(beta) Running the compiled optimizer with an LR Scheduler

Author: Michael Lazos

The optimizer is a key algorithm for training any deep learning model. In this example, we will show how to pair the optimizer, which has been compiled using torch.compile, with the LR schedulers to accelerate training convergence.

Note

This tutorial requires PyTorch 2.3.0 or later.

Model Setup

For this example, we’ll use a simple sequence of linear layers.

import torch

# Create simple model
model = torch.nn.Sequential(
    *[torch.nn.Linear(1024, 1024, False, device="cuda") for _ in range(10)]
)
input = torch.rand(1024, device="cuda")

# run forward pass
output = model(input)

# run backward to populate the grads for our optimizer below
output.sum().backward()

Setting up and running the compiled optimizer with LR Scheduler

In this section, we’ll use the Adam optimizer with LinearLR Scheduler and create a helper function to wrap the step() call for each of them in torch.compile().

Note

torch.compile is only supported on CUDA devices that have a compute capability of 7.0 or higher.

# exit cleanly if we are on a device that doesn't support ``torch.compile``
if torch.cuda.get_device_capability() < (7, 0):
    print("Exiting because torch.compile is not supported on this device.")
    import sys
    sys.exit(0)

# !!! IMPORTANT !!! Wrap the lr in a Tensor if we are pairing the
# the optimizer with an LR Scheduler.
# Without this, torch.compile will recompile as the value of the LR
# changes.
opt = torch.optim.Adam(model.parameters(), lr=torch.tensor(0.01))
sched = torch.optim.lr_scheduler.LinearLR(opt, total_iters=5)

@torch.compile(fullgraph=False)
def fn():
    opt.step()
    sched.step()


# Warmup runs to compile the function
for _ in range(5):
    fn()
    print(opt.param_groups[0]["lr"])

tensor(0.0047)
tensor(0.0060)
tensor(0.0073)
tensor(0.0087)
tensor(0.0100)

Extension: What happens with a non-tensor LR?

For the curious, we will show how to peek into what happens with torch.compile when we don’t wrap the LR in a tensor.

# No longer wrap the LR in a tensor here
opt = torch.optim.Adam(model.parameters(), lr=0.01)
sched = torch.optim.lr_scheduler.LinearLR(opt, total_iters=5)

@torch.compile(fullgraph=False)
def fn():
    opt.step()
    sched.step()

# Setup logging to view recompiles
torch._logging.set_logs(recompiles=True)

# Warmup runs to compile the function
# We will now recompile on each iteration
# as the value of the lr is mutated.
for _ in range(5):
    fn()

[rank0]:V1126 21:24:08.593000 619 torch/_dynamo/guards.py:2813] [7/2] [__recompiles] Recompiling function step in /usr/local/lib/python3.10/dist-packages/torch/optim/adam.py:189
[rank0]:V1126 21:24:08.593000 619 torch/_dynamo/guards.py:2813] [7/2] [__recompiles]     triggered by the following guard failure(s):
[rank0]:V1126 21:24:08.593000 619 torch/_dynamo/guards.py:2813] [7/2] [__recompiles]     - 7/1: L['self'].param_groups[0]['lr'] == 0.003333333333333333
[rank0]:V1126 21:24:11.094000 619 torch/_dynamo/guards.py:2813] [7/3] [__recompiles] Recompiling function step in /usr/local/lib/python3.10/dist-packages/torch/optim/adam.py:189
[rank0]:V1126 21:24:11.094000 619 torch/_dynamo/guards.py:2813] [7/3] [__recompiles]     triggered by the following guard failure(s):
[rank0]:V1126 21:24:11.094000 619 torch/_dynamo/guards.py:2813] [7/3] [__recompiles]     - 7/2: L['self'].param_groups[0]['lr'] == 0.004666666666666667
[rank0]:V1126 21:24:11.094000 619 torch/_dynamo/guards.py:2813] [7/3] [__recompiles]     - 7/1: L['self'].param_groups[0]['lr'] == 0.003333333333333333
[rank0]:V1126 21:24:13.483000 619 torch/_dynamo/guards.py:2813] [7/4] [__recompiles] Recompiling function step in /usr/local/lib/python3.10/dist-packages/torch/optim/adam.py:189
[rank0]:V1126 21:24:13.483000 619 torch/_dynamo/guards.py:2813] [7/4] [__recompiles]     triggered by the following guard failure(s):
[rank0]:V1126 21:24:13.483000 619 torch/_dynamo/guards.py:2813] [7/4] [__recompiles]     - 7/3: L['self'].param_groups[0]['lr'] == 0.006000000000000001
[rank0]:V1126 21:24:13.483000 619 torch/_dynamo/guards.py:2813] [7/4] [__recompiles]     - 7/2: L['self'].param_groups[0]['lr'] == 0.004666666666666667
[rank0]:V1126 21:24:13.483000 619 torch/_dynamo/guards.py:2813] [7/4] [__recompiles]     - 7/1: L['self'].param_groups[0]['lr'] == 0.003333333333333333
[rank0]:V1126 21:24:15.856000 619 torch/_dynamo/guards.py:2813] [7/5] [__recompiles] Recompiling function step in /usr/local/lib/python3.10/dist-packages/torch/optim/adam.py:189
[rank0]:V1126 21:24:15.856000 619 torch/_dynamo/guards.py:2813] [7/5] [__recompiles]     triggered by the following guard failure(s):
[rank0]:V1126 21:24:15.856000 619 torch/_dynamo/guards.py:2813] [7/5] [__recompiles]     - 7/4: L['self'].param_groups[0]['lr'] == 0.007333333333333335
[rank0]:V1126 21:24:15.856000 619 torch/_dynamo/guards.py:2813] [7/5] [__recompiles]     - 7/3: L['self'].param_groups[0]['lr'] == 0.006000000000000001
[rank0]:V1126 21:24:15.856000 619 torch/_dynamo/guards.py:2813] [7/5] [__recompiles]     - 7/2: L['self'].param_groups[0]['lr'] == 0.004666666666666667
[rank0]:V1126 21:24:15.856000 619 torch/_dynamo/guards.py:2813] [7/5] [__recompiles]     - 7/1: L['self'].param_groups[0]['lr'] == 0.003333333333333333

With this example, we can see that we recompile the optimizer a few times due to the guard failure on the lr in param_groups[0].

Conclusion

In this tutorial we showed how to pair the optimizer compiled with torch.compile with an LR Scheduler to accelerate training convergence. We used a model consisting of a simple sequence of linear layers with the Adam optimizer paired with a LinearLR scheduler to demonstrate the LR changing across iterations.

See also:

• Compiled optimizer tutorial - an intro into the compiled optimizer.

• Compiling the optimizer with PT2 - deeper technical details on the compiled optimizer.