lucid.no_grad¶

lucid.no_grad¶: alias of _NoGrad

Overview¶

no_grad temporarily disables gradient calculation. It can be used either

as a context manager

with lucid.no_grad():
    output = model(x)

or as a decorator

@lucid.no_grad()
def inference(x):
    return model(x)

Using it during inference (or any code that does not require gradients) saves memory and speeds up computation.

Function Signature¶

class no_grad:
    def __enter__(self): ...
    def __exit__(self, exc_type, exc_value, traceback): ...
    def __call__(self, fn): ...

Parameters¶

no_grad takes no arguments. Calling it—either in a with statement or as @lucid.no_grad()—creates a new instance that manages gradient state for the enclosed block or function.

Returns¶

When used as a context manager - returns the instance itself so that nested usage is safe.

When used as a decorator - returns a wrapped function that will execute with gradient tracking disabled.

Examples¶

Context-manager usage:

with lucid.no_grad():
    preds = model(inputs)  # gradients are *not* stored

Decorator usage:

@lucid.no_grad()
def evaluate(batch):
    return model(batch).argmax(-1)

Note

The previous global state is restored even if an exception occurs.
Avoid wrapping training steps with no_grad; doing so prevents parameter updates.

Benefits¶

Reduced memory usage - tensors skip gradient bookkeeping.
Faster execution - no backward graph is built.
Clean syntax - one line disables and then restores gradients.

Caution

Because it toggles a global flag, be mindful when mixing code that both requires and disables gradients.