Source code for jax._src.profiler

# Copyright 2020 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from contextlib import contextmanager
from functools import wraps
import threading
from typing import Callable, Optional
import warnings

from jax.lib import xla_bridge
from jax.lib import xla_client


[docs]def start_server(port: int): """Starts a profiler server on port `port`. Using the "TensorFlow profiler" feature in `TensorBoard <https://www.tensorflow.org/tensorboard>`_ 2.2 or newer, you can connect to the profiler server and sample execution traces that show CPU, GPU, and/or TPU device activity. Returns a profiler server object. The server remains alive and listening until the server object is destroyed. """ return xla_client.profiler.start_server(port)
class _ProfileState(object): def __init__(self): self.profile_session = None self.log_dir = None self.lock = threading.Lock() _profile_state = _ProfileState()
[docs]def start_trace(log_dir): """Starts a profiler trace. The trace will capture CPU, GPU, and/or TPU activity, including Python functions and JAX on-device operations. Use ``stop_trace()`` to end the trace and save the results to ``log_dir``. The resulting trace can be viewed with TensorBoard. Note that TensorBoard doesn't need to be running when collecting the trace. Only once trace may be collected a time. A RuntimeError will be raised if ``start_trace()`` is called while another trace is running. Args: log_dir: The directory to save the profiler trace to (usually the TensorBoard log directory). """ with _profile_state.lock: if _profile_state.profile_session is not None: raise RuntimeError("Profile has already been started. " "Only one profile may be run at a time.") _profile_state.profile_session = xla_client.profiler.ProfilerSession() _profile_state.log_dir = log_dir
[docs]def stop_trace(): """Stops the currently-running profiler trace. The trace will be saved to the ``log_dir`` passed to the corresponding ``start_trace()`` call. Raises a RuntimeError if a trace hasn't been started. """ with _profile_state.lock: if _profile_state.profile_session is None: raise RuntimeError("No profile started") _profile_state.profile_session.stop_and_export(_profile_state.log_dir) _profile_state.profile_session = None _profile_state.log_dir = None
[docs]@contextmanager def trace(log_dir): """Context manager to take a profiler trace. The trace will capture CPU, GPU, and/or TPU activity, including Python functions and JAX on-device operations. The resulting trace can be viewed with TensorBoard. Note that TensorBoard doesn't need to be running when collecting the trace. Only once trace may be collected a time. A RuntimeError will be raised if a trace is started while another trace is running. Args: log_dir: The directory to save the profiler trace to (usually the TensorBoard log directory). """ start_trace(log_dir) try: yield finally: stop_trace()
[docs]class TraceAnnotation(xla_client.profiler.TraceMe): """Context manager that generates a trace event in the profiler. The trace event spans the duration of the code enclosed by the context. For example: >>> x = jnp.ones((1000, 1000)) >>> with jax.profiler.TraceAnnotation("my_label"): ... result = jnp.dot(x, x.T).block_until_ready() This will cause a "my_label" event to show up on the trace timeline if the event occurs while the process is being traced. """ pass
# TODO: remove this sometime after jax 0.2.11 is released
[docs]class TraceContext(TraceAnnotation):
[docs] def __init__(self, *args, **kwargs): warnings.warn( "TraceContext has been renamed to TraceAnnotation. This alias " "will eventually be removed; please update your code.") super().__init__(*args, **kwargs)
[docs]class StepTraceAnnotation(TraceAnnotation): """Context manager that generates a step trace event in the profiler. The step trace event spans the duration of the code enclosed by the context. The profiler will provide the performance analysis for each step trace event. For example, it can be used to mark training steps and enable the profiler to provide the performance analysis per step: >>> while global_step < NUM_STEPS: # doctest: +SKIP ... with jax.profiler.StepTraceAnnotation("train", step_num=global_step): # doctest: +SKIP ... train_step() # doctest: +SKIP ... global_step += 1 # doctest: +SKIP This will cause a "train xx" event to show up on the trace timeline if the event occurs while the process is being traced by TensorBoard. In addition, if using accelerators, the device trace timeline will also show a "train xx" event. Note that "step_num" can be set as a keyword argument to pass the global step number to the profiler. """
[docs] def __init__(self, name: str, **kwargs): super().__init__(name, _r=1, **kwargs)
# TODO: remove this sometime after jax 0.2.11 is released
[docs]class StepTraceContext(StepTraceAnnotation):
[docs] def __init__(self, *args, **kwargs): warnings.warn( "StepTraceContext has been renamed to StepTraceAnnotation. This alias " "will eventually be removed; please update your code.") super().__init__(*args, **kwargs)
[docs]def annotate_function(func: Callable, name: str = None, **kwargs): """Decorator that generates a trace event for the execution of a function. For example: >>> @jax.profiler.annotate_function ... def f(x): ... return jnp.dot(x, x.T).block_until_ready() >>> >>> result = f(jnp.ones((1000, 1000))) This will cause an "f" event to show up on the trace timeline if the function execution occurs while the process is being traced by TensorBoard. Arguments can be passed to the decorator via :py:func:`functools.partial`. >>> from functools import partial >>> @partial(jax.profiler.annotate_function, name="event_name") ... def f(x): ... return jnp.dot(x, x.T).block_until_ready() >>> result = f(jnp.ones((1000, 1000))) """ name = name or getattr(func, '__qualname__', None) name = name or func.__name__ @wraps(func) def wrapper(*args, **kwargs): with TraceAnnotation(name, **kwargs): return func(*args, **kwargs) return wrapper return wrapper
# TODO: remove this sometime after jax 0.2.11 is released
[docs]def trace_function(*args, **kwargs): warnings.warn( "trace_function has been renamed to annotate_function. This alias " "will eventually be removed; please update your code.") return annotate_function(*args, **kwargs)
[docs]def device_memory_profile(backend: Optional[str] = None) -> bytes: """Captures a JAX device memory profile as ``pprof``-format protocol buffer. A device memory profile is a snapshot of the state of memory, that describes the JAX :class:`jax.DeviceArray` and executable objects present in memory and their allocation sites. For more information how to use the device memory profiler, see :doc:`/device_memory_profiling`. The profiling system works by instrumenting JAX on-device allocations, capturing a Python stack trace for each allocation. The instrumentation is always enabled; :func:`device_memory_profile` provides an API to capture it. The output of :func:`device_memory_profile` is a binary protocol buffer that can be interpreted and visualized by the `pprof tool <https://github.com/google/pprof>`_. Args: backend: optional; the name of the JAX backend for which the device memory profile should be collected. Returns: A byte string containing a binary `pprof`-format protocol buffer. """ return xla_client.heap_profile(xla_bridge.get_backend(backend))
[docs]def save_device_memory_profile(filename, backend: Optional[str] = None): """Collects a device memory profile and writes it to a file. :func:`save_device_memory_profile` is a convenience wrapper around :func:`device_memory_profile` that saves its output to a ``filename``. See the :func:`device_memory_profile` documentation for more information. Args: filename: the filename to which the profile should be written. backend: optional; the name of the JAX backend for which the device memory profile should be collected. """ profile = device_memory_profile(backend) with open(filename, "wb") as f: f.write(profile)