torch.js has not been released yet.

torch.masked_fill

function masked_fill(input: unknown, mask: unknown, value: unknown): Tensor

Fills elements of input tensor with value where mask is True.

Creates a new tensor where positions indicated by the boolean mask are replaced with the specified value, while other positions retain original values. Essential for:

Attention masking: Setting padding tokens to -inf before softmax
Conditional updates: Selective tensor modification based on conditions
Loss masking: Zeroing out gradients for ignored positions
NaN/Inf handling: Replacing invalid values with valid ones

\begin{aligned} \text{output}_i = \begin{cases} \text{value} & \text{if mask}_i = \text{True} \\ \text{input}_i & \text{otherwise} \end{cases} \end{aligned}

Gradient: Gradient is 0 where mask is True, unchanged elsewhere
Broadcasting: Mask is broadcast to input shape if smaller
Memory: Creates new tensor; use masked_fill_ for in-place

Parameters

inputunknown: The input tensor (any shape, any dtype)
maskunknown: Boolean mask tensor (broadcastable to input shape)
valueunknown: The value to fill at masked positions

Returns

Tensor– New tensor with masked positions filled with value

Examples

// Basic masking
const x = torch.tensor([1, 2, 3, 4, 5]);
const mask = torch.tensor([true, false, true, false, true]);
torch.masked_fill(x, mask, 0);  // [0, 2, 0, 4, 0]

// Attention masking (set padding to -inf)
const scores = torch.randn(4, 4);
const padding_mask = torch.tensor([[false, false, true, true],
                                   [false, false, false, true],
                                   [false, false, false, false],
                                   [false, false, false, false]]);
const masked_scores = torch.masked_fill(scores, padding_mask, -Infinity);
const attention = torch.softmax(masked_scores, -1);  // Padding positions become 0

// Replace NaN values
const data = torch.tensor([1, NaN, 3, NaN, 5]);
const nan_mask = torch.isnan(data);
torch.masked_fill(data, nan_mask, 0);  // [1, 0, 3, 0, 5]

See Also

PyTorch torch.masked_fill()
where - Conditional element selection between two tensors
masked_scatter - Scatter values from source at masked positions
masked_select - Select elements where mask is True