Position encodings

For training from scratch, configure position encoding under model.embeddings. Eole propagates relative attention settings to its Transformer configuration. For converted checkpoints, preserve the converter's architecture-specific values.

model:
  architecture: transformer_lm
  embeddings:
    position_encoding: true
    position_encoding_type: SinusoidalInterleaved
training:
  param_init_method: xavier_uniform

SinusoidalConcat is another absolute sinusoidal layout. Learned positions use position_encoding_type: Learned and require n_positions. Shaw-style relative positions use Relative with n_positions specifying the maximum relative distance; rotary positions use Rotary, and ALiBi uses Alibi. These are different model architectures, not inference switches to change on an existing checkpoint. See the model configuration reference for rotary scaling, position limits, and validation rules.