attn_dist_mode: v0
attn_positional_bias: rope
attn_positional_bias_n_spatial: 16
causal_norm: quiet_softmax
coord_dim: 2
d_model: 320
dropout: 0.0
feat_dim: 7
feat_embed_per_dim: 8
nhead: 4
num_decoder_layers: 6
num_encoder_layers: 6
pos_embed_per_dim: 32
spatial_pos_cutoff: 256
window: 4