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.device_count() > 0
logging.disable(logging.ERROR)


def system_info():
    obs = {}
    obs["processor"] = platform.processor()
    obs["cores"] = multiprocessing.cpu_count()
    try:
        obs["cuda"] = 1 if torch.cuda.device_count() > 0 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"], dynamo=False)


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")
Memory Consumption of the Export model size=0 Mb, Memory peak (Mb), Memory peak - memory begin (Mb), Memory average - memory begin (Mb), GPU Memory peak (Mb), GPU Memory peak - memory begin (Mb), GPU Memory average - memory begin (Mb)
          peak         mean   n  ...    gpu0_end  nodes    export
0  1067.714844  1066.507457  11  ...  320.617188   12.0    script
1  1072.089844  1068.715588  84  ...  320.617188   12.0    dynamo
2  1073.964844  1072.393364  87  ...  320.617188   12.0    dynopt
3  1073.964844  1073.964844  14  ...  320.617188   12.0    cus_p0
4  1073.964844  1073.964844  33  ...  320.617188   12.0    cus_p2
5  1073.964844  1073.964844  16  ...  320.617188    NaN  torch.fx

[6 rows x 12 columns]

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
     export      time       min  ...      last       std  nodes
0    script  0.049330  0.027718  ...  0.032687  0.022521     12
1    dynamo  0.882212  0.781171  ...  0.873998  0.139963     12
2    dynopt  1.009805  0.843696  ...  1.181878  0.155271     12
3    cus_p0  0.088279  0.075285  ...  0.079296  0.010623     12
4    cus_p2  0.079832  0.075047  ...  0.083776  0.002993     12
5  torch.fx  0.056789  0.054247  ...  0.054247  0.001807     12

[6 rows x 8 columns]

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 0x7bf6bfd6f740>
         698259 function calls (685660 primitive calls) in 0.827 seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
    80/20    0.001    0.000    0.499    0.025 ~/vv/this312/lib/python3.12/site-packages/torch/nn/functional.py:1709(relu)
        5    0.000    0.000    0.323    0.065 ~/vv/this312/lib/python3.12/site-packages/torch/export/_trace.py:1916(forward)
        5    0.000    0.000    0.302    0.060 ~/github/experimental-experiment/_doc/examples/plot_torch_export_201.py:191(forward)
        5    0.001    0.000    0.238    0.048 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:5262(to_onnx)
     1425    0.006    0.000    0.212    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1417(__torch_function__)
        5    0.001    0.000    0.205    0.041 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:5959(optimize)
     1425    0.006    0.000    0.200    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1447(__torch_function__)
     2300    0.006    0.000    0.195    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_export/non_strict_utils.py:1100(__torch_function__)
       60    0.002    0.000    0.178    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_ops.py:947(handler)
       60    0.008    0.000    0.174    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_library/utils.py:298(handle_dispatch_mode)
        5    0.000    0.000    0.173    0.035 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:6336(optimize_with_patterns)
        5    0.001    0.000    0.172    0.034 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:1367(optimize)
   300/60    0.001    0.000    0.164    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_stats.py:24(wrapper)
       60    0.001    0.000    0.164    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1540(__torch_dispatch__)
       60    0.003    0.000    0.163    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:879(proxy_call)
    40/10    0.000    0.000    0.152    0.015 ~/vv/this312/lib/python3.12/site-packages/torch/_jit_internal.py:609(fn)
       25    0.010    0.000    0.137    0.005 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:1180(_optimize_matching_step)
     1770    0.033    0.000    0.124    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:148(enumerate_matches)
     20/5    0.000    0.000    0.113    0.023 {built-in method torch.flatten}
      115    0.001    0.000    0.102    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/proxy.py:263(create_proxy)
  390/130    0.002    0.000    0.101    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/overrides.py:1681(handle_torch_function)
      120    0.001    0.000    0.098    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:2036(create_node)
      120    0.001    0.000    0.096    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1189(create_node)
      120    0.003    0.000    0.093    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/proxy.py:152(create_node)
 1685/265    0.005    0.000    0.076    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1197(unflatten)
  340/175    0.000    0.000    0.074    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_ops.py:835(__call__)
        5    0.000    0.000    0.072    0.014 ~/vv/this312/lib/python3.12/site-packages/torch/export/_trace.py:534(_produce_aten_artifact)
      135    0.001    0.000    0.070    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_traceback.py:174(summary)
       65    0.000    0.000    0.066    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:661(track_tensor_tree)
   120/65    0.001    0.000    0.065    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:683(wrap_with_proxy)
       30    0.000    0.000    0.063    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:815(recompile)
       20    0.000    0.000    0.059    0.003 {built-in method torch.relu}
      185    0.001    0.000    0.059    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1341(tree_map)
      240    0.001    0.000    0.059    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1376(__torch_dispatch__)
      240    0.002    0.000    0.058    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:2124(dispatch)
3270/3250    0.003    0.000    0.055    0.000 {built-in method builtins.next}
       95    0.001    0.000    0.054    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1469(_cached_dispatch_impl)
        5    0.001    0.000    0.052    0.010 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:4974(process)
      120    0.002    0.000    0.051    0.000 ~/github/experimental-experiment/experimental_experiment/torch_interpreter/interpreter.py:173(run_node)
      115    0.001    0.000    0.050    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:552(set_meta)
       30    0.000    0.000    0.049    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1677(python_code)
       15    0.000    0.000    0.048    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:130(forward)
    60/15    0.001    0.000    0.048    0.003 {built-in method torch._C._nn.linear}
        5    0.000    0.000    0.045    0.009 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:635(create_args_for_root)
      135    0.010    0.000    0.044    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_traceback.py:252(_extract_symbolized_tb)
      440    0.000    0.000    0.043    0.000 {method 'extend' of 'list' objects}
       60    0.000    0.000    0.043    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:701(<genexpr>)
       55    0.000    0.000    0.043    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:698(proxy_placeholder)
       55    0.000    0.000    0.043    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:903(_proxy_placeholder)
       55    0.000    0.000    0.041    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:907(replace_ph)
       55    0.001    0.000    0.040    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:877(meta_tensor)
       30    0.000    0.000    0.037    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1756(_python_code)
       30    0.004    0.000    0.037    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:431(_gen_python_code)
      720    0.001    0.000    0.037    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:994(match)
       60    0.001    0.000    0.036    0.001 ~/github/experimental-experiment/experimental_experiment/torch_interpreter/interpreter.py:1388(call_function)
       55    0.002    0.000    0.036    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/autograd/grad_mode.py:293(__exit__)
      720    0.001    0.000    0.035    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:383(_get_match_pattern)
     30/6    0.001    0.000    0.035    0.006 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:1868(__call__)
1545/1535    0.001    0.000    0.034    0.000 /usr/lib/python3.12/contextlib.py:132(__enter__)
      205    0.005    0.000    0.034    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:3982(make_node)
       10    0.000    0.000    0.033    0.003 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:326(_build_pattern)
     25/5    0.000    0.000    0.033    0.007 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:3045(from_tensor)
      185    0.001    0.000    0.032    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder_opset.py:115(make_node)
       10    0.000    0.000    0.032    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/conv.py:549(forward)
       10    0.000    0.000    0.032    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/conv.py:531(_conv_forward)
    40/10    0.000    0.000    0.032    0.003 {built-in method torch.conv2d}
     9155    0.004    0.000    0.032    0.000 /usr/lib/python3.12/traceback.py:265(__init__)
     25/5    0.001    0.000    0.031    0.006 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:357(from_real_tensor)
    40/10    0.000    0.000    0.031    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/nn/functional.py:810(_max_pool2d)
      115    0.002    0.000    0.029    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/passes/shape_prop.py:40(_extract_tensor_metadata)
       10    0.000    0.000    0.029    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:2027(tree_map_with_path)
        5    0.000    0.000    0.028    0.006 ~/vv/this312/lib/python3.12/site-packages/torch/_export/non_strict_utils.py:379(make_fake_inputs)
    10155    0.007    0.000    0.028    0.000 /usr/lib/python3.12/traceback.py:318(line)
       10    0.000    0.000    0.028    0.003 {built-in method torch.max_pool2d}
100810/99610    0.024    0.000    0.028    0.000 {built-in method builtins.isinstance}
       95    0.001    0.000    0.027    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:2053(_output_from_cache_entry)
       20    0.000    0.000    0.026    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:2059(<genexpr>)
      105    0.003    0.000    0.026    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1980(_get_output_tensor_from_cache_entry)
       95    0.001    0.000    0.026    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1573(_cache_key)
      135    0.025    0.000    0.025    0.000 {built-in method torch._C._profiler.symbolize_tracebacks}
       55    0.003    0.000    0.024    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/autograd/grad_mode.py:289(__enter__)
1545/1535    0.002    0.000    0.024    0.000 /usr/lib/python3.12/contextlib.py:141(__exit__)
   440/95    0.005    0.000    0.024    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1693(_prep_args_for_hash)
  185/155    0.001    0.000    0.024    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/module.py:1973(__setattr__)
       10    0.000    0.000    0.024    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:452(__init__)
      375    0.001    0.000    0.023    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1270(tree_flatten)
       10    0.000    0.000    0.023    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/_export/passes/replace_with_hop_pass_util.py:160(_replace_with_hop_pass_helper)
        5    0.001    0.000    0.022    0.004 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:4524(_build_initializers)
 2345/375    0.005    0.000    0.022    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1278(helper)
       10    0.000    0.000    0.021    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:571(graph)
      720    0.003    0.000    0.021    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:636(emit_node)
       50    0.001    0.000    0.020    0.000 ~/github/onnx-diagnostic/onnx_diagnostic/helpers/mini_onnx_builder.py:17(proto_from_array)
       15    0.002    0.000    0.019    0.001 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns/__init__.py:139(get_default_patterns)
      115    0.000    0.000    0.019    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:452(extract_val)
      135    0.000    0.000    0.019    0.000 /usr/lib/python3.12/inspect.py:3308(signature)
      115    0.000    0.000    0.019    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:426(snapshot_fake)
      135    0.000    0.000    0.019    0.000 /usr/lib/python3.12/inspect.py:3050(from_callable)
  265/135    0.005    0.000    0.018    0.000 /usr/lib/python3.12/inspect.py:2470(_signature_from_callable)
      115    0.002    0.000    0.018    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_impls.py:1425(fast_detach)
     9155    0.008    0.000    0.018    0.000 /usr/lib/python3.12/linecache.py:26(getline)
       55    0.003    0.000    0.018    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:278(describe_tensor)
      275    0.005    0.000    0.017    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:722(__new__)
        5    0.000    0.000    0.017    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_export/non_strict_utils.py:454(<lambda>)
      115    0.000    0.000    0.017    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_prims_common/__init__.py:431(is_contiguous_for_memory_format_or_false)
        5    0.000    0.000    0.017    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:715(apply_runtime_assertion_pass)
      115    0.000    0.000    0.017    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_prims_common/__init__.py:394(is_contiguous_for_memory_format)
        5    0.000    0.000    0.017    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:1052(__init__)
        5    0.000    0.000    0.017    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_export/non_strict_utils.py:161(fakify)
     2515    0.006    0.000    0.016    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:904(__setattr__)
     1185    0.004    0.000    0.016    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns_api.py:124(__init__)
      115    0.001    0.000    0.016    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_prims_common/__init__.py:297(is_contiguous)
       10    0.001    0.000    0.016    0.002 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:6515(constant_folding)
       50    0.015    0.000    0.015    0.000 {method 'clone' of 'torch._C.TensorBase' objects}
        5    0.001    0.000    0.015    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:977(placeholder_naming_pass)
     1350    0.001    0.000    0.015    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:195(is_sparse_any)
     5765    0.015    0.000    0.015    0.000 {built-in method builtins.setattr}
      115    0.002    0.000    0.014    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:580(track_tensor)
       15    0.000    0.000    0.014    0.001 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:64(__init__)
        5    0.000    0.000    0.013    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/export/_trace.py:493(_replace_unbacked_bindings)
       30    0.000    0.000    0.013    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:99(_forward_from_src)
       30    0.000    0.000    0.013    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:105(_method_from_src)
       30    0.000    0.000    0.013    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:94(_exec_with_source)
        5    0.000    0.000    0.013    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:1674(_create_graph_module_for_export)
       25    0.001    0.000    0.012    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:1282(_optimize_apply_step)
      145    0.003    0.000    0.012    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1048(_flatten_into)
 1050/115    0.006    0.000    0.012    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/symbolic_shapes.py:1097(_free_unbacked_symbols_with_path)
       30    0.012    0.000    0.012    0.000 {built-in method builtins.compile}
      115    0.001    0.000    0.012    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_prims_common/__init__.py:252(check_contiguous_sizes_strides)
      240    0.001    0.000    0.012    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:7590(_make_node_set_type_shape)
        5    0.000    0.000    0.012    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/_export/passes/replace_set_grad_with_hop_pass.py:110(replace_set_grad_with_hop_pass)
     3985    0.003    0.000    0.012    0.000 <frozen _collections_abc>:804(get)
        5    0.000    0.000    0.012    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_lazy_graph_module.py:57(_make_graph_module)
       50    0.006    0.000    0.011    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:6451(remove_unused)
      240    0.001    0.000    0.011    0.000 ~/github/experimental-experiment/experimental_experiment/xshape/shape_type_compute.py:1534(set_shape_type_op_any)
        5    0.000    0.000    0.011    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/_export/passes/replace_autocast_with_hop_pass.py:178(replace_autocast_with_hop_pass)
      120    0.001    0.000    0.011    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1214(create_node)
    50/40    0.000    0.000    0.011    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:1149(compute_constant)
      420    0.003    0.000    0.011    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:1810(set_shape)
    50/40    0.002    0.000    0.011    0.000 ~/github/experimental-experiment/experimental_experiment/xshape/_inference_runtime.py:289(compute_constant)
       60    0.001    0.000    0.011    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1732(override_node_repr)
       90    0.001    0.000    0.010    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:2703(add_initializer)
  510/210    0.002    0.000    0.010    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/__init__.py:862(sym_max)
      135    0.004    0.000    0.009    0.000 /usr/lib/python3.12/inspect.py:2366(_signature_from_function)
        5    0.000    0.000    0.009    0.002 ~/github/experimental-experiment/experimental_experiment/torch_interpreter/_aten_functions.py:3433(aten_flatten_using_ints)
       30    0.001    0.000    0.009    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:788(apply_match)
       10    0.002    0.000    0.009    0.001 {built-in method torch._ops.aten.}
        5    0.000    0.000    0.009    0.002 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns/onnx_functions.py:137(match_pattern)
       15    0.000    0.000    0.009    0.001 ~/github/experimental-experiment/experimental_experiment/xbuilder/optimization_options.py:62(__init__)
       15    0.000    0.000    0.009    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_logging/_internal.py:1307(trace_structured)
     9155    0.006    0.000    0.008    0.000 /usr/lib/python3.12/linecache.py:36(getlines)
     3985    0.004    0.000    0.008    0.000 <frozen os>:680(__getitem__)
       35    0.004    0.000    0.008    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:6782(remove_identity_nodes)
        5    0.000    0.000    0.008    0.002 ~/github/experimental-experiment/experimental_experiment/xoptim/__init__.py:101(get_pattern_list)
        5    0.000    0.000    0.008    0.002 ~/github/experimental-experiment/experimental_experiment/xoptim/__init__.py:14(get_pattern)
      145    0.002    0.000    0.008    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1070(extract_tensor_metadata)
      110    0.001    0.000    0.008    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:1124(_check_graph)
       55    0.001    0.000    0.008    0.000 ~/github/experimental-experiment/experimental_experiment/torch_interpreter/interpreter.py:485(placeholder)
       15    0.000    0.000    0.008    0.001 ~/github/experimental-experiment/experimental_experiment/torch_interpreter/_aten_functions.py:6060(aten_linear)
1675/1590    0.002    0.000    0.008    0.000 {built-in method builtins.all}
     1570    0.003    0.000    0.008    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:190(is_sparse_compressed)
     3740    0.002    0.000    0.008    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1014(_get_node_type)
     1375    0.003    0.000    0.008    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:632(__set__)
39845/39795    0.007    0.000    0.007    0.000 {built-in method builtins.len}
    29795    0.007    0.000    0.007    0.000 {method 'append' of 'list' objects}
     2750    0.002    0.000    0.007    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1026(tree_is_leaf)
       10    0.000    0.000    0.007    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_functorch/_aot_autograd/frontend_utils.py:190(_detect_attribute_assignment)
       50    0.005    0.000    0.007    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:143(_build)
  385/235    0.001    0.000    0.007    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1153(create_arg)
     2350    0.001    0.000    0.007    0.000 {built-in method builtins.repr}
      435    0.001    0.000    0.007    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:3780(verify_shape)
      250    0.001    0.000    0.007    0.000 ~/github/experimental-experiment/experimental_experiment/xshape/_inference_runtime.py:84(_make_node_set_type_shape_constant)
       60    0.000    0.000    0.007    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:362(_call_method_with_signature_check)
      110    0.006    0.000    0.007    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:1023(_check_graph_nodes)
     2370    0.003    0.000    0.006    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:157(create_name)
    24555    0.006    0.000    0.006    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/graph_builder_optim.py:204(is_used)
     1350    0.002    0.000    0.006    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:176(is_sparse_coo)
       95    0.001    0.000    0.006    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:6671(_refresh_values_cache)
       10    0.000    0.000    0.006    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:1111(patch_method)
     1325    0.003    0.000    0.006    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:844(make_key)
      250    0.002    0.000    0.006    0.000 ~/github/onnx/onnx/helper.py:129(make_node)
        5    0.000    0.000    0.006    0.001 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:5223(_update_metadata_props)
       10    0.000    0.000    0.006    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:1079(patch)
  365/295    0.002    0.000    0.006    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/recording.py:247(wrapper)
      120    0.001    0.000    0.006    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:375(prepend)
      845    0.002    0.000    0.006    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1074(__post_init__)
       55    0.001    0.000    0.006    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:394(mk_fake_tensor)
     2300    0.004    0.000    0.006    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_export/non_strict_utils.py:1000(_override)
       55    0.001    0.000    0.006    0.000 ~/github/experimental-experiment/experimental_experiment/xbuilder/graph_builder.py:2565(make_initializer)
      360    0.001    0.000    0.006    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns/onnx_reshape.py:970(match)
       10    0.000    0.000    0.006    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:1164(_new_patcher)
     2615    0.001    0.000    0.006    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:571(__repr__)
      360    0.003    0.000    0.006    0.000 ~/github/experimental-experiment/experimental_experiment/xoptim/patterns/onnx_conv.py:15(match)
        5    0.000    0.000    0.006    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:1132(revert_all_patches)
       10    0.000    0.000    0.005    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:1076(revert)
     1545    0.002    0.000    0.005    0.000 /usr/lib/python3.12/contextlib.py:299(helper)
done.
profile custom2: <function export_cus_p2 at 0x7bf6bfd6f7e0>
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 0x7bf6bfd6f600>
         10185218 function calls (9989987 primitive calls) in 9.387 seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        5    0.012    0.002    3.991    0.798 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_registration.py:159(from_torchlib)
        5    0.068    0.014    2.823    0.565 ~/github/onnxscript/onnxscript/_framework_apis/torch_2_5.py:84(get_torchlib_ops)
     2220    0.026    0.000    2.743    0.001 ~/github/onnxscript/onnxscript/values.py:630(function_ir)
       10    0.173    0.017    1.912    0.191 ~/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:191(_override_composite_implicit_decomp)
     2220    0.017    0.000    1.516    0.001 ~/github/onnxscript/onnxscript/_internal/ast_utils.py:13(get_src_and_ast)
    64635    0.090    0.000    1.305    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_ops.py:122(py_impl)
       10    0.001    0.000    1.303    0.130 ~/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1318(_collect_all_valid_cia_ops)
      260    0.012    0.000    1.301    0.005 ~/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1301(_collect_all_valid_cia_ops_for_namespace)
   130290    0.207    0.000    1.219    0.000 <frozen _collections_abc>:469(__new__)
      260    0.433    0.002    1.200    0.005 ~/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1236(_materialize_cpp_cia_ops)
     2220    0.005    0.000    1.114    0.001 ~/github/onnxscript/onnxscript/converter.py:1458(translate_function_signature)
     2525    0.027    0.000    1.105    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_registration.py:61(__post_init__)
     2220    0.069    0.000    1.102    0.000 ~/github/onnxscript/onnxscript/converter.py:1373(_translate_function_signature_common)
     2525    0.080    0.000    1.064    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_schemas.py:432(from_function)
     2220    0.004    0.000    1.060    0.000 /usr/lib/python3.12/inspect.py:1272(getsource)
     2220    0.093    0.000    1.053    0.000 /usr/lib/python3.12/inspect.py:1251(getsourcelines)
   129270    0.143    0.000    0.910    0.000 <frozen _collections_abc>:511(_is_param_expr)
    110/6    0.002    0.000    0.871    0.145 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:1868(__call__)
     35/5    0.001    0.000    0.869    0.174 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:3045(from_tensor)
    100/5    0.003    0.000    0.868    0.174 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:357(from_real_tensor)
     2220    0.222    0.000    0.789    0.000 /usr/lib/python3.12/inspect.py:1232(getblock)
        5    0.008    0.002    0.786    0.157 ~/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:275(_split_decomp_table_to_cia_and_python_decomp)
136235/136220    0.053    0.000    0.776    0.000 {built-in method builtins.any}
88695/15685    0.150    0.000    0.714    0.000 ~/github/onnxscript/onnxscript/type_annotation.py:146(is_value_type)
   129270    0.704    0.000    0.704    0.000 <frozen _collections_abc>:521(<genexpr>)
        5    0.008    0.002    0.677    0.135 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_decomp.py:39(create_onnx_friendly_decomposition_table)
    23080    0.663    0.000    0.663    0.000 {built-in method builtins.compile}
        5    0.000    0.000    0.618    0.124 ~/vv/this312/lib/python3.12/site-packages/torch/export/decomp_utils.py:141(items)
        5    0.000    0.000    0.618    0.124 ~/vv/this312/lib/python3.12/site-packages/torch/export/decomp_utils.py:158(_materialize_if_needed)
        5    0.002    0.000    0.618    0.124 ~/vv/this312/lib/python3.12/site-packages/torch/export/decomp_utils.py:145(materialize)
    80/20    0.001    0.000    0.532    0.027 ~/vv/this312/lib/python3.12/site-packages/torch/nn/functional.py:1709(relu)
   259910    0.291    0.000    0.500    0.000 /usr/lib/python3.12/tokenize.py:569(_generate_tokens_from_c_tokenizer)
     3550    0.011    0.000    0.478    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1417(__torch_function__)
1727680/1720250    0.377    0.000    0.475    0.000 {built-in method builtins.isinstance}
      110    0.001    0.000    0.463    0.004 ~/vv/this312/lib/python3.12/site-packages/torch/_functorch/_aot_autograd/functional_utils.py:33(to_fun)
     9080    0.007    0.000    0.457    0.000 ~/github/onnxscript/onnxscript/type_annotation.py:187(is_valid_type)
   799720    0.419    0.000    0.426    0.000 {built-in method builtins.getattr}
27870/4195    0.120    0.000    0.378    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_schemas.py:268(_get_allowed_types_from_type_annotation)
    64635    0.301    0.000    0.377    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_ops.py:131(inner)
     2240    0.008    0.000    0.377    0.000 /usr/lib/python3.12/ast.py:34(parse)
     2525    0.047    0.000    0.356    0.000 /usr/lib/python3.12/typing.py:2186(get_type_hints)
      230    0.104    0.000    0.340    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:358(__torch_dispatch__)
        5    0.000    0.000    0.334    0.067 ~/vv/this312/lib/python3.12/site-packages/torch/export/_trace.py:1916(forward)
      230    0.002    0.000    0.324    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1540(__torch_dispatch__)
        5    0.000    0.000    0.313    0.063 ~/github/experimental-experiment/_doc/examples/plot_torch_export_201.py:191(forward)
      120    0.018    0.000    0.310    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:879(proxy_call)
    88695    0.073    0.000    0.294    0.000 ~/github/onnxscript/onnxscript/type_annotation.py:138(_is_tensor_type)
      960    0.015    0.000    0.287    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:2124(dispatch)
     6605    0.004    0.000    0.268    0.000 ~/github/onnxscript/onnxscript/type_annotation.py:183(is_attr_type)
      440    0.004    0.000    0.267    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1469(_cached_dispatch_impl)
10540/758    0.012    0.000    0.261    0.000 {built-in method builtins.next}
 4605/364    0.005    0.000    0.255    0.001 /usr/lib/python3.12/contextlib.py:132(__enter__)
   135/10    0.001    0.000    0.238    0.024 ~/github/ir-py/src/onnx_ir/passes/_pass_infra.py:111(__call__)
    20/10    0.000    0.000    0.238    0.024 ~/github/ir-py/src/onnx_ir/passes/_pass_infra.py:232(call)
        5    0.000    0.000    0.225    0.045 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_onnx_program.py:315(optimize)
        5    0.000    0.000    0.225    0.045 ~/github/onnxscript/onnxscript/_framework_apis/torch_2_8.py:27(optimize)
        5    0.000    0.000    0.225    0.045 ~/github/onnxscript/onnxscript/optimizer/_optimizer.py:17(optimize_ir)
     1425    0.006    0.000    0.217    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1447(__torch_function__)
       95    0.002    0.000    0.214    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:815(recompile)
        5    0.000    0.000    0.212    0.042 ~/github/ir-py/src/onnx_ir/passes/_pass_infra.py:273(call)
     2300    0.009    0.000    0.211    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_export/non_strict_utils.py:1100(__torch_function__)
   257690    0.111    0.000    0.209    0.000 /usr/lib/python3.12/collections/__init__.py:447(_make)
    91755    0.059    0.000    0.202    0.000 ~/github/onnxscript/onnxscript/type_annotation.py:85(_remove_annotation)
    11165    0.027    0.000    0.201    0.000 ~/github/onnxscript/onnxscript/converter.py:444(_eval_constant_expr)
       60    0.002    0.000    0.190    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_ops.py:947(handler)
       65    0.010    0.000    0.188    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_library/utils.py:298(handle_dispatch_mode)
     2775    0.003    0.000    0.187    0.000 /usr/lib/python3.12/inspect.py:3308(signature)
       70    0.001    0.000    0.187    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/_higher_order_ops/utils.py:36(autograd_not_implemented_inner)
      260    0.187    0.001    0.187    0.001 {built-in method torch._C._dispatch_get_registrations_for_dispatch_key}
     2775    0.004    0.000    0.184    0.000 /usr/lib/python3.12/inspect.py:3050(from_callable)
3015/2775    0.028    0.000    0.180    0.000 /usr/lib/python3.12/inspect.py:2470(_signature_from_callable)
      100    0.001    0.000    0.175    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1677(python_code)
   130430    0.098    0.000    0.172    0.000 /usr/lib/python3.12/typing.py:2310(get_origin)
    40/10    0.000    0.000    0.163    0.016 ~/vv/this312/lib/python3.12/site-packages/torch/_jit_internal.py:609(fn)
      230    0.001    0.000    0.160    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/proxy.py:263(create_proxy)
     6540    0.005    0.000    0.156    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1222(_is_preservable_cia_op)
     2220    0.025    0.000    0.155    0.000 /usr/lib/python3.12/inspect.py:1063(findsource)
      240    0.002    0.000    0.149    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:2036(create_node)
      240    0.001    0.000    0.147    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:1189(create_node)
   485885    0.146    0.000    0.146    0.000 {method 'split' of 'str' objects}
48875/48805    0.027    0.000    0.146    0.000 {built-in method builtins.repr}
       10    0.000    0.000    0.143    0.014 ~/github/onnxscript/onnxscript/rewriter/__init__.py:76(call)
       10    0.000    0.000    0.143    0.014 ~/github/onnxscript/onnxscript/rewriter/_rewrite_rule.py:752(apply_to_model)
      240    0.006    0.000    0.139    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/proxy.py:152(create_node)
    36485    0.026    0.000    0.138    0.000 ~/github/ir-py/src/onnx_ir/_core.py:2021(__hash__)
     9580    0.016    0.000    0.136    0.000 /usr/lib/python3.12/typing.py:885(__init__)
       10    0.004    0.000    0.136    0.014 ~/github/onnxscript/onnxscript/rewriter/_rewrite_rule.py:626(_apply_to_graph_or_function)
1079800/1079260    0.135    0.000    0.135    0.000 {built-in method builtins.len}
      100    0.001    0.000    0.134    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:1756(_python_code)
      100    0.014    0.000    0.133    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:431(_gen_python_code)
     5740    0.006    0.000    0.131    0.000 ~/github/onnxscript/onnxscript/rewriter/_rewrite_rule.py:296(try_rewrite)
21735/9580    0.024    0.000    0.131    0.000 /usr/lib/python3.12/typing.py:406(_eval_type)
     6540    0.073    0.000    0.131    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:1270(_check_valid_to_preserve)
      130    0.000    0.000    0.124    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:661(track_tensor_tree)
     20/5    0.000    0.000    0.123    0.025 {built-in method torch.flatten}
     2775    0.044    0.000    0.122    0.000 /usr/lib/python3.12/inspect.py:2366(_signature_from_function)
17155/540    0.036    0.000    0.121    0.000 /usr/lib/python3.12/copy.py:118(deepcopy)
     9580    0.026    0.000    0.121    0.000 /usr/lib/python3.12/typing.py:909(_evaluate)
     5740    0.007    0.000    0.121    0.000 ~/github/onnxscript/onnxscript/rewriter/_rewrite_rule.py:94(match)
  240/130    0.001    0.000    0.121    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:683(wrap_with_proxy)
      980    0.002    0.000    0.117    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/interpreter.py:231(_set_current_node)
        5    0.000    0.000    0.117    0.023 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_fx_passes.py:22(insert_type_promotion_nodes)
      665    0.005    0.000    0.116    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/traceback.py:64(__init__)
      980    0.003    0.000    0.112    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/traceback.py:336(set_current_meta)
        5    0.000    0.000    0.112    0.022 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/fx/_pass.py:227(run)
        5    0.000    0.000    0.112    0.022 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/fx/passes/type_promotion.py:1650(_run)
      120    0.001    0.000    0.109    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/fx/passes/type_promotion.py:1570(run_node)
       10    0.001    0.000    0.109    0.011 ~/vv/this312/lib/python3.12/site-packages/torch/export/_trace.py:534(_produce_aten_artifact)
 1905/535    0.003    0.000    0.109    0.000 /usr/lib/python3.12/copy.py:191(_deepcopy_list)
      170    0.004    0.000    0.108    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/meta_utils.py:877(meta_tensor)
     5740    0.012    0.000    0.108    0.000 ~/github/onnxscript/onnxscript/rewriter/_matcher.py:345(match)
  510/250    0.004    0.000    0.108    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/overrides.py:1681(handle_torch_function)
      440    0.004    0.000    0.106    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1573(_cache_key)
      510    0.007    0.000    0.104    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1551(tree_map_only)
   393035    0.103    0.000    0.103    0.000 {built-in method __new__ of type object at 0xa44b40}
 1560/730    0.008    0.000    0.102    0.000 /usr/lib/python3.12/copy.py:247(_reconstruct)
        5    0.000    0.000    0.101    0.020 ~/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:1386(module)
        5    0.001    0.000    0.101    0.020 ~/vv/this312/lib/python3.12/site-packages/torch/export/_unlift.py:744(_unlift_exported_program_lifted_states)
      205    0.002    0.000    0.098    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_traceback.py:174(summary)
 1710/440    0.020    0.000    0.097    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1693(_prep_args_for_hash)
      365    0.003    0.000    0.096    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:2053(_output_from_cache_entry)
      230    0.002    0.000    0.094    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:552(set_meta)
      385    0.011    0.000    0.093    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1980(_get_output_tensor_from_cache_entry)
 1390/560    0.034    0.000    0.090    0.000 /usr/lib/python3.12/copy.py:217(_deepcopy_dict)
    40100    0.041    0.000    0.090    0.000 ~/github/ir-py/src/onnx_ir/_core.py:2029(__repr__)
    87765    0.038    0.000    0.088    0.000 {built-in method builtins.issubclass}
      170    0.005    0.000    0.083    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/autograd/grad_mode.py:293(__exit__)
       30    0.001    0.000    0.082    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:452(__init__)
       10    0.000    0.000    0.081    0.008 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:635(create_args_for_root)
     1835    0.001    0.000    0.080    0.000 {method 'extend' of 'list' objects}
      120    0.000    0.000    0.078    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:701(<genexpr>)
      110    0.000    0.000    0.078    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:698(proxy_placeholder)
      110    0.000    0.000    0.078    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:903(_proxy_placeholder)
   216526    0.060    0.000    0.076    0.000 {built-in method builtins.hasattr}
  740/640    0.004    0.000    0.076    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/module.py:1973(__setattr__)
        5    0.001    0.000    0.075    0.015 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_core.py:1028(_exported_program_to_onnx_program)
     2415    0.011    0.000    0.075    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph.py:636(emit_node)
      110    0.001    0.000    0.075    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/fx/_symbolic_trace.py:907(replace_ph)
   159990    0.049    0.000    0.075    0.000 /usr/lib/python3.12/inspect.py:295(isclass)
     5600    0.006    0.000    0.074    0.000 ~/github/onnxscript/onnxscript/rewriter/_matcher.py:286(_match_single_output_node)
    95810    0.035    0.000    0.072    0.000 <frozen abc>:117(__instancecheck__)
    20770    0.052    0.000    0.070    0.000 {built-in method builtins.eval}
       30    0.000    0.000    0.068    0.002 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:571(graph)
     1205    0.004    0.000    0.068    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1270(tree_flatten)
   257690    0.067    0.000    0.067    0.000 /usr/lib/python3.12/inspect.py:1189(tokeneater)
5365/1205    0.018    0.000    0.065    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1278(helper)
5730/5600    0.011    0.000    0.063    0.000 ~/github/onnxscript/onnxscript/rewriter/_matcher.py:132(_match_node)
        5    0.000    0.000    0.062    0.012 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_core.py:757(_translate_fx_graph)
      205    0.014    0.000    0.061    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_traceback.py:252(_extract_symbolized_tb)
       10    0.000    0.000    0.061    0.006 ~/github/onnxscript/onnxscript/optimizer/_constant_folding.py:1300(call)
       10    0.000    0.000    0.061    0.006 ~/github/onnxscript/onnxscript/optimizer/_constant_folding.py:1276(visit_graph)
     2220    0.013    0.000    0.061    0.000 /usr/lib/python3.12/textwrap.py:419(dedent)
      110    0.001    0.000    0.061    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:211(to_functional)
      145    0.001    0.000    0.060    0.000 ~/github/onnxscript/onnxscript/optimizer/_constant_folding.py:1266(visit_node)
       20    0.000    0.000    0.060    0.003 {built-in method torch.relu}
       60    0.002    0.000    0.059    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_core.py:505(_handle_call_function_node_with_lowering)
       75    0.001    0.000    0.059    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:2343(_dispatch_impl)
      145    0.003    0.000    0.058    0.000 ~/github/onnxscript/onnxscript/optimizer/_constant_folding.py:1088(process_node)
     2220    0.013    0.000    0.054    0.000 /usr/lib/python3.12/inspect.py:944(getsourcefile)
     8580    0.019    0.000    0.053    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/node.py:904(__setattr__)
    95130    0.032    0.000    0.053    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_ops.py:845(__hash__)
      170    0.008    0.000    0.052    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/autograd/grad_mode.py:289(__enter__)
       15    0.000    0.000    0.051    0.003 ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/linear.py:130(forward)
    60/15    0.001    0.000    0.051    0.003 {built-in method torch._C._nn.linear}
    73850    0.023    0.000    0.050    0.000 <frozen abc>:121(__subclasscheck__)
      565    0.012    0.000    0.049    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:1048(_flatten_into)
3695/1240    0.003    0.000    0.048    0.000 ~/github/ir-py/src/onnx_ir/serde.py:96(wrapper)
      785    0.015    0.000    0.048    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_tensor.py:722(__new__)
       95    0.000    0.000    0.047    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:99(_forward_from_src)
       95    0.000    0.000    0.047    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:105(_method_from_src)
       95    0.000    0.000    0.047    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/graph_module.py:94(_exec_with_source)
      110    0.001    0.000    0.047    0.000 {built-in method torch._to_functional_tensor}
      230    0.004    0.000    0.046    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/passes/shape_prop.py:40(_extract_tensor_metadata)
      125    0.005    0.000    0.045    0.000 ~/github/onnxscript/onnxscript/optimizer/_constant_folding.py:986(_do_inference)
       45    0.002    0.000    0.045    0.001 ~/vv/this312/lib/python3.12/site-packages/torch/_logging/_internal.py:1307(trace_structured)
     9610    0.027    0.000    0.045    0.000 /usr/lib/python3.12/typing.py:175(_type_check)
    12765    0.006    0.000    0.044    0.000 /usr/lib/python3.12/traceback.py:265(__init__)
    40100    0.018    0.000    0.044    0.000 ~/github/ir-py/src/onnx_ir/_enums.py:366(__repr__)
      230    0.000    0.000    0.043    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:452(extract_val)
       60    0.000    0.000    0.043    0.001 ~/github/onnxscript/onnxscript/values.py:624(__call__)
      230    0.001    0.000    0.042    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/fx/experimental/proxy_tensor.py:426(snapshot_fake)
      720    0.003    0.000    0.041    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/utils/_pytree.py:1490(wrapped)
      230    0.006    0.000    0.041    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/_subclasses/fake_impls.py:1425(fast_detach)
158925/157275    0.038    0.000    0.041    0.000 {built-in method builtins.hash}
     2220    0.031    0.000    0.041    0.000 /usr/lib/python3.12/inspect.py:1599(getclosurevars)
       80    0.001    0.000    0.040    0.000 ~/github/onnxscript/onnxscript/values.py:300(__call__)
    14115    0.010    0.000    0.039    0.000 /usr/lib/python3.12/traceback.py:318(line)
       10    0.000    0.000    0.039    0.004 ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/conv.py:549(forward)
       10    0.000    0.000    0.039    0.004 ~/vv/this312/lib/python3.12/site-packages/torch/nn/modules/conv.py:531(_conv_forward)
    40/10    0.001    0.000    0.039    0.004 {built-in method torch.conv2d}
       80    0.000    0.000    0.039    0.000 ~/vv/this312/lib/python3.12/site-packages/torch/onnx/_internal/exporter/_building.py:598(eval)
        5    0.000    0.000    0.039    0.008 ~/vv/this312/lib/python3.12/site-packages/torch/export/exported_program.py:319(default_decompositions)
        5    0.003    0.001    0.038    0.008 ~/vv/this312/lib/python3.12/site-packages/torch/export/decomp_utils.py:48(__init__)
       10    0.001    0.000    0.038    0.004 ~/vv/this312/lib/python3.12/site-packages/torch/_export/utils.py:715(apply_runtime_assertion_pass)
     2310    0.006    0.000    0.037    0.000 /usr/lib/python3.12/linecache.py:52(checkcache)
done.
profile dynopt: <function export_dynopt at 0x7bf6bfd6f6a0>
done.

Benchmark exported models with ORT

def benchmark(shape):
    from onnxruntime import InferenceSession, SessionOptions, GraphOptimizationLevel

    providers = [["CPUExecutionProvider"]]
    if has_cuda:
        providers.append(["CUDAExecutionProvider", "CPUExecutionProvider"])

    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")],
            providers,
            ["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

6.585910958862882e-05 plot_torch_export_dynamo.onnx ['CPUExecutionProvider']:   0%|          | 0/20 [00:01<?, ?it/s]
6.585910958862882e-05 plot_torch_export_dynamo.onnx ['CPUExecutionProvider']:   5%|▌         | 1/20 [00:01<00:29,  1.56s/it]
5.613043661566299e-05 plot_torch_export_dynamo.onnx ['CPUExecutionProvider']:   5%|▌         | 1/20 [00:02<00:29,  1.56s/it]
5.613043661566299e-05 plot_torch_export_dynamo.onnx ['CPUExecutionProvider']:  10%|█         | 2/20 [00:02<00:18,  1.00s/it]
0.0008543609078353889 plot_torch_export_dynamo.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  10%|█         | 2/20 [00:03<00:18,  1.00s/it]
0.0008543609078353889 plot_torch_export_dynamo.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  15%|█▌        | 3/20 [00:03<00:20,  1.19s/it]
0.0008889594918117405 plot_torch_export_dynamo.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  15%|█▌        | 3/20 [00:04<00:20,  1.19s/it]
0.0008889594918117405 plot_torch_export_dynamo.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  20%|██        | 4/20 [00:04<00:16,  1.03s/it]
6.676405104745221e-05 plot_torch_export_cus_p0.onnx ['CPUExecutionProvider']:  20%|██        | 4/20 [00:04<00:16,  1.03s/it]
6.676405104745221e-05 plot_torch_export_cus_p0.onnx ['CPUExecutionProvider']:  25%|██▌       | 5/20 [00:05<00:13,  1.10it/s]
6.36155665702836e-05 plot_torch_export_cus_p0.onnx ['CPUExecutionProvider']:  25%|██▌       | 5/20 [00:05<00:13,  1.10it/s]
6.36155665702836e-05 plot_torch_export_cus_p0.onnx ['CPUExecutionProvider']:  30%|███       | 6/20 [00:05<00:11,  1.27it/s]
0.0008678395782653194 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  30%|███       | 6/20 [00:06<00:11,  1.27it/s]
0.0008678395782653194 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  35%|███▌      | 7/20 [00:06<00:09,  1.32it/s]
0.0010038928376472432 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  35%|███▌      | 7/20 [00:06<00:09,  1.32it/s]
0.0010038928376472432 plot_torch_export_cus_p0.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  40%|████      | 8/20 [00:07<00:09,  1.32it/s]
6.462736041040783e-05 plot_torch_export_script.onnx ['CPUExecutionProvider']:  40%|████      | 8/20 [00:07<00:09,  1.32it/s]
6.462736041040783e-05 plot_torch_export_script.onnx ['CPUExecutionProvider']:  45%|████▌     | 9/20 [00:07<00:08,  1.35it/s]
7.335691779210316e-05 plot_torch_export_script.onnx ['CPUExecutionProvider']:  45%|████▌     | 9/20 [00:08<00:08,  1.35it/s]
7.335691779210316e-05 plot_torch_export_script.onnx ['CPUExecutionProvider']:  50%|█████     | 10/20 [00:08<00:07,  1.41it/s]
0.0008220590635097497 plot_torch_export_script.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  50%|█████     | 10/20 [00:08<00:07,  1.41it/s]
0.0008220590635097497 plot_torch_export_script.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  55%|█████▌    | 11/20 [00:09<00:06,  1.41it/s]
0.0007300038644330349 plot_torch_export_script.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  55%|█████▌    | 11/20 [00:09<00:06,  1.41it/s]
0.0007300038644330349 plot_torch_export_script.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  60%|██████    | 12/20 [00:09<00:06,  1.33it/s]
6.1033217242375595e-05 plot_torch_export_dynopt.onnx ['CPUExecutionProvider']:  60%|██████    | 12/20 [00:10<00:06,  1.33it/s]
6.1033217242375595e-05 plot_torch_export_dynopt.onnx ['CPUExecutionProvider']:  65%|██████▌   | 13/20 [00:10<00:05,  1.36it/s]
7.227429223004098e-05 plot_torch_export_dynopt.onnx ['CPUExecutionProvider']:  65%|██████▌   | 13/20 [00:11<00:05,  1.36it/s]
7.227429223004098e-05 plot_torch_export_dynopt.onnx ['CPUExecutionProvider']:  70%|███████   | 14/20 [00:11<00:04,  1.42it/s]
0.0008511455761743262 plot_torch_export_dynopt.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  70%|███████   | 14/20 [00:11<00:04,  1.42it/s]
0.0008511455761743262 plot_torch_export_dynopt.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  75%|███████▌  | 15/20 [00:11<00:03,  1.45it/s]
0.000906193115655214 plot_torch_export_dynopt.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  75%|███████▌  | 15/20 [00:12<00:03,  1.45it/s]
0.000906193115655214 plot_torch_export_dynopt.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  80%|████████  | 16/20 [00:12<00:02,  1.41it/s]
6.109760394827254e-05 plot_torch_export_cus_p2.onnx ['CPUExecutionProvider']:  80%|████████  | 16/20 [00:13<00:02,  1.41it/s]
6.109760394827254e-05 plot_torch_export_cus_p2.onnx ['CPUExecutionProvider']:  85%|████████▌ | 17/20 [00:13<00:02,  1.41it/s]
6.34724912800066e-05 plot_torch_export_cus_p2.onnx ['CPUExecutionProvider']:  85%|████████▌ | 17/20 [00:13<00:02,  1.41it/s]
6.34724912800066e-05 plot_torch_export_cus_p2.onnx ['CPUExecutionProvider']:  90%|█████████ | 18/20 [00:13<00:01,  1.50it/s]
0.0007315974304605189 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  90%|█████████ | 18/20 [00:14<00:01,  1.50it/s]
0.0007315974304605189 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  95%|█████████▌| 19/20 [00:14<00:00,  1.50it/s]
0.0011498848292742368 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']:  95%|█████████▌| 19/20 [00:15<00:00,  1.50it/s]
0.0011498848292742368 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 100%|██████████| 20/20 [00:15<00:00,  1.43it/s]
0.0011498848292742368 plot_torch_export_cus_p2.onnx ['CUDAExecutionProvider', 'CPUExecutionProvider']: 100%|██████████| 20/20 [00:15<00:00,  1.30it/s]
                             name  ... warmup_time
0   plot_torch_export_dynamo.onnx  ...    0.000406
1   plot_torch_export_dynamo.onnx  ...    0.000615
2   plot_torch_export_dynamo.onnx  ...    0.002523
3   plot_torch_export_dynamo.onnx  ...    0.002792
4   plot_torch_export_cus_p0.onnx  ...    0.000545
5   plot_torch_export_cus_p0.onnx  ...    0.000352
6   plot_torch_export_cus_p0.onnx  ...    0.001793
7   plot_torch_export_cus_p0.onnx  ...    0.001892
8   plot_torch_export_script.onnx  ...    0.000398
9   plot_torch_export_script.onnx  ...    0.000582
10  plot_torch_export_script.onnx  ...    0.001712
11  plot_torch_export_script.onnx  ...    0.003696
12  plot_torch_export_dynopt.onnx  ...    0.000560
13  plot_torch_export_dynopt.onnx  ...    0.000804
14  plot_torch_export_dynopt.onnx  ...    0.001777
15  plot_torch_export_dynopt.onnx  ...    0.001868
16  plot_torch_export_cus_p2.onnx  ...    0.000366
17  plot_torch_export_cus_p2.onnx  ...    0.000347
18  plot_torch_export_cus_p2.onnx  ...    0.001732
19  plot_torch_export_cus_p2.onnx  ...    0.001814

[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")
Compares onnxruntime time on exported models, CPU, CUDA
compute       CPU                CUDA
aot             0         1         0         1
export
cus_p0   0.000064  0.000067  0.001004  0.000868
cus_p2   0.000063  0.000061  0.001150  0.000732
dynamo   0.000056  0.000066  0.000889  0.000854
dynopt   0.000072  0.000061  0.000906  0.000851
script   0.000073  0.000065  0.000730  0.000822

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")
Compares onnxruntime time on exported models Hide dynamo without AOT, CPU, CUDA

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",
)
Compares onnxruntime loading time and first run on exported models, CPU, CUDA
compute       CPU                CUDA
aot             0         1         0         1
export
cus_p0   0.004913  0.005725  0.028890  0.028765
cus_p2   0.004755  0.005781  0.026039  0.024075
dynamo   0.004361  0.020478  0.027552  0.043551
dynopt   0.006196  0.005832  0.026476  0.025074
script   0.005761  0.004581  0.024671  0.026444

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 Consumption of onnxruntime loading time running on CPU, Memory peak (Mb), Memory peak - memory begin (Mb), Memory average - memory begin (Mb), GPU Memory peak (Mb), GPU Memory peak - memory begin (Mb), GPU Memory average - memory begin (Mb)
  • Memory Consumption of onnxruntime loading time running on CUDA, Memory peak (Mb), Memory peak - memory begin (Mb), Memory average - memory begin (Mb), GPU Memory peak (Mb), GPU Memory peak - memory begin (Mb), GPU Memory average - memory begin (Mb)

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 Consumption of onnxruntime first running time running on CPU, Memory peak (Mb), Memory peak - memory begin (Mb), Memory average - memory begin (Mb), GPU Memory peak (Mb), GPU Memory peak - memory begin (Mb), GPU Memory average - memory begin (Mb)
  • Memory Consumption of onnxruntime first running time running on CUDA, Memory peak (Mb), Memory peak - memory begin (Mb), Memory average - memory begin (Mb), GPU Memory peak (Mb), GPU Memory peak - memory begin (Mb), GPU Memory average - memory begin (Mb)

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")
  • Memory Consumption of onnxruntime running time running on CPU, Memory peak (Mb), Memory peak - memory begin (Mb), Memory average - memory begin (Mb), GPU Memory peak (Mb), GPU Memory peak - memory begin (Mb), GPU Memory average - memory begin (Mb)
  • Memory Consumption of onnxruntime running time running on CUDA, Memory peak (Mb), Memory peak - memory begin (Mb), Memory average - memory begin (Mb), GPU Memory peak (Mb), GPU Memory peak - memory begin (Mb), GPU Memory average - memory begin (Mb)

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
input: name='input' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='_onx_concat_max_pool2d_1::Shape:1' type=int64 shape=(2,) -- array([ 1, -1])-- GraphBuilder.constant_folding.from/fold(init7_s1_-1,max_pool2d_1::Shape:1)##max_pool2d_1::Shape:1/##init7_s1_-1/Opset.make_node.1/Shape
init: name='GemmTransposePattern--p_fc1_weight::T10' type=float32 shape=(512, 16)
init: name='GemmTransposePattern--p_fc2_weight::T10' type=float32 shape=(128, 512)
init: name='GemmTransposePattern--p_fc3_weight::T10' type=float32 shape=(10, 128)
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_2' 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,)
init: name='fc2.bias' type=float32 shape=(128,)
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
  MaxPool[com.microsoft.nchwc](reorder_token_0, auto_pad=b'NOTSET', storage_order=0, ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> reorder_token_1
    Conv[com.microsoft.nchwc](reorder_token_1, reorder_token_2, 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
      MaxPool[com.microsoft.nchwc](reorder_token_3, auto_pad=b'NOTSET', storage_order=0, ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> reorder_token_4
        ReorderOutput[com.microsoft.nchwc](reorder_token_4, channels_last=0, channels=16) -> max_pool2d_1
          Reshape(max_pool2d_1, _onx_concat_max_pool2d_1::Shape:1, allowzero=0) -> flatten
            FusedGemm[com.microsoft](flatten, GemmTransposePattern--p_fc1_weight::T10, fc1.bias, transA=0, alpha=1.00, activation=b'Relu', transB=1, beta=1.00) -> relu_2
              FusedGemm[com.microsoft](relu_2, GemmTransposePattern--p_fc2_weight::T10, fc2.bias, transA=0, alpha=1.00, activation=b'Relu', transB=1, beta=1.00) -> relu_3
                Gemm(relu_3, GemmTransposePattern--p_fc3_weight::T10, fc3.bias, transA=0, alpha=1.00, transB=1, beta=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
input: name='input' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='_onx_concat_max_pool2d_1::Shape:1' type=int64 shape=(2,) -- array([ 1, -1])-- GraphBuilder.constant_folding.from/fold(init7_s1_-1,max_pool2d_1::Shape:1)##max_pool2d_1::Shape:1/##init7_s1_-1/Opset.make_node.1/Shape
init: name='GemmTransposePattern--p_fc1_weight::T10' type=float32 shape=(512, 16)
init: name='GemmTransposePattern--p_fc2_weight::T10' type=float32 shape=(128, 512)
init: name='GemmTransposePattern--p_fc3_weight::T10' type=float32 shape=(10, 128)
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_2' 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,)
init: name='fc2.bias' type=float32 shape=(128,)
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
  MaxPool[com.microsoft.nchwc](reorder_token_0, auto_pad=b'NOTSET', storage_order=0, ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> reorder_token_1
    Conv[com.microsoft.nchwc](reorder_token_1, reorder_token_2, 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
      MaxPool[com.microsoft.nchwc](reorder_token_3, auto_pad=b'NOTSET', storage_order=0, ceil_mode=0, dilations=[1,1], kernel_shape=[2,2], pads=[0,0,0,0], strides=[2,2]) -> reorder_token_4
        ReorderOutput[com.microsoft.nchwc](reorder_token_4, channels_last=0, channels=16) -> max_pool2d_1
          Reshape(max_pool2d_1, _onx_concat_max_pool2d_1::Shape:1, allowzero=0) -> flatten
            FusedGemm[com.microsoft](flatten, GemmTransposePattern--p_fc1_weight::T10, fc1.bias, transA=0, alpha=1.00, activation=b'Relu', transB=1, beta=1.00) -> relu_2
              FusedGemm[com.microsoft](relu_2, GemmTransposePattern--p_fc2_weight::T10, fc2.bias, transA=0, alpha=1.00, activation=b'Relu', transB=1, beta=1.00) -> relu_3
                Gemm(relu_3, GemmTransposePattern--p_fc3_weight::T10, fc3.bias, transA=0, alpha=1.00, transB=1, beta=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='' version=20
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.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_5' 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
  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_5, allowzero=1) -> view
            FusedGemm[com.microsoft](view, fc1.weight, fc1.bias, transA=0, alpha=1.00, activation=b'Relu', transB=1, beta=1.00) -> relu_2
              FusedGemm[com.microsoft](relu_2, fc2.weight, fc2.bias, transA=0, alpha=1.00, activation=b'Relu', transB=1, beta=1.00) -> relu_3
                Gemm(relu_3, fc3.weight, fc3.bias, transA=0, alpha=1.00, transB=1, beta=1.00) -> linear_2
output: name='linear_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='' version=20
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.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_5' 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
  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_5, allowzero=1) -> view
            FusedGemm[com.microsoft](view, fc1.weight, fc1.bias, transA=0, alpha=1.00, activation=b'Relu', transB=1, beta=1.00) -> relu_2
              FusedGemm[com.microsoft](relu_2, fc2.weight, fc2.bias, transA=0, alpha=1.00, activation=b'Relu', transB=1, beta=1.00) -> relu_3
                Gemm(relu_3, fc3.weight, fc3.bias, transA=0, alpha=1.00, transB=1, beta=1.00) -> linear_2
output: name='linear_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
input: name='input' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='_onx_concat_max_pool2d_1::Shape:1' type=int64 shape=(2,) -- array([ 1, -1])-- GraphBuilder.constant_folding.from/fold(init7_s1_-1,max_pool2d_1::Shape:1)##max_pool2d_1::Shape:1/##init7_s1_-1/Opset.make_node.1/Shape
init: name='GemmTransposePattern--p_fc1_weight::T10' type=float32 shape=(512, 16)
init: name='GemmTransposePattern--p_fc2_weight::T10' type=float32 shape=(128, 512)
init: name='GemmTransposePattern--p_fc3_weight::T10' type=float32 shape=(10, 128)
init: name='conv1.weight' type=float32 shape=(16, 1, 5, 5)
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)
init: name='conv2.bias' type=float32 shape=(16,)                      -- DynamoInterpret.placeholder.1/P(conv2.bias)
init: name='fc1.bias' type=float32 shape=(512,)
init: name='fc2.bias' type=float32 shape=(128,)
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]) -> max_pool2d
      Conv(max_pool2d, 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]) -> max_pool2d_1
            Reshape(max_pool2d_1, _onx_concat_max_pool2d_1::Shape:1) -> flatten
              Gemm(flatten, GemmTransposePattern--p_fc1_weight::T10, fc1.bias, transB=1) -> linear
                Relu(linear) -> relu_2
                  Gemm(relu_2, GemmTransposePattern--p_fc2_weight::T10, fc2.bias, transB=1) -> linear_1
                    Relu(linear_1) -> relu_3
                      Gemm(relu_3, GemmTransposePattern--p_fc3_weight::T10, 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
input: name='input' type=dtype('float32') shape=[1, 1, 16, 16]
init: name='_onx_concat_max_pool2d_1::Shape:1' type=int64 shape=(2,) -- array([ 1, -1])-- GraphBuilder.constant_folding.from/fold(init7_s1_-1,max_pool2d_1::Shape:1)##max_pool2d_1::Shape:1/##init7_s1_-1/Opset.make_node.1/Shape
init: name='GemmTransposePattern--p_fc1_weight::T10' type=float32 shape=(512, 16)
init: name='GemmTransposePattern--p_fc2_weight::T10' type=float32 shape=(128, 512)
init: name='GemmTransposePattern--p_fc3_weight::T10' type=float32 shape=(10, 128)
init: name='conv1.weight' type=float32 shape=(16, 1, 5, 5)
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)
init: name='conv2.bias' type=float32 shape=(16,)                      -- DynamoInterpret.placeholder.1/P(conv2.bias)
init: name='fc1.bias' type=float32 shape=(512,)
init: name='fc2.bias' type=float32 shape=(128,)
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]) -> max_pool2d
      Conv(max_pool2d, 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]) -> max_pool2d_1
            Reshape(max_pool2d_1, _onx_concat_max_pool2d_1::Shape:1) -> flatten
              Gemm(flatten, GemmTransposePattern--p_fc1_weight::T10, fc1.bias, transB=1) -> linear
                Relu(linear) -> relu_2
                  Gemm(relu_2, GemmTransposePattern--p_fc2_weight::T10, fc2.bias, transB=1) -> linear_1
                    Relu(linear_1) -> relu_3
                      Gemm(relu_3, GemmTransposePattern--p_fc3_weight::T10, 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='' version=20
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_5' type=int64 shape=(2,) -- array([ 1, 16])
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_5, allowzero=1) -> view
              Gemm(view, fc1.weight, fc1.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> linear
                Relu(linear) -> relu_2
                  Gemm(relu_2, fc2.weight, fc2.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> linear_1
                    Relu(linear_1) -> relu_3
                      Gemm(relu_3, fc3.weight, fc3.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> linear_2
output: name='linear_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='' version=20
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_5' type=int64 shape=(2,) -- array([ 1, 16])
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_5, allowzero=1) -> view
              Gemm(view, fc1.weight, fc1.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> linear
                Relu(linear) -> relu_2
                  Gemm(relu_2, fc2.weight, fc2.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> linear_1
                    Relu(linear_1) -> relu_3
                      Gemm(relu_3, fc3.weight, fc3.bias, beta=1.00, transB=1, alpha=1.00, transA=0) -> linear_2
output: name='linear_2' type=dtype('float32') shape=[1, 10]

Total running time of the script: (1 minutes 15.211 seconds)

Related examples

101: Profile an existing model with onnxruntime

101: Profile an existing model with onnxruntime

101: Onnx Model Optimization based on Pattern Rewriting

101: Onnx Model Optimization based on Pattern Rewriting

201: Use torch to export a scikit-learn model into ONNX

201: Use torch to export a scikit-learn model into ONNX

101: A custom backend for torch

101: A custom backend for torch

101: Some dummy examples with torch.export.export

101: Some dummy examples with torch.export.export

Gallery generated by Sphinx-Gallery