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

LogitsPostprocessMixin

`LogitsPostprocessMixin`

class max.nn.transformer.LogitsPostprocessMixin

source

Mixin providing logits postprocessing for single-device models.

Subclasses must define norm, lm_head, and return_logits. The return_hidden_states and logits_scaling attributes are optional and have defaults.

lm_head: Callable[[TensorValue], TensorValue]

source

The language model head projection.

logits_scaling: float = 1.0

source

A divisor applied to logits after projection. Logits are divided by this value before returning. Defaults to 1.0 (no scaling).

norm: Callable[[TensorValue], TensorValue]

source

The normalization function applied before the language model head.

return_hidden_states: ReturnHiddenStates = 'none'

source

Which hidden states to return. Defaults to ReturnHiddenStates.NONE.

return_logits: ReturnLogits

source

Which logits to return.