For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python class

EncoderInputs

`EncoderInputs`

class max.pipelines.lib.registration_encoder.EncoderInputs(*, tokens, attention_mask=None, input_row_offsets=None, return_n_logits=None)

source

Bases: Generic[TensorT]

The inputs of an ENCODER module’s __call__.

A field that is None is not a graph input: a padded encoder takes attention_mask, and a ragged one input_row_offsets and return_n_logits.

Parameters:

attention_mask: TensorT | None = None

source

1.0 for a real token and 0.0 for padding, [batch_size, seq_len], for a padded encoder.

input_row_offsets: TensorT | None = None

source

Each sequence’s start in tokens, for a ragged encoder.

return_n_logits: TensorT | None = None

source

The ragged batch’s logit count, which an encoder ignores; the graph takes it so a ragged encoder’s inputs match a decoder’s.

tokens: TensorT

source

The token IDs: [batch_size, seq_len] padded, or [total_seq_len] ragged.