CUDA Automatic Mixed Precision examples

Ordinarily, “automatic mixed precision training” means training with torch.autocast and torch.cuda.amp.GradScaler together.

Instances of torch.autocast enable autocasting for chosen regions. Autocasting automatically chooses the precision for GPU operations to improve performance while maintaining accuracy.

Instances of torch.cuda.amp.GradScaler help perform the steps of gradient scaling conveniently. Gradient scaling improves convergence for networks with float16 gradients by minimizing gradient underflow, as explained here.

torch.autocast and torch.cuda.amp.GradScaler are modular. In the samples below, each is used as its individual documentation suggests.

(Samples here are illustrative. See the Automatic Mixed Precision recipe for a runnable walkthrough.)

Typical Mixed Precision Training

# Creates model and optimizer in default precision
model = Net().cuda()
optimizer = optim.SGD(model.parameters(), ...)

# Creates a GradScaler once at the beginning of training.
scaler = GradScaler()

for epoch in epochs:
    for input, target in data:
        optimizer.zero_grad()

        # Runs the forward pass with autocasting.
        with autocast(device_type='cuda', dtype=torch.float16):
            output = model(input)
            loss = loss_fn(output, target)

        # Scales loss.  Calls backward() on scaled loss to create scaled gradients.
        # Backward passes under autocast are not recommended.
        # Backward ops run in the same dtype autocast chose for corresponding forward ops.
        scaler.scale(loss).backward()

        # scaler.step() first unscales the gradients of the optimizer's assigned params.
        # If these gradients do not contain infs or NaNs, optimizer.step() is then called,
        # otherwise, optimizer.step() is skipped.
        scaler.step(optimizer)

        # Updates the scale for next iteration.
        scaler.update()

Working with Unscaled Gradients

All gradients produced by scaler.scale(loss).backward() are scaled. If you wish to modify or inspect the parameters’ .grad attributes between backward() and scaler.step(optimizer), you should unscale them first. For example, gradient clipping manipulates a set of gradients such that their global norm (see torch.nn.utils.clip_grad_norm_()) or maximum magnitude (see torch.nn.utils.clip_grad_value_()) is $<=$ some user-imposed threshold. If you attempted to clip without unscaling, the gradients’ norm/maximum magnitude would also be scaled, so your requested threshold (which was meant to be the threshold for unscaled gradients) would be invalid.

scaler.unscale_(optimizer) unscales gradients held by optimizer’s assigned parameters. If your model or models contain other parameters that were assigned to another optimizer (say optimizer2), you may call scaler.unscale_(optimizer2) separately to unscale those parameters’ gradients as well.

Gradient clipping

Calling scaler.unscale_(optimizer) before clipping enables you to clip unscaled gradients as usual:

scaler = GradScaler()

for epoch in epochs:
    for input, target in data:
        optimizer.zero_grad()
        with autocast(device_type='cuda', dtype=torch.float16):
            output = model(input)
            loss = loss_fn(output, target)
        scaler.scale(loss).backward()

        # Unscales the gradients of optimizer's assigned params in-place
        scaler.unscale_(optimizer)

        # Since the gradients of optimizer's assigned params are unscaled, clips as usual:
        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)

        # optimizer's gradients are already unscaled, so scaler.step does not unscale them,
        # although it still skips optimizer.step() if the gradients contain infs or NaNs.
        scaler.step(optimizer)

        # Updates the scale for next iteration.
        scaler.update()

scaler records that scaler.unscale_(optimizer) was already called for this optimizer this iteration, so scaler.step(optimizer) knows not to redundantly unscale gradients before (internally) calling optimizer.step().

Warning

unscale_ should only be called once per optimizer per step call, and only after all gradients for that optimizer’s assigned parameters have been accumulated. Calling unscale_ twice for a given optimizer between each step triggers a RuntimeError.

Working with Scaled Gradients

Gradient accumulation

Gradient accumulation adds gradients over an effective batch of size batch_per_iter * iters_to_accumulate (* num_procs if distributed). The scale should be calibrated for the effective batch, which means inf/NaN checking, step skipping if inf/NaN grads are found, and scale updates should occur at effective-batch granularity. Also, grads should remain scaled, and the scale factor should remain constant, while grads for a given effective batch are accumulated. If grads are unscaled (or the scale factor changes) before accumulation is complete, the next backward pass will add scaled grads to unscaled grads (or grads scaled by a different factor) after which it’s impossible to recover the accumulated unscaled grads step must apply.

Therefore, if you want to unscale_ grads (e.g., to allow clipping unscaled grads), call unscale_ just before step, after all (scaled) grads for the upcoming step have been accumulated. Also, only call update at the end of iterations where you called step for a full effective batch:

scaler = GradScaler()

for epoch in epochs:
    for i, (input, target) in enumerate(data):
        with autocast(device_type='cuda', dtype=torch.float16):
            output = model(input)
            loss = loss_fn(output, target)
            loss = loss / iters_to_accumulate

        # Accumulates scaled gradients.
        scaler.scale(loss).backward()

        if (i + 1) % iters_to_accumulate == 0:
            # may unscale_ here if desired (e.g., to allow clipping unscaled gradients)

            scaler.step(optimizer)
            scaler.update()
            optimizer.zero_grad()

Gradient penalty

A gradient penalty implementation commonly creates gradients using torch.autograd.grad(), combines them to create the penalty value, and adds the penalty value to the loss.

Here’s an ordinary example of an L2 penalty without gradient scaling or autocasting:

for epoch in epochs:
    for input, target in data:
        optimizer.zero_grad()
        output = model(input)
        loss = loss_fn(output, target)

        # Creates gradients
        grad_params = torch.autograd.grad(outputs=loss,
                                          inputs=model.parameters(),
                                          create_graph=True)

        # Computes the penalty term and adds it to the loss
        grad_norm = 0
        for grad in grad_params:
            grad_norm += grad.pow(2).sum()
        grad_norm = grad_norm.sqrt()
        loss = loss + grad_norm

        loss.backward()

        # clip gradients here, if desired

        optimizer.step()

To implement a gradient penalty with gradient scaling, the outputs Tensor(s) passed to torch.autograd.grad() should be scaled. The resulting gradients will therefore be scaled, and should be unscaled before being combined to create the penalty value.

Also, the penalty term computation is part of the forward pass, and therefore should be inside an autocast context.

Here’s how that looks for the same L2 penalty:

scaler = GradScaler()

for epoch in epochs:
    for input, target in data:
        optimizer.zero_grad()
        with autocast(device_type='cuda', dtype=torch.float16):
            output = model(input)
            loss = loss_fn(output, target)

        # Scales the loss for autograd.grad's backward pass, producing scaled_grad_params
        scaled_grad_params = torch.autograd.grad(outputs=scaler.scale(loss),
                                                 inputs=model.parameters(),
                                                 create_graph=True)

        # Creates unscaled grad_params before computing the penalty. scaled_grad_params are
        # not owned by any optimizer, so ordinary division is used instead of scaler.unscale_:
        inv_scale = 1./scaler.get_scale()
        grad_params = [p * inv_scale for p in scaled_grad_params]

        # Computes the penalty term and adds it to the loss
        with autocast(device_type='cuda', dtype=torch.float16):
            grad_norm = 0
            for grad in grad_params:
                grad_norm += grad.pow(2).sum()
            grad_norm = grad_norm.sqrt()
            loss = loss + grad_norm

        # Applies scaling to the backward call as usual.
        # Accumulates leaf gradients that are correctly scaled.
        scaler.scale(loss).backward()

        # may unscale_ here if desired (e.g., to allow clipping unscaled gradients)

        # step() and update() proceed as usual.
        scaler.step(optimizer)
        scaler.update()

Working with Multiple Models, Losses, and Optimizers

If your network has multiple losses, you must call scaler.scale on each of them individually. If your network has multiple optimizers, you may call scaler.unscale_ on any of them individually, and you must call scaler.step on each of them individually.

However, scaler.update should only be called once, after all optimizers used this iteration have been stepped:

scaler = torch.cuda.amp.GradScaler()

for epoch in epochs:
    for input, target in data:
        optimizer0.zero_grad()
        optimizer1.zero_grad()
        with autocast(device_type='cuda', dtype=torch.float16):
            output0 = model0(input)
            output1 = model1(input)
            loss0 = loss_fn(2 * output0 + 3 * output1, target)
            loss1 = loss_fn(3 * output0 - 5 * output1, target)

        # (retain_graph here is unrelated to amp, it's present because in this
        # example, both backward() calls share some sections of graph.)
        scaler.scale(loss0).backward(retain_graph=True)
        scaler.scale(loss1).backward()

        # You can choose which optimizers receive explicit unscaling, if you
        # want to inspect or modify the gradients of the params they own.
        scaler.unscale_(optimizer0)

        scaler.step(optimizer0)
        scaler.step(optimizer1)

        scaler.update()

Each optimizer checks its gradients for infs/NaNs and makes an independent decision whether or not to skip the step. This may result in one optimizer skipping the step while the other one does not. Since step skipping occurs rarely (every several hundred iterations) this should not impede convergence. If you observe poor convergence after adding gradient scaling to a multiple-optimizer model, please report a bug.

Working with Multiple GPUs

The issues described here only affect autocast. GradScaler‘s usage is unchanged.

DataParallel in a single process

Even if torch.nn.DataParallel spawns threads to run the forward pass on each device. The autocast state is propagated in each one and the following will work:

model = MyModel()
dp_model = nn.DataParallel(model)

# Sets autocast in the main thread
with autocast(device_type='cuda', dtype=torch.float16):
    # dp_model's internal threads will autocast.
    output = dp_model(input)
    # loss_fn also autocast
    loss = loss_fn(output)

DistributedDataParallel, one GPU per process

torch.nn.parallel.DistributedDataParallel’s documentation recommends one GPU per process for best performance. In this case, DistributedDataParallel does not spawn threads internally, so usages of autocast and GradScaler are not affected.

DistributedDataParallel, multiple GPUs per process

Here torch.nn.parallel.DistributedDataParallel may spawn a side thread to run the forward pass on each device, like torch.nn.DataParallel. The fix is the same: apply autocast as part of your model’s forward method to ensure it’s enabled in side threads.

Autocast and Custom Autograd Functions

If your network uses custom autograd functions (subclasses of torch.autograd.Function), changes are required for autocast compatibility if any function

• takes multiple floating-point Tensor inputs,

• wraps any autocastable op (see the Autocast Op Reference), or

• requires a particular dtype (for example, if it wraps CUDA extensions that were only compiled for dtype).

In all cases, if you’re importing the function and can’t alter its definition, a safe fallback is to disable autocast and force execution in float32 ( or dtype) at any points of use where errors occur:

with autocast(device_type='cuda', dtype=torch.float16):
    ...
    with autocast(device_type='cuda', dtype=torch.float16, enabled=False):
        output = imported_function(input1.float(), input2.float())

If you’re the function’s author (or can alter its definition) a better solution is to use the torch.cuda.amp.custom_fwd() and torch.cuda.amp.custom_bwd() decorators as shown in the relevant case below.

Functions with multiple inputs or autocastable ops

Apply custom_fwd and custom_bwd (with no arguments) to forward and backward respectively. These ensure forward executes with the current autocast state and backward executes with the same autocast state as forward (which can prevent type mismatch errors):

class MyMM(torch.autograd.Function):
    @staticmethod
    @custom_fwd
    def forward(ctx, a, b):
        ctx.save_for_backward(a, b)
        return a.mm(b)
    @staticmethod
    @custom_bwd
    def backward(ctx, grad):
        a, b = ctx.saved_tensors
        return grad.mm(b.t()), a.t().mm(grad)

Now MyMM can be invoked anywhere, without disabling autocast or manually casting inputs:

mymm = MyMM.apply

with autocast(device_type='cuda', dtype=torch.float16):
    output = mymm(input1, input2)

Functions that need a particular dtype

Consider a custom function that requires torch.float32 inputs. Apply custom_fwd(cast_inputs=torch.float32) to forward and custom_bwd (with no arguments) to backward. If forward runs in an autocast-enabled region, the decorators cast floating-point CUDA Tensor inputs to float32, and locally disable autocast during forward and backward:

class MyFloat32Func(torch.autograd.Function):
    @staticmethod
    @custom_fwd(cast_inputs=torch.float32)
    def forward(ctx, input):
        ctx.save_for_backward(input)
        ...
        return fwd_output
    @staticmethod
    @custom_bwd
    def backward(ctx, grad):
        ...

Now MyFloat32Func can be invoked anywhere, without manually disabling autocast or casting inputs:

func = MyFloat32Func.apply

with autocast(device_type='cuda', dtype=torch.float16):
    # func will run in float32, regardless of the surrounding autocast state
    output = func(input)