0.000 0.109 0.001 /usr/lib/python3.12/copy.py:217(_deepcopy_dict)
25 0.000 0.000 0.108 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/utils.py:2085(<lambda>)
70 0.000 0.000 0.107 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1569(python_code)
1215 0.001 0.000 0.106 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1260(__torch_function__)
60 0.001 0.000 0.104 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_ops.py:829(handler)
60 0.006 0.000 0.101 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_library/utils.py:280(handle_dispatch_mode)
60 0.001 0.000 0.100 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:792(recompile)
60 0.000 0.000 0.095 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1329(__torch_dispatch__)
60 0.002 0.000 0.094 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:762(proxy_call)
70 0.001 0.000 0.091 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1646(_python_code)
250 0.002 0.000 0.090 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:2695(__torch_function__)
70 0.009 0.000 0.090 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:408(_gen_python_code)
5 0.000 0.000 0.086 0.017 /home/xadupre/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:4854(optimize)
5 0.000 0.000 0.081 0.016 /home/xadupre/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:5143(optimize_with_patterns)
485 0.002 0.000 0.080 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1722(_output_from_cache_entry)
5 0.005 0.001 0.080 0.016 /home/xadupre/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:1033(optimize)
515 0.009 0.000 0.078 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1656(_get_output_tensor_from_cache_entry)
11280/10990 0.006 0.000 0.073 0.000 {built-in method builtins.next}
485 0.002 0.000 0.072 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1391(_cache_key)
135/26 0.001 0.000 0.072 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/interpreter.py:212(run_node)
45 0.000 0.000 0.069 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:124(forward)
75/45 0.005 0.000 0.069 0.002 {built-in method torch._C._nn.linear}
2070/535 0.009 0.000 0.067 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1467(_prep_args_for_hash)
5 0.000 0.000 0.064 0.013 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:2690(__init__)
5 0.001 0.000 0.064 0.013 /home/xadupre/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:3956(_build_initializers)
150 0.001 0.000 0.064 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:764(__torch_dispatch__)
50 0.000 0.000 0.062 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:490(call_module)
50 0.003 0.000 0.061 0.001 /home/xadupre/github/experimental-experiment/experimental_experiment/mini_onnx_builder.py:104(proto_from_array)
25 0.000 0.000 0.061 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:811(module_call_wrapper)
25 0.000 0.000 0.061 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1741(call_module)
25 0.000 0.000 0.057 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:813(forward)
60 0.000 0.000 0.056 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/utils.py:2529(<lambda>)
60 0.000 0.000 0.056 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/utils.py:2619(run_node)
20/4 0.000 0.000 0.056 0.014 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/interpreter.py:336(call_module)
100/60 0.000 0.000 0.056 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/functional.py:1693(relu)
1030 0.015 0.000 0.056 0.000 /home/xadupre/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:120(enumerate_matches)
60 0.003 0.000 0.055 0.001 {built-in method torch.relu}
35 0.000 0.000 0.055 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/interpreter.py:290(call_function)
5 0.000 0.000 0.049 0.010 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/_trace.py:412(_produce_aten_artifact)
35 0.001 0.000 0.048 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/torch.py:876(call_function)
50 0.000 0.000 0.047 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/parameter.py:40(__new__)
175/125 0.002 0.000 0.047 0.000 {method 'detach' of 'torch._C.TensorBase' objects}
3700/3555 0.003 0.000 0.046 0.000 /usr/lib/python3.12/contextlib.py:132(__enter__)
60/30 0.000 0.000 0.045 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/overrides.py:1669(handle_torch_function)
1330 0.013 0.000 0.044 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:634(emit_node)
5 0.000 0.000 0.044 0.009 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:3049(RETURN_VALUE)
5 0.000 0.000 0.044 0.009 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:3021(_return)
5 0.000 0.000 0.043 0.009 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:957(compile_subgraph)
5 0.000 0.000 0.042 0.008 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_functorch/functional_call.py:11(functional_call)
5 0.000 0.000 0.042 0.008 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/utils/stateless.py:264(_functional_call)
110 0.000 0.000 0.041 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/base.py:445(build)
5 0.001 0.000 0.041 0.008 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1299(compile_and_call_fx_graph)
10 0.000 0.000 0.041 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/lazy.py:61(realize)
5 0.000 0.000 0.041 0.008 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/_lazy_graph_module.py:115(_lazy_forward)
105 0.001 0.000 0.040 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:836(meta_tensor)
110 0.001 0.000 0.040 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:369(__call__)
565/495 0.002 0.000 0.040 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/module.py:1932(__setattr__)
25 0.001 0.000 0.040 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:437(__init__)
198940/196470 0.033 0.000 0.040 0.000 {built-in method builtins.isinstance}
55/5 0.001 0.000 0.038 0.008 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/lazy.py:104(realize_all)
5 0.000 0.000 0.038 0.008 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:1841(SHAPE_ENV)
5 0.034 0.007 0.038 0.008 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:1255(add_python_lambda_leaf_guard_to_root)
65 0.002 0.000 0.038 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:528(_wrap)
25 0.000 0.000 0.036 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:548(graph)
105 0.002 0.000 0.036 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/autograd/grad_mode.py:273(__exit__)
5 0.000 0.000 0.035 0.007 /home/xadupre/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:4379(process)
240 0.001 0.000 0.034 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/proxy.py:209(create_proxy)
120 0.001 0.000 0.034 0.000 /home/xadupre/github/experimental-experiment/experimental_experiment/torch_interpreter/interpreter.py:171(run_node)
5 0.000 0.000 0.034 0.007 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:821(call_wrapped)
5 0.000 0.000 0.034 0.007 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:382(__call__)
5 0.000 0.000 0.033 0.007 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:2602(build_guard_function)
30 0.000 0.000 0.032 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/conv.py:553(forward)
30 0.000 0.000 0.032 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/conv.py:536(_conv_forward)
50/30 0.003 0.000 0.032 0.001 {built-in method torch.conv2d}
3700/3555 0.003 0.000 0.031 0.000 /usr/lib/python3.12/contextlib.py:141(__exit__)
5 0.000 0.000 0.030 0.006 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py:1162(rewrite_signature)
735 0.008 0.000 0.029 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:678(__new__)
10 0.000 0.000 0.029 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/lazy.py:20(realize)
635 0.009 0.000 0.028 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:974(extract_tensor_metadata)
635 0.009 0.000 0.028 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:954(_flatten_into)
2920/135 0.003 0.000 0.027 0.000 /usr/lib/python3.12/ast.py:403(visit)
50/30 0.000 0.000 0.027 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_jit_internal.py:614(fn)
10 0.000 0.000 0.026 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:908(print_readable)
50/30 0.000 0.000 0.026 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/functional.py:807(_max_pool2d)
10 0.000 0.000 0.026 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:297(_print_readable)
30 0.002 0.000 0.026 0.001 {built-in method torch.max_pool2d}
65 0.000 0.000 0.026 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:594(track_tensor_tree)
60 0.001 0.000 0.026 0.000 /home/xadupre/github/experimental-experiment/experimental_experiment/torch_interpreter/interpreter.py:1305(call_function)
5 0.001 0.000 0.025 0.005 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/bytecode_transformation.py:1414(cleaned_instructions)
250 0.002 0.000 0.025 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/proxy.py:143(create_node)
120/65 0.000 0.000 0.025 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:616(wrap_with_proxy)
220 0.004 0.000 0.023 0.000 /home/xadupre/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:3327(make_node)
190 0.001 0.000 0.023 0.000 /home/xadupre/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder_opset.py:113(make_node)
105 0.004 0.000 0.022 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/autograd/grad_mode.py:269(__enter__)
5 0.000 0.000 0.022 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/metrics_context.py:37(__exit__)
5 0.000 0.000 0.022 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/utils.py:1010(record_compilation_metrics)
8495/5035 0.013 0.000 0.022 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:899(map_aggregate)
10 0.000 0.000 0.022 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/passes/replace_with_hop_pass_util.py:157(_replace_with_hop_pass_helper)
4480 0.003 0.000 0.021 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:890(map_arg)
260 0.002 0.000 0.021 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1111(create_node)
4910 0.008 0.000 0.021 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:869(__setattr__)
5 0.001 0.000 0.021 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/utils.py:969(_scrubbed_inductor_config_for_logging)
5 0.000 0.000 0.021 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py:1103(transform)
5 0.000 0.000 0.021 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/interpreter.py:565(transform)
65 0.000 0.000 0.020 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:1799(LOAD_ATTR)
5 0.000 0.000 0.020 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1362(<lambda>)
105 0.000 0.000 0.020 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:1252(get_guard_manager)
5 0.000 0.000 0.019 0.004 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:1533(wrap_tensor)
65 0.000 0.000 0.019 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:1792(_load_attr)
4365 0.004 0.000 0.019 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:177(is_sparse_any)
15 0.000 0.000 0.019 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/_lazy_graph_module.py:57(_make_graph_module)
360 0.000 0.000 0.018 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:443(_ast_unparse)
60 0.000 0.000 0.018 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py:1769(LOAD_METHOD)
960 0.001 0.000 0.018 0.000 /home/xadupre/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:913(match)
360 0.000 0.000 0.018 0.000 /usr/lib/python3.12/ast.py:1789(unparse)
65 0.000 0.000 0.018 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builtin.py:987(call_function)
115 0.001 0.000 0.018 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:487(set_meta)
360 0.000 0.000 0.018 0.000 /usr/lib/python3.12/ast.py:855(visit)
65 0.000 0.000 0.017 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builtin.py:850(builtin_dispatch)
65 0.000 0.000 0.017 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builtin.py:770(call_self_handler)
65 0.001 0.000 0.017 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builtin.py:1632(call_getattr)
210/105 0.003 0.000 0.017 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:929(get_guard_manager_from_source)
1845/360 0.003 0.000 0.017 0.000 /usr/lib/python3.12/ast.py:845(traverse)
960 0.001 0.000 0.017 0.000 /home/xadupre/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:336(_get_match_pattern)
150 0.016 0.000 0.016 0.000 {built-in method builtins.compile}
65 0.002 0.000 0.016 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:1939(create_proxy)
60 0.000 0.000 0.016 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:91(_forward_from_src)
60 0.000 0.000 0.016 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:97(_method_from_src)
60 0.000 0.000 0.016 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:86(_exec_with_source)
3955 0.002 0.000 0.016 0.000 /usr/lib/python3.12/re/__init__.py:174(search)
6320 0.003 0.000 0.016 0.000 {built-in method builtins.repr}
105 0.004 0.000 0.015 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:260(describe_tensor)
5 0.000 0.000 0.015 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:2212(count)
37410 0.012 0.000 0.015 0.000 {built-in method builtins.getattr}
10 0.000 0.000 0.015 0.002 /home/xadupre/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:286(_build_pattern)
1570 0.002 0.000 0.015 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/symbolic_shapes.py:2982(_suppress_guards)
1655/1585 0.006 0.000 0.015 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/recording.py:238(wrapper)
2940 0.008 0.000 0.015 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:597(__set__)
60 0.000 0.000 0.015 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/output_graph.py:576(create_proxy)
705/45 0.002 0.000 0.014 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:2159(visit)
140 0.002 0.000 0.014 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1623(override_node_repr)
705/45 0.001 0.000 0.014 0.000 /usr/lib/python3.12/ast.py:409(generic_visit)
45 0.000 0.000 0.014 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:2608(replace)
5 0.000 0.000 0.014 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:797(__init__)
45 0.000 0.000 0.014 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:2221(replace)
4560 0.007 0.000 0.014 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:154(create_name)
825 0.005 0.000 0.014 0.000 /usr/lib/python3.12/dis.py:434(_get_instructions_bytes)
120 0.001 0.000 0.013 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1777(create_node)
85 0.005 0.000 0.013 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:774(getitem_on_dict_mgr)
4815/4710 0.004 0.000 0.013 0.000 {method 'join' of 'str' objects}
4320 0.002 0.000 0.013 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:605(__repr__)
5 0.000 0.000 0.012 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:2833(wrap_to_fake_tensor_and_record)
5 0.000 0.000 0.012 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:538(apply_runtime_assertion_pass)
3955 0.012 0.000 0.012 0.000 {method 'search' of 're.Pattern' objects}
840 0.001 0.000 0.012 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:555(_format_args)
5 0.000 0.000 0.012 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:2870(<lambda>)
115 0.000 0.000 0.012 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:391(extract_val)
70 0.000 0.000 0.011 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py:1083(run_node)
5 0.000 0.000 0.011 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:774(placeholder_naming_pass)
115 0.000 0.000 0.011 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:365(snapshot_fake)
115 0.002 0.000 0.011 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_impls.py:991(fast_detach)
5 0.000 0.000 0.011 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/passes/replace_autocast_with_hop_pass.py:178(replace_autocast_with_hop_pass)
5 0.000 0.000 0.011 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/passes/replace_set_grad_with_hop_pass.py:110(replace_set_grad_with_hop_pass)
30 0.000 0.000 0.011 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/nn_module.py:249(var_getattr)
455/45 0.001 0.000 0.011 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/guards.py:2175(visit)
415/45 0.001 0.000 0.011 0.000 /usr/lib/python3.12/ast.py:477(generic_visit)
30 0.000 0.000 0.011 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_dynamo/variables/builder.py:1387(wrap_module)
10 0.000 0.000 0.011 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_config_module.py:525(get_config_copy)
10 0.001 0.000 0.011 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_config_module.py:413(_get_dict)
3700 0.004 0.000 0.010 0.000 /usr/lib/python3.12/contextlib.py:299(helper)
3745 0.002 0.000 0.010 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1620(node_repr)
done.
profile custom2: <function export_cus_p2 at 0x7f622b732980>
done.
Same with dynamo-exporter.
profile_function("dynamo", export_dynamo, verbose=True)
if "dynopt" in supported_exporters:
profile_function("dynopt", export_dynopt)
profile dynamo: <function export_dynamo at 0x7f62d447d580>
7570581 function calls (7444505 primitive calls) in 5.082 seconds
Ordered by: cumulative time
ncalls tottime percall cumtime percall filename:lineno(function)
5 0.023 0.005 2.084 0.417 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_registration.py:115(from_torchlib)
5 0.024 0.005 1.272 0.254 /home/xadupre/github/onnxscript/onnxscript/_framework_apis/torch_2_5.py:99(get_torchlib_ops)
920 0.007 0.000 1.242 0.001 /home/xadupre/github/onnxscript/onnxscript/values.py:640(function_ir)
5 0.008 0.002 0.893 0.179 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_decomp.py:42(create_onnx_friendly_decomposition_table)
5 0.000 0.000 0.846 0.169 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/decomp_utils.py:125(items)
5 0.000 0.000 0.846 0.169 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/decomp_utils.py:142(_materialize_if_needed)
5 0.001 0.000 0.846 0.169 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/decomp_utils.py:129(materialize)
10 0.001 0.000 0.840 0.084 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1088(_collect_all_valid_cia_ops)
250 0.009 0.000 0.839 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1071(_collect_all_valid_cia_ops_for_namespace)
250 0.291 0.001 0.763 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1006(_materialize_cpp_cia_ops)
2900 0.054 0.000 0.699 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_schemas.py:431(from_function)
121745/120665 0.030 0.000 0.528 0.000 {built-in method builtins.next}
920 0.005 0.000 0.490 0.001 /home/xadupre/github/onnxscript/onnxscript/_internal/ast_utils.py:16(get_src_and_ast)
5 0.005 0.001 0.484 0.097 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:302(_split_decomp_table_to_cia_and_python_decomp)
4980/4445 0.003 0.000 0.452 0.000 /usr/lib/python3.12/contextlib.py:132(__enter__)
35/5 0.001 0.000 0.447 0.089 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:2594(from_tensor)
100/5 0.001 0.000 0.447 0.089 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:331(from_real_tensor)
105/5 0.002 0.000 0.443 0.089 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:1794(__call__)
5780 0.440 0.000 0.440 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1113(_get_decomp_for_cia)
20 0.102 0.005 0.402 0.020 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:195(_override_composite_implicit_decomp)
920 0.002 0.000 0.389 0.000 /home/xadupre/github/onnxscript/onnxscript/converter.py:1463(translate_function_signature)
920 0.026 0.000 0.385 0.000 /home/xadupre/github/onnxscript/onnxscript/converter.py:1378(_translate_function_signature_common)
720/660 0.095 0.000 0.345 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:352(__torch_dispatch__)
920 0.001 0.000 0.343 0.000 /usr/lib/python3.12/inspect.py:1279(getsource)
80/20 0.000 0.000 0.342 0.017 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/functional.py:1693(relu)
920 0.024 0.000 0.341 0.000 /usr/lib/python3.12/inspect.py:1606(getclosurevars)
920 0.028 0.000 0.340 0.000 /usr/lib/python3.12/inspect.py:1258(getsourcelines)
3250 0.007 0.000 0.309 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1231(__torch_function__)
41240 0.118 0.000 0.298 0.000 /usr/lib/python3.12/dis.py:434(_get_instructions_bytes)
18620 0.266 0.000 0.266 0.000 {built-in method builtins.compile}
920 0.075 0.000 0.264 0.000 /usr/lib/python3.12/inspect.py:1239(getblock)
63195 0.199 0.000 0.247 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_ops.py:106(inner)
2900 0.036 0.000 0.244 0.000 /usr/lib/python3.12/typing.py:2215(get_type_hints)
785285 0.239 0.000 0.242 0.000 {built-in method builtins.getattr}
28010/5170 0.074 0.000 0.231 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_schemas.py:267(_get_allowed_types_from_type_annotation)
37845/9080 0.046 0.000 0.228 0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:131(is_value_type)
685 0.003 0.000 0.219 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1329(__torch_dispatch__)
1915 0.013 0.000 0.200 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1788(dispatch)
135 0.006 0.000 0.199 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:762(proxy_call)
1051930/1046135 0.162 0.000 0.197 0.000 {built-in method builtins.isinstance}
35/5 0.000 0.000 0.186 0.037 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/module.py:1743(_call_impl)
5 0.000 0.000 0.184 0.037 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/_trace.py:1672(forward)
525 0.004 0.000 0.181 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1348(_cached_dispatch_impl)
85 0.001 0.000 0.180 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:792(recompile)
5 0.000 0.000 0.177 0.035 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_export_201.py:191(forward)
133950 0.089 0.000 0.166 0.000 /usr/lib/python3.12/tokenize.py:569(_generate_tokens_from_c_tokenizer)
90 0.001 0.000 0.159 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1569(python_code)
5110 0.003 0.000 0.148 0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:172(is_valid_type)
1215 0.003 0.000 0.134 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1260(__torch_function__)
1625 0.003 0.000 0.131 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/non_strict_utils.py:563(__torch_function__)
90 0.001 0.000 0.128 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1646(_python_code)
90 0.014 0.000 0.127 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:408(_gen_python_code)
250 0.124 0.000 0.124 0.000 {built-in method torch._C._dispatch_get_registrations_for_dispatch_key}
60 0.002 0.000 0.123 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_ops.py:829(handler)
60 0.008 0.000 0.119 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_library/utils.py:280(handle_dispatch_mode)
920 0.001 0.000 0.118 0.000 /usr/lib/python3.12/ast.py:34(parse)
2940 0.002 0.000 0.114 0.000 /usr/lib/python3.12/inspect.py:3343(signature)
2940 0.002 0.000 0.112 0.000 /usr/lib/python3.12/inspect.py:3081(from_callable)
2970/2940 0.015 0.000 0.110 0.000 /usr/lib/python3.12/inspect.py:2501(_signature_from_callable)
1425 0.003 0.000 0.109 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1152(tree_map_only)
40/10 0.000 0.000 0.109 0.011 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_jit_internal.py:614(fn)
7010 0.004 0.000 0.106 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:992(_is_preservable_cia_op)
460380 0.105 0.000 0.105 0.000 {method 'split' of 'str' objects}
51615/51545 0.021 0.000 0.102 0.000 {built-in method builtins.repr}
37845 0.024 0.000 0.093 0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:123(_is_tensor_type)
23960/10730 0.016 0.000 0.093 0.000 /usr/lib/python3.12/typing.py:407(_eval_type)
7010 0.046 0.000 0.087 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1040(_check_valid_to_preserve)
34100 0.015 0.000 0.087 0.000 /home/xadupre/github/onnxscript/onnxscript/ir/_core.py:1417(__hash__)
10730 0.018 0.000 0.086 0.000 /usr/lib/python3.12/typing.py:916(_evaluate)
10730 0.011 0.000 0.085 0.000 /usr/lib/python3.12/typing.py:892(__init__)
3970 0.002 0.000 0.084 0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:168(is_attr_type)
6885 0.013 0.000 0.084 0.000 /home/xadupre/github/onnxscript/onnxscript/converter.py:451(_eval_constant_expr)
20/5 0.000 0.000 0.082 0.016 {built-in method torch.flatten}
265/175 0.001 0.000 0.082 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_ops.py:756(decompose)
5 0.000 0.000 0.081 0.016 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_fx_passes.py:23(insert_type_promotion_nodes)
170 0.002 0.000 0.081 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:836(meta_tensor)
40 0.000 0.000 0.080 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1124(_special_op_to_decompose_cia)
10 0.001 0.000 0.079 0.008 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/_trace.py:412(_produce_aten_artifact)
2940 0.027 0.000 0.078 0.000 /usr/lib/python3.12/inspect.py:2397(_signature_from_function)
133030 0.043 0.000 0.078 0.000 /usr/lib/python3.12/collections/__init__.py:447(_make)
140/5 0.002 0.000 0.076 0.015 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/diagnostics/infra/decorator.py:66(wrapper)
5 0.000 0.000 0.076 0.015 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:1178(module)
5 0.000 0.000 0.076 0.015 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/fx/_pass.py:240(run)
5 0.000 0.000 0.076 0.015 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/_unlift.py:369(_unlift_exported_program_lifted_states)
5 0.000 0.000 0.076 0.015 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/fx/passes/type_promotion.py:1696(_run)
30 0.001 0.000 0.074 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:437(__init__)
2910 0.008 0.000 0.073 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_registration.py:54(_get_overload)
625/525 0.002 0.000 0.072 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/module.py:1932(__setattr__)
525 0.002 0.000 0.072 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1391(_cache_key)
245 0.001 0.000 0.069 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/proxy.py:209(create_proxy)
769505/769270 0.068 0.000 0.068 0.000 {built-in method builtins.len}
135 0.001 0.000 0.068 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/fx/passes/type_promotion.py:1601(run_node)
70905 0.039 0.000 0.067 0.000 /usr/lib/python3.12/typing.py:2340(get_origin)
82480 0.057 0.000 0.067 0.000 /usr/lib/python3.12/dis.py:623(_unpack_opargs)
30 0.000 0.000 0.067 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:548(graph)
1965/525 0.010 0.000 0.066 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1467(_prep_args_for_hash)
40225 0.020 0.000 0.066 0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:70(_remove_annotation)
450 0.002 0.000 0.065 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1722(_output_from_cache_entry)
170 0.004 0.000 0.064 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/autograd/grad_mode.py:273(__exit__)
470 0.007 0.000 0.063 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1656(_get_output_tensor_from_cache_entry)
145 0.001 0.000 0.063 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:594(track_tensor_tree)
90/30 0.000 0.000 0.061 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/overrides.py:1669(handle_torch_function)
5 0.001 0.000 0.060 0.012 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_core.py:929(_exported_program_to_onnx_program)
255/145 0.001 0.000 0.060 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:616(wrap_with_proxy)
1130 0.060 0.000 0.060 0.000 {built-in method posix.stat}
4980/4445 0.004 0.000 0.059 0.000 /usr/lib/python3.12/contextlib.py:141(__exit__)
255 0.003 0.000 0.058 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1777(create_node)
2280 0.008 0.000 0.057 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:634(emit_node)
5 0.013 0.003 0.057 0.011 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_decomp.py:15(get_onnx_implemented_overloads)
37125 0.026 0.000 0.057 0.000 /home/xadupre/github/onnxscript/onnxscript/ir/_core.py:1425(__repr__)
920 0.018 0.000 0.056 0.000 /usr/lib/python3.12/dis.py:647(findlabels)
5 0.000 0.000 0.055 0.011 <frozen importlib.util>:70(find_spec)
5 0.000 0.000 0.055 0.011 <frozen importlib._bootstrap>:1240(_find_spec)
5 0.000 0.000 0.055 0.011 <frozen importlib._bootstrap_external>:1524(find_spec)
5 0.000 0.000 0.055 0.011 <frozen importlib._bootstrap_external>:1495(_get_spec)
110 0.001 0.000 0.055 0.000 <frozen importlib._bootstrap_external>:1597(find_spec)
2065 0.003 0.000 0.054 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:874(tree_flatten)
130 0.000 0.000 0.053 0.000 <frozen importlib._bootstrap_external>:140(_path_stat)
5 0.000 0.000 0.053 0.011 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_core.py:658(_translate_fx_graph)
75 0.001 0.000 0.051 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_core.py:448(_handle_call_function_node_with_lowering)
7175/2065 0.013 0.000 0.050 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:882(helper)
17635 0.035 0.000 0.044 0.000 {built-in method builtins.eval}
920 0.007 0.000 0.044 0.000 /usr/lib/python3.12/inspect.py:1070(findsource)
166531 0.033 0.000 0.044 0.000 {built-in method builtins.hasattr}
245 0.001 0.000 0.043 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:487(set_meta)
1635 0.001 0.000 0.042 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1101(wrapped)
95064 0.025 0.000 0.041 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_ops.py:729(__hash__)
8460 0.015 0.000 0.041 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:869(__setattr__)
170 0.006 0.000 0.040 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/autograd/grad_mode.py:269(__enter__)
75 0.001 0.000 0.039 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1974(_dispatch_impl)
20 0.000 0.000 0.038 0.002 {built-in method torch.relu}
8795 0.006 0.000 0.038 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:890(map_arg)
216030 0.038 0.000 0.038 0.000 {built-in method __new__ of type object at 0xa20960}
16445/10005 0.019 0.000 0.036 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:899(map_aggregate)
103800 0.023 0.000 0.035 0.000 /usr/lib/python3.12/inspect.py:302(isclass)
885 0.010 0.000 0.034 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:678(__new__)
85/80 0.001 0.000 0.032 0.000 /home/xadupre/github/onnxscript/onnxscript/values.py:295(__call__)
85/80 0.000 0.000 0.032 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_building.py:558(eval)
430 0.002 0.000 0.032 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1111(create_node)
85 0.000 0.000 0.031 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_traceback.py:171(summary)
152489/149479 0.027 0.000 0.031 0.000 {built-in method builtins.hash}
15015 0.005 0.000 0.031 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_registration.py:261(is_registered)
35 0.001 0.000 0.031 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_logging/_internal.py:1156(trace_structured)
15 0.000 0.000 0.031 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:124(forward)
60/15 0.000 0.000 0.031 0.002 {built-in method torch._C._nn.linear}
85 0.000 0.000 0.030 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:91(_forward_from_src)
85 0.000 0.000 0.030 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:97(_method_from_src)
650 0.009 0.000 0.030 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:954(_flatten_into)
85 0.000 0.000 0.030 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:86(_exec_with_source)
10730 0.018 0.000 0.029 0.000 /usr/lib/python3.12/typing.py:175(_type_check)
170 0.006 0.000 0.029 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:260(describe_tensor)
55 0.000 0.000 0.028 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_functorch/_aot_autograd/dispatch_and_compile_graph.py:66(_detach_and_copy_item_memo)
10 0.000 0.000 0.028 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:797(__init__)
105 0.002 0.000 0.028 0.000 {method 'detach' of 'torch._C.TensorBase' objects}
37125 0.011 0.000 0.028 0.000 /home/xadupre/github/onnxscript/onnxscript/ir/_enums.py:95(__repr__)
10 0.000 0.000 0.027 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:538(apply_runtime_assertion_pass)
180 0.003 0.000 0.027 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1623(override_node_repr)
245 0.000 0.000 0.027 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:391(extract_val)
5 0.000 0.000 0.027 0.005 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/_unlift.py:172(_unlift)
255 0.002 0.000 0.027 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/proxy.py:143(create_node)
15090 0.014 0.000 0.026 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_registration.py:239(get_decomps)
5 0.000 0.000 0.026 0.005 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:346(default_decompositions)
5 0.001 0.000 0.026 0.005 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/decomp_utils.py:33(__init__)
34380/34110 0.012 0.000 0.026 0.000 {built-in method builtins.issubclass}
245 0.001 0.000 0.026 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:365(snapshot_fake)
50/40 0.000 0.000 0.026 0.001 /home/xadupre/github/onnxscript/onnxscript/values.py:634(__call__)
3615/160 0.006 0.000 0.026 0.000 /usr/lib/python3.12/copy.py:118(deepcopy)
10 0.001 0.000 0.026 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:774(placeholder_naming_pass)
10 0.000 0.000 0.026 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/conv.py:553(forward)
10 0.000 0.000 0.026 0.003 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/modules/conv.py:536(_conv_forward)
40/10 0.000 0.000 0.026 0.003 {built-in method torch.conv2d}
650 0.008 0.000 0.025 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:974(extract_tensor_metadata)
245 0.003 0.000 0.025 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_impls.py:991(fast_detach)
5515 0.004 0.000 0.025 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:177(is_sparse_any)
98754 0.023 0.000 0.025 0.000 {method 'get' of 'dict' objects}
75/15 0.000 0.000 0.025 0.002 {built-in method torch._to_functional_tensor}
1305/820 0.010 0.000 0.024 0.000 {built-in method torch._ops.prim.}
920 0.004 0.000 0.024 0.000 /usr/lib/python3.12/textwrap.py:419(dedent)
44590 0.011 0.000 0.024 0.000 <frozen abc>:117(__instancecheck__)
2575/16 0.004 0.000 0.023 0.001 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_stats.py:15(wrapper)
6650/6405 0.006 0.000 0.023 0.000 {method 'join' of 'str' objects}
133030 0.023 0.000 0.023 0.000 /usr/lib/python3.12/inspect.py:1196(tokeneater)
7540 0.011 0.000 0.022 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:154(create_name)
10 0.000 0.000 0.022 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/_lazy_graph_module.py:57(_make_graph_module)
40/10 0.000 0.000 0.022 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/nn/functional.py:807(_max_pool2d)
7190 0.019 0.000 0.021 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_decomp/__init__.py:56(_should_decompose_because_unsafe_op)
8015 0.012 0.000 0.021 0.000 /usr/lib/python3.12/inspect.py:2743(__init__)
13230/10730 0.013 0.000 0.021 0.000 /usr/lib/python3.12/typing.py:2315(_strip_annotations)
10 0.000 0.000 0.021 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:1449(_create_graph_module_for_export)
10 0.000 0.000 0.021 0.002 {built-in method torch.max_pool2d}
1200 0.002 0.000 0.020 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:555(_format_args)
1020 0.001 0.000 0.020 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/fx/interpreter.py:205(_set_current_node)
85 0.004 0.000 0.020 0.000 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/utils/_traceback.py:247(_extract_symbolized_tb)
10 0.000 0.000 0.019 0.002 /home/xadupre/vv/this312/lib/python3.12/site-packages/torch/_export/passes/replace_with_hop_pass_util.py:157(_replace_with_hop_pass_helper)
159255 0.019 0.000 0.019 0.000 /usr/lib/python3.12/dis.py:195(_deoptop)
done.
profile dynopt: <function export_dynopt at 0x7f62d447dd00>
done.
Benchmark exported models with ORT¶
def benchmark(shape):
from onnxruntime import InferenceSession, SessionOptions, GraphOptimizationLevel
data = []
data1 = []
data_mem_load = []
data_mem_first_run = []
data_mem_run = []
confs = list(
itertools.product(
[_ for _ in os.listdir(".") if ".onnx" in _ and _.startswith("plot_torch")],
[
["CPUExecutionProvider"],
["CUDAExecutionProvider", "CPUExecutionProvider"],
],
["0", "1"],
)
)
loop = tqdm(confs)
print(f"number of experiments: {len(loop)}")
for name, ps, aot in loop:
root = os.path.split(name)[-1]
_, ext = os.path.splitext(root)
if ext != ".onnx":
continue
obs = {} # system_info()
obs["name"] = name
obs["providers"] = ",".join(ps)
p = "CUDA" if "CUDA" in obs["providers"] else "CPU"
obs["compute"] = p
obs["aot"] = 1 if aot == "0" else 0
obs["export"] = name.replace("plot_torch_export_", "").replace(".onnx", "")
if not has_cuda and p == "CUDA":
continue
onx = onnx.load(name)
obs["n_nodes"] = len(onx.graph.node)
obs["n_function"] = len(onx.functions or [])
obs["n_sub"] = len([n for n in onx.graph.node if n.op_type == "Sub"])
obs1 = obs.copy()
short_obs = dict(
name=obs["name"],
aot=obs["aot"],
providers=obs["providers"],
export=obs["export"],
compute=obs["compute"],
)
opts = SessionOptions()
opts.add_session_config_entry("session.disable_aot_function_inlining", aot)
opts.graph_optimization_level = GraphOptimizationLevel.ORT_ENABLE_ALL
opts.optimized_model_filepath = (
f"ort-{name.replace('.onnx', '')}-{p.lower()}-aot{1 if aot == '0' else 0}.onnx"
)
try:
InferenceSession(name, opts, providers=ps)
except Exception as e:
loop.set_description(f"ERROR-load: {name} {e}")
obs.update({"error": e, "step": "run"})
data.append(obs)
continue
opts = SessionOptions()
opts.add_session_config_entry("session.disable_aot_function_inlining", aot)
opts.graph_optimization_level = GraphOptimizationLevel.ORT_ENABLE_ALL
stat = start_spying_on(cuda=1 if has_cuda else 0)
sess = InferenceSession(name, opts, providers=ps)
memobs = flatten(stat.stop())
memobs.update(short_obs)
data_mem_load.append(memobs)
input_name = sess.get_inputs()[0].name
feeds = {input_name: np.random.rand(*shape).astype(np.float32)}
stat = start_spying_on(cuda=1 if has_cuda else 0)
try:
sess.run(None, feeds)
except Exception as e:
loop.set_description(f"ERROR-run: {name} {e}")
obs.update({"error": e, "step": "load"})
data.append(obs)
stat.stop()
continue
memobs = flatten(stat.stop())
memobs.update(short_obs)
data_mem_first_run.append(memobs)
# memory consumption
stat = start_spying_on(cuda=1 if has_cuda else 0)
for _ in range(0, script_args.warmup):
sess.run(None, feeds)
memobs = flatten(stat.stop())
memobs.update(short_obs)
data_mem_run.append(memobs)
obs.update(
measure_time(
lambda sess=sess, feeds=feeds: sess.run(None, feeds),
max_time=script_args.maxtime,
repeat=script_args.repeat,
number=1,
)
)
loop.set_description(f"{obs['average']} {name} {ps}")
data.append(obs)
# check first run
obs1.update(
measure_time(
lambda name=name, opts=opts, ps=ps, feeds=feeds: InferenceSession(
name, opts, providers=ps
).run(None, feeds),
max_time=script_args.maxtime,
repeat=max(1, script_args.repeat // 2),
number=1,
)
)
data1.append(obs1)
df = pandas.DataFrame(data)
df.to_csv("plot_torch_export_ort_time.csv", index=False)
df.to_excel("plot_torch_export_ort_time.xlsx", index=False)
df1 = pandas.DataFrame(data1)
df1.to_csv("plot_torch_export_ort_time1_init.csv", index=False)
df1.to_excel("plot_torch_export_ort_time1_init.xlsx", index=False)
dfmem = pandas.DataFrame(data_mem_load)
dfmem.to_csv("plot_torch_export_ort_load_mem.csv", index=False)
dfmem.to_excel("plot_torch_export_ort_load_mem.xlsx", index=False)
dfmemr = pandas.DataFrame(data_mem_run)
dfmemr.to_csv("plot_torch_export_ort_run_mem.csv", index=False)
dfmemr.to_excel("plot_torch_export_ort_run_mem.xlsx", index=False)
dfmemfr = pandas.DataFrame(data_mem_first_run)
dfmemfr.to_csv("plot_torch_export_ort_first_run_mem.csv", index=False)
dfmemfr.to_excel("plot_torch_export_ort_first_run_mem.xlsx", index=False)
return df, df1, dfmem, dfmemfr, dfmemr
df, df_init, dfmem, dfmemfr, dfmemr = benchmark(list(input_tensor.shape))
print(df)
0%| | 0/20 [00:00<?, ?it/s]number of experiments: 20
0.00021816104420614694 plot_torch_export_cus_p2.onnx ['CPUExecutionProvider']: 0%| | 0/20 [00:00<?, ?it/s]
0.00021816104420614694 plot_torch_export_cus_p2.onnx ['CPUExecutionProvider']: 5%|▌ | 1/20 [00:00<00:17, 1.07it/s]
4.772552724445772e-05 plot_torch_export_cus_p2.onnx ['CPUExecutionProvider']: 5%|▌ | 1/20 [00:01<00:17, 1.07it/s]
4.772552724445772e-05 plot_torch_export_cus_p2.onnx ['CPUExecutionProvider']: 10%|█ | 2/20 [00:01<00:16, 1.08it/s]
0.0006552248373983217 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 10%|█ | 2/20 [00:02<00:16, 1.08it/s]
0.0006552248373983217 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 15%|█▌ | 3/20 [00:02<00:16, 1.05it/s]
0.0006638402703643554 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 15%|█▌ | 3/20 [00:03<00:16, 1.05it/s]
0.0006638402703643554 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 20%|██ | 4/20 [00:03<00:15, 1.05it/s]
0.00013941977999395433 plot_torch_export_dynopt.onnx ['CPUExecutionProvider']: 20%|██ | 4/20 [00:04<00:15, 1.05it/s]
0.00013941977999395433 plot_torch_export_dynopt.onnx ['CPUExecutionProvider']: 25%|██▌ | 5/20 [00:04<00:14, 1.04it/s]
0.00021697416701787838 plot_torch_export_dynopt.onnx ['CPUExecutionProvider']: 25%|██▌ | 5/20 [00:05<00:14, 1.04it/s]
0.00021697416701787838 plot_torch_export_dynopt.onnx ['CPUExecutionProvider']: 30%|███ | 6/20 [00:05<00:13, 1.07it/s]
0.0007045571522676195 plot_torch_export_dynopt.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 30%|███ | 6/20 [00:06<00:13, 1.07it/s]
0.0007045571522676195 plot_torch_export_dynopt.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 35%|███▌ | 7/20 [00:06<00:13, 1.07s/it]
0.0006713529356703577 plot_torch_export_dynopt.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 35%|███▌ | 7/20 [00:07<00:13, 1.07s/it]
0.0006713529356703577 plot_torch_export_dynopt.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 40%|████ | 8/20 [00:07<00:12, 1.03s/it]
0.00011099798357383748 plot_torch_export_dynamo.onnx ['CPUExecutionProvider']: 40%|████ | 8/20 [00:08<00:12, 1.03s/it]
0.00011099798357383748 plot_torch_export_dynamo.onnx ['CPUExecutionProvider']: 45%|████▌ | 9/20 [00:09<00:11, 1.04s/it]
0.00010369685928426969 plot_torch_export_dynamo.onnx ['CPUExecutionProvider']: 45%|████▌ | 9/20 [00:09<00:11, 1.04s/it]
0.00010369685928426969 plot_torch_export_dynamo.onnx ['CPUExecutionProvider']: 50%|█████ | 10/20 [00:10<00:10, 1.04s/it]
0.00066768406872932 plot_torch_export_dynamo.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 50%|█████ | 10/20 [00:10<00:10, 1.04s/it]
0.00066768406872932 plot_torch_export_dynamo.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 55%|█████▌ | 11/20 [00:10<00:08, 1.03it/s]
0.0005236236300087456 plot_torch_export_dynamo.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 55%|█████▌ | 11/20 [00:11<00:08, 1.03it/s]
0.0005236236300087456 plot_torch_export_dynamo.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 60%|██████ | 12/20 [00:11<00:07, 1.03it/s]
7.582991399394285e-05 plot_torch_export_script.onnx ['CPUExecutionProvider']: 60%|██████ | 12/20 [00:12<00:07, 1.03it/s]
7.582991399394285e-05 plot_torch_export_script.onnx ['CPUExecutionProvider']: 65%|██████▌ | 13/20 [00:12<00:06, 1.03it/s]
0.00011427159782254856 plot_torch_export_script.onnx ['CPUExecutionProvider']: 65%|██████▌ | 13/20 [00:13<00:06, 1.03it/s]
0.00011427159782254856 plot_torch_export_script.onnx ['CPUExecutionProvider']: 70%|███████ | 14/20 [00:13<00:05, 1.04it/s]
0.0007412413836866119 plot_torch_export_script.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 70%|███████ | 14/20 [00:14<00:05, 1.04it/s]
0.0007412413836866119 plot_torch_export_script.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 75%|███████▌ | 15/20 [00:14<00:04, 1.06it/s]
0.0006610327643687905 plot_torch_export_script.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 75%|███████▌ | 15/20 [00:15<00:04, 1.06it/s]
0.0006610327643687905 plot_torch_export_script.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 80%|████████ | 16/20 [00:15<00:03, 1.08it/s]
5.10916548248871e-05 plot_torch_export_cus_p0.onnx ['CPUExecutionProvider']: 80%|████████ | 16/20 [00:16<00:03, 1.08it/s]
5.10916548248871e-05 plot_torch_export_cus_p0.onnx ['CPUExecutionProvider']: 85%|████████▌ | 17/20 [00:16<00:02, 1.04it/s]
8.172642746880395e-05 plot_torch_export_cus_p0.onnx ['CPUExecutionProvider']: 85%|████████▌ | 17/20 [00:17<00:02, 1.04it/s]
8.172642746880395e-05 plot_torch_export_cus_p0.onnx ['CPUExecutionProvider']: 90%|█████████ | 18/20 [00:17<00:01, 1.04it/s]
0.0006202865654894599 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 90%|█████████ | 18/20 [00:18<00:01, 1.04it/s]
0.0006202865654894599 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 95%|█████████▌| 19/20 [00:18<00:00, 1.03it/s]
0.0006668382455158455 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 95%|█████████▌| 19/20 [00:19<00:00, 1.03it/s]
0.0006668382455158455 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 100%|██████████| 20/20 [00:19<00:00, 1.06it/s]
0.0006668382455158455 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 100%|██████████| 20/20 [00:19<00:00, 1.03it/s]
name providers compute ... ttime context_size warmup_time
0 plot_torch_export_cus_p2.onnx CPUExecutionProvider CPU ... 0.118461 64 0.001059
1 plot_torch_export_cus_p2.onnx CPUExecutionProvider CPU ... 0.112966 64 0.000480
2 plot_torch_export_cus_p2.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.133011 64 0.001523
3 plot_torch_export_cus_p2.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.105551 64 0.001979
4 plot_torch_export_dynopt.onnx CPUExecutionProvider CPU ... 0.126733 64 0.000455
5 plot_torch_export_dynopt.onnx CPUExecutionProvider CPU ... 0.101327 64 0.000965
6 plot_torch_export_dynopt.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.106388 64 0.001838
7 plot_torch_export_dynopt.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.114801 64 0.001775
8 plot_torch_export_dynamo.onnx CPUExecutionProvider CPU ... 0.101341 64 0.000903
9 plot_torch_export_dynamo.onnx CPUExecutionProvider CPU ... 0.103904 64 0.000556
10 plot_torch_export_dynamo.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.126192 64 0.001695
11 plot_torch_export_dynamo.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.142949 64 0.001249
12 plot_torch_export_script.onnx CPUExecutionProvider CPU ... 0.125195 64 0.000309
13 plot_torch_export_script.onnx CPUExecutionProvider CPU ... 0.105130 64 0.000703
14 plot_torch_export_script.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.117857 64 0.002100
15 plot_torch_export_script.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.126257 64 0.001982
16 plot_torch_export_cus_p0.onnx CPUExecutionProvider CPU ... 0.102132 64 0.000341
17 plot_torch_export_cus_p0.onnx CPUExecutionProvider CPU ... 0.106490 64 0.001241
18 plot_torch_export_cus_p0.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.118475 64 0.001771
19 plot_torch_export_cus_p0.onnx CUDAExecutionProvider,CPUExecutionProvider CUDA ... 0.111362 64 0.001935
[20 rows x 17 columns]
Other view
def view_time(df, title, suffix="time"):
piv = pandas.pivot_table(df, index="export", columns=["compute", "aot"], values="average")
print(piv)
piv.to_csv(f"plot_torch_export_ort_{suffix}_compute.csv")
piv.to_excel(f"plot_torch_export_ort_{suffix}_compute.xlsx")
piv_cpu = pandas.pivot_table(
df[df.compute == "CPU"],
index="export",
columns=["compute", "aot"],
values="average",
)
fig, ax = plt.subplots(1, 2, figsize=(12, 4))
fig.suptitle(title)
piv_cpu.plot.barh(ax=ax[0], title="CPU")
if has_cuda:
piv_gpu = pandas.pivot_table(
df[df.compute == "CUDA"],
index="export",
columns=["compute", "aot"],
values="average",
)
piv_gpu.plot.barh(ax=ax[1], title="CUDA")
fig.tight_layout()
fig.savefig(f"plot_torch_export_ort_{suffix}.png")
return ax
view_time(df, "Compares onnxruntime time on exported models")
compute CPU CUDA
aot 0 1 0 1
export
cus_p0 0.000082 0.000051 0.000667 0.000620
cus_p2 0.000048 0.000218 0.000664 0.000655
dynamo 0.000104 0.000111 0.000524 0.000668
dynopt 0.000217 0.000139 0.000671 0.000705
script 0.000114 0.000076 0.000661 0.000741
array([<Axes: title={'center': 'CPU'}, ylabel='export'>,
<Axes: title={'center': 'CUDA'}, ylabel='export'>], dtype=object)
New graph without the very long times.
piv_cpu = pandas.pivot_table(
df[
(df.compute == "CPU")
& ((df.aot == 1) | ((df.export != "dynamo") & (df.export != "dynopt")))
],
index="export",
columns=["compute", "aot"],
values="average",
)
fig, ax = plt.subplots(1, 2, figsize=(12, 4))
fig.suptitle("Compares onnxruntime time on exported models\nHide dynamo without AOT")
piv_cpu.plot.barh(ax=ax[0], title="CPU")
if has_cuda:
piv_gpu = pandas.pivot_table(
df[df.compute == "CUDA"],
index="export",
columns=["compute", "aot"],
values="average",
)
piv_gpu.plot.barh(ax=ax[1], title="CUDA")
fig.tight_layout()
fig.savefig("plot_torch_export_ort_time_2.png")
Let’s do the same with the loading time + the first run.
view_time(
df_init,
"Compares onnxruntime loading time and first run on exported models",
suffix="time1_init",
)
compute CPU CUDA
aot 0 1 0 1
export
cus_p0 0.005571 0.004664 0.035513 0.031205
cus_p2 0.005076 0.003848 0.023408 0.032264
dynamo 0.004979 0.006685 0.027405 0.021185
dynopt 0.004020 0.004537 0.033564 0.029715
script 0.006355 0.005371 0.026566 0.030412
array([<Axes: title={'center': 'CPU'}, ylabel='export'>,
<Axes: title={'center': 'CUDA'}, ylabel='export'>], dtype=object)
Memory Loading Time (ORT)¶
for compute in ["CPU", "CUDA"]:
if not has_cuda and compute == "CUDA":
continue
ax = memory_peak_plot(
dfmem[dfmem.compute == compute],
("export", "aot"),
suptitle=f"Memory Consumption of onnxruntime loading time\nrunning on {compute}",
bars=[model_size * i / 2**20 for i in range(1, 3)],
figsize=(18, 6),
)
get_figure(ax).savefig(f"plot_torch_export_ort_load_mem_{compute}.png")
Memory First Running Time (ORT)¶
for compute in ["CPU", "CUDA"]:
if not has_cuda and compute == "CUDA":
continue
ax = memory_peak_plot(
dfmemfr[dfmemfr.compute == compute],
("export", "aot"),
suptitle=f"Memory Consumption of onnxruntime first running time"
f"\nrunning on {compute}",
bars=[model_size * i / 2**20 for i in range(1, 3)],
figsize=(18, 6),
)
get_figure(ax).savefig(f"plot_torch_export_ort_first_run_mem_{compute}.png")
Memory Running Time (ORT)¶
for compute in ["CPU", "CUDA"]:
if not has_cuda and compute == "CUDA":
continue
ax = memory_peak_plot(
dfmemr[dfmemr.compute == compute],
("export", "aot"),
suptitle=f"Memory Consumption of onnxruntime running time\nrunning on {compute}",
bars=[model_size * i / 2**20 for i in range(1, 3)],
figsize=(18, 6),
)
get_figure(ax).savefig(f"plot_torch_export_ort_run_mem_{compute}.png")
Show the interesting models for CPU¶
script¶
model = "ort-plot_torch_export_cus_p2-cpu-aot0.onnx"
if os.path.exists(model):
print(pretty_onnx(onnx.load(model)))
opset: domain='' version=18
opset: domain='ai.onnx.ml' version=5
opset: domain='ai.onnx.training' version=1
opset: domain='ai.onnx.preview.training' version=1
opset: domain='com.microsoft' version=1
opset: domain='com.microsoft.experimental' version=1
opset: domain='com.microsoft.nchwc' version=1
opset: domain='org.pytorch.aten' version=1
doc_string: large_model=False, inline=False, external_threshold=102...
input: name='input' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='_onx_concat0' type=int64 shape=(2,) -- array([ 1, -1]) -- GraphBuilder.constant_folding.from/fold(_onx_gatherelements0,init7_s1_-1)##_onx_gatherelements0/GraphBuilder.constant_folding.from/fold(_onx_shape0,init7_s1_0)##_onx_shape0/##init7_s1_0/Opset.make_node.1/Shape##init7_s1_-1/Opset.make_node.1/Shape
init: name='GemmTransposePattern--_onx_transpose0' type=float32 shape=(512, 16)-- GraphBuilder.constant_folding.from/fold(_onx_transpose0)##_onx_transpose0/GraphBuilder.constant_folding.from/fold(p_fc1_weight)##p_fc1_weight/DynamoInterpret.placeholder.1/P(fc1.weight)
init: name='GemmTransposePattern--_onx_transpose02' type=float32 shape=(128, 512)-- GraphBuilder.constant_folding.from/fold(_onx_transpose02)##_onx_transpose02/GraphBuilder.constant_folding.from/fold(p_fc2_weight)##p_fc2_weight/DynamoInterpret.placeholder.1/P(fc2.weight)
init: name='GemmTransposePattern--_onx_transpose03' type=float32 shape=(10, 128)-- GraphBuilder.constant_folding.from/fold(_onx_transpose03)##_onx_transpose03/GraphBuilder.constant_folding.from/fold(p_fc3_weight)##p_fc3_weight/DynamoInterpret.placeholder.1/P(fc3.weight)
init: name='reorder' type=float32 shape=(16, 1, 5, 5)
init: name='conv1.bias' type=float32 shape=(16,) -- DynamoInterpret.placeholder.1/P(conv1.bias)
init: name='reorder_token_1' type=float32 shape=(16, 16, 5, 5)
init: name='conv2.bias' type=float32 shape=(16,) -- DynamoInterpret.placeholder.1/P(conv2.bias)
init: name='fc1.bias' type=float32 shape=(512,) -- DynamoInterpret.placeholder.1/P(fc1.bias)
init: name='fc2.bias' type=float32 shape=(128,) -- DynamoInterpret.placeholder.1/P(fc2.bias)
init: name='fc3.bias' type=float32 shape=(10,) -- DynamoInterpret.placeholder.1/P(fc3.bias)
Conv[com.microsoft.nchwc](input, reorder, conv1.bias, activation=b'Relu', dilations=[1,1], group=1, strides=[1,1], pads=[0,0,0,0], auto_pad=b'NOTSET') -> reorder_token_0
ReorderOutput[com.microsoft.nchwc](reorder_token_0, channels_last=0, channels=16) -> relu
MaxPool(relu, storage_order=0, auto_pad=b'NOTSET', ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> _onx_maxpool0, _onx_maxpool1
ReorderInput[com.microsoft.nchwc](_onx_maxpool0, channels_last=0) -> reorder_token_2
Conv[com.microsoft.nchwc](reorder_token_2, reorder_token_1, conv2.bias, activation=b'Relu', dilations=[1,1], group=1, strides=[1,1], pads=[0,0,0,0], auto_pad=b'NOTSET') -> reorder_token_3
ReorderOutput[com.microsoft.nchwc](reorder_token_3, channels_last=0, channels=16) -> relu_1
MaxPool(relu_1, storage_order=0, auto_pad=b'NOTSET', ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> _onx_maxpool02, _onx_maxpool12
Reshape(_onx_maxpool02, _onx_concat0, allowzero=0) -> flatten
FusedGemm[com.microsoft](flatten, GemmTransposePattern--_onx_transpose0, fc1.bias, transA=0, beta=1.00, activation=b'Relu', transB=1, alpha=1.00) -> relu_2
FusedGemm[com.microsoft](relu_2, GemmTransposePattern--_onx_transpose02, fc2.bias, transA=0, beta=1.00, activation=b'Relu', transB=1, alpha=1.00) -> relu_3
Gemm(relu_3, GemmTransposePattern--_onx_transpose03, fc3.bias, transA=0, beta=1.00, transB=1, alpha=1.00) -> output_0
output: name='output_0' type=dtype('float32') shape=[1, 10]
cus_p2¶
model = "ort-plot_torch_export_cus_p2-cpu-aot0.onnx"
if os.path.exists(model):
print(pretty_onnx(onnx.load(model)))
opset: domain='' version=18
opset: domain='ai.onnx.ml' version=5
opset: domain='ai.onnx.training' version=1
opset: domain='ai.onnx.preview.training' version=1
opset: domain='com.microsoft' version=1
opset: domain='com.microsoft.experimental' version=1
opset: domain='com.microsoft.nchwc' version=1
opset: domain='org.pytorch.aten' version=1
doc_string: large_model=False, inline=False, external_threshold=102...
input: name='input' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='_onx_concat0' type=int64 shape=(2,) -- array([ 1, -1]) -- GraphBuilder.constant_folding.from/fold(_onx_gatherelements0,init7_s1_-1)##_onx_gatherelements0/GraphBuilder.constant_folding.from/fold(_onx_shape0,init7_s1_0)##_onx_shape0/##init7_s1_0/Opset.make_node.1/Shape##init7_s1_-1/Opset.make_node.1/Shape
init: name='GemmTransposePattern--_onx_transpose0' type=float32 shape=(512, 16)-- GraphBuilder.constant_folding.from/fold(_onx_transpose0)##_onx_transpose0/GraphBuilder.constant_folding.from/fold(p_fc1_weight)##p_fc1_weight/DynamoInterpret.placeholder.1/P(fc1.weight)
init: name='GemmTransposePattern--_onx_transpose02' type=float32 shape=(128, 512)-- GraphBuilder.constant_folding.from/fold(_onx_transpose02)##_onx_transpose02/GraphBuilder.constant_folding.from/fold(p_fc2_weight)##p_fc2_weight/DynamoInterpret.placeholder.1/P(fc2.weight)
init: name='GemmTransposePattern--_onx_transpose03' type=float32 shape=(10, 128)-- GraphBuilder.constant_folding.from/fold(_onx_transpose03)##_onx_transpose03/GraphBuilder.constant_folding.from/fold(p_fc3_weight)##p_fc3_weight/DynamoInterpret.placeholder.1/P(fc3.weight)
init: name='reorder' type=float32 shape=(16, 1, 5, 5)
init: name='conv1.bias' type=float32 shape=(16,) -- DynamoInterpret.placeholder.1/P(conv1.bias)
init: name='reorder_token_1' type=float32 shape=(16, 16, 5, 5)
init: name='conv2.bias' type=float32 shape=(16,) -- DynamoInterpret.placeholder.1/P(conv2.bias)
init: name='fc1.bias' type=float32 shape=(512,) -- DynamoInterpret.placeholder.1/P(fc1.bias)
init: name='fc2.bias' type=float32 shape=(128,) -- DynamoInterpret.placeholder.1/P(fc2.bias)
init: name='fc3.bias' type=float32 shape=(10,) -- DynamoInterpret.placeholder.1/P(fc3.bias)
Conv[com.microsoft.nchwc](input, reorder, conv1.bias, activation=b'Relu', dilations=[1,1], group=1, strides=[1,1], pads=[0,0,0,0], auto_pad=b'NOTSET') -> reorder_token_0
ReorderOutput[com.microsoft.nchwc](reorder_token_0, channels_last=0, channels=16) -> relu
MaxPool(relu, storage_order=0, auto_pad=b'NOTSET', ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> _onx_maxpool0, _onx_maxpool1
ReorderInput[com.microsoft.nchwc](_onx_maxpool0, channels_last=0) -> reorder_token_2
Conv[com.microsoft.nchwc](reorder_token_2, reorder_token_1, conv2.bias, activation=b'Relu', dilations=[1,1], group=1, strides=[1,1], pads=[0,0,0,0], auto_pad=b'NOTSET') -> reorder_token_3
ReorderOutput[com.microsoft.nchwc](reorder_token_3, channels_last=0, channels=16) -> relu_1
MaxPool(relu_1, storage_order=0, auto_pad=b'NOTSET', ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> _onx_maxpool02, _onx_maxpool12
Reshape(_onx_maxpool02, _onx_concat0, allowzero=0) -> flatten
FusedGemm[com.microsoft](flatten, GemmTransposePattern--_onx_transpose0, fc1.bias, transA=0, beta=1.00, activation=b'Relu', transB=1, alpha=1.00) -> relu_2
FusedGemm[com.microsoft](relu_2, GemmTransposePattern--_onx_transpose02, fc2.bias, transA=0, beta=1.00, activation=b'Relu', transB=1, alpha=1.00) -> relu_3
Gemm(relu_3, GemmTransposePattern--_onx_transpose03, fc3.bias, transA=0, beta=1.00, transB=1, alpha=1.00) -> output_0
output: name='output_0' type=dtype('float32') shape=[1, 10]
dynopt¶
model = "ort-plot_torch_export_dynopt-cpu-aot1.onnx"
if os.path.exists(model):
print(pretty_onnx(onnx.load(model)))
opset: domain='pkg.onnxscript.torch_lib.common' version=1
opset: domain='' version=18
opset: domain='ai.onnx.ml' version=5
opset: domain='ai.onnx.training' version=1
opset: domain='ai.onnx.preview.training' version=1
opset: domain='com.microsoft' version=1
opset: domain='com.microsoft.experimental' version=1
opset: domain='com.microsoft.nchwc' version=1
opset: domain='org.pytorch.aten' version=1
input: name='x' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='reorder' type=float32 shape=(16, 1, 5, 5)
init: name='conv1.bias' type=float32 shape=(16,)
init: name='reorder_token_2' type=float32 shape=(16, 16, 5, 5)
init: name='conv2.bias' type=float32 shape=(16,)
init: name='fc1.bias' type=float32 shape=(512,)
init: name='fc2.bias' type=float32 shape=(128,)
init: name='fc3.bias' type=float32 shape=(10,)
init: name='val_3' type=int64 shape=(2,) -- array([ 1, 16])
init: name='t' type=float32 shape=(16, 512)
init: name='t_1' type=float32 shape=(512, 128)
init: name='t_2' type=float32 shape=(128, 10)
Conv[com.microsoft.nchwc](x, reorder, conv1.bias, activation=b'Relu', group=1, strides=[1,1], pads=[0,0,0,0], auto_pad=b'NOTSET', dilations=[1,1]) -> reorder_token_0
MaxPool[com.microsoft.nchwc](reorder_token_0, pads=[0,0,0,0], kernel_shape=[2,2], ceil_mode=0, auto_pad=b'NOTSET', dilations=[1,1], strides=[2,2], storage_order=0) -> reorder_token_1
Conv[com.microsoft.nchwc](reorder_token_1, reorder_token_2, conv2.bias, activation=b'Relu', group=1, strides=[1,1], pads=[0,0,0,0], auto_pad=b'NOTSET', dilations=[1,1]) -> reorder_token_3
MaxPool[com.microsoft.nchwc](reorder_token_3, pads=[0,0,0,0], kernel_shape=[2,2], ceil_mode=0, auto_pad=b'NOTSET', dilations=[1,1], strides=[2,2], storage_order=0) -> reorder_token_4
ReorderOutput[com.microsoft.nchwc](reorder_token_4, channels_last=0, channels=16) -> max_pool2d_1
Reshape(max_pool2d_1, val_3, allowzero=0) -> view
FusedGemm[com.microsoft](view, t, fc1.bias, transA=0, alpha=1.00, activation=b'Relu', transB=0, beta=1.00) -> relu_2
FusedGemm[com.microsoft](relu_2, t_1, fc2.bias, transA=0, alpha=1.00, activation=b'Relu', transB=0, beta=1.00) -> relu_3
Gemm(relu_3, t_2, fc3.bias, transA=0, alpha=1.00, transB=0, beta=1.00) -> addmm_2
output: name='addmm_2' type=dtype('float32') shape=[1, 10]
dynamo¶
model = "ort-plot_torch_export_dynamo-cpu-aot1.onnx"
if os.path.exists(model):
print(pretty_onnx(onnx.load(model)))
opset: domain='pkg.onnxscript.torch_lib.common' version=1
opset: domain='' version=18
opset: domain='ai.onnx.ml' version=5
opset: domain='ai.onnx.training' version=1
opset: domain='ai.onnx.preview.training' version=1
opset: domain='com.microsoft' version=1
opset: domain='com.microsoft.experimental' version=1
opset: domain='com.microsoft.nchwc' version=1
opset: domain='org.pytorch.aten' version=1
input: name='x' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='reorder' type=float32 shape=(16, 1, 5, 5)
init: name='conv1.bias' type=float32 shape=(16,)
init: name='reorder_token_1' type=float32 shape=(16, 16, 5, 5)
init: name='conv2.bias' type=float32 shape=(16,)
init: name='fc1.weight' type=float32 shape=(512, 16)
init: name='fc1.bias' type=float32 shape=(512,)
init: name='fc2.weight' type=float32 shape=(128, 512)
init: name='fc2.bias' type=float32 shape=(128,)
init: name='fc3.weight' type=float32 shape=(10, 128)
init: name='fc3.bias' type=float32 shape=(10,)
init: name='val_2' type=int64 shape=(2,) -- array([ 1, 16])
Conv[com.microsoft.nchwc](x, reorder, conv1.bias, activation=b'Relu', group=1, strides=[1,1], pads=[0,0,0,0], auto_pad=b'NOTSET', dilations=[1,1]) -> reorder_token_0
ReorderOutput[com.microsoft.nchwc](reorder_token_0, channels_last=0, channels=16) -> relu
MaxPool(relu, pads=[0,0,0,0], kernel_shape=[2,2], ceil_mode=0, auto_pad=b'NOTSET', dilations=[1,1], strides=[2,2], storage_order=0) -> max_pool2d, val_0
ReorderInput[com.microsoft.nchwc](max_pool2d, channels_last=0) -> reorder_token_2
Conv[com.microsoft.nchwc](reorder_token_2, reorder_token_1, conv2.bias, activation=b'Relu', group=1, strides=[1,1], pads=[0,0,0,0], auto_pad=b'NOTSET', dilations=[1,1]) -> reorder_token_3
ReorderOutput[com.microsoft.nchwc](reorder_token_3, channels_last=0, channels=16) -> relu_1
MaxPool(relu_1, pads=[0,0,0,0], kernel_shape=[2,2], ceil_mode=0, auto_pad=b'NOTSET', dilations=[1,1], strides=[2,2], storage_order=0) -> max_pool2d_1, val_1
Reshape(max_pool2d_1, val_2, allowzero=0) -> view
FusedGemm[com.microsoft](view, fc1.weight, fc1.bias, activation=b'Relu', beta=1.00, transB=1, alpha=1.00, transA=0) -> relu_2
FusedGemm[com.microsoft](relu_2, fc2.weight, fc2.bias, activation=b'Relu', beta=1.00, transB=1, alpha=1.00, transA=0) -> relu_3
Gemm(relu_3, fc3.weight, fc3.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> addmm_2
output: name='addmm_2' type=dtype('float32') shape=[1, 10]
Show the interesting models for CUDA¶
script¶
model = "ort-plot_torch_export_cus_p2-cuda-aot0.onnx"
if os.path.exists(model):
print(pretty_onnx(onnx.load(model)))
opset: domain='' version=18
opset: domain='ai.onnx.ml' version=5
opset: domain='ai.onnx.training' version=1
opset: domain='ai.onnx.preview.training' version=1
opset: domain='com.microsoft' version=1
opset: domain='com.microsoft.experimental' version=1
opset: domain='com.microsoft.nchwc' version=1
opset: domain='org.pytorch.aten' version=1
doc_string: large_model=False, inline=False, external_threshold=102...
input: name='input' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='_onx_concat0' type=int64 shape=(2,) -- array([ 1, -1]) -- GraphBuilder.constant_folding.from/fold(_onx_gatherelements0,init7_s1_-1)##_onx_gatherelements0/GraphBuilder.constant_folding.from/fold(_onx_shape0,init7_s1_0)##_onx_shape0/##init7_s1_0/Opset.make_node.1/Shape##init7_s1_-1/Opset.make_node.1/Shape
init: name='GemmTransposePattern--_onx_transpose0' type=float32 shape=(512, 16)-- GraphBuilder.constant_folding.from/fold(_onx_transpose0)##_onx_transpose0/GraphBuilder.constant_folding.from/fold(p_fc1_weight)##p_fc1_weight/DynamoInterpret.placeholder.1/P(fc1.weight)
init: name='GemmTransposePattern--_onx_transpose02' type=float32 shape=(128, 512)-- GraphBuilder.constant_folding.from/fold(_onx_transpose02)##_onx_transpose02/GraphBuilder.constant_folding.from/fold(p_fc2_weight)##p_fc2_weight/DynamoInterpret.placeholder.1/P(fc2.weight)
init: name='GemmTransposePattern--_onx_transpose03' type=float32 shape=(10, 128)-- GraphBuilder.constant_folding.from/fold(_onx_transpose03)##_onx_transpose03/GraphBuilder.constant_folding.from/fold(p_fc3_weight)##p_fc3_weight/DynamoInterpret.placeholder.1/P(fc3.weight)
init: name='conv1.weight' type=float32 shape=(16, 1, 5, 5) -- DynamoInterpret.placeholder.1/P(conv1.weight)
init: name='conv1.bias' type=float32 shape=(16,) -- DynamoInterpret.placeholder.1/P(conv1.bias)
init: name='conv2.weight' type=float32 shape=(16, 16, 5, 5) -- DynamoInterpret.placeholder.1/P(conv2.weight)
init: name='conv2.bias' type=float32 shape=(16,) -- DynamoInterpret.placeholder.1/P(conv2.bias)
init: name='fc1.bias' type=float32 shape=(512,) -- DynamoInterpret.placeholder.1/P(fc1.bias)
init: name='fc2.bias' type=float32 shape=(128,) -- DynamoInterpret.placeholder.1/P(fc2.bias)
init: name='fc3.bias' type=float32 shape=(10,) -- DynamoInterpret.placeholder.1/P(fc3.bias)
Conv(input, conv1.weight, conv1.bias, dilations=[1,1], group=1, pads=[0,0,0,0], strides=[1,1]) -> conv2d
Relu(conv2d) -> relu
MaxPool(relu, ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> _onx_maxpool0, _onx_maxpool1
Conv(_onx_maxpool0, conv2.weight, conv2.bias, dilations=[1,1], group=1, pads=[0,0,0,0], strides=[1,1]) -> conv2d_1
Relu(conv2d_1) -> relu_1
MaxPool(relu_1, ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> _onx_maxpool02, _onx_maxpool12
Reshape(_onx_maxpool02, _onx_concat0) -> flatten
Gemm(flatten, GemmTransposePattern--_onx_transpose0, fc1.bias, transB=1) -> linear
Relu(linear) -> relu_2
Gemm(relu_2, GemmTransposePattern--_onx_transpose02, fc2.bias, transB=1) -> linear_1
Relu(linear_1) -> relu_3
Gemm(relu_3, GemmTransposePattern--_onx_transpose03, fc3.bias, transB=1) -> output_0
output: name='output_0' type=dtype('float32') shape=[1, 10]
cus_p2¶
model = "ort-plot_torch_export_cus_p2-cuda-aot0.onnx"
if os.path.exists(model):
print(pretty_onnx(onnx.load(model)))
opset: domain='' version=18
opset: domain='ai.onnx.ml' version=5
opset: domain='ai.onnx.training' version=1
opset: domain='ai.onnx.preview.training' version=1
opset: domain='com.microsoft' version=1
opset: domain='com.microsoft.experimental' version=1
opset: domain='com.microsoft.nchwc' version=1
opset: domain='org.pytorch.aten' version=1
doc_string: large_model=False, inline=False, external_threshold=102...
input: name='input' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='_onx_concat0' type=int64 shape=(2,) -- array([ 1, -1]) -- GraphBuilder.constant_folding.from/fold(_onx_gatherelements0,init7_s1_-1)##_onx_gatherelements0/GraphBuilder.constant_folding.from/fold(_onx_shape0,init7_s1_0)##_onx_shape0/##init7_s1_0/Opset.make_node.1/Shape##init7_s1_-1/Opset.make_node.1/Shape
init: name='GemmTransposePattern--_onx_transpose0' type=float32 shape=(512, 16)-- GraphBuilder.constant_folding.from/fold(_onx_transpose0)##_onx_transpose0/GraphBuilder.constant_folding.from/fold(p_fc1_weight)##p_fc1_weight/DynamoInterpret.placeholder.1/P(fc1.weight)
init: name='GemmTransposePattern--_onx_transpose02' type=float32 shape=(128, 512)-- GraphBuilder.constant_folding.from/fold(_onx_transpose02)##_onx_transpose02/GraphBuilder.constant_folding.from/fold(p_fc2_weight)##p_fc2_weight/DynamoInterpret.placeholder.1/P(fc2.weight)
init: name='GemmTransposePattern--_onx_transpose03' type=float32 shape=(10, 128)-- GraphBuilder.constant_folding.from/fold(_onx_transpose03)##_onx_transpose03/GraphBuilder.constant_folding.from/fold(p_fc3_weight)##p_fc3_weight/DynamoInterpret.placeholder.1/P(fc3.weight)
init: name='conv1.weight' type=float32 shape=(16, 1, 5, 5) -- DynamoInterpret.placeholder.1/P(conv1.weight)
init: name='conv1.bias' type=float32 shape=(16,) -- DynamoInterpret.placeholder.1/P(conv1.bias)
init: name='conv2.weight' type=float32 shape=(16, 16, 5, 5) -- DynamoInterpret.placeholder.1/P(conv2.weight)
init: name='conv2.bias' type=float32 shape=(16,) -- DynamoInterpret.placeholder.1/P(conv2.bias)
init: name='fc1.bias' type=float32 shape=(512,) -- DynamoInterpret.placeholder.1/P(fc1.bias)
init: name='fc2.bias' type=float32 shape=(128,) -- DynamoInterpret.placeholder.1/P(fc2.bias)
init: name='fc3.bias' type=float32 shape=(10,) -- DynamoInterpret.placeholder.1/P(fc3.bias)
Conv(input, conv1.weight, conv1.bias, dilations=[1,1], group=1, pads=[0,0,0,0], strides=[1,1]) -> conv2d
Relu(conv2d) -> relu
MaxPool(relu, ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> _onx_maxpool0, _onx_maxpool1
Conv(_onx_maxpool0, conv2.weight, conv2.bias, dilations=[1,1], group=1, pads=[0,0,0,0], strides=[1,1]) -> conv2d_1
Relu(conv2d_1) -> relu_1
MaxPool(relu_1, ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> _onx_maxpool02, _onx_maxpool12
Reshape(_onx_maxpool02, _onx_concat0) -> flatten
Gemm(flatten, GemmTransposePattern--_onx_transpose0, fc1.bias, transB=1) -> linear
Relu(linear) -> relu_2
Gemm(relu_2, GemmTransposePattern--_onx_transpose02, fc2.bias, transB=1) -> linear_1
Relu(linear_1) -> relu_3
Gemm(relu_3, GemmTransposePattern--_onx_transpose03, fc3.bias, transB=1) -> output_0
output: name='output_0' type=dtype('float32') shape=[1, 10]
dynopt¶
model = "ort-plot_torch_export_dynopt-cuda-aot1.onnx"
if os.path.exists(model):
print(pretty_onnx(onnx.load(model)))
opset: domain='pkg.onnxscript.torch_lib.common' version=1
opset: domain='' version=18
opset: domain='ai.onnx.ml' version=5
opset: domain='ai.onnx.training' version=1
opset: domain='ai.onnx.preview.training' version=1
opset: domain='com.microsoft' version=1
opset: domain='com.microsoft.experimental' version=1
opset: domain='com.microsoft.nchwc' version=1
opset: domain='org.pytorch.aten' version=1
input: name='x' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='conv1.weight' type=float32 shape=(16, 1, 5, 5)
init: name='conv1.bias' type=float32 shape=(16,)
init: name='conv2.weight' type=float32 shape=(16, 16, 5, 5)
init: name='conv2.bias' type=float32 shape=(16,)
init: name='fc1.bias' type=float32 shape=(512,)
init: name='fc2.bias' type=float32 shape=(128,)
init: name='fc3.bias' type=float32 shape=(10,)
init: name='val_3' type=int64 shape=(2,) -- array([ 1, 16])
init: name='t' type=float32 shape=(16, 512)
init: name='t_1' type=float32 shape=(512, 128)
init: name='t_2' type=float32 shape=(128, 10)
Conv(x, conv1.weight, conv1.bias, group=1, pads=[0,0,0,0], auto_pad=b'NOTSET', strides=[1,1], dilations=[1,1]) -> conv2d
Relu(conv2d) -> relu
MaxPool(relu, storage_order=0, dilations=[1,1], ceil_mode=0, pads=[0,0,0,0], auto_pad=b'NOTSET', strides=[2,2], kernel_shape=[2,2]) -> max_pool2d
Conv(max_pool2d, conv2.weight, conv2.bias, group=1, pads=[0,0,0,0], auto_pad=b'NOTSET', strides=[1,1], dilations=[1,1]) -> conv2d_1
Relu(conv2d_1) -> relu_1
MaxPool(relu_1, storage_order=0, dilations=[1,1], ceil_mode=0, pads=[0,0,0,0], auto_pad=b'NOTSET', strides=[2,2], kernel_shape=[2,2]) -> max_pool2d_1
Reshape(max_pool2d_1, val_3, allowzero=0) -> view
Gemm(view, t, fc1.bias, beta=1.00, transB=0, alpha=1.00, transA=0) -> addmm
Relu(addmm) -> relu_2
Gemm(relu_2, t_1, fc2.bias, beta=1.00, transB=0, alpha=1.00, transA=0) -> addmm_1
Relu(addmm_1) -> relu_3
Gemm(relu_3, t_2, fc3.bias, beta=1.00, transB=0, alpha=1.00, transA=0) -> addmm_2
output: name='addmm_2' type=dtype('float32') shape=[1, 10]
dynamo¶
model = "ort-plot_torch_export_dynamo-cuda-aot1.onnx"
if os.path.exists(model):
print(pretty_onnx(onnx.load(model)))
opset: domain='pkg.onnxscript.torch_lib.common' version=1
opset: domain='' version=18
opset: domain='ai.onnx.ml' version=5
opset: domain='ai.onnx.training' version=1
opset: domain='ai.onnx.preview.training' version=1
opset: domain='com.microsoft' version=1
opset: domain='com.microsoft.experimental' version=1
opset: domain='com.microsoft.nchwc' version=1
opset: domain='org.pytorch.aten' version=1
input: name='x' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='conv1.weight' type=float32 shape=(16, 1, 5, 5)
init: name='conv1.bias' type=float32 shape=(16,)
init: name='conv2.weight' type=float32 shape=(16, 16, 5, 5)
init: name='conv2.bias' type=float32 shape=(16,)
init: name='fc1.weight' type=float32 shape=(512, 16)
init: name='fc1.bias' type=float32 shape=(512,)
init: name='fc2.weight' type=float32 shape=(128, 512)
init: name='fc2.bias' type=float32 shape=(128,)
init: name='fc3.weight' type=float32 shape=(10, 128)
init: name='fc3.bias' type=float32 shape=(10,)
init: name='val_2' type=int64 shape=(2,) -- array([ 1, 16])
Conv(x, conv1.weight, conv1.bias, dilations=[1,1], auto_pad=b'NOTSET', pads=[0,0,0,0], strides=[1,1], group=1) -> conv2d
Relu(conv2d) -> relu
MaxPool(relu, pads=[0,0,0,0], kernel_shape=[2,2], ceil_mode=0, auto_pad=b'NOTSET', dilations=[1,1], strides=[2,2], storage_order=0) -> max_pool2d, val_0
Conv(max_pool2d, conv2.weight, conv2.bias, dilations=[1,1], auto_pad=b'NOTSET', pads=[0,0,0,0], strides=[1,1], group=1) -> conv2d_1
Relu(conv2d_1) -> relu_1
MaxPool(relu_1, pads=[0,0,0,0], kernel_shape=[2,2], ceil_mode=0, auto_pad=b'NOTSET', dilations=[1,1], strides=[2,2], storage_order=0) -> max_pool2d_1, val_1
Reshape(max_pool2d_1, val_2, allowzero=0) -> view
Gemm(view, fc1.weight, fc1.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> addmm
Relu(addmm) -> relu_2
Gemm(relu_2, fc2.weight, fc2.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> addmm_1
Relu(addmm_1) -> relu_3
Gemm(relu_3, fc3.weight, fc3.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> addmm_2
output: name='addmm_2' type=dtype('float32') shape=[1, 10]
Total running time of the script: (0 minutes 51.057 seconds)
Related examples
