Note
Go to the end to download the full example code
201: Evaluate DORT¶
It compares DORT to eager mode and onnxrt backend.
To run the script:
python _doc/examples/plot_torch_dort --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': 8, 'cuda': 0, 'cuda_count': 0, 'processor': 'x86_64'}
Scripts arguments
script_args = get_parsed_args(
"plot_torch_dort",
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(MyModelClass, self).__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)
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=0.5401992797851562 Mb
Backends¶
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)
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")
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=False)
optimized_mod = torch.compile(model, backend=local_aot_ort, fullgraph=True)
optimized_mod(*args)
return optimized_mod
def get_torch_opti(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)
optimized_mod(*args)
return optimized_mod
Let’s check they are working.
export_functions = [
get_torch_eager,
get_torch_default,
get_torch_dort,
# get_torch_opti,
]
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_dort_{k}.onnx"
torch._dynamo.reset()
model, input_tensor = create_model_and_input()
try:
v(model, input_tensor)
except Exception as e:
print(f"skipped due to {str(e)[:1000]}")
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, v in supported_exporters.items():
print(f"run compile for memory {k} on cpu")
filename = f"plot_torch_dort_{k}.onnx"
if has_cuda:
torch.cuda.set_device(0)
torch._dynamo.reset()
# CPU
model, input_tensor = create_model_and_input()
stat = start_spying_on(cuda=1 if has_cuda else 0)
v(model, input_tensor)
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
if k in {"torch_default"}:
print(f"skip compile for memory {k} on cuda")
continue
torch._dynamo.reset()
# CUDA
model, input_tensor = create_model_and_input()
model = model.cuda()
input_tensor = input_tensor.cuda()
print(f"run compile for memory {k} on cuda")
stat = start_spying_on(cuda=1 if has_cuda else 0)
v(model, input_tensor)
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_default on cpu
done.
run compile for memory torch_dort on cpu
done.
The result.
df1 = pandas.DataFrame(data)
df1.to_csv("plot_torch_dort_1_memory.csv", index=False)
df1.to_excel("plot_torch_dort_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_dort_1_memory_{p}.png")
peak mean ... export p
0 1057.968750 1057.968750 ... torch_eager cpu
1 1057.945312 1057.945312 ... torch_default cpu
2 1057.945312 1057.945312 ... torch_dort cpu
[3 rows x 7 columns]
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
dort first iteration speed¶
data = []
for k, v in supported_exporters.items():
print(f"run dort cpu {k}: {script_args.repeat1}")
times = []
for i in range(int(script_args.repeat1)):
model, input_tensor = create_model_and_input()
torch._dynamo.reset()
begin = time.perf_counter()
v(model, input_tensor)
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
if k in {"torch_dort", "torch_default"}:
print(f"skip dort cuda {k}: {script_args.repeat1}")
continue
print(f"run dort cuda {k}: {script_args.repeat1}")
times = []
for i in range(int(script_args.repeat1)):
model, input_tensor = create_model_and_input()
model = model.cuda()
input_tensor = input_tensor.cuda()
torch._dynamo.reset()
begin = time.perf_counter()
v(model, input_tensor)
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: 0.02599510000072769
run dort cpu torch_default: 1
done: 1.8705952000000252
run dort cpu torch_dort: 1
done: 0.28137710000009974
The result.
df1 = pandas.DataFrame(data)
df1.to_csv("plot_torch_dort_1_time.csv", index=False)
df1.to_excel("plot_torch_dort_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_dort_1_time.png")
export time min ... last std p
0 torch_eager 0.025995 0.025995 ... 0.025995 0.0 cpu
1 torch_default 1.870595 1.870595 ... 1.870595 0.0 cpu
2 torch_dort 0.281377 0.281377 ... 0.281377 0.0 cpu
[3 rows x 8 columns]
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_tensor = create_model_and_input()
pr = cProfile.Profile()
pr.enable()
for i in range(int(script_args.repeat1)):
export_function(model, input_tensor)
pr.disable()
else:
pr = cProfile.Profile()
pr.enable()
for i 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_dort_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_dort_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_dort_profile_{name}_{suffix}_h.txt", "w") as f:
f.write(text)
if verbose:
print("done.")
model, input_tensor = create_model_and_input()
def function_to_profile(model=model, input_tensor=input_tensor):
return get_torch_dort(model, input_tensor)
profile_function("dort", function_to_profile, verbose=True, suffix="1")
profile dort: <function function_to_profile at 0x7f30e664c9d0>
525623 function calls (509414 primitive calls) in 0.824 seconds
Ordered by: cumulative time
ncalls tottime percall cumtime percall filename:lineno(function)
1 0.000 0.000 0.884 0.884 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_dort_201.py:510(function_to_profile)
1 0.000 0.000 0.884 0.884 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_dort_201.py:238(get_torch_dort)
16/1 0.000 0.000 0.613 0.613 /home/xadupre/.local/lib/python3.10/site-packages/torch/nn/modules/module.py:1523(_wrapped_call_impl)
16/1 0.000 0.000 0.613 0.613 /home/xadupre/.local/lib/python3.10/site-packages/torch/nn/modules/module.py:1529(_call_impl)
4/1 0.000 0.000 0.613 0.613 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py:367(_fn)
1 0.000 0.000 0.546 0.546 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:887(catch_errors)
1 0.000 0.000 0.546 0.546 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:288(_convert_frame_assert)
2/1 0.000 0.000 0.546 0.546 /usr/lib/python3.10/contextlib.py:76(inner)
1 0.000 0.000 0.545 0.545 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:444(_compile)
3/1 0.000 0.000 0.544 0.544 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/utils.py:258(time_wrapper)
1 0.000 0.000 0.544 0.544 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:527(compile_inner)
1 0.000 0.000 0.533 0.533 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/bytecode_transformation.py:1028(transform_code_object)
1 0.000 0.000 0.530 0.530 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:150(_fn)
1 0.000 0.000 0.530 0.530 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py:480(transform)
1 0.000 0.000 0.527 0.527 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:2115(run)
1 0.000 0.000 0.527 0.527 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:835(run)
44 0.000 0.000 0.527 0.012 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:727(step)
1 0.000 0.000 0.488 0.488 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:2221(RETURN_VALUE)
1 0.000 0.000 0.488 0.488 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/output_graph.py:813(compile_subgraph)
1 0.000 0.000 0.487 0.487 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/output_graph.py:1075(compile_and_call_fx_graph)
1 0.000 0.000 0.481 0.481 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/output_graph.py:1160(call_user_compiler)
2/1 0.000 0.000 0.481 0.481 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/repro/after_dynamo.py:59(debug_wrapper)
1 0.000 0.000 0.481 0.481 /home/xadupre/.local/lib/python3.10/site-packages/torch/__init__.py:1777(__call__)
1 0.000 0.000 0.481 0.481 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.py:1089(__call__)
1 0.000 0.000 0.481 0.481 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/backends/common.py:18(compiler_fn)
1 0.000 0.000 0.481 0.481 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py:804(aot_module_simplified)
1 0.000 0.000 0.480 0.480 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py:411(create_aot_dispatcher_function)
1 0.000 0.000 0.411 0.411 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:411(aot_wrapper_dedupe)
1 0.000 0.000 0.411 0.411 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:630(aot_wrapper_synthetic_base)
1 0.000 0.000 0.410 0.410 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py:233(aot_dispatch_autograd)
3/2 0.000 0.000 0.288 0.144 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/external_utils.py:34(inner)
1 0.000 0.000 0.270 0.270 /home/xadupre/github/experimental-experiment/experimental_experiment/torch_models/training_helper.py:5(make_aot_ort)
1 0.000 0.000 0.270 0.270 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.py:722(__init__)
2 0.047 0.024 0.259 0.130 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/decomposition_table.py:18(_create_onnx_supports_op_overload_table)
285 0.012 0.000 0.244 0.001 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/functional_tensor.py:268(__torch_dispatch__)
1 0.000 0.000 0.228 0.228 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/dispatch_and_compile_graph.py:129(aot_dispatch_autograd_graph)
1 0.000 0.000 0.222 0.222 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/dispatch_and_compile_graph.py:33(_create_graph)
1 0.000 0.000 0.222 0.222 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:1080(wrapped)
1 0.000 0.000 0.221 0.221 /home/xadupre/.local/lib/python3.10/site-packages/torch/_compile.py:20(inner)
1 0.000 0.000 0.221 0.221 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:627(dispatch_trace)
1 0.000 0.000 0.220 0.220 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:663(trace)
1 0.000 0.000 0.215 0.215 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:650(flatten_fn)
1 0.000 0.000 0.215 0.215 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:643(wrapped)
1 0.000 0.000 0.198 0.198 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:520(joint_helper)
1 0.000 0.000 0.198 0.198 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:350(_functionalized_f_helper)
1 0.000 0.000 0.178 0.178 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:251(inner_fn_with_anomaly)
1 0.000 0.000 0.178 0.178 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:186(inner_fn)
871/482 0.002 0.000 0.159 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_stats.py:15(wrapper)
1 0.000 0.000 0.136 0.136 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/exporter.py:357(__init__)
15108 0.018 0.000 0.127 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/exporter.py:251(is_registered_op)
1 0.000 0.000 0.127 0.127 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/decomposition_table.py:78(create_onnx_friendly_decomposition_table)
1 0.001 0.001 0.127 0.127 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/partitioners.py:637(min_cut_rematerialization_partition)
263/242 0.002 0.000 0.113 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:727(__torch_dispatch__)
1 0.000 0.000 0.112 0.112 /home/xadupre/.local/lib/python3.10/site-packages/torch/autograd/__init__.py:278(grad)
1 0.000 0.000 0.111 0.111 /home/xadupre/.local/lib/python3.10/site-packages/torch/autograd/graph.py:739(_engine_run_backward)
1 0.002 0.002 0.111 0.111 {method 'run_backward' of 'torch._C._EngineBase' objects}
3 0.000 0.000 0.110 0.037 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/interpreter.py:106(run)
15123 0.027 0.000 0.109 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/exporter.py:228(get_op_functions)
263/242 0.001 0.000 0.106 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:758(inner_torch_dispatch)
49 0.000 0.000 0.104 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/interpreter.py:184(run_node)
2 0.000 0.000 0.099 0.049 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:670(functional_call)
69/54 0.003 0.000 0.098 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:298(proxy_call)
22 0.000 0.000 0.096 0.004 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:652(run_node)
584/580 0.002 0.000 0.084 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:886(__torch_dispatch__)
592/414 0.002 0.000 0.083 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:859(tree_map)
4513/4477 0.008 0.000 0.081 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/node.py:724(map_arg)
584/580 0.006 0.000 0.081 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1215(dispatch)
9626/4486 0.033 0.000 0.071 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/node.py:732(map_aggregate)
271 0.002 0.000 0.070 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:940(_cached_dispatch_impl)
572 0.002 0.000 0.069 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:1066(tree_map_only)
4 0.002 0.000 0.069 0.017 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/partitioners.py:59(_extract_graph_with_inputs_outputs)
1 0.000 0.000 0.067 0.067 /home/xadupre/github/experimental-experiment/_doc/examples/plot_torch_dort_201.py:163(forward)
1 0.000 0.000 0.067 0.067 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py:913(forward)
3/1 0.000 0.000 0.067 0.067 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py:88(g)
1 0.000 0.000 0.067 0.067 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py:77(runtime_wrapper)
2/1 0.000 0.000 0.067 0.067 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py:105(call_func_at_runtime_with_args)
1 0.000 0.000 0.067 0.067 /home/xadupre/.local/lib/python3.10/site-packages/torch/autograd/function.py:590(apply)
1 0.000 0.000 0.067 0.067 {built-in method apply}
1 0.000 0.000 0.067 0.067 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py:534(forward)
1 0.000 0.000 0.067 0.067 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/_lazy_graph_module.py:112(_lazy_forward)
1 0.000 0.000 0.065 0.065 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/traced_function_transforms.py:102(inner_fn)
2/1 0.000 0.000 0.064 0.064 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph_module.py:736(call_wrapped)
1 0.000 0.000 0.064 0.064 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph_module.py:299(__call__)
1 0.000 0.000 0.064 0.064 <eval_with_key>.49:4(forward)
1 0.000 0.000 0.064 0.064 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.py:837(_ort_acclerated_call)
421 0.003 0.000 0.064 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph.py:886(create_node)
2409/495 0.011 0.000 0.060 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:734(unflatten)
1 0.000 0.000 0.055 0.055 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/partitioners.py:157(_extract_fwd_bwd_modules)
20069 0.024 0.000 0.054 0.000 {method 'get' of 'dict' objects}
1 0.000 0.000 0.054 0.054 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/collect_metadata_analysis.py:128(inner)
270 0.002 0.000 0.053 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph.py:1228(node_copy)
27 0.000 0.000 0.053 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/interpreter.py:256(call_function)
721/614 0.001 0.000 0.050 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_ops.py:597(__call__)
850 0.001 0.000 0.050 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:799(tree_flatten)
2865/850 0.010 0.000 0.049 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:778(_tree_flatten_helper)
8 0.000 0.000 0.047 0.006 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/interpreter.py:298(call_module)
5870/5734 0.005 0.000 0.045 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/node.py:738(<genexpr>)
430 0.005 0.000 0.044 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/node.py:166(__init__)
71498/70536 0.037 0.000 0.043 0.000 {built-in method builtins.isinstance}
8857 0.024 0.000 0.043 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/registration.py:55(from_qualified_name)
61 0.000 0.000 0.042 0.001 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:212(track_tensor_tree)
76/61 0.000 0.000 0.042 0.001 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:213(wrap_with_proxy)
94/86 0.002 0.000 0.041 0.000 {method 'detach' of 'torch._C.TensorBase' objects}
9 0.000 0.000 0.039 0.004 /home/xadupre/.local/lib/python3.10/site-packages/torch/nn/modules/linear.py:115(forward)
9 0.002 0.000 0.039 0.004 {built-in method torch._C._nn.linear}
74 0.000 0.000 0.036 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:172(set_meta)
6 0.000 0.000 0.036 0.006 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph.py:1291(python_code)
79/74 0.000 0.000 0.034 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:139(extract_val)
87/3 0.002 0.000 0.034 0.011 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/diagnostics/infra/decorator.py:71(wrapper)
6 0.000 0.000 0.034 0.006 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph.py:1353(_python_code)
76 0.000 0.000 0.034 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:136(snapshot_fake)
6 0.003 0.000 0.034 0.006 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph.py:380(_gen_python_code)
19 0.000 0.000 0.032 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/functional_utils.py:21(to_fun)
19 0.000 0.000 0.032 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/functional_tensor.py:172(to_functional)
94 0.001 0.000 0.032 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/proxy.py:173(create_proxy)
15123 0.018 0.000 0.032 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/registration.py:44(from_name_parts)
9 0.000 0.000 0.032 0.004 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:474(wrapper)
271 0.005 0.000 0.032 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:975(_cache_key)
9 0.000 0.000 0.032 0.004 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:1198(CALL_FUNCTION)
9 0.000 0.000 0.032 0.004 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py:673(call_function)
6 0.000 0.000 0.031 0.005 /home/xadupre/.local/lib/python3.10/site-packages/torch/_logging/_internal.py:1026(trace_structured)
1503/640 0.003 0.000 0.031 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:792(<listcomp>)
4 0.000 0.000 0.030 0.007 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph_module.py:820(print_readable)
9 0.000 0.000 0.030 0.003 /home/xadupre/.local/lib/python3.10/site-packages/torch/nn/functional.py:1489(relu)
9 0.001 0.000 0.030 0.003 {built-in method torch.relu}
4 0.000 0.000 0.029 0.007 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:761(module_call_wrapper)
4 0.000 0.000 0.029 0.007 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:552(call_module)
4 0.000 0.000 0.029 0.007 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py:763(forward)
1 0.000 0.000 0.028 0.028 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/onnxruntime.py:1035(compile)
1 0.000 0.000 0.027 0.027 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/passes/infra/partitioner.py:326(partition_and_fuse)
10 0.000 0.000 0.026 0.003 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/variables/builder.py:1332(wrap_fx_proxy)
10 0.000 0.000 0.026 0.003 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/variables/builder.py:1392(wrap_fx_proxy_cls)
1 0.000 0.000 0.026 0.026 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/passes/infra/partitioner.py:265(fuse_partitions)
1 0.000 0.000 0.026 0.026 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/passes/utils/fuser_utils.py:218(fuse_by_partitions)
559 0.004 0.000 0.025 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/node.py:461(__update_args_kwargs)
4 0.000 0.000 0.024 0.006 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/variables/nn_module.py:249(call_function)
11 0.003 0.000 0.024 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph.py:1395(lint)
6 0.001 0.000 0.024 0.004 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph.py:1466(eliminate_dead_code)
14 0.000 0.000 0.023 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/utils.py:1207(wrap_fake_exception)
9 0.000 0.000 0.023 0.003 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/utils.py:1630(get_fake_value)
1 0.001 0.001 0.022 0.022 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/compile_utils.py:25(fx_graph_cse)
1 0.000 0.000 0.022 0.022 /home/xadupre/.local/lib/python3.10/site-packages/networkx/algorithms/flow/maxflow.py:304(minimum_cut)
1 0.000 0.000 0.020 0.020 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/fx_onnx_interpreter.py:495(run)
96 0.001 0.000 0.020 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/proxy.py:117(create_node)
1 0.000 0.000 0.020 0.020 /home/xadupre/.local/lib/python3.10/site-packages/networkx/algorithms/flow/preflowpush.py:291(preflow_push)
1 0.001 0.001 0.020 0.020 /home/xadupre/.local/lib/python3.10/site-packages/networkx/algorithms/flow/preflowpush.py:22(preflow_push_impl)
246 0.003 0.000 0.019 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1144(_output_from_cache_entry)
311/269 0.003 0.000 0.018 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1035(_prep_args_for_hash)
4313 0.005 0.000 0.018 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:608(_is_leaf)
1 0.000 0.000 0.018 0.018 /home/xadupre/github/onnxruntime/build/linux_cuda/Release/onnxruntime/capi/onnxruntime_inference_collection.py:358(__init__)
27 0.000 0.000 0.018 0.001 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/fx_onnx_interpreter.py:413(run_node)
1 0.017 0.017 0.018 0.018 /home/xadupre/github/onnxruntime/build/linux_cuda/Release/onnxruntime/capi/onnxruntime_inference_collection.py:436(_create_inference_session)
26658/25731 0.016 0.000 0.017 0.000 {built-in method builtins.hash}
6083 0.007 0.000 0.017 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:601(_get_node_type)
8272 0.008 0.000 0.017 0.000 {method 'add' of 'set' objects}
1503 0.002 0.000 0.017 0.000 <string>:2(__init__)
1 0.000 0.000 0.016 0.016 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/partitioners.py:692(classify_nodes)
129/69 0.000 0.000 0.016 0.000 /usr/lib/python3.10/copy.py:259(_reconstruct)
4 0.000 0.000 0.016 0.004 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/utils.py:1218(deepcopy_to_fake_tensor)
4 0.000 0.000 0.016 0.004 /home/xadupre/.local/lib/python3.10/site-packages/torch/_dynamo/utils.py:1220(<lambda>)
216/4 0.001 0.000 0.016 0.004 /usr/lib/python3.10/copy.py:128(deepcopy)
17 0.000 0.000 0.016 0.001 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/fx_onnx_interpreter.py:647(call_function)
217 0.003 0.000 0.016 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/graph.py:536(emit_node)
8920 0.010 0.000 0.016 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_ops.py:602(__hash__)
11509/11161 0.009 0.000 0.015 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/node.py:714(__setattr__)
69/54 0.000 0.000 0.015 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py:1206(maybe_handle_decomp)
25 0.001 0.000 0.015 0.001 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1242(_dispatch_impl)
31 0.000 0.000 0.015 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:338(__call__)
4 0.000 0.000 0.015 0.004 /usr/lib/python3.10/copy.py:227(_deepcopy_dict)
31 0.000 0.000 0.015 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:260(from_real_tensor)
5 0.000 0.000 0.015 0.003 /home/xadupre/.local/lib/python3.10/site-packages/torch/_prims_common/wrappers.py:242(_fn)
3 0.000 0.000 0.015 0.005 /home/xadupre/.local/lib/python3.10/site-packages/torch/_decomp/__init__.py:115(_fn)
3 0.000 0.000 0.015 0.005 /home/xadupre/.local/lib/python3.10/site-packages/torch/_decomp/decompositions.py:209(threshold_backward)
9 0.000 0.000 0.015 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/collect_metadata_analysis.py:118(_to_fun)
3341/3185 0.004 0.000 0.015 0.000 {built-in method builtins.next}
1503 0.004 0.000 0.014 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_pytree.py:629(__post_init__)
31 0.001 0.000 0.014 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/meta_utils.py:863(__call__)
4496 0.006 0.000 0.014 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/node.py:730(<lambda>)
2 0.000 0.000 0.014 0.007 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/_pass.py:240(run)
31 0.002 0.000 0.013 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/meta_utils.py:184(meta_tensor)
8 0.000 0.000 0.013 0.002 /home/xadupre/.local/lib/python3.10/site-packages/torch/nn/parameter.py:55(__deepcopy__)
1 0.000 0.000 0.013 0.013 /home/xadupre/.local/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py:254(<lambda>)
1 0.000 0.000 0.013 0.013 /home/xadupre/.local/lib/python3.10/site-packages/torch/onnx/_internal/fx/passes/type_promotion.py:1716(_run)
19 0.000 0.000 0.013 0.001 {built-in method torch._to_functional_tensor}
85 0.002 0.000 0.013 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/functional_tensor.py:80(__new__)
40 0.000 0.000 0.013 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:1757(__torch_function__)
296 0.004 0.000 0.012 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py:695(extract_tensor_metadata)
20033 0.012 0.000 0.012 0.000 {method 'split' of 'str' objects}
149/140 0.001 0.000 0.012 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/_ops.py:628(decompose)
38 0.001 0.000 0.012 0.000 {built-in method torch._mirror_autograd_meta_to}
15985 0.011 0.000 0.012 0.000 {built-in method builtins.getattr}
1959 0.002 0.000 0.011 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/node.py:742(<genexpr>)
1 0.000 0.000 0.011 0.011 /home/xadupre/.local/lib/python3.10/site-packages/torch/fx/passes/utils/fuser_utils.py:91(fuse_as_graphmodule)
63 0.001 0.000 0.011 0.000 /home/xadupre/.local/lib/python3.10/site-packages/torch/utils/_python_dispatch.py:505(return_and_correct_aliasing)
3 0.001 0.000 0.011 0.004 {built-in method torch.flatten}
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_tensor = create_model_and_input()
if p == "CUDA":
if not has_cuda:
continue
model = model.cuda()
input_tensor = input_tensor.cuda()
try:
exported_model = export_fct(model, input_tensor)
except torch._dynamo.exc.BackendCompilerFailed as e:
# Triton only supports devices of CUDA Capability >= 7.0, but your device is of CUDA capability 6.1
obs["error"] = str(e)
data.append(obs)
continue
def call_model(
export_fct=export_fct,
exported_model=exported_model,
input_tensor=input_tensor,
):
res = exported_model(input_tensor).sum()
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 i 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_dort_ort_time.csv", index=False)
df.to_excel("plot_torch_dort_ort_time.xlsx", index=False)
dfmemr = pandas.DataFrame(data_mem_run)
dfmemr.to_csv("plot_torch_dort_ort_run_mem.csv", index=False)
dfmemr.to_excel("plot_torch_dort_ort_run_mem.xlsx", index=False)
dfmemfr = pandas.DataFrame(data_mem_first_run)
dfmemfr.to_csv("plot_torch_dort_ort_first_run_mem.csv", index=False)
dfmemfr.to_excel("plot_torch_dort_ort_first_run_mem.xlsx", index=False)
return df, dfmemfr, dfmemr
df, dfmemfr, dfmemr = benchmark(list(input_tensor.shape))
print(df)
0%| | 0/6 [00:00<?, ?it/s]number of experiments: 6
0.0003008790960442341 eager CPU: 0%| | 0/6 [00:00<?, ?it/s]
0.0003008790960442341 eager CPU: 17%|█▋ | 1/6 [00:01<00:08, 1.76s/it]
0.0003699560493798883 default CPU: 17%|█▋ | 1/6 [00:03<00:08, 1.76s/it]
0.0003699560493798883 default CPU: 50%|█████ | 3/6 [00:05<00:05, 1.72s/it]
0.00044833927124921094 dort CPU: 50%|█████ | 3/6 [00:05<00:05, 1.72s/it]
0.00044833927124921094 dort CPU: 83%|████████▎ | 5/6 [00:07<00:01, 1.38s/it]
0.00044833927124921094 dort CPU: 100%|██████████| 6/6 [00:07<00:00, 1.22s/it]
name compute ... context_size warmup_time
0 eager CPU ... 64 0.001304
1 default CPU ... 64 0.001084
2 dort CPU ... 64 0.001165
[3 rows x 12 columns]
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_dort_{suffix}_compute.csv")
piv.to_excel(f"plot_torch_dort_{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_dort_{suffix}.png")
return ax
view_time(df, "Compares processing time on backends")
compute CPU
export
default 0.000370
dort 0.000448
eager 0.000301
array([<Axes: title={'center': 'CPU'}, ylabel='export'>, <Axes: >],
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_dort_first_run_mem_{compute}.png")
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
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"
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_dort_run_mem_{compute}.png")
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
/home/xadupre/github/experimental-experiment/experimental_experiment/plotting/memory.py:68: UserWarning: set_ticklabels() should only be used with a fixed number of ticks, i.e. after set_ticks() or using a FixedLocator.
ax[i, j].set_xticklabels(ls, ha="right")
Total running time of the script: (0 minutes 32.212 seconds)