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 function

logits_outputs

logits_outputs()

max.nn.transformer.logits_outputs(next_token_logits, logits, logit_offsets, return_hidden_states, hidden_states)

source

Names a decoder’s outputs, in the order the graph returns them.

Parameters:

• next_token_logits (TensorValue | None) – Each sequence’s last-token logits, or None when the caller only wants the variable logits.
• logits (TensorValue | None) – The requested logits beyond the last token, if any.
• logit_offsets (TensorValue | None) – Where each sequence’s rows start in logits.
• return_hidden_states (ReturnHiddenStates) – Which hidden states hidden_states holds, which names them.
• hidden_states (Sequence[TensorValue]) – The hidden states extract_hs() returned.

Returns:

The logits group, then a hidden_states group when there are any.

Return type:

DecoderOutputs