New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Debugging for lightning data loader and fix for simple profiler. #3253
Conversation
@@ -99,7 +101,7 @@ def train_dataloader(self): | |||
# To avoid loading too much data in memory, need to calculate the queue size | |||
# dynamicaly, and limit the data loaded in queue. | |||
# Add 1 in size for storing the None in the end of each epoch. | |||
kwargs['async_loader_queue_size'] = max(1, min(100000 // kwargs['batch_size'], kwargs['limit_step_per_epoch'] // 4)) + 1 | |||
kwargs['async_loader_queue_size'] = max(1, min(2000000 // kwargs['batch_size'], kwargs['limit_step_per_epoch'] // 4)) + 1 |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
How is this 100000 or 2000000 calculated based on? This looks like a significant increase. If this is not dynamically determined, maybe we can define some constants.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It estimated based on deepETA's data. Only an estimation because the real data size in MB is unknown here.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
hmm ok, seems arbitrary. if possible, would be good to pass this in as an argument from our side so that we dont need to modify the horovod part again if anything changes.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Modify the code to let user pass the queue size.
Unit Test Results (with flaky tests) 753 files - 33 753 suites - 33 5h 41m 34s ⏱️ - 15m 9s Results for commit 7037d8c. ± Comparison against base commit 49678ce. ♻️ This comment has been updated with latest results. |
horovod/data/data_loader_base.py
Outdated
except Empty: | ||
break | ||
print(f"close_async_loader[{self.async_loader_queue_size}], joining...") |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
make these prints all under debug_data_loader=True
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yes, will add this.
self.queue.put(None) | ||
except Exception as ex: | ||
self.queue.put(ex) | ||
self.queue.put(None) | ||
finally: | ||
self.queue.put(None) | ||
print(f"_async_worker[{self.async_loader_queue_size}], stoped") |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
put this print under debug_data_loader
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Like to keep this outside, it is only 1 time printing in the end. We saw some data loader hang issue during close, it helps to detect that.
@@ -207,7 +210,7 @@ def on_epoch_end(self, trainer: "pl.Trainer", pl_module: "pl.LightningModule") - | |||
print("Creating trainer with: \n ", kwargs) | |||
trainer = Trainer(**kwargs) | |||
|
|||
if trainer.profiler: | |||
if profiler != 'simple' and trainer.profiler: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
why simple_profiler
doesn't need to set dirpath
here?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
it will be easier to just print the simple profiler log to stdout. It is not big, no need to put in file.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
overall it looks good to me. just a few nits to put prints under debug=True and a question of simple profiler.
Signed-off-by: Peng Zhang <pengz@uber.com>
Signed-off-by: Peng Zhang <pengz@uber.com>
Signed-off-by: Peng Zhang <pengz@uber.com>
Signed-off-by: Peng Zhang <pengz@uber.com>
76bdbaa
to
07428f9
Compare
3f6f99d
to
7037d8c
Compare
…ovod#3253) add debugging flag for lightning data loader , make async data loader queue size configurable Signed-off-by: weihanmines <weihan13@amd.com>
- Fixes issue when start_epoch != 0 Signed-off-by: Dinesh Ramasamy <89654805+iitmdinesh@users.noreply.github.com> Signed-off-by: weihanmines <weihan13@amd.com> fix torch op handles lazy release which may cause oom in elastic scenario (horovod#3110) * fix torch op handles lazy release which may cause oom in elastic scenario Signed-off-by: guoze.lin <guozelin@tencent.com> * Update mpi_ops.py Co-authored-by: guoze.lin <guozelin@tencent.com> Co-authored-by: Travis Addair <tgaddair@gmail.com> Signed-off-by: weihanmines <weihan13@amd.com> Added support for extraction of storage options from url. (horovod#3137) * Added support for extraction of storage options from url. Signed-off-by: Manjur Ansari <maansar@microsoft.com> * mock fsspec.utils Signed-off-by: Manjur Ansari <maansar@microsoft.com> * Added missing comma Co-authored-by: Travis Addair <tgaddair@gmail.com> Signed-off-by: weihanmines <weihan13@amd.com> Make RayExecutor use the current placement group if one exists (horovod#3134) Signed-off-by: weihanmines <weihan13@amd.com> Fix the mapping btw pyspark and numpy (horovod#3146) Signed-off-by: Haoyang Chen <haoyang@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Add tests for Keras callbacks: MetricAverageCallback, LearningRateScheduleCallback and LearningRateWarmupCallback (horovod#3102) There were no tests for MetricAverageCallback, LearningRateScheduleCallback and LearningRateWarmupCallback from hvd as noted in horovod#2659. This PR adds testing to verify the callback works. Signed-off-by: Moses Lee <14leeyuchieh@gmail.com> Co-authored-by: Moses Lee <molee@molee-ld4.linkedin.biz> Signed-off-by: weihanmines <weihan13@amd.com> Split gpu tests in head and non-head versions (horovod#3155) Signed-off-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Allow caller to customize the Tensorboard callback (horovod#3153) * Keras Estimator: Allow user to pass in TensorBoard callback Signed-off-by: Rich Porter <rich.porter@uber.com> * Remove callback from other processes on the same machine Signed-off-by: Rich Porter <rich.porter@uber.com> * Allow other ranks to profile as well. Doesn't seem to conflict Signed-off-by: Rich Porter <rich.porter@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> test_torch.py: add explicit join() for testing duplicated name errors (horovod#3159) For torch nightly >=10.0, we need to add an explict join() call to avoid hanging when testing duplicated name errors. Signed-off-by: Chongxiao Cao <chongxiaoc@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Disable TF2.6.0 XLA support on OSX (horovod#3133) Related to issue#3132 Signed-off-by: Chongxiao Cao <chongxiaoc@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Fix linking _pywrap_tensorflow_internal.so and re-enable XLA on macOS (horovod#3173) Signed-off-by: weihanmines <weihan13@amd.com> Spark/Lightning: fix the usage of checkpoint callback (horovod#3186) Signed-off-by: Chongxiao Cao <chongxiaoc@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Fix Cometlogger experiment key lost issue (horovod#3184) * test Signed-off-by: Peng Zhang <pengz@uber.com> * test Signed-off-by: Peng Zhang <pengz@uber.com> * fix_logger Signed-off-by: Peng Zhang <pengz@uber.com> * fix_logger Signed-off-by: Peng Zhang <pengz@uber.com> * recreate_loger Signed-off-by: Peng Zhang <pengz@uber.com> * fix_var Signed-off-by: Peng Zhang <pengz@uber.com> * test Signed-off-by: Peng Zhang <pengz@uber.com> * test Signed-off-by: Peng Zhang <pengz@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Updated torch c++ to use new aten api (horovod#3175) Signed-off-by: weihanmines <weihan13@amd.com> Spark/Keras: remove bare Keras support (horovod#3191) Signed-off-by: weihanmines <weihan13@amd.com> Make fork PRs publish test change stats (horovod#3185) Signed-off-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Support for nccl on cuda 11.4 (horovod#3182) Signed-off-by: Evan Brossard <evanb@maka-ars.com> Signed-off-by: weihanmines <weihan13@amd.com> Fix MPICH support (horovod#3148) * fix MPICH implementation * enable tests for MPICH and Intel MPI Signed-off-by: Jinzhe Zeng <jinzhe.zeng@rutgers.edu> Signed-off-by: weihanmines <weihan13@amd.com> Increase build timeout to 40m on Buildkite (horovod#3192) Signed-off-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Change CMake syntax to be compatible with old versions of CMake (horovod#3196) Signed-off-by: Max H. Gerlach <git@maxgerlach.de> Signed-off-by: weihanmines <weihan13@amd.com> Reinit every torch test (horovod#3194) Signed-off-by: weihanmines <weihan13@amd.com> Add barrier call to torch module to support easy synchronization for process sets (horovod#3139) * Added barrier call to torch module Signed-off-by: TJ <tix@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Bump version to 0.23.0 (horovod#3200) Signed-off-by: Travis Addair <tgaddair@gmail.com> Co-authored-by: Max H. Gerlach <git@maxgerlach.de> Signed-off-by: weihanmines <weihan13@amd.com> Increase Parallel PyTest timeout to 10m (horovod#3198) * Increase MPI and Gloo Parallel PyTest timeout to 10m Signed-off-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Spark/Lightning: don't overwrite model with checkpoint by default (horovod#3201) Lightning estimator saves model by default if there is no specified checkpoint callback. However, model is not overwritten with checkpoint file in that case. Signed-off-by: Chongxiao Cao <chongxiaoc@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Spark/Lightning: fix checkpoint callback dirpath typo (horovod#3204) Signed-off-by: Chongxiao Cao <chongxiaoc@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Rework events in CI workflows (horovod#3202) Signed-off-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Allow for concurrent schedule and master build, document concurrency (horovod#3206) Signed-off-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Ray: fix RayExecutor to fail when num_workers=0 and num_hosts=None (horovod#3210) Signed-off-by: Travis Addair <tgaddair@gmail.com> Signed-off-by: weihanmines <weihan13@amd.com> add_history_in_lightning_estimator (horovod#3214) Signed-off-by: Peng Zhang <pengz@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Allow buildkite building merge commits on forks (horovod#3215) Signed-off-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Fix json output in ci-results.yaml (horovod#3217) Signed-off-by: weihanmines <weihan13@amd.com> Spark/Lightning: fix history metrics for estimator serialization (horovod#3216) Save metrics inside the checkpoint dict , which will be load with map_location=torch.device('cpu') Signed-off-by: Peng Zhang <pengz@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> patch python source files on macCI (horovod#3220) * patch python source files on macCI * Trigger build and test CI Signed-off-by: TJ <tix@uber.com> Co-authored-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Updated examples of torch and tf to include mixed precision training (horovod#3222) * Added mixed precision example for pytorch * added mixed precision for keras Signed-off-by: TJ <tix@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Job buildkite-heads accesses ci-workflow outputs, add it to the needs (horovod#3225) Signed-off-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Fixes race condition for ray scale up down tests (horovod#3205) Ensure that at least one host from the previous set of hosts have been registered. Without this, the discovery script will "discover" the new set of hosts before the current set can register. This would result in a race condition. Consider a discovery schedule: ``` discovery_schedule = [ (10, ['host-1:2']), (30, ['host-1:2', 'host-2:1', 'host-3:1']), (None, ['host-2:1']), ] ``` The initial set is: ['host-1:2']. Before this is registered in the driver, the discovery script discovers the set: ['host-1:2', 'host-2:1', 'host-3:1'], and adds ['host-2:1', 'host-3:1']. However, since ['host-1:2'] has not registered, there is no coordinator to notify the workers. When host-1 and host-3 are removed, driver.resume will call _activate_workers, which will update the host assignments. It has a check to see if the intersection between the previous and current set of hosts. It finds that the previous set is ['host-1:2'], and the current set is ['host-2:1'], since there was no notification for the added and removed hosts. This ensures that the previous set of hosts can register before the current set is discovered. Signed-off-by: Abin Shahab <ashahab@linkedin.com> Signed-off-by: weihanmines <weihan13@amd.com> Removed a case of the default mutable argument pitfall (horovod#3227) Signed-off-by: Naelson Douglas <naelson17@gmail.com> Signed-off-by: weihanmines <weihan13@amd.com> Updates to TSC members (horovod#3234) Signed-off-by: Travis Addair <tgaddair@gmail.com> Signed-off-by: weihanmines <weihan13@amd.com> Add in-place broadcast for TensorFlow (horovod#3128) * Update comment in FindTensorflow.cmake Signed-off-by: Max H. Gerlach <git@maxgerlach.de> * Add in-place broadcast_() and broadcast_variables() for TF Signed-off-by: Max H. Gerlach <git@maxgerlach.de> * Include source files from TF in build to avoid missing symbol errors Signed-off-by: Max H. Gerlach <git@maxgerlach.de> * Limit build and test to TF 2.6+ Signed-off-by: Max H. Gerlach <git@maxgerlach.de> * Remove source files copied from TensorFlow The missing symbols are resolved by linking against _pywrap_tensorflow_internal.so, which was introduced to Horovod with PR horovod#3053. Signed-off-by: Max H. Gerlach <git@maxgerlach.de> * Fix possible type attribute values for HorovodBroadcastInplace Signed-off-by: Max H. Gerlach <git@maxgerlach.de> * Add reference variables to test Signed-off-by: Max H. Gerlach <git@maxgerlach.de> * Update comments, doc strings, changelog Signed-off-by: Max H. Gerlach <git@maxgerlach.de> Signed-off-by: weihanmines <weihan13@amd.com> [Elastic Horovod] Fix the bug for ElasticSampler and hvd.elastic.state (horovod#3144) Co-authored-by: gethinhu <gethinhu@tencent.com> Signed-off-by: weihanmines <weihan13@amd.com> a better way to handle nccl error under elastic scenario (horovod#3112) Signed-off-by: guoze.lin <guozelin@tencent.com> Signed-off-by: weihanmines <weihan13@amd.com> check torch version for mixed precision example (horovod#3238) Signed-off-by: weihanmines <weihan13@amd.com> Lightning: set limit_train_batches and limit_val_batches (horovod#3237) Tell Lightning trainer that how many batches a single epoch needs. Signed-off-by: Chongxiao Cao <chongxiaoc@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Spark/Lightning: reduce memory footprint of async dataloader (horovod#3239) Limit async data loader queue size. Signed-off-by: Peng Zhang <pengz@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Change default fusion threshold from 64MB to 128MB in docs (horovod#3241) Signed-off-by: weihanmines <weihan13@amd.com> fix the example of pytorch_lightning_mnist.py (horovod#3245) - remove unused arg parameters - fix model test issue on GPU Signed-off-by: Chongxiao Cao <chongxiaoc@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> CI: use latest pytorch_lightning with torchhead (horovod#3243) Signed-off-by: weihanmines <weihan13@amd.com> test_gradient_aggregation with real gradient instead of a constant (horovod#3176) This fixes issue horovod#2664 by performing gradient aggregation with a real gradient instead of a constant. PR: horovod#2647 shifts the gradient allreduce when the gradient is computed (both through the DistributedOptimizer or through the DistributedGradientTape). Which means that this unittest, by design in TF2.4, doesn't call allreduce in _aggregate_gradients(). Since this unittest provide a gradient as constant (without effectively computing it), the gradient will never be allreduced. The current change ensure that instead of a constant a real gradient is computed from a loss-function. Note: The current loss-function intentionally evaluates to zero. A future PR should convert it to a real loss function(e.g. MeanSquaredError) and compute gradients from that to test gradient aggregation. Signed-off-by: Abin Shahab <ashahab@linkedin.com> Signed-off-by: weihanmines <weihan13@amd.com> Remove MetricAverageCallback warning on tf >= 2.5 (horovod#3050) Signed-off-by: Henrique Mendonça <henrique.mendonca@cscs.ch> Signed-off-by: weihanmines <weihan13@amd.com> Fix Horovod pyarrow IndexError: list index out of range (horovod#3255) Signed-off-by: Weichen Xu <weichen.xu@databricks.com> Signed-off-by: weihanmines <weihan13@amd.com> Fixing up current CI test failures. (horovod#3259) Signed-off-by: Josh Romero <joshr@nvidia.com> Co-authored-by: Travis Addair <tgaddair@gmail.com> Co-authored-by: Enrico Minack <github@enrico.minack.dev> Signed-off-by: weihanmines <weihan13@amd.com> Revert "Fix Horovod pyarrow IndexError: list index out of range (horovod#3255)" (horovod#3265) This reverts commit 3efc229. Signed-off-by: Travis Addair <tgaddair@gmail.com> Signed-off-by: weihanmines <weihan13@amd.com> Debugging for lightning data loader and fix for simple profiler. (horovod#3253) add debugging flag for lightning data loader , make async data loader queue size configurable Signed-off-by: weihanmines <weihan13@amd.com> Call process_set._setup in init() to point to the correct native lib path (horovod#3258) * call setup for common process_set in remote trainers moved _setup call to init() Signed-off-by: TJ <tix@uber.com> Signed-off-by: weihanmines <weihan13@amd.com> Add support for MXNet async dependency engine. (horovod#3242) Signed-off-by: Josh Romero <joshr@nvidia.com> Signed-off-by: weihanmines <weihan13@amd.com>
Description