Source code for torch.profiler.profiler
# mypy: allow-untyped-defs
import gzip
import json
import os
import shutil
import tempfile
from abc import ABC, abstractmethod
from enum import Enum
from functools import partial
from typing import Any, Callable, Dict, Iterable, List, Optional, Tuple
from typing_extensions import Self
from warnings import warn
import torch
import torch.autograd.profiler as prof
from torch._C import _get_privateuse1_backend_name
from torch._C._profiler import (
_add_execution_trace_observer,
_disable_execution_trace_observer,
_enable_execution_trace_observer,
_ExperimentalConfig,
_remove_execution_trace_observer,
)
from torch.autograd import kineto_available, ProfilerActivity
from torch.profiler._memory_profiler import MemoryProfile, MemoryProfileTimeline
__all__ = [
"supported_activities",
"ProfilerAction",
"schedule",
"tensorboard_trace_handler",
"profile",
"ExecutionTraceObserver",
]
PROFILER_STEP_NAME = "ProfilerStep"
class _NumpyEncoder(json.JSONEncoder):
"""
Json encoder for numpy types (np.int, np.float, np.array etc.)
Returns default encoder if numpy is not available
"""
def default(self, obj):
"""Encode NumPy types to JSON"""
try:
import numpy as np
except ImportError:
return json.JSONEncoder.default(self, obj)
if isinstance(obj, np.integer):
return int(obj)
elif isinstance(obj, np.floating):
return float(obj)
elif isinstance(obj, np.ndarray):
return obj.tolist()
else:
return json.JSONEncoder.default(self, obj)
def supported_activities():
"""
Returns a set of supported profiler tracing activities.
Note: profiler uses CUPTI library to trace on-device CUDA kernels.
In case when CUDA is enabled but CUPTI is not available, passing
``ProfilerActivity.CUDA`` to profiler results in using the legacy CUDA
profiling code (same as in the legacy ``torch.autograd.profiler``).
This, in turn, results in including CUDA time in the profiler table output,
but not in the JSON trace.
"""
return torch.autograd._supported_activities()
class _ITraceObserver(ABC):
"""Abstract interface for a Trace observer.
This satisfies 3 methods: start, stop and cleanup"""
@abstractmethod
def start(self):
pass
@abstractmethod
def stop(self):
pass
@abstractmethod
def cleanup(self):
pass
[docs]class _KinetoProfile:
"""Low-level profiler wrap the autograd profile
Args:
activities (iterable): list of activity groups (CPU, CUDA) to use in profiling, supported values:
``torch.profiler.ProfilerActivity.CPU``, ``torch.profiler.ProfilerActivity.CUDA``,
``torch.profiler.ProfilerActivity.XPU``.
Default value: ProfilerActivity.CPU and (when available) ProfilerActivity.CUDA
or (when available) ProfilerActivity.XPU.
record_shapes (bool): save information about operator's input shapes.
profile_memory (bool): track tensor memory allocation/deallocation (see ``export_memory_timeline``
for more details).
with_stack (bool): record source information (file and line number) for the ops.
with_flops (bool): use formula to estimate the FLOPS of specific operators
(matrix multiplication and 2D convolution).
with_modules (bool): record module hierarchy (including function names)
corresponding to the callstack of the op. e.g. If module A's forward call's
module B's forward which contains an aten::add op,
then aten::add's module hierarchy is A.B
Note that this support exist, at the moment, only for TorchScript models
and not eager mode models.
experimental_config (_ExperimentalConfig) : A set of experimental options
used by profiler libraries like Kineto. Note, backward compatibility is not guaranteed.
execution_trace_observer (ExecutionTraceObserver) : A PyTorch Execution Trace Observer object.
`PyTorch Execution Traces <https://arxiv.org/pdf/2305.14516.pdf>`__ offer a graph based
representation of AI/ML workloads and enable replay benchmarks, simulators, and emulators.
When this argument is included the observer start() and stop() will be called for the
same time window as PyTorch profiler.
acc_events (bool): Enable the accumulation of FunctionEvents across multiple profiling cycles
.. note::
This API is experimental and subject to change in the future.
Enabling shape and stack tracing results in additional overhead.
When record_shapes=True is specified, profiler will temporarily hold references to the tensors;
that may further prevent certain optimizations that depend on the reference count and introduce
extra tensor copies.
"""
def __init__(
self,
*,
activities: Optional[Iterable[ProfilerActivity]] = None,
record_shapes: bool = False,
profile_memory: bool = False,
with_stack: bool = False,
with_flops: bool = False,
with_modules: bool = False,
experimental_config: Optional[_ExperimentalConfig] = None,
execution_trace_observer: Optional[_ITraceObserver] = None,
acc_events: bool = False,
custom_trace_id_callback: Optional[Callable[[], str]] = None,
):
self.activities = set(activities) if activities else supported_activities()
self.record_shapes = record_shapes
self.with_flops = with_flops
self.profile_memory = profile_memory
self.with_stack = with_stack
self.with_modules = with_modules
self.experimental_config = experimental_config
self.execution_trace_observer = execution_trace_observer
self.acc_events = acc_events
self.custom_trace_id_callback = custom_trace_id_callback
self.profiler: Optional[prof.profile] = None
self.mem_tl: Optional[MemoryProfileTimeline] = None
self.use_device = None
if ProfilerActivity.CUDA in self.activities:
self.use_device = "cuda"
elif ProfilerActivity.XPU in self.activities:
self.use_device = "xpu"
elif ProfilerActivity.MTIA in self.activities:
self.use_device = "mtia"
elif ProfilerActivity.PrivateUse1 in self.activities:
self.use_device = _get_privateuse1_backend_name()
# user-defined metadata to be amended to the trace
self.preset_metadata: Dict[str, str] = {}
def start(self):
self.prepare_trace()
self.start_trace()
def stop(self):
self.stop_trace()
def prepare_trace(self):
if (self.profiler is None) or (not self.acc_events):
self.profiler = prof.profile(
use_cpu=(ProfilerActivity.CPU in self.activities),
use_device=self.use_device,
record_shapes=self.record_shapes,
with_flops=self.with_flops,
profile_memory=self.profile_memory,
with_stack=self.with_stack,
with_modules=self.with_modules,
use_kineto=True,
experimental_config=self.experimental_config,
acc_events=self.acc_events,
custom_trace_id_callback=self.custom_trace_id_callback,
)
self.profiler._prepare_trace()
def start_trace(self):
if self.execution_trace_observer:
self.execution_trace_observer.start()
assert self.profiler is not None
self.profiler._start_trace()
if self.profile_memory:
self.add_metadata_json("profile_memory", "1")
if self.with_stack:
self.add_metadata_json("with_stack", "1")
if self.record_shapes:
self.add_metadata_json("record_shapes", "1")
if self.with_modules:
self.add_metadata_json("with_modules", "1")
if self.with_flops:
self.add_metadata_json("with_flops", "1")
if kineto_available():
dist_info = self._get_distributed_info()
if dist_info:
self.add_metadata_json(
"distributedInfo", json.dumps(dist_info, cls=_NumpyEncoder)
)
if hasattr(torch, "_inductor"):
import torch._inductor.config as inductor_config
if inductor_config.triton.cudagraphs:
os.environ["DISABLE_CUPTI_LAZY_REINIT"] = "1"
self.add_metadata_json("DISABLE_CUPTI_LAZY_REINIT", "1")
# FIXME: CUDA Graph does not work well with CUPTI teardown.
# 1) crashes on 1st lazy CUPTI re-init after teardown (CUDA 11)
# 2) crashes on 2nd non-lazy CUPTI re-init after teardown (CUDA 12)
# Workaround: turn off CUPTI teardown when using CUDA Graphs.
os.environ["TEARDOWN_CUPTI"] = "0"
# Insert the preset user metadata to the trace
for k, v in self.preset_metadata.items():
self.add_metadata_json(k, v)
def stop_trace(self):
if self.execution_trace_observer:
self.execution_trace_observer.stop()
assert self.profiler is not None
self.profiler.__exit__(None, None, None)
[docs] def export_chrome_trace(self, path: str):
"""
Exports the collected trace in Chrome JSON format. If kineto is enabled, only
last cycle in schedule is exported.
"""
assert self.profiler
if path.endswith(".gz"):
fp = tempfile.NamedTemporaryFile("w+b", suffix=".json", delete=False)
fp.close()
retvalue = self.profiler.export_chrome_trace(fp.name)
with open(fp.name, "rb") as fin:
with gzip.open(path, "wb") as fout:
fout.writelines(fin)
os.remove(fp.name)
return retvalue
else:
return self.profiler.export_chrome_trace(path)
[docs] def export_stacks(self, path: str, metric: str = "self_cpu_time_total"):
"""Save stack traces to a file
Args:
path (str): save stacks file to this location;
metric (str): metric to use: "self_cpu_time_total" or "self_cuda_time_total"
"""
assert self.profiler
return self.profiler.export_stacks(path, metric)
[docs] def toggle_collection_dynamic(
self, enable: bool, activities: Iterable[ProfilerActivity]
):
"""Toggle collection of activities on/off at any point of collection. Currently supports toggling Torch Ops
(CPU) and CUDA activity supported in Kineto
Args:
activities (iterable): list of activity groups to use in profiling, supported values:
``torch.profiler.ProfilerActivity.CPU``, ``torch.profiler.ProfilerActivity.CUDA``
Examples:
.. code-block:: python
with torch.profiler.profile(
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA,
]
) as p:
code_to_profile_0()
// turn off collection of all CUDA activity
p.toggle_collection_dynamic(False, [torch.profiler.ProfilerActivity.CUDA])
code_to_profile_1()
// turn on collection of all CUDA activity
p.toggle_collection_dynamic(True, [torch.profiler.ProfilerActivity.CUDA])
code_to_profile_2()
print(p.key_averages().table(
sort_by="self_cuda_time_total", row_limit=-1))
"""
if not self.profiler:
return
self.profiler.toggle_collection_dynamic(enable, activities)
[docs] def key_averages(
self, group_by_input_shape: bool = False, group_by_stack_n: int = 0
):
"""Averages events, grouping them by operator name and (optionally) input shapes and
stack.
.. note::
To use shape/stack functionality make sure to set record_shapes/with_stack
when creating profiler context manager.
"""
assert self.profiler
return self.profiler.key_averages(group_by_input_shape, group_by_stack_n)
[docs] def events(self):
"""
Returns the list of unaggregated profiler events,
to be used in the trace callback or after the profiling is finished
"""
assert self.profiler
return self.profiler.function_events
[docs] def add_metadata(self, key: str, value: str):
"""
Adds a user defined metadata with a string key and a string value
into the trace file
"""
wrapped_value = '"' + value.replace('"', '\\"') + '"'
torch.autograd._add_metadata_json(key, wrapped_value)
[docs] def add_metadata_json(self, key: str, value: str):
"""
Adds a user defined metadata with a string key and a valid json value
into the trace file
"""
torch.autograd._add_metadata_json(key, value)
[docs] def preset_metadata_json(self, key: str, value: str):
"""
Preset a user defined metadata when the profiler is not started
and added into the trace file later.
Metadata is in the format of a string key and a valid json value
"""
self.preset_metadata[key] = value
def _get_distributed_info(self):
import torch.distributed as dist
if not dist.is_available() or not dist.is_initialized():
return None
backend = dist.get_backend()
dist_info = {
"backend": backend,
"rank": dist.get_rank(),
"world_size": dist.get_world_size(),
"pg_count": dist.get_pg_count(),
"pg_config": dist.distributed_c10d._get_all_pg_configs(),
}
if backend == "nccl":
nccl_version = torch.cuda.nccl.version()
dist_info["nccl_version"] = ".".join(str(v) for v in nccl_version)
return dist_info
def _memory_profile(self) -> MemoryProfile:
required = ("record_shapes", "profile_memory", "with_stack")
missing = [f"{i}=True" for i in required if not getattr(self, i)]
if missing:
raise ValueError(f"{', '.join(missing)} required for memory profiling.")
assert self.profiler is not None and self.profiler.kineto_results is not None
return MemoryProfile(self.profiler.kineto_results)
[docs] def export_memory_timeline(self, path: str, device: Optional[str] = None) -> None:
"""Export memory event information from the profiler collected
tree for a given device, and export a timeline plot. There are 3
exportable files using ``export_memory_timeline``, each controlled by the
``path``'s suffix.
- For an HTML compatible plot, use the suffix ``.html``, and a memory timeline
plot will be embedded as a PNG file in the HTML file.
- For plot points consisting of ``[times, [sizes by category]]``, where
``times`` are timestamps and ``sizes`` are memory usage for each category.
The memory timeline plot will be saved a JSON (``.json``) or gzipped JSON
(``.json.gz``) depending on the suffix.
- For raw memory points, use the suffix ``.raw.json.gz``. Each raw memory
event will consist of ``(timestamp, action, numbytes, category)``, where
``action`` is one of ``[PREEXISTING, CREATE, INCREMENT_VERSION, DESTROY]``,
and ``category`` is one of the enums from
``torch.profiler._memory_profiler.Category``.
Output: Memory timeline written as gzipped JSON, JSON, or HTML.
"""
# Default to device 0, if unset. Fallback on cpu.
if device is None and self.use_device and self.use_device != "cuda":
device = self.use_device + ":0"
if device is None:
device = "cuda:0" if torch.cuda.is_available() else "cpu"
# Construct the memory timeline plot data
self.mem_tl = MemoryProfileTimeline(self._memory_profile())
# Depending on the file suffix, save the data as json.gz or json.
# For html, we can embed the image into an HTML file.
if path.endswith(".html"):
self.mem_tl.export_memory_timeline_html(path, device)
elif path.endswith(".gz"):
fp = tempfile.NamedTemporaryFile("w+t", suffix=".json", delete=False)
fp.close()
if path.endswith("raw.json.gz"):
self.mem_tl.export_memory_timeline_raw(fp.name, device)
else:
self.mem_tl.export_memory_timeline(fp.name, device)
with open(fp.name) as fin:
with gzip.open(path, "wt") as fout:
fout.writelines(fin)
os.remove(fp.name)
else:
self.mem_tl.export_memory_timeline(path, device)
[docs]class ProfilerAction(Enum):
"""
Profiler actions that can be taken at the specified intervals
"""
NONE = 0
WARMUP = 1
RECORD = 2
RECORD_AND_SAVE = 3
[docs]def schedule(
*, wait: int, warmup: int, active: int, repeat: int = 0, skip_first: int = 0
) -> Callable:
"""
Returns a callable that can be used as profiler ``schedule`` argument. The profiler will skip
the first ``skip_first`` steps, then wait for ``wait`` steps, then do the warmup for the next ``warmup`` steps,
then do the active recording for the next ``active`` steps and then repeat the cycle starting with ``wait`` steps.
The optional number of cycles is specified with the ``repeat`` parameter, the zero value means that
the cycles will continue until the profiling is finished.
"""
def schedule_fn(step: int) -> ProfilerAction:
assert step >= 0
if step < skip_first:
return ProfilerAction.NONE
else:
step -= skip_first
num_steps = wait + warmup + active
if repeat > 0 and step / num_steps >= repeat:
return ProfilerAction.NONE
mod_step = step % num_steps
if mod_step < wait:
return ProfilerAction.NONE
elif mod_step < wait + warmup:
return ProfilerAction.WARMUP
else:
return (
ProfilerAction.RECORD
if mod_step < num_steps - 1
else ProfilerAction.RECORD_AND_SAVE
)
assert (
wait >= 0 and warmup >= 0 and active > 0 and repeat >= 0 and skip_first >= 0
), "Invalid profiler schedule arguments"
if warmup == 0:
warn("Profiler won't be using warmup, this can skew profiler results")
return schedule_fn
def _default_schedule_fn(_: int) -> ProfilerAction:
"""
Default profiler behavior - immediately starts recording the events,
keeps doing it on every profiler step.
"""
return ProfilerAction.RECORD
[docs]def tensorboard_trace_handler(
dir_name: str, worker_name: Optional[str] = None, use_gzip: bool = False
):
"""
Outputs tracing files to directory of ``dir_name``, then that directory can be
directly delivered to tensorboard as logdir.
``worker_name`` should be unique for each worker in distributed scenario,
it will be set to '[hostname]_[pid]' by default.
"""
import os
import socket
import time
def handler_fn(prof) -> None:
nonlocal worker_name
if not os.path.isdir(dir_name):
try:
os.makedirs(dir_name, exist_ok=True)
except Exception as e:
raise RuntimeError("Can't create directory: " + dir_name) from e
if not worker_name:
worker_name = f"{socket.gethostname()}_{os.getpid()}"
# Use nanosecond here to avoid naming clash when exporting the trace
file_name = f"{worker_name}.{time.time_ns()}.pt.trace.json"
if use_gzip:
file_name = file_name + ".gz"
prof.export_chrome_trace(os.path.join(dir_name, file_name))
return handler_fn
[docs]class profile(_KinetoProfile):
"""Profiler context manager.
Args:
activities (iterable): list of activity groups (CPU, CUDA) to use in profiling, supported values:
``torch.profiler.ProfilerActivity.CPU``, ``torch.profiler.ProfilerActivity.CUDA``,
``torch.profiler.ProfilerActivity.XPU``.
Default value: ProfilerActivity.CPU and (when available) ProfilerActivity.CUDA
or (when available) ProfilerActivity.XPU.
schedule (Callable): callable that takes step (int) as a single parameter and returns
``ProfilerAction`` value that specifies the profiler action to perform at each step.
on_trace_ready (Callable): callable that is called at each step when ``schedule``
returns ``ProfilerAction.RECORD_AND_SAVE`` during the profiling.
record_shapes (bool): save information about operator's input shapes.
profile_memory (bool): track tensor memory allocation/deallocation.
with_stack (bool): record source information (file and line number) for the ops.
with_flops (bool): use formula to estimate the FLOPs (floating point operations) of specific operators
(matrix multiplication and 2D convolution).
with_modules (bool): record module hierarchy (including function names)
corresponding to the callstack of the op. e.g. If module A's forward call's
module B's forward which contains an aten::add op,
then aten::add's module hierarchy is A.B
Note that this support exist, at the moment, only for TorchScript models
and not eager mode models.
experimental_config (_ExperimentalConfig) : A set of experimental options
used for Kineto library features. Note, backward compatibility is not guaranteed.
execution_trace_observer (ExecutionTraceObserver) : A PyTorch Execution Trace Observer object.
`PyTorch Execution Traces <https://arxiv.org/pdf/2305.14516.pdf>`__ offer a graph based
representation of AI/ML workloads and enable replay benchmarks, simulators, and emulators.
When this argument is included the observer start() and stop() will be called for the
same time window as PyTorch profiler. See the examples section below for a code sample.
acc_events (bool): Enable the accumulation of FunctionEvents across multiple profiling cycles
use_cuda (bool):
.. deprecated:: 1.8.1
use ``activities`` instead.
.. note::
Use :func:`~torch.profiler.schedule` to generate the callable schedule.
Non-default schedules are useful when profiling long training jobs
and allow the user to obtain multiple traces at the different iterations
of the training process.
The default schedule simply records all the events continuously for the
duration of the context manager.
.. note::
Use :func:`~torch.profiler.tensorboard_trace_handler` to generate result files for TensorBoard:
``on_trace_ready=torch.profiler.tensorboard_trace_handler(dir_name)``
After profiling, result files can be found in the specified directory. Use the command:
``tensorboard --logdir dir_name``
to see the results in TensorBoard.
For more information, see
`PyTorch Profiler TensorBoard Plugin <https://github.com/pytorch/kineto/tree/master/tb_plugin>`__
.. note::
Enabling shape and stack tracing results in additional overhead.
When record_shapes=True is specified, profiler will temporarily hold references to the tensors;
that may further prevent certain optimizations that depend on the reference count and introduce
extra tensor copies.
Examples:
.. code-block:: python
with torch.profiler.profile(
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA,
]
) as p:
code_to_profile()
print(p.key_averages().table(
sort_by="self_cuda_time_total", row_limit=-1))
Using the profiler's ``schedule``, ``on_trace_ready`` and ``step`` functions:
.. code-block:: python
# Non-default profiler schedule allows user to turn profiler on and off
# on different iterations of the training loop;
# trace_handler is called every time a new trace becomes available
def trace_handler(prof):
print(prof.key_averages().table(
sort_by="self_cuda_time_total", row_limit=-1))
# prof.export_chrome_trace("/tmp/test_trace_" + str(prof.step_num) + ".json")
with torch.profiler.profile(
activities=[
torch.profiler.ProfilerActivity.CPU,
torch.profiler.ProfilerActivity.CUDA,
],
# In this example with wait=1, warmup=1, active=2, repeat=1,
# profiler will skip the first step/iteration,
# start warming up on the second, record
# the third and the forth iterations,
# after which the trace will become available
# and on_trace_ready (when set) is called;
# the cycle repeats starting with the next step
schedule=torch.profiler.schedule(
wait=1,
warmup=1,
active=2,
repeat=1),
on_trace_ready=trace_handler
# on_trace_ready=torch.profiler.tensorboard_trace_handler('./log')
# used when outputting for tensorboard
) as p:
for iter in range(N):
code_iteration_to_profile(iter)
# send a signal to the profiler that the next iteration has started
p.step()
The following sample shows how to setup up an Execution Trace Observer (`execution_trace_observer`)
.. code-block:: python
with torch.profiler.profile(
...
execution_trace_observer=(
ExecutionTraceObserver().register_callback("./execution_trace.json")
),
) as p:
for iter in range(N):
code_iteration_to_profile(iter)
p.step()
You can also refer to test_execution_trace_with_kineto() in tests/profiler/test_profiler.py.
Note: One can also pass any object satisfying the _ITraceObserver interface.
"""
def __init__(
self,
*,
activities: Optional[Iterable[ProfilerActivity]] = None,
schedule: Optional[Callable[[int], ProfilerAction]] = None,
on_trace_ready: Optional[Callable[..., Any]] = None,
record_shapes: bool = False,
profile_memory: bool = False,
with_stack: bool = False,
with_flops: bool = False,
with_modules: bool = False,
experimental_config: Optional[_ExperimentalConfig] = None,
execution_trace_observer: Optional[_ITraceObserver] = None,
acc_events: bool = False,
# deprecated:
use_cuda: Optional[bool] = None,
custom_trace_id_callback: Optional[Callable[[], str]] = None,
):
activities_set = set(activities) if activities else supported_activities()
if use_cuda is not None:
warn(
"`use_cuda` is deprecated, use `activities` argument instead",
FutureWarning,
stacklevel=2,
)
if use_cuda:
activities_set.add(ProfilerActivity.CUDA)
elif ProfilerActivity.CUDA in activities_set:
activities_set.remove(ProfilerActivity.CUDA)
assert len(activities_set) > 0, "No valid profiler activities found"
super().__init__(
activities=activities,
record_shapes=record_shapes,
profile_memory=profile_memory,
with_stack=with_stack,
with_flops=with_flops,
with_modules=with_modules,
experimental_config=experimental_config,
execution_trace_observer=execution_trace_observer,
acc_events=acc_events,
custom_trace_id_callback=custom_trace_id_callback,
)
if schedule:
self.schedule = schedule
# add step markers into the trace and table view
self.record_steps = True
else:
self.schedule = _default_schedule_fn
self.record_steps = False
self.on_trace_ready = on_trace_ready
self.step_num = 0
self.current_action = self.schedule(self.step_num)
self.step_rec_fn: Optional[prof.record_function] = None
self.action_map: Dict[
Tuple[ProfilerAction, Optional[ProfilerAction]], List[Any]
] = {
# key is (prev_action, current_action), value is action list corresponding to the state pair.
(ProfilerAction.NONE, ProfilerAction.NONE): [],
(ProfilerAction.NONE, ProfilerAction.WARMUP): [self.prepare_trace],
(ProfilerAction.NONE, ProfilerAction.RECORD): [
self.prepare_trace,
self.start_trace,
],
(ProfilerAction.NONE, ProfilerAction.RECORD_AND_SAVE): [
self.prepare_trace,
self.start_trace,
],
(ProfilerAction.WARMUP, ProfilerAction.NONE): [
partial(warn, "Incorrect schedule: WARMUP followed by NONE"),
self.start_trace,
self.stop_trace,
],
(ProfilerAction.WARMUP, ProfilerAction.WARMUP): [],
(ProfilerAction.WARMUP, ProfilerAction.RECORD): [self.start_trace],
(ProfilerAction.WARMUP, ProfilerAction.RECORD_AND_SAVE): [self.start_trace],
(ProfilerAction.RECORD, ProfilerAction.NONE): [
partial(warn, "Incorrect schedule: RECORD followed by NONE"),
self.stop_trace,
],
(ProfilerAction.RECORD, ProfilerAction.WARMUP): [
partial(warn, "Incorrect schedule: RECORD followed by WARMUP"),
self.stop_trace,
],
(ProfilerAction.RECORD, ProfilerAction.RECORD): [],
(ProfilerAction.RECORD, ProfilerAction.RECORD_AND_SAVE): [],
(ProfilerAction.RECORD_AND_SAVE, ProfilerAction.NONE): [
self.stop_trace,
self._trace_ready,
],
(ProfilerAction.RECORD_AND_SAVE, ProfilerAction.WARMUP): [
self.stop_trace,
self._trace_ready,
self.prepare_trace,
],
(ProfilerAction.RECORD_AND_SAVE, ProfilerAction.RECORD): [
self.stop_trace,
self._trace_ready,
self.prepare_trace,
self.start_trace,
],
(ProfilerAction.RECORD_AND_SAVE, ProfilerAction.RECORD_AND_SAVE): [
self.stop_trace,
self._trace_ready,
self.prepare_trace,
self.start_trace,
],
# used for exit action
(ProfilerAction.WARMUP, None): [self.start_trace, self.stop_trace],
(ProfilerAction.RECORD, None): [self.stop_trace, self._trace_ready],
(ProfilerAction.RECORD_AND_SAVE, None): [
self.stop_trace,
self._trace_ready,
],
}
# Start tracking increments to profiler step, this will be used
# by Kineto
prof.KinetoStepTracker.init_step_count(PROFILER_STEP_NAME)
def __enter__(self):
self.start()
return self
def __exit__(self, exc_type, exc_val, exc_tb):
self.stop()
prof.KinetoStepTracker.erase_step_count(PROFILER_STEP_NAME)
if self.execution_trace_observer:
self.execution_trace_observer.cleanup()
def start(self):
self._transit_action(ProfilerAction.NONE, self.current_action)
if self.record_steps:
self.step_rec_fn = prof.record_function(
"ProfilerStep#" + str(self.step_num)
)
self.step_rec_fn.__enter__()
def stop(self):
if self.record_steps and self.step_rec_fn:
self.step_rec_fn.__exit__(None, None, None)
self._transit_action(self.current_action, None)
[docs] def step(self):
"""
Signals the profiler that the next profiling step has started.
"""
if self.record_steps and self.step_rec_fn:
self.step_rec_fn.__exit__(None, None, None)
prev_action = self.current_action
self.step_num += 1
self.current_action = self.schedule(self.step_num)
self._transit_action(prev_action, self.current_action)
prof.KinetoStepTracker.increment_step(PROFILER_STEP_NAME)
if self.record_steps:
self.step_rec_fn = prof.record_function(
"ProfilerStep#" + str(self.step_num)
)
self.step_rec_fn.__enter__()
[docs] def set_custom_trace_id_callback(self, callback):
"""
Sets a callback to be called when a new trace ID is generated.
"""
self.custom_trace_id_callback = callback
[docs] def get_trace_id(self):
"""
Returns the current trace ID.
"""
if self.profiler is None:
return None
return self.profiler.trace_id
def _trace_ready(self):
if self.on_trace_ready:
self.on_trace_ready(self)
def _transit_action(self, prev_action, current_action):
action_list = self.action_map.get((prev_action, current_action))
if action_list:
for action in action_list:
action()
def _stats(self) -> Optional[prof._ProfilerStats]:
if self.profiler is None:
return None
return self.profiler._stats
class ExecutionTraceObserver(_ITraceObserver):
"""Execution Trace Observer
Each process can have a single ExecutionTraceObserver instance. The observer
can be added to record function callbacks via calling register_callback()
explicitly. Without calling unregister_callback(), repeated calls to
register_callback() will not add additional observers to record function
callbacks. Once an ExecutionTraceObserver is created, the start() and stop()
methods control when the event data is recorded.
Deleting or calling unregister_callback() will remove the observer from the
record function callbacks, finalize the output file, and will stop
incurring any overheads.
"""
def __init__(self) -> None:
"""
Initializes the default states.
"""
self._registered = False
self._execution_trace_running = False
def __del__(self):
"""
Calls unregister_callback() to make sure to finalize outputs.
"""
self.unregister_callback()
def register_callback(self, output_file_path: str) -> Self:
"""
Adds ET observer to record function callbacks. The data will be
written to output_file_path.
"""
if not self._registered:
self._output_file_path = output_file_path
self._registered = _add_execution_trace_observer(output_file_path)
return self
def unregister_callback(self):
"""
Removes ET observer from record function callbacks.
"""
def _save_triton_kernels():
# Save the kernel paths for the generated kernels
from torch._inductor.codecache import PyCodeCache as PyCodeCache
kernel_files = [
v.__file__
for v in PyCodeCache.modules
if getattr(v, "__file__", None) is not None
]
work_dir, file_name = os.path.split(self._output_file_path)
resource_dir = os.path.join(
work_dir, os.path.splitext(file_name)[0] + "_resources"
)
if not os.path.exists(resource_dir):
os.mkdir(resource_dir)
for kernel_file in kernel_files:
if kernel_file is None:
continue
name = os.path.basename(kernel_file)
dst = os.path.join(resource_dir, name)
shutil.copyfile(kernel_file, dst)
if self._registered:
self.stop()
try:
_save_triton_kernels()
except Exception as e:
warn(f"Execution trace failed to save kernels: {e}")
_remove_execution_trace_observer()
self._registered = False
@property
def is_registered(self):
"""
Returns True if the execution trace observer is registered, otherwise False.
"""
return self._registered
def is_running(self):
"""
Returns True if the observer is running, otherwise False.
"""
return self._execution_trace_running
def start(self):
"""
Starts to capture.
"""
if self._registered and not self._execution_trace_running:
_enable_execution_trace_observer()
self._execution_trace_running = True
self._record_pg_config()
def stop(self):
"""
Stops to capture.
"""
if self._execution_trace_running:
_disable_execution_trace_observer()
self._execution_trace_running = False
def cleanup(self):
"""
Calls unregister_callback() to make sure to finalize outputs.
"""
self.unregister_callback()
def get_output_file_path(self) -> str:
"""
Returns the output file name.
"""
if self.is_registered:
return self._output_file_path
else:
raise RuntimeError(
"A callback to the ET profiler needs to be registered "
"first before getting the output file path"
)
def _record_pg_config(self) -> None:
# Records the PG config info to the trace as node:
# ## process_group:init ##
if (
self.is_registered
and torch.distributed.is_available()
and torch.distributed.is_initialized()
):
pg_config_info = torch.distributed.distributed_c10d._world.pg_config_info
torch.autograd._record_function_with_args_enter(
"## process_group:init ##",
json.dumps(pg_config_info, cls=_NumpyEncoder),
)
