201: Evaluate DORT Training

It compares DORT to eager mode and onnxrt backend.

To run the script:

python _doc/examples/plot_torch_aot --help

Some helpers

import warnings

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 torch._dynamo
import contextlib
import itertools
import os
import gc
import platform

# import pickle
import pprint
import multiprocessing
import time
import cProfile
import pstats
import io
import logging
from pstats import SortKey

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.plotting.memory import memory_peak_plot
from experimental_experiment.ext_test_case import measure_time, get_figure
from experimental_experiment.args import get_parsed_args
from experimental_experiment.memory_peak import start_spying_on
from experimental_experiment.torch_models.training_helper import make_aot_ort
from tqdm import tqdm

has_cuda = has_cuda and torch.cuda.is_available()
logging.disable(logging.ERROR)


def system_info():
    obs = {}
    obs["processor"] = platform.processor()
    obs["cores"] = multiprocessing.cpu_count()
    try:
        obs["cuda"] = 1 if torch.cuda.is_available() else 0
        obs["cuda_count"] = torch.cuda.device_count()
        obs["cuda_name"] = torch.cuda.get_device_name()
        obs["cuda_capa"] = torch.cuda.get_device_capability()
    except (RuntimeError, AssertionError):
        # no cuda
        pass
    return obs


pprint.pprint(system_info())
{'cores': 20,
 'cuda': 1,
 'cuda_capa': (8, 9),
 'cuda_count': 1,
 'cuda_name': 'NVIDIA GeForce RTX 4060 Laptop GPU',
 'processor': 'x86_64'}

Scripts arguments

script_args = get_parsed_args(
    "plot_torch_aot",
    description=__doc__,
    scenarios={
        "small": "small model to test",
        "middle": "55Mb model",
        "large": "1Gb model",
    },
    warmup=5,
    repeat=5,
    repeat1=(1, "repeat for the first iteration"),
    maxtime=(
        2,
        "maximum time to run a model to measure the computation time, "
        "it is 0.1 when scenario is small",
    ),
    expose="scenarios,repeat,repeat1,warmup",
)

if script_args.scenario in (None, "small"):
    script_args.maxtime = 0.1
print(f"scenario={script_args.scenario or 'small'}")
print(f"warmup={script_args.warmup}")
print(f"repeat={script_args.repeat}")
print(f"repeat1={script_args.repeat1}")
print(f"maxtime={script_args.maxtime}")
scenario=small
warmup=5
repeat=5
repeat1=1
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, 32, 5)
            # self.conv2 = nn.Conv2d(128, 16, 5)
            self.fc1 = nn.Linear(30752, 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(144, 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, 32, 5)
            # self.conv2 = nn.Conv2d(128, 16, 5)
            self.fc1 = nn.Linear(30752, 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)
            # 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)), (4, 4))
        # 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))
            # end of the loop
        x = F.relu(self.fc2(x))
        y = self.fc3(x)
        return y


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)
    y = torch.rand((1, 10)).to(torch.float32)
    model = MyModelClass(scenario=scenario)
    assert model(input_tensor) is not None
    return model, (input_tensor, y)


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_tensors = create_model_and_input()
model_size = torch_model_size(model)
print(f"model size={model_size / 2 ** 20} Mb")
model size=0.5401992797851562 Mb

Backends

def run(model, tensor_x, tensor_y):
    tensor_x = tensor_x.detach()
    tensor_y = tensor_y.detach()
    for param in model.parameters():
        param.grad = None
    try:
        output = model(tensor_x)
    except Exception as e:
        raise AssertionError(f"issue with {type(tensor_x)}") from e
    loss = F.mse_loss(output, tensor_y)

    # return loss
    def _backward_():
        loss.backward()

    _backward_()
    return loss, (param.grad for param in model.parameters())


def get_torch_eager(model, *args):
    def my_compiler(gm, example_inputs):
        return gm.forward

    with contextlib.redirect_stdout(io.StringIO()):
        with warnings.catch_warnings():
            warnings.simplefilter("ignore")
            optimized_mod = torch.compile(model, fullgraph=True, backend=my_compiler)
            assert run(optimized_mod, *args)
            return optimized_mod


def get_torch_default(model, *args):
    with contextlib.redirect_stdout(io.StringIO()):
        with warnings.catch_warnings():
            warnings.simplefilter("ignore")
            optimized_mod = torch.compile(model, fullgraph=True, mode="reduce-overhead")
            assert run(optimized_mod, *args)
            return optimized_mod


def get_torch_dort(model, *args):
    with contextlib.redirect_stdout(io.StringIO()):
        with warnings.catch_warnings():
            warnings.simplefilter("ignore")
            local_aot_ort, _ = make_aot_ort(dynamic=True, rewrite=True)
            optimized_mod = torch.compile(model, backend=local_aot_ort, fullgraph=True)
            run(optimized_mod, *args)
            assert run(optimized_mod, *args)
            return optimized_mod

Let’s check they are working.

export_functions = [
    get_torch_eager,
    get_torch_default,
    get_torch_dort,
]

exporters = {f.__name__.replace("get_", ""): f for f in export_functions}

supported_exporters = {}
for k, v in exporters.items():
    print(f"run function {k}")
    filename = f"plot_torch_aot_{k}.onnx"
    torch._dynamo.reset()
    model, input_tensors = create_model_and_input()
    try:
        run(model, *input_tensors)
    except Exception as e:
        print(f"skipped due to {str(e)[:1000]}")  # noqa: F821
        continue
    supported_exporters[k] = v
    del model
    gc.collect()
    time.sleep(1)
run function torch_eager
run function torch_default
run function torch_dort

Compile and 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 in supported_exporters:
    print(f"run compile for memory {k} on cpu")
    filename = f"plot_torch_aot_{k}.onnx"
    if has_cuda:
        torch.cuda.set_device(0)
    torch._dynamo.reset()
    # CPU
    model, input_tensors = create_model_and_input()
    stat = start_spying_on(cuda=1 if has_cuda else 0)
    run(model, *input_tensors)
    obs = flatten(stat.stop())
    print("done.")
    obs.update(dict(export=k, p="cpu"))
    data.append(obs)
    del model
    gc.collect()
    time.sleep(1)

    if not has_cuda:
        continue
    torch._dynamo.reset()
    # CUDA
    model, input_tensors = create_model_and_input()
    model = model.cuda()
    input_tensors = [i.cuda() for i in input_tensors]
    print(f"run compile for memory {k} on cuda")
    stat = start_spying_on(cuda=1 if has_cuda else 0)
    run(model, *input_tensors)
    obs = flatten(stat.stop())
    print("done.")
    obs.update(dict(export=k, p="cuda"))
    data.append(obs)
    del model
    gc.collect()
    time.sleep(1)
run compile for memory torch_eager on cpu
done.
run compile for memory torch_eager on cuda
done.
run compile for memory torch_default on cpu
done.
run compile for memory torch_default on cuda
done.
run compile for memory torch_dort on cpu
done.
run compile for memory torch_dort on cuda
done.

The result.

df1 = pandas.DataFrame(data)
df1.to_csv("plot_torch_aot_1_memory.csv", index=False)
df1.to_excel("plot_torch_aot_1_memory.xlsx", index=False)
print(df1)

for p in ["cpu", "cuda"]:
    if not has_cuda and p == "cuda":
        continue
    ax = memory_peak_plot(
        df1[df1["p"] == p],
        key=("export",),
        bars=[model_size * i / 2**20 for i in range(1, 5)],
        suptitle=f"Memory Consumption of the Compilation on {p}\n"
        f"model size={model_size / 2**20:1.0f} Mb",
    )
    get_figure(ax).savefig(f"plot_torch_aot_1_memory_{p}.png")
  • Memory Consumption of the Compilation on cpu model size=1 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)
  • Memory Consumption of the Compilation on cuda model size=1 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        begin          end   gpu0_peak   gpu0_mean  gpu0_n  gpu0_begin    gpu0_end         export     p
0  3103.039062  3101.729167   3  3101.089844  3101.058594  552.617188  552.617188       3  552.617188  552.617188    torch_eager   cpu
1  3197.410156  3140.345459  48  3101.042969  3197.410156  560.617188  553.950521      48  552.617188  560.617188    torch_eager  cuda
2  3199.574219  3198.843750   3  3199.394531  3199.574219  560.617188  560.617188       3  560.617188  560.617188  torch_default   cpu
3  3197.902344  3197.750000   2  3197.597656  3197.902344  560.617188  560.617188       2  560.617188  560.617188  torch_default  cuda
4  3197.898438  3197.895833   3  3197.898438  3197.898438  560.617188  560.617188       3  560.617188  560.617188     torch_dort   cpu
5  3197.902344  3197.902344   2  3197.902344  3197.902344  560.617188  560.617188       2  560.617188  560.617188     torch_dort  cuda

dort first iteration speed

data = []

for k in supported_exporters:
    print(f"run dort cpu {k}: {script_args.repeat1}")
    times = []
    for _ in range(int(script_args.repeat1)):
        model, input_tensors = create_model_and_input()
        torch._dynamo.reset()
        begin = time.perf_counter()
        run(model, *input_tensors)
        duration = time.perf_counter() - begin
        times.append(duration)
        del model
        gc.collect()
        time.sleep(1)

    print(f"done: {times[-1]}")
    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),
            p="cpu",
        )
    )

    if not has_cuda:
        continue

    print(f"run dort cuda {k}: {script_args.repeat1}")
    times = []
    for i in range(int(script_args.repeat1)):
        model, input_tensors = create_model_and_input()
        model = model.cuda()
        input_tensors = [i.cuda() for i in input_tensors]
        torch._dynamo.reset()
        begin = time.perf_counter()
        run(model, *input_tensors)
        duration = time.perf_counter() - begin
        times.append(duration)
        del model
        gc.collect()
        time.sleep(1)

    print(f"done: {times[-1]}")
    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),
            p="cuda",
        )
    )
run dort cpu torch_eager: 1
done: 0.014885390999552328
run dort cuda torch_eager: 1
done: 0.003067197998461779
run dort cpu torch_default: 1
done: 0.007820370999979787
run dort cuda torch_default: 1
done: 0.0023014689977571834
run dort cpu torch_dort: 1
done: 0.006999388999247458
run dort cuda torch_dort: 1
done: 0.004341561001638183

The result.

df1 = pandas.DataFrame(data)
df1.to_csv("plot_torch_aot_1_time.csv", index=False)
df1.to_excel("plot_torch_aot_1_time.xlsx", index=False)
print(df1)

fig, ax = plt.subplots(1, 1)
dfi = df1[["export", "p", "time", "std"]].set_index(["export", "p"])
dfi["time"].plot.bar(ax=ax, title="Compilation time", yerr=dfi["std"], rot=30)
fig.tight_layout()
fig.savefig("plot_torch_aot_1_time.png")
Compilation time
          export      time       min       max     first      last  std     p
0    torch_eager  0.014885  0.014885  0.014885  0.014885  0.014885  0.0   cpu
1    torch_eager  0.003067  0.003067  0.003067  0.003067  0.003067  0.0  cuda
2  torch_default  0.007820  0.007820  0.007820  0.007820  0.007820  0.0   cpu
3  torch_default  0.002301  0.002301  0.002301  0.002301  0.002301  0.0  cuda
4     torch_dort  0.006999  0.006999  0.006999  0.006999  0.006999  0.0   cpu
5     torch_dort  0.004342  0.004342  0.004342  0.004342  0.004342  0.0  cuda

Compilation 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, with_args=True, verbose=False, suffix="export"):
    if verbose:
        print(f"profile {name}: {export_function}")
    if with_args:
        model, input_tensors = create_model_and_input()
        export_function(model, input_tensors)
        pr = cProfile.Profile()
        pr.enable()
        for _ in range(int(script_args.repeat1)):
            export_function(model, input_tensors)
        pr.disable()
    else:
        pr = cProfile.Profile()
        pr.enable()
        for _ in range(int(script_args.repeat1)):
            export_function()
        pr.disable()
    s = io.StringIO()
    sortby = SortKey.CUMULATIVE
    ps = pstats.Stats(pr, stream=s).sort_stats(sortby)
    ps.print_stats()
    # with open(f"plot_torch_aot_profile_{name}_{suffix}.pickle", "wb") as f:
    #     pickle.dump(ps, f)

    raw = s.getvalue()
    text = "\n".join(raw.split("\n")[:200])
    if verbose:
        print(text)
    with open(f"plot_torch_aot_profile_{name}_{suffix}.txt", "w") as f:
        f.write(raw)

    root, nodes = profile2graph(ps, clean_text=clean_text)
    text = root.to_text()
    with open(f"plot_torch_aot_profile_{name}_{suffix}_h.txt", "w") as f:
        f.write(text)
    if verbose:
        print("done.")


model, input_tensors = create_model_and_input()


def function_to_profile(model=model, input_tensors=input_tensors):
    return get_torch_dort(model, *input_tensors)


profile_function("dort", function_to_profile, verbose=True, suffix="1")
profile dort: <function function_to_profile at 0x7f00a0713130>
         1382362 function calls (1346683 primitive calls) in 1.256 seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    1.388    1.388 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_aot_201.py:512(function_to_profile)
        1    0.000    0.000    1.388    1.388 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_aot_201.py:260(get_torch_dort)
        2    0.000    0.000    0.935    0.467 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_aot_201.py:220(run)
     11/4    0.000    0.000    0.836    0.209 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/nn/modules/module.py:1732(_wrapped_call_impl)
     11/4    0.001    0.000    0.836    0.209 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/nn/modules/module.py:1740(_call_impl)
      7/6    0.000    0.000    0.684    0.114 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py:715(_fn)
        3    0.000    0.000    0.529    0.176 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/autograd/graph.py:816(_engine_run_backward)
        3    0.004    0.001    0.529    0.176 {method 'run_backward' of 'torch._C._EngineBase' objects}
        2    0.000    0.000    0.491    0.245 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py:523(_fn)
        1    0.000    0.000    0.452    0.452 /home/xadupre/github/experimental-experiment/experimental_experiment/torch_models/training_helper.py:7(make_aot_ort)
        1    0.000    0.000    0.451    0.451 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.py:763(__init__)
      6/4    0.000    0.000    0.447    0.112 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py:116(call_func_at_runtime_with_args)
     12/4    0.001    0.000    0.443    0.111 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph_module.py:821(call_wrapped)
        4    0.000    0.000    0.443    0.111 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph_module.py:382(__call__)
        2    0.000    0.000    0.443    0.221 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_aot_201.py:232(_backward_)
        2    0.000    0.000    0.443    0.221 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_tensor.py:568(backward)
        2    0.000    0.000    0.442    0.221 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/autograd/__init__.py:242(backward)
        2    0.000    0.000    0.441    0.220 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/autograd/function.py:292(apply)
        2    0.000    0.000    0.441    0.220 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:1645(backward)
        2    0.000    0.000    0.440    0.220 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:1912(call_compiled_backward)
        1    0.000    0.000    0.392    0.392 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/_exporter_legacy.py:308(__init__)
        1    0.000    0.000    0.388    0.388 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:1331(__call__)
        1    0.000    0.000    0.388    0.388 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:449(__call__)
        1    0.000    0.000    0.388    0.388 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:597(_compile)
        1    0.000    0.000    0.387    0.387 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:689(compile_inner)
        1    0.000    0.000    0.386    0.386 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_utils_internal.py:89(wrapper_function)
        1    0.000    0.000    0.386    0.386 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:709(_compile_inner)
        1    0.000    0.000    0.363    0.363 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/bytecode_transformation.py:1329(transform_code_object)
        1    0.000    0.000    0.361    0.361 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:204(_fn)
        1    0.000    0.000    0.360    0.360 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:632(transform)
        8    0.002    0.000    0.359    0.045 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.py:884(_ort_acclerated_call)
        1    0.000    0.000    0.359    0.359 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:2907(run)
      6/1    0.000    0.000    0.359    0.359 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:1110(run)
   100/44    0.000    0.000    0.358    0.008 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:998(step)
        2    0.000    0.000    0.345    0.172 <eval_with_key>.244:4(forward)
        1    0.000    0.000    0.334    0.334 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/_exporter_legacy.py:95(__init__)
        1    0.001    0.001    0.334    0.334 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/_exporter_legacy.py:123(_initiate_registry_from_torchlib)
        1    0.007    0.007    0.331    0.331 /home/xadupre/github/onnxscript/onnxscript/_framework_apis/torch_2_5.py:107(get_torchlib_ops)
      184    0.002    0.000    0.323    0.002 /home/xadupre/github/onnxscript/onnxscript/values.py:640(function_ir)
        1    0.000    0.000    0.314    0.314 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:3098(RETURN_VALUE)
        1    0.000    0.000    0.314    0.314 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:3070(_return)
        1    0.000    0.000    0.314    0.314 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/output_graph.py:989(compile_subgraph)
        1    0.000    0.000    0.313    0.313 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/output_graph.py:1324(compile_and_call_fx_graph)
        1    0.000    0.000    0.307    0.307 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/output_graph.py:1444(call_user_compiler)
        1    0.000    0.000    0.307    0.307 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/output_graph.py:1450(_call_user_compiler)
      2/1    0.000    0.000    0.307    0.307 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/repro/after_dynamo.py:73(__call__)
        1    0.000    0.000    0.307    0.307 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/__init__.py:2318(__call__)
        1    0.000    0.000    0.307    0.307 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.py:1153(__call__)
        1    0.000    0.000    0.307    0.307 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/backends/common.py:23(__call__)
        1    0.000    0.000    0.306    0.306 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py:940(aot_module_simplified)
        1    0.000    0.000    0.301    0.301 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py:1061(dispatch_and_compile)
        1    0.000    0.000    0.301    0.301 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py:516(create_aot_dispatcher_function)
        1    0.000    0.000    0.301    0.301 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py:529(_create_aot_dispatcher_function)
        1    0.000    0.000    0.251    0.251 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py:337(aot_dispatch_autograd)
      184    0.001    0.000    0.215    0.001 /home/xadupre/github/onnxscript/onnxscript/_internal/ast_utils.py:16(get_src_and_ast)
      185    0.000    0.000    0.179    0.001 /usr/lib/python3.10/inspect.py:1133(getsource)
      185    0.005    0.000    0.178    0.001 /usr/lib/python3.10/inspect.py:1112(getsourcelines)
  593/496    0.001    0.000    0.164    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_ops.py:722(__call__)
  890/570    0.002    0.000    0.160    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/utils/_stats.py:16(wrapper)
        1    0.000    0.000    0.160    0.160 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/dispatch_and_compile_graph.py:234(aot_dispatch_autograd_graph)
      184    0.026    0.000    0.159    0.001 /usr/lib/python3.10/inspect.py:1101(getblock)
      155    0.007    0.000    0.148    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_subclasses/functional_tensor.py:372(__torch_dispatch__)
        1    0.000    0.000    0.145    0.145 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/dispatch_and_compile_graph.py:46(_create_graph)
        1    0.000    0.000    0.145    0.145 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:2170(wrapped)
        1    0.000    0.000    0.145    0.145 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:2108(trace)
        1    0.000    0.000    0.145    0.145 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:1999(_trace_inner)
        1    0.000    0.000    0.144    0.144 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_compile.py:22(inner)
        1    0.000    0.000    0.144    0.144 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:1131(dispatch_trace)
        1    0.000    0.000    0.141    0.141 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:711(trace)
        4    0.000    0.000    0.140    0.035 /home/xadupre/github/onnxscript/onnxscript/optimizer/__init__.py:15(optimize)
        4    0.000    0.000    0.140    0.035 /home/xadupre/github/onnxscript/onnxscript/optimizer/_legacy/_optimizer.py:24(optimize)
        1    0.000    0.000    0.136    0.136 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:698(flatten_fn)
   305/12    0.006    0.000    0.135    0.011 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/diagnostics/infra/decorator.py:66(wrapper)
        1    0.000    0.000    0.135    0.135 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:1181(wrapped)
        1    0.000    0.000    0.132    0.132 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:663(inner_fn)
        1    0.000    0.000    0.131    0.131 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:643(joint_helper)
        1    0.000    0.000    0.131    0.131 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:396(_functionalized_f_helper)
        6    0.001    0.000    0.120    0.020 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/interpreter.py:117(run)
      429    0.001    0.000    0.120    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:1230(__torch_function__)
    26710    0.070    0.000    0.119    0.000 /usr/lib/python3.10/tokenize.py:431(_tokenize)
        1    0.000    0.000    0.119    0.119 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:276(inner_fn_with_anomaly)
        1    0.000    0.000    0.119    0.119 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:193(inner_fn)
        2    0.026    0.013    0.112    0.056 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/decomposition_table.py:14(_create_onnx_supports_op_overload_table)
        2    0.000    0.000    0.110    0.055 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.py:1099(compile)
        2    0.000    0.000    0.109    0.055 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/passes/infra/partitioner.py:370(partition_and_fuse)
  591/586    0.002    0.000    0.106    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1242(__torch_dispatch__)
        2    0.000    0.000    0.104    0.052 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/passes/infra/partitioner.py:283(fuse_partitions)
        2    0.000    0.000    0.104    0.052 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/passes/utils/fuser_utils.py:244(fuse_by_partitions)
  591/586    0.005    0.000    0.103    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1768(dispatch)
        2    0.000    0.000    0.102    0.051 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_aot_201.py:165(forward)
        2    0.000    0.000    0.102    0.051 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py:1113(forward)
        2    0.000    0.000    0.102    0.051 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:284(runtime_wrapper)
      4/2    0.000    0.000    0.101    0.051 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py:99(g)
        2    0.000    0.000    0.101    0.051 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/autograd/function.py:559(apply)
        2    0.000    0.000    0.101    0.050 {built-in method apply}
        2    0.000    0.000    0.101    0.050 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:1520(forward)
        5    0.000    0.000    0.100    0.020 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/overrides.py:1668(handle_torch_function)
        2    0.000    0.000    0.100    0.050 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:474(wrapper)
        2    0.000    0.000    0.100    0.050 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:659(inner_fn)
      184    0.000    0.000    0.098    0.001 /home/xadupre/github/onnxscript/onnxscript/converter.py:1463(translate_function_signature)
        2    0.000    0.000    0.098    0.049 <eval_with_key>.240:4(forward)
      130    0.001    0.000    0.098    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/interpreter.py:210(run_node)
      184    0.007    0.000    0.097    0.001 /home/xadupre/github/onnxscript/onnxscript/converter.py:1378(_translate_function_signature_common)
      212    0.002    0.000    0.096    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1326(_cached_dispatch_impl)
        1    0.000    0.000    0.094    0.094 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/backends/common.py:49(_wrapped_bw_compiler)
        1    0.000    0.000    0.093    0.093 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/_lazy_graph_module.py:115(_lazy_forward)
       82    0.000    0.000    0.092    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/interpreter.py:288(call_function)
      2/1    0.000    0.000    0.087    0.087 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/autograd/__init__.py:358(grad)
  297/276    0.001    0.000    0.084    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:1328(__torch_dispatch__)
    69/54    0.003    0.000    0.075    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:761(proxy_call)
186085/180763    0.035    0.000    0.074    0.000 {built-in method builtins.isinstance}
        8    0.000    0.000    0.069    0.009 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/_pass.py:240(run)
        4    0.000    0.000    0.066    0.017 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/passes/type_promotion.py:1696(_run)
       32    0.000    0.000    0.066    0.002 /home/xadupre/github/onnxscript/onnxscript/_legacy_ir/visitor.py:786(visit_model)
        4    0.000    0.000    0.065    0.016 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/fx_onnx_interpreter.py:463(run)
       12    0.000    0.000    0.060    0.005 /home/xadupre/github/onnxscript/onnxscript/rewriter/__init__.py:28(rewrite)
       89    0.001    0.000    0.060    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/fx_onnx_interpreter.py:388(run_node)
       86    0.001    0.000    0.058    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/passes/type_promotion.py:1601(run_node)
        1    0.001    0.001    0.057    0.057 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/decomposition_table.py:73(create_onnx_friendly_decomposition_table)
7575/1820    0.010    0.000    0.057    0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:131(is_value_type)
        1    0.000    0.000    0.056    0.056 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/partitioners.py:1779(min_cut_rematerialization_partition)
        2    0.000    0.000    0.056    0.028 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:848(functional_call)
       38    0.000    0.000    0.054    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/experimental/symbolic_shapes.py:6471(run_node)
       61    0.001    0.000    0.054    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/fx_onnx_interpreter.py:604(call_function)
       32    0.001    0.000    0.054    0.002 /home/xadupre/github/onnxscript/onnxscript/_legacy_ir/visitor.py:646(visit_graph)
        4    0.001    0.000    0.054    0.013 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/passes/utils/fuser_utils.py:95(fuse_as_graphmodule)
      590    0.005    0.000    0.053    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph.py:1104(create_node)
       11    0.000    0.000    0.050    0.005 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph.py:1562(python_code)
    16562    0.008    0.000    0.050    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/_exporter_legacy.py:210(is_registered_op)
     1573    0.050    0.000    0.050    0.000 {built-in method builtins.compile}
  331/147    0.001    0.000    0.049    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/utils/_pytree.py:923(tree_map)
      397    0.002    0.000    0.047    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph.py:1493(node_copy)
  593/536    0.002    0.000    0.047    0.000 /home/xadupre/github/onnxscript/onnxscript/_legacy_ir/visitor.py:799(visit_node)
  133/104    0.001    0.000    0.045    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/nn/modules/module.py:1935(__setattr__)
        8    0.000    0.000    0.045    0.006 /home/xadupre/github/onnxscript/onnxscript/optimizer/_legacy/constant_folding.py:270(fold_constants)
        8    0.000    0.000    0.045    0.006 /home/xadupre/github/onnxscript/onnxscript/optimizer/_legacy/constant_folding.py:264(visit_model)
        8    0.000    0.000    0.045    0.006 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph_module.py:792(recompile)
  5942/63    0.003    0.000    0.044    0.001 /home/xadupre/github/onnxscript/onnxscript/ir/serde.py:94(wrapper)
       13    0.000    0.000    0.043    0.003 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph_module.py:548(graph)
 1513/224    0.004    0.000    0.043    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/utils/_pytree.py:801(unflatten)
    16621    0.011    0.000    0.043    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/_exporter_legacy.py:188(get_op_functions)
      198    0.001    0.000    0.043    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1701(_output_from_cache_entry)
11524/5330    0.020    0.000    0.042    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/node.py:883(map_aggregate)
      202    0.005    0.000    0.041    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1635(_get_output_tensor_from_cache_entry)
       11    0.000    0.000    0.041    0.004 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph.py:1639(_python_code)
       11    0.004    0.000    0.041    0.004 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph.py:408(_gen_python_code)
     14/9    0.000    0.000    0.039    0.004 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:627(wrapper)
      212    0.001    0.000    0.039    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1369(_cache_key)
     14/9    0.000    0.000    0.039    0.004 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:1728(CALL_FUNCTION)
     14/9    0.000    0.000    0.039    0.004 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:941(call_function)
     6208    0.013    0.000    0.038    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/node.py:854(__setattr__)
      314    0.001    0.000    0.038    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/utils/_pytree.py:1130(tree_map_only)
        1    0.000    0.000    0.038    0.038 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/collect_metadata_analysis.py:171(inner)
      176    0.002    0.000    0.038    0.000 /home/xadupre/github/onnxscript/onnxscript/optimizer/_legacy/constant_folding.py:165(process_node)
     1024    0.001    0.000    0.037    0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:172(is_valid_type)
       10    0.000    0.000    0.037    0.004 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph_module.py:437(__init__)
7059/6619    0.004    0.000    0.036    0.000 {built-in method builtins.next}
  804/219    0.005    0.000    0.035    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1445(_prep_args_for_hash)
     3697    0.004    0.000    0.034    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/node.py:874(map_arg)
     17/6    0.000    0.000    0.033    0.006 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/variables/lazy.py:166(realize_and_forward)
        4    0.000    0.000    0.033    0.008 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/variables/nn_module.py:850(call_function)
    25891    0.032    0.000    0.032    0.000 {method 'match' of 're.Pattern' objects}
      5/4    0.000    0.000    0.031    0.008 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py:325(call_function)
      5/4    0.000    0.000    0.031    0.008 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py:119(call_function)
      5/4    0.000    0.000    0.031    0.008 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:963(inline_user_function_return)
      5/4    0.000    0.000    0.031    0.008 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:3120(inline_call)
        1    0.000    0.000    0.031    0.031 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:109(inner_fn)
  897/801    0.002    0.000    0.031    0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:150(<listcomp>)
      5/4    0.000    0.000    0.031    0.008 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:3157(inline_call_)
2800/2665    0.003    0.000    0.030    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/node.py:889(<listcomp>)
        4    0.000    0.000    0.030    0.007 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/passes/utils/common.py:27(lift_subgraph_as_module)
      184    0.000    0.000    0.028    0.000 /usr/lib/python3.10/ast.py:33(parse)
        4    0.000    0.000    0.028    0.007 /home/xadupre/github/onnxruntime/build/linux_cuda/Release/onnxruntime/capi/onnxruntime_inference_collection.py:406(__init__)
       20    0.000    0.000    0.028    0.001 /home/xadupre/github/onnxscript/onnxscript/ir/serde.py:461(deserialize_model)
        4    0.027    0.007    0.027    0.007 /home/xadupre/github/onnxruntime/build/linux_cuda/Release/onnxruntime/capi/onnxruntime_inference_collection.py:484(_create_inference_session)
       12    0.000    0.000    0.027    0.002 /home/xadupre/github/onnxscript/onnxscript/rewriter/pattern.py:1445(apply_to_model)
       12    0.001    0.000    0.025    0.002 /home/xadupre/github/onnxscript/onnxscript/rewriter/pattern.py:1412(_apply_to_graph_or_function)
      112    0.001    0.000    0.025    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/proxy.py:209(create_proxy)
       20    0.000    0.000    0.024    0.001 /home/xadupre/github/onnxscript/onnxscript/ir/serde.py:551(_deserialize_graph)
        4    0.001    0.000    0.024    0.006 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/partitioners.py:153(_extract_graph_with_inputs_outputs)
     12/9    0.002    0.000    0.023    0.003 {built-in method torch._C._nn.linear}
      614    0.004    0.000    0.023    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/node.py:367(prepend)
       59    0.000    0.000    0.023    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/onnx/_internal/fx/onnxfunction_dispatcher.py:96(dispatch)
        1    0.000    0.000    0.023    0.023 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/guards.py:2107(__init__)
     7575    0.007    0.000    0.023    0.000 /home/xadupre/github/onnxscript/onnxscript/type_annotation.py:123(_is_tensor_type)
     2553    0.002    0.000    0.022    0.000 /home/xadupre/github/onnxscript/onnxscript/rewriter/pattern.py:1284(try_rewrite)
       28    0.000    0.000    0.022    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:1869(LOAD_ATTR)
       28    0.000    0.000    0.022    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:1862(_load_attr)
       20    0.000    0.000    0.022    0.001 /home/xadupre/github/onnxscript/onnxscript/optimizer/_remove_unused_function.py:64(remove_unused_functions)
     12/9    0.000    0.000    0.022    0.002 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/nn/functional.py:1693(relu)
      370    0.004    0.000    0.022    0.000 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/fx/graph.py:634(emit_node)
2873/2654    0.002    0.000    0.022    0.000 /usr/lib/python3.10/contextlib.py:130(__enter__)
        9    0.001    0.000    0.022    0.002 {built-in method torch.relu}
    31641    0.010    0.000    0.022    0.000 {method 'get' of 'dict' objects}
    31/29    0.000    0.000    0.021    0.001 /home/xadupre/vv/this/lib/python3.10/site-packages/torch/_dynamo/variables/builtin.py:980(call_function)
done.

Benchmark exported models with ORT

def benchmark(shape):
    data = []
    data_mem_first_run = []
    data_mem_run = []
    confs = list(
        itertools.product(
            export_functions,
            ["CPU", "CUDA"],
        )
    )
    loop = tqdm(confs)
    print(f"number of experiments: {len(loop)}")
    for export_fct, p in loop:
        name = export_fct.__name__.replace("get_torch_", "")
        obs = {}  # system_info()
        obs["name"] = name
        obs["compute"] = p
        obs["export"] = name

        model, input_tensors = create_model_and_input()
        if p == "CUDA":
            if not has_cuda:
                continue
            model = model.cuda()
            input_tensors = [i.cuda() for i in input_tensors]
        try:
            exported_model = export_fct(model, *input_tensors)
        except Exception as e:
            obs["error"] = str(e)
            data.append(obs)
            continue

        def call_model(
            export_fct=export_fct,
            exported_model=exported_model,
            input_tensors=input_tensors,
        ):
            res = run(exported_model, *input_tensors)
            return res

        stat = start_spying_on(cuda=1 if has_cuda else 0)
        try:
            call_model()
        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(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):
            call_model()
        memobs = flatten(stat.stop())
        memobs.update(obs)
        data_mem_run.append(memobs)

        obs.update(
            measure_time(
                call_model,
                max_time=script_args.maxtime,
                repeat=script_args.repeat,
                number=1,
            )
        )

        profile_function(name, call_model, with_args=False, suffix=f"run_{p}")

        loop.set_description(f"{obs['average']} {name} {p}")
        data.append(obs)
        del model
        del exported_model
        gc.collect()
        time.sleep(1)

    df = pandas.DataFrame(data)
    df.to_csv("plot_torch_aot_ort_time.csv", index=False)
    df.to_excel("plot_torch_aot_ort_time.xlsx", index=False)
    dfmemr = pandas.DataFrame(data_mem_run)
    dfmemr.to_csv("plot_torch_aot_ort_run_mem.csv", index=False)
    dfmemr.to_excel("plot_torch_aot_ort_run_mem.xlsx", index=False)
    dfmemfr = pandas.DataFrame(data_mem_first_run)
    dfmemfr.to_csv("plot_torch_aot_ort_first_run_mem.csv", index=False)
    dfmemfr.to_excel("plot_torch_aot_ort_first_run_mem.xlsx", index=False)
    return df, dfmemfr, dfmemr


df, dfmemfr, dfmemr = benchmark(list(input_tensors[0].shape))
print(df)
  0%|          | 0/6 [00:00<?, ?it/s]number of experiments: 6

0.009931689636604542 eager CPU:   0%|          | 0/6 [00:00<?, ?it/s]
0.009931689636604542 eager CPU:  17%|█▋        | 1/6 [00:02<00:10,  2.10s/it]
0.0015548287599813193 eager CUDA:  17%|█▋        | 1/6 [00:02<00:10,  2.10s/it]
0.0015548287599813193 eager CUDA:  33%|███▎      | 2/6 [00:03<00:07,  1.96s/it]
0.009628557636460755 default CPU:  33%|███▎      | 2/6 [00:16<00:07,  1.96s/it]
0.009628557636460755 default CPU:  50%|█████     | 3/6 [00:17<00:21,  7.22s/it]
0.0009324981768500316 default CUDA:  50%|█████     | 3/6 [00:23<00:21,  7.22s/it]
0.0009324981768500316 default CUDA:  67%|██████▋   | 4/6 [00:24<00:14,  7.28s/it]/home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py:130: UserWarning: Your compiler for AOTAutograd is returning a function that doesn't take boxed arguments. Please wrap it with functorch.compile.make_boxed_func or handle the boxed arguments yourself. See https://github.com/pytorch/pytorch/pull/83137#issuecomment-1211320670 for rationale.
  warnings.warn(

0.005942785190459939 dort CPU:  67%|██████▋   | 4/6 [00:26<00:14,  7.28s/it]
0.005942785190459939 dort CPU:  83%|████████▎ | 5/6 [00:27<00:05,  5.72s/it]/home/xadupre/vv/this/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py:130: UserWarning: Your compiler for AOTAutograd is returning a function that doesn't take boxed arguments. Please wrap it with functorch.compile.make_boxed_func or handle the boxed arguments yourself. See https://github.com/pytorch/pytorch/pull/83137#issuecomment-1211320670 for rationale.
  warnings.warn(

0.0029184219165472314 dort CUDA:  83%|████████▎ | 5/6 [00:29<00:05,  5.72s/it]
0.0029184219165472314 dort CUDA: 100%|██████████| 6/6 [00:30<00:00,  4.74s/it]
0.0029184219165472314 dort CUDA: 100%|██████████| 6/6 [00:30<00:00,  5.10s/it]
      name compute   export   average  deviation  min_exec  max_exec  repeat  number     ttime  context_size  warmup_time
0    eager     CPU    eager  0.009932   0.000668  0.008584  0.010456       1    11.0  0.109249            64     0.008224
1    eager    CUDA    eager  0.001555   0.000574  0.001247  0.003799       1    75.0  0.116612            64     0.004597
2  default     CPU  default  0.009629   0.001138  0.008059  0.010996       1    11.0  0.105914            64     0.006512
3  default    CUDA  default  0.000932   0.000089  0.000859  0.001395       1   147.0  0.137077            64     0.002328
4     dort     CPU     dort  0.005943   0.002426  0.003015  0.010610       1    21.0  0.124798            64     0.003301
5     dort    CUDA     dort  0.002918   0.000118  0.002435  0.002954       1    36.0  0.105063            64     0.003967

Other view

def view_time(df, title, suffix="time"):
    piv = pandas.pivot_table(df, index="export", columns=["compute"], values="average")
    print(piv)
    piv.to_csv(f"plot_torch_aot_{suffix}_compute.csv")
    piv.to_excel(f"plot_torch_aot_{suffix}_compute.xlsx")

    piv_cpu = pandas.pivot_table(
        df[df.compute == "CPU"],
        index="export",
        columns=["compute"],
        values="average",
    )

    fig, ax = plt.subplots(1, 2, figsize=(12, 4))
    fig.suptitle(title)
    piv_cpu.plot.barh(ax=ax[0], title="CPU", logx=True)

    if has_cuda:
        piv_gpu = pandas.pivot_table(
            df[df.compute == "CUDA"],
            index="export",
            columns=["compute"],
            values="average",
        )
        piv_gpu.plot.barh(ax=ax[1], title="CUDA", logx=True)

    fig.tight_layout()
    fig.savefig(f"plot_torch_aot_{suffix}.png")
    return ax


view_time(df, "Compares processing time on backends")
Compares processing time on backends, CPU, CUDA
compute       CPU      CUDA
export
default  0.009629  0.000932
dort     0.005943  0.002918
eager    0.009932  0.001555

array([<Axes: title={'center': 'CPU'}, ylabel='export'>,
       <Axes: title={'center': 'CUDA'}, ylabel='export'>], dtype=object)

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",),
        suptitle=f"Memory Consumption of backend, 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_aot_first_run_mem_{compute}.png")
  • Memory Consumption of backend, 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 backend, 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",),
        suptitle=f"Memory Consumption of backens, 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_aot_run_mem_{compute}.png")
  • Memory Consumption of backens, 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 backens, 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)

Total running time of the script: (0 minutes 59.956 seconds)

Gallery generated by Sphinx-Gallery