config.yaml

# GridTok (2D) AR d36 - Text-to-Image (T2I) Inference Configuration (No MuP)
# Model: 36 blocks, embed_dim=2304, 36 heads
# Uses 2D grid tokenizer (16x16) instead of 1D (256x1)
# Uses nomup checkpoint (muP scaling baked into weights)

model:
  modality: image
  
  # For 2D grid, the AR checkpoint does NOT contain image_tokenizer weights.
  # The tokenizer is loaded from a separate HuggingFace repo.
  remove_image_tokenizer: true
  
  meta_model:
    _target_: l3m.model.meta_models.InSeriesMetaModels
    models:
      image_tokenizer:
        _target_: flextok_ar.model.integration.ImageResamplerTokenizer
        # 2D grid tokenizer hosted separately on HuggingFace
        model_id: ZhitongGao/GridAR_256
        force_vae_encode: true
        sample_posterior: true
        image_dims: [256, 256]
        read_key: image
        write_key: image_token_ids
        vae_latent_read_key: vae_latents
        del_decoders: false
        token_grid_size: [16, 16]
      
      text_encoder:
        _target_: flextok_ar.model.text_encoder.T5EmbedderWithMLP
        t5_embedder:
          _target_: flextok_ar.model.text_encoder.T5Embedder
          read_key: text
          text_embeddings_write_key: text_embeddings
          text_embeddings_mask_write_key: cross_attn_mask
          hf_hub_path: google/flan-t5-xl
          encoder_seqlen_max: 128
          decoder_seqlen: 256
          cond_dropout_p: 0.1
        mlp:
          _target_: flextok_ar.model.text_encoder.TextToEmbedMLP
          text_dim: 2048
          embed_dim: 2304
          act_layer:
            _target_: torch.nn.GELU
            _partial_: true
            approximate: tanh
          use_bias: false
      
      ar_image_model:
        _target_: l3m.model.meta_models.MetaModel
        preprocessor:
          _target_: flextok_ar.model.preprocessors.ARImageEmbedPreprocessor
          read_key: image_token_ids
          write_key:
          - input_embeddings
          - target_token_ids
          inference_read_key: pred_image_token_ids
          token_grid_size: [16, 16]
          codebook_size: 64000
          embed_dim: 2304
          pos_embed_type: absolute
          num_classes: null
        
        trunk:
          _target_: l3m.model.trunks.transformer_decoder.TransformerDecoder
          read_key: input_embeddings
          write_key: output_embeddings
          encoder_output_key: text_embeddings
          self_attn_mask_read_key: cross_attn_mask
          embed_dim: 2304
          num_blocks: 36
          mlp_ratio: 4
          norm_layer:
            _target_: l3m.model.layers.normalization.LayerNormFP32
            _partial_: true
            eps: 1.0e-05
          ffn_target:
            _target_: l3m.model.layers.ffn.SwiGLUFFN
            _partial_: true
          self_attn_target:
            _target_: l3m.model.layers.attention.EfficientAttention
            _partial_: true
            dim: 2304
            num_heads: 36
            qkv_bias: false
            is_causal: true
            qk_norm:
              _target_: l3m.model.layers.normalization.LayerNormFP32
              _partial_: true
              eps: 1.0e-05
            rope_pos_embed: null
          cross_attn_target:
            _target_: flextok_ar.model.attention.GeneralizedAttentionWithMask
            _partial_: true
            dim: 2304
            encoder_dim: 2304
            num_heads: 36
            qkv_bias: false
            is_causal: false
            qk_norm:
              _target_: l3m.model.layers.normalization.LayerNormFP32
              _partial_: true
              eps: 1.0e-05
          weight_init_style: jax
          post_trunk_norm: true
          use_bias: false
        
        postprocessor:
          _target_: torch.nn.Identity
        
        head:
          _target_: l3m.model.heads.classifier.LinearClassifier
          read_key: output_embeddings
          write_key: token_preds
          in_features: 2304
          out_features: 64000

generation:
  model_type: ar_text_to_image_model
  sample: true
  temperature: 1.0
  top_k: 0
  top_p: 0.0
  cfg_factor: 3.0
  num_keep_tokens: 256
  num_samples: 1
  timesteps: 25
  tokenizer_cfg_factor: 5.0
  tokenizer_perform_norm_guidance: true

# VAE image sizes for 2D grid decode (f8 VAE: 256/8 = 32)
decode:
  vae_image_sizes: 32

image:
  mean: [0.5, 0.5, 0.5]
  std: [0.5, 0.5, 0.5]
  size: 256

device: cuda
seed: 42