Fast performance tips¶
Lightning builds in all the micro-optimizations we can find to increase your performance. But we can only automate so much.
Here are some additional things you can do to increase your performance.
When building your DataLoader set
num_workers > 0 and
pin_memory=True (only for GPUs).
Dataloader(dataset, num_workers=8, pin_memory=True)
The question of how many
num_workers is tricky. Here’s a summary of
some references, , and our suggestions.
num_workers=0means ONLY the main process will load batches (that can be a bottleneck).
num_workers=1means ONLY one worker (just not the main process) will load data but it will still be slow.
num_workersdepends on the batch size and your machine.
A general place to start is to set
num_workersequal to the number of CPUs on that machine.
num_workers will ALSO increase your CPU memory consumption.
The best thing to do is to increase the
num_workers slowly and stop once you see no more improvement in your training speed.
accelerator=ddp_spawn (the ddp default) or TPU training, the way multiple GPUs/TPU cores are used is by calling
.spawn() under the hood.
The problem is that PyTorch has issues with
num_workers > 0 when using
.spawn(). For this reason we recommend you
accelerator=ddp so you can increase the
num_workers, however your script has to be callable like so:
python my_program.py --gpus X
.item(), .numpy(), .cpu()¶
.item() anywhere in your code. Use
.detach() instead to remove the connected graph calls. Lightning
takes a great deal of care to be optimized for this.
Don’t call this unnecessarily! Every time you call this ALL your GPUs have to wait to sync.
Construct tensors directly on the device¶
LightningModules know what device they are on! Construct tensors on the device directly to avoid CPU->Device transfer.
# bad t = torch.rand(2, 2).cuda() # good (self is LightningModule) t = torch.rand(2, 2, device=self.device)
For tensors that need to be model attributes, it is best practice to register them as buffers in the modules’s
# bad self.t = torch.rand(2, 2, device=self.device) # good self.register_buffer("t", torch.rand(2, 2))
Use DDP not DP¶
DP performs three GPU transfers for EVERY batch:
Copy model to device.
Copy data to device.
Copy outputs of each device back to master.
Whereas DDP only performs 1 transfer to sync gradients. Because of this, DDP is MUCH faster than DP.
Use 16-bit to decrease the memory consumption (and thus increase your batch size). On certain GPUs (V100s, 2080tis), 16-bit calculations are also faster. However, know that 16-bit and multi-processing (any DDP) can have issues. Here are some common problems.
- CUDA error: an illegal memory access was encountered.
The solution is likely setting a specific CUDA, CUDNN, PyTorch version combination.
CUDA error: device-side assert triggered. This is a general catch-all error. To see the actual error run your script like so:
# won't see what the error is python main.py # will see what the error is CUDA_LAUNCH_BLOCKING=1 python main.py
We also recommend using 16-bit native found in PyTorch 1.6. Just install this version and Lightning will automatically use it.
Use Sharded DDP for GPU memory and scaling optimization¶
When training on multiple GPUs sharded DDP can assist to increase memory efficiency substantially, and in some cases performance on multi-node is better than traditional DDP. This is due to efficient communication and parallelization under the hood.
To use Optimizer Sharded Training, refer to Model Parallelism [BETA].
Sharded DDP can work across all DDP variants by adding the additional
--plugins ddp_sharded flag.
Refer to the distributed computing guide for more details.
Sequential Model Parallelism with Checkpointing¶
PyTorch Lightning integration for Sequential Model Parallelism using FairScale. Sequential Model Parallelism splits a sequential module onto multiple GPUs, reducing peak GPU memory requirements substantially.
For more information, refer to Sequential Model Parallelism with Checkpointing.
Preload Data Into RAM¶
When your training or preprocessing requires many operations to be performed on entire dataset(s) it can
sometimes be beneficial to store all data in RAM given there is enough space.
However, loading all data at the beginning of the training script has the disadvantage that it can take a long
time and hence it slows down the development process. Another downside is that in multiprocessing (e.g. DDP)
the data would get copied in each process.
One can overcome these problems by copying the data into RAM in advance.
Most UNIX-based operating systems provide direct access to tmpfs through a mount point typically named
Increase shared memory if necessary. Refer to the documentation of your OS how to do this.
Copy training data to shared memory:
cp -r /path/to/data/on/disk /dev/shm/
Refer to the new data root in your script or command line arguments:
datamodule = MyDataModule(data_root="/dev/shm/my_data")