diff --git a/ferritin-esm/Readme.md b/ferritin-esm/Readme.md
index 4301c1b1..f9adf544 100644
--- a/ferritin-esm/Readme.md
+++ b/ferritin-esm/Readme.md
@@ -28,6 +28,31 @@ for (name, tensor) in pth.tensor_infos() {
 ```
 
 
+```text
+embed.weight: TensorInfo { name: "embed.weight", dtype: F32, layout: Layout { shape: [64, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/301", storage_size: 61440 }
+
+transformer.norm.weight: TensorInfo { name: "transformer.norm.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/300", storage_size: 960 }
+
+sequence_head.0.bias: TensorInfo { name: "sequence_head.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/303", storage_size: 960 }
+sequence_head.0.weight: TensorInfo { name: "sequence_head.0.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/302", storage_size: 921600 }
+sequence_head.2.bias: TensorInfo { name: "sequence_head.2.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/305", storage_size: 960 }
+sequence_head.2.weight: TensorInfo { name: "sequence_head.2.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/304", storage_size: 960 }
+sequence_head.3.bias: TensorInfo { name: "sequence_head.3.bias", dtype: F32, layout: Layout { shape: [64], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/307", storage_size: 64 }
+sequence_head.3.weight: TensorInfo { name: "sequence_head.3.weight", dtype: F32, layout: Layout { shape: [64, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/306", storage_size: 61440 }
+
+transformer.blocks.0.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.0.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/4", storage_size: 960 }
+transformer.blocks.0.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.0.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/1", storage_size: 960 }
+transformer.blocks.0.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.0.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/0", storage_size: 960 }
+transformer.blocks.0.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.0.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/2", storage_size: 2764800 }
+transformer.blocks.0.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.0.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/5", storage_size: 921600 }
+transformer.blocks.0.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.0.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/3", storage_size: 960 }
+transformer.blocks.0.ffn.0.bias: TensorInfo { name: "transformer.blocks.0.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/7", storage_size: 960 }
+transformer.blocks.0.ffn.0.weight: TensorInfo { name: "transformer.blocks.0.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/6", storage_size: 960 }
+transformer.blocks.0.ffn.1.weight: TensorInfo { name: "transformer.blocks.0.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/8", storage_size: 4915200 }
+transformer.blocks.0.ffn.3.weight: TensorInfo { name: "transformer.blocks.0.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/9", storage_size: 2457600 }
+```
+
+
 ```text
 embed.weight: TensorInfo { name: "embed.weight", dtype: F32, layout: Layout { shape: [64, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/301", storage_size: 61440 }
 sequence_head.0.bias: TensorInfo { name: "sequence_head.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/303", storage_size: 960 }
@@ -56,6 +81,106 @@ transformer.blocks.1.ffn.0.bias: TensorInfo { name: "transformer.blocks.1.ffn.0.
 transformer.blocks.1.ffn.0.weight: TensorInfo { name: "transformer.blocks.1.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/16", storage_size: 960 }
 transformer.blocks.1.ffn.1.weight: TensorInfo { name: "transformer.blocks.1.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/18", storage_size: 4915200 }
 transformer.blocks.1.ffn.3.weight: TensorInfo { name: "transformer.blocks.1.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/19", storage_size: 2457600 }
+transformer.blocks.10.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.10.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/104", storage_size: 960 }
+transformer.blocks.10.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.10.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/101", storage_size: 960 }
+transformer.blocks.10.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.10.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/100", storage_size: 960 }
+transformer.blocks.10.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.10.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/102", storage_size: 2764800 }
+transformer.blocks.10.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.10.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/105", storage_size: 921600 }
+transformer.blocks.10.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.10.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/103", storage_size: 960 }
+transformer.blocks.10.ffn.0.bias: TensorInfo { name: "transformer.blocks.10.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/107", storage_size: 960 }
+transformer.blocks.10.ffn.0.weight: TensorInfo { name: "transformer.blocks.10.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/106", storage_size: 960 }
+transformer.blocks.10.ffn.1.weight: TensorInfo { name: "transformer.blocks.10.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/108", storage_size: 4915200 }
+transformer.blocks.10.ffn.3.weight: TensorInfo { name: "transformer.blocks.10.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/109", storage_size: 2457600 }
+transformer.blocks.11.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.11.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/114", storage_size: 960 }
+transformer.blocks.11.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.11.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/111", storage_size: 960 }
+transformer.blocks.11.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.11.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/110", storage_size: 960 }
+transformer.blocks.11.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.11.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/112", storage_size: 2764800 }
+transformer.blocks.11.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.11.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/115", storage_size: 921600 }
+transformer.blocks.11.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.11.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/113", storage_size: 960 }
+transformer.blocks.11.ffn.0.bias: TensorInfo { name: "transformer.blocks.11.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/117", storage_size: 960 }
+transformer.blocks.11.ffn.0.weight: TensorInfo { name: "transformer.blocks.11.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/116", storage_size: 960 }
+transformer.blocks.11.ffn.1.weight: TensorInfo { name: "transformer.blocks.11.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/118", storage_size: 4915200 }
+transformer.blocks.11.ffn.3.weight: TensorInfo { name: "transformer.blocks.11.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/119", storage_size: 2457600 }
+transformer.blocks.12.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.12.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/124", storage_size: 960 }
+transformer.blocks.12.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.12.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/121", storage_size: 960 }
+transformer.blocks.12.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.12.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/120", storage_size: 960 }
+transformer.blocks.12.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.12.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/122", storage_size: 2764800 }
+transformer.blocks.12.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.12.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/125", storage_size: 921600 }
+transformer.blocks.12.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.12.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/123", storage_size: 960 }
+transformer.blocks.12.ffn.0.bias: TensorInfo { name: "transformer.blocks.12.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/127", storage_size: 960 }
+transformer.blocks.12.ffn.0.weight: TensorInfo { name: "transformer.blocks.12.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/126", storage_size: 960 }
+transformer.blocks.12.ffn.1.weight: TensorInfo { name: "transformer.blocks.12.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/128", storage_size: 4915200 }
+transformer.blocks.12.ffn.3.weight: TensorInfo { name: "transformer.blocks.12.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/129", storage_size: 2457600 }
+transformer.blocks.13.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.13.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/134", storage_size: 960 }
+transformer.blocks.13.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.13.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/131", storage_size: 960 }
+transformer.blocks.13.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.13.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/130", storage_size: 960 }
+transformer.blocks.13.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.13.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/132", storage_size: 2764800 }
+transformer.blocks.13.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.13.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/135", storage_size: 921600 }
+transformer.blocks.13.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.13.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/133", storage_size: 960 }
+transformer.blocks.13.ffn.0.bias: TensorInfo { name: "transformer.blocks.13.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/137", storage_size: 960 }
+transformer.blocks.13.ffn.0.weight: TensorInfo { name: "transformer.blocks.13.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/136", storage_size: 960 }
+transformer.blocks.13.ffn.1.weight: TensorInfo { name: "transformer.blocks.13.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/138", storage_size: 4915200 }
+transformer.blocks.13.ffn.3.weight: TensorInfo { name: "transformer.blocks.13.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/139", storage_size: 2457600 }
+transformer.blocks.14.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.14.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/144", storage_size: 960 }
+transformer.blocks.14.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.14.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/141", storage_size: 960 }
+transformer.blocks.14.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.14.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/140", storage_size: 960 }
+transformer.blocks.14.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.14.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/142", storage_size: 2764800 }
+transformer.blocks.14.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.14.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/145", storage_size: 921600 }
+transformer.blocks.14.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.14.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/143", storage_size: 960 }
+transformer.blocks.14.ffn.0.bias: TensorInfo { name: "transformer.blocks.14.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/147", storage_size: 960 }
+transformer.blocks.14.ffn.0.weight: TensorInfo { name: "transformer.blocks.14.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/146", storage_size: 960 }
+transformer.blocks.14.ffn.1.weight: TensorInfo { name: "transformer.blocks.14.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/148", storage_size: 4915200 }
+transformer.blocks.14.ffn.3.weight: TensorInfo { name: "transformer.blocks.14.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/149", storage_size: 2457600 }
+transformer.blocks.15.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.15.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/154", storage_size: 960 }
+transformer.blocks.15.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.15.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/151", storage_size: 960 }
+transformer.blocks.15.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.15.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/150", storage_size: 960 }
+transformer.blocks.15.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.15.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/152", storage_size: 2764800 }
+transformer.blocks.15.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.15.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/155", storage_size: 921600 }
+transformer.blocks.15.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.15.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/153", storage_size: 960 }
+transformer.blocks.15.ffn.0.bias: TensorInfo { name: "transformer.blocks.15.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/157", storage_size: 960 }
+transformer.blocks.15.ffn.0.weight: TensorInfo { name: "transformer.blocks.15.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/156", storage_size: 960 }
+transformer.blocks.15.ffn.1.weight: TensorInfo { name: "transformer.blocks.15.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/158", storage_size: 4915200 }
+transformer.blocks.15.ffn.3.weight: TensorInfo { name: "transformer.blocks.15.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/159", storage_size: 2457600 }
+transformer.blocks.16.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.16.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/164", storage_size: 960 }
+transformer.blocks.16.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.16.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/161", storage_size: 960 }
+transformer.blocks.16.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.16.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/160", storage_size: 960 }
+transformer.blocks.16.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.16.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/162", storage_size: 2764800 }
+transformer.blocks.16.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.16.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/165", storage_size: 921600 }
+transformer.blocks.16.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.16.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/163", storage_size: 960 }
+transformer.blocks.16.ffn.0.bias: TensorInfo { name: "transformer.blocks.16.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/167", storage_size: 960 }
+transformer.blocks.16.ffn.0.weight: TensorInfo { name: "transformer.blocks.16.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/166", storage_size: 960 }
+transformer.blocks.16.ffn.1.weight: TensorInfo { name: "transformer.blocks.16.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/168", storage_size: 4915200 }
+transformer.blocks.16.ffn.3.weight: TensorInfo { name: "transformer.blocks.16.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/169", storage_size: 2457600 }
+transformer.blocks.17.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.17.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/174", storage_size: 960 }
+transformer.blocks.17.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.17.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/171", storage_size: 960 }
+transformer.blocks.17.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.17.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/170", storage_size: 960 }
+transformer.blocks.17.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.17.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/172", storage_size: 2764800 }
+transformer.blocks.17.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.17.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/175", storage_size: 921600 }
+transformer.blocks.17.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.17.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/173", storage_size: 960 }
+transformer.blocks.17.ffn.0.bias: TensorInfo { name: "transformer.blocks.17.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/177", storage_size: 960 }
+transformer.blocks.17.ffn.0.weight: TensorInfo { name: "transformer.blocks.17.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/176", storage_size: 960 }
+transformer.blocks.17.ffn.1.weight: TensorInfo { name: "transformer.blocks.17.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/178", storage_size: 4915200 }
+transformer.blocks.17.ffn.3.weight: TensorInfo { name: "transformer.blocks.17.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/179", storage_size: 2457600 }
+transformer.blocks.18.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.18.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/184", storage_size: 960 }
+transformer.blocks.18.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.18.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/181", storage_size: 960 }
+transformer.blocks.18.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.18.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/180", storage_size: 960 }
+transformer.blocks.18.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.18.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/182", storage_size: 2764800 }
+transformer.blocks.18.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.18.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/185", storage_size: 921600 }
+transformer.blocks.18.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.18.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/183", storage_size: 960 }
+transformer.blocks.18.ffn.0.bias: TensorInfo { name: "transformer.blocks.18.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/187", storage_size: 960 }
+transformer.blocks.18.ffn.0.weight: TensorInfo { name: "transformer.blocks.18.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/186", storage_size: 960 }
+transformer.blocks.18.ffn.1.weight: TensorInfo { name: "transformer.blocks.18.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/188", storage_size: 4915200 }
+transformer.blocks.18.ffn.3.weight: TensorInfo { name: "transformer.blocks.18.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/189", storage_size: 2457600 }
+transformer.blocks.19.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.19.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/194", storage_size: 960 }
+transformer.blocks.19.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.19.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/191", storage_size: 960 }
+transformer.blocks.19.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.19.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/190", storage_size: 960 }
+transformer.blocks.19.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.19.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/192", storage_size: 2764800 }
+transformer.blocks.19.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.19.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/195", storage_size: 921600 }
+transformer.blocks.19.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.19.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/193", storage_size: 960 }
+transformer.blocks.19.ffn.0.bias: TensorInfo { name: "transformer.blocks.19.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/197", storage_size: 960 }
+transformer.blocks.19.ffn.0.weight: TensorInfo { name: "transformer.blocks.19.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/196", storage_size: 960 }
+transformer.blocks.19.ffn.1.weight: TensorInfo { name: "transformer.blocks.19.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/198", storage_size: 4915200 }
+transformer.blocks.19.ffn.3.weight: TensorInfo { name: "transformer.blocks.19.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/199", storage_size: 2457600 }
 transformer.blocks.2.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.2.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/24", storage_size: 960 }
 transformer.blocks.2.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.2.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/21", storage_size: 960 }
 transformer.blocks.2.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.2.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/20", storage_size: 960 }
@@ -66,6 +191,106 @@ transformer.blocks.2.ffn.0.bias: TensorInfo { name: "transformer.blocks.2.ffn.0.
 transformer.blocks.2.ffn.0.weight: TensorInfo { name: "transformer.blocks.2.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/26", storage_size: 960 }
 transformer.blocks.2.ffn.1.weight: TensorInfo { name: "transformer.blocks.2.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/28", storage_size: 4915200 }
 transformer.blocks.2.ffn.3.weight: TensorInfo { name: "transformer.blocks.2.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/29", storage_size: 2457600 }
+transformer.blocks.20.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.20.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/204", storage_size: 960 }
+transformer.blocks.20.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.20.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/201", storage_size: 960 }
+transformer.blocks.20.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.20.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/200", storage_size: 960 }
+transformer.blocks.20.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.20.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/202", storage_size: 2764800 }
+transformer.blocks.20.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.20.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/205", storage_size: 921600 }
+transformer.blocks.20.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.20.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/203", storage_size: 960 }
+transformer.blocks.20.ffn.0.bias: TensorInfo { name: "transformer.blocks.20.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/207", storage_size: 960 }
+transformer.blocks.20.ffn.0.weight: TensorInfo { name: "transformer.blocks.20.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/206", storage_size: 960 }
+transformer.blocks.20.ffn.1.weight: TensorInfo { name: "transformer.blocks.20.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/208", storage_size: 4915200 }
+transformer.blocks.20.ffn.3.weight: TensorInfo { name: "transformer.blocks.20.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/209", storage_size: 2457600 }
+transformer.blocks.21.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.21.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/214", storage_size: 960 }
+transformer.blocks.21.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.21.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/211", storage_size: 960 }
+transformer.blocks.21.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.21.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/210", storage_size: 960 }
+transformer.blocks.21.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.21.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/212", storage_size: 2764800 }
+transformer.blocks.21.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.21.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/215", storage_size: 921600 }
+transformer.blocks.21.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.21.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/213", storage_size: 960 }
+transformer.blocks.21.ffn.0.bias: TensorInfo { name: "transformer.blocks.21.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/217", storage_size: 960 }
+transformer.blocks.21.ffn.0.weight: TensorInfo { name: "transformer.blocks.21.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/216", storage_size: 960 }
+transformer.blocks.21.ffn.1.weight: TensorInfo { name: "transformer.blocks.21.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/218", storage_size: 4915200 }
+transformer.blocks.21.ffn.3.weight: TensorInfo { name: "transformer.blocks.21.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/219", storage_size: 2457600 }
+transformer.blocks.22.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.22.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/224", storage_size: 960 }
+transformer.blocks.22.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.22.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/221", storage_size: 960 }
+transformer.blocks.22.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.22.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/220", storage_size: 960 }
+transformer.blocks.22.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.22.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/222", storage_size: 2764800 }
+transformer.blocks.22.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.22.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/225", storage_size: 921600 }
+transformer.blocks.22.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.22.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/223", storage_size: 960 }
+transformer.blocks.22.ffn.0.bias: TensorInfo { name: "transformer.blocks.22.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/227", storage_size: 960 }
+transformer.blocks.22.ffn.0.weight: TensorInfo { name: "transformer.blocks.22.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/226", storage_size: 960 }
+transformer.blocks.22.ffn.1.weight: TensorInfo { name: "transformer.blocks.22.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/228", storage_size: 4915200 }
+transformer.blocks.22.ffn.3.weight: TensorInfo { name: "transformer.blocks.22.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/229", storage_size: 2457600 }
+transformer.blocks.23.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.23.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/234", storage_size: 960 }
+transformer.blocks.23.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.23.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/231", storage_size: 960 }
+transformer.blocks.23.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.23.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/230", storage_size: 960 }
+transformer.blocks.23.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.23.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/232", storage_size: 2764800 }
+transformer.blocks.23.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.23.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/235", storage_size: 921600 }
+transformer.blocks.23.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.23.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/233", storage_size: 960 }
+transformer.blocks.23.ffn.0.bias: TensorInfo { name: "transformer.blocks.23.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/237", storage_size: 960 }
+transformer.blocks.23.ffn.0.weight: TensorInfo { name: "transformer.blocks.23.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/236", storage_size: 960 }
+transformer.blocks.23.ffn.1.weight: TensorInfo { name: "transformer.blocks.23.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/238", storage_size: 4915200 }
+transformer.blocks.23.ffn.3.weight: TensorInfo { name: "transformer.blocks.23.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/239", storage_size: 2457600 }
+transformer.blocks.24.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.24.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/244", storage_size: 960 }
+transformer.blocks.24.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.24.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/241", storage_size: 960 }
+transformer.blocks.24.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.24.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/240", storage_size: 960 }
+transformer.blocks.24.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.24.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/242", storage_size: 2764800 }
+transformer.blocks.24.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.24.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/245", storage_size: 921600 }
+transformer.blocks.24.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.24.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/243", storage_size: 960 }
+transformer.blocks.24.ffn.0.bias: TensorInfo { name: "transformer.blocks.24.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/247", storage_size: 960 }
+transformer.blocks.24.ffn.0.weight: TensorInfo { name: "transformer.blocks.24.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/246", storage_size: 960 }
+transformer.blocks.24.ffn.1.weight: TensorInfo { name: "transformer.blocks.24.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/248", storage_size: 4915200 }
+transformer.blocks.24.ffn.3.weight: TensorInfo { name: "transformer.blocks.24.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/249", storage_size: 2457600 }
+transformer.blocks.25.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.25.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/254", storage_size: 960 }
+transformer.blocks.25.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.25.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/251", storage_size: 960 }
+transformer.blocks.25.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.25.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/250", storage_size: 960 }
+transformer.blocks.25.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.25.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/252", storage_size: 2764800 }
+transformer.blocks.25.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.25.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/255", storage_size: 921600 }
+transformer.blocks.25.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.25.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/253", storage_size: 960 }
+transformer.blocks.25.ffn.0.bias: TensorInfo { name: "transformer.blocks.25.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/257", storage_size: 960 }
+transformer.blocks.25.ffn.0.weight: TensorInfo { name: "transformer.blocks.25.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/256", storage_size: 960 }
+transformer.blocks.25.ffn.1.weight: TensorInfo { name: "transformer.blocks.25.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/258", storage_size: 4915200 }
+transformer.blocks.25.ffn.3.weight: TensorInfo { name: "transformer.blocks.25.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/259", storage_size: 2457600 }
+transformer.blocks.26.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.26.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/264", storage_size: 960 }
+transformer.blocks.26.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.26.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/261", storage_size: 960 }
+transformer.blocks.26.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.26.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/260", storage_size: 960 }
+transformer.blocks.26.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.26.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/262", storage_size: 2764800 }
+transformer.blocks.26.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.26.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/265", storage_size: 921600 }
+transformer.blocks.26.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.26.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/263", storage_size: 960 }
+transformer.blocks.26.ffn.0.bias: TensorInfo { name: "transformer.blocks.26.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/267", storage_size: 960 }
+transformer.blocks.26.ffn.0.weight: TensorInfo { name: "transformer.blocks.26.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/266", storage_size: 960 }
+transformer.blocks.26.ffn.1.weight: TensorInfo { name: "transformer.blocks.26.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/268", storage_size: 4915200 }
+transformer.blocks.26.ffn.3.weight: TensorInfo { name: "transformer.blocks.26.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/269", storage_size: 2457600 }
+transformer.blocks.27.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.27.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/274", storage_size: 960 }
+transformer.blocks.27.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.27.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/271", storage_size: 960 }
+transformer.blocks.27.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.27.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/270", storage_size: 960 }
+transformer.blocks.27.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.27.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/272", storage_size: 2764800 }
+transformer.blocks.27.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.27.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/275", storage_size: 921600 }
+transformer.blocks.27.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.27.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/273", storage_size: 960 }
+transformer.blocks.27.ffn.0.bias: TensorInfo { name: "transformer.blocks.27.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/277", storage_size: 960 }
+transformer.blocks.27.ffn.0.weight: TensorInfo { name: "transformer.blocks.27.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/276", storage_size: 960 }
+transformer.blocks.27.ffn.1.weight: TensorInfo { name: "transformer.blocks.27.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/278", storage_size: 4915200 }
+transformer.blocks.27.ffn.3.weight: TensorInfo { name: "transformer.blocks.27.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/279", storage_size: 2457600 }
+transformer.blocks.28.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.28.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/284", storage_size: 960 }
+transformer.blocks.28.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.28.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/281", storage_size: 960 }
+transformer.blocks.28.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.28.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/280", storage_size: 960 }
+transformer.blocks.28.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.28.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/282", storage_size: 2764800 }
+transformer.blocks.28.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.28.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/285", storage_size: 921600 }
+transformer.blocks.28.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.28.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/283", storage_size: 960 }
+transformer.blocks.28.ffn.0.bias: TensorInfo { name: "transformer.blocks.28.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/287", storage_size: 960 }
+transformer.blocks.28.ffn.0.weight: TensorInfo { name: "transformer.blocks.28.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/286", storage_size: 960 }
+transformer.blocks.28.ffn.1.weight: TensorInfo { name: "transformer.blocks.28.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/288", storage_size: 4915200 }
+transformer.blocks.28.ffn.3.weight: TensorInfo { name: "transformer.blocks.28.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/289", storage_size: 2457600 }
+transformer.blocks.29.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.29.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/294", storage_size: 960 }
+transformer.blocks.29.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.29.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/291", storage_size: 960 }
+transformer.blocks.29.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.29.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/290", storage_size: 960 }
+transformer.blocks.29.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.29.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/292", storage_size: 2764800 }
+transformer.blocks.29.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.29.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/295", storage_size: 921600 }
+transformer.blocks.29.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.29.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/293", storage_size: 960 }
+transformer.blocks.29.ffn.0.bias: TensorInfo { name: "transformer.blocks.29.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/297", storage_size: 960 }
+transformer.blocks.29.ffn.0.weight: TensorInfo { name: "transformer.blocks.29.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/296", storage_size: 960 }
+transformer.blocks.29.ffn.1.weight: TensorInfo { name: "transformer.blocks.29.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/298", storage_size: 4915200 }
+transformer.blocks.29.ffn.3.weight: TensorInfo { name: "transformer.blocks.29.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/299", storage_size: 2457600 }
 transformer.blocks.3.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.3.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/34", storage_size: 960 }
 transformer.blocks.3.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.3.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/31", storage_size: 960 }
 transformer.blocks.3.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.3.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/30", storage_size: 960 }
@@ -114,4 +339,28 @@ transformer.blocks.7.attn.out_proj.weight: TensorInfo { name: "transformer.block
 transformer.blocks.7.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.7.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/73", storage_size: 960 }
 transformer.blocks.7.ffn.0.bias: TensorInfo { name: "transformer.blocks.7.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/77", storage_size: 960 }
 transformer.blocks.7.ffn.0.weight: TensorInfo { name: "transformer.blocks.7.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/76", storage_size: 960 }
-transformer.blocks.7.ffn.1.weight: TensorInfo { name: "transformer.blocks.7.
+transformer.blocks.7.ffn.1.weight: TensorInfo { name: "transformer.blocks.7.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/78", storage_size: 4915200 }
+transformer.blocks.7.ffn.3.weight: TensorInfo { name: "transformer.blocks.7.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/79", storage_size: 2457600 }
+transformer.blocks.8.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.8.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/84", storage_size: 960 }
+transformer.blocks.8.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.8.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/81", storage_size: 960 }
+transformer.blocks.8.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.8.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/80", storage_size: 960 }
+transformer.blocks.8.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.8.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/82", storage_size: 2764800 }
+transformer.blocks.8.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.8.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/85", storage_size: 921600 }
+transformer.blocks.8.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.8.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/83", storage_size: 960 }
+transformer.blocks.8.ffn.0.bias: TensorInfo { name: "transformer.blocks.8.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/87", storage_size: 960 }
+transformer.blocks.8.ffn.0.weight: TensorInfo { name: "transformer.blocks.8.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/86", storage_size: 960 }
+transformer.blocks.8.ffn.1.weight: TensorInfo { name: "transformer.blocks.8.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/88", storage_size: 4915200 }
+transformer.blocks.8.ffn.3.weight: TensorInfo { name: "transformer.blocks.8.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/89", storage_size: 2457600 }
+transformer.blocks.9.attn.k_ln.weight: TensorInfo { name: "transformer.blocks.9.attn.k_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/94", storage_size: 960 }
+transformer.blocks.9.attn.layernorm_qkv.0.bias: TensorInfo { name: "transformer.blocks.9.attn.layernorm_qkv.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/91", storage_size: 960 }
+transformer.blocks.9.attn.layernorm_qkv.0.weight: TensorInfo { name: "transformer.blocks.9.attn.layernorm_qkv.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/90", storage_size: 960 }
+transformer.blocks.9.attn.layernorm_qkv.1.weight: TensorInfo { name: "transformer.blocks.9.attn.layernorm_qkv.1.weight", dtype: F32, layout: Layout { shape: [2880, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/92", storage_size: 2764800 }
+transformer.blocks.9.attn.out_proj.weight: TensorInfo { name: "transformer.blocks.9.attn.out_proj.weight", dtype: F32, layout: Layout { shape: [960, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/95", storage_size: 921600 }
+transformer.blocks.9.attn.q_ln.weight: TensorInfo { name: "transformer.blocks.9.attn.q_ln.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/93", storage_size: 960 }
+transformer.blocks.9.ffn.0.bias: TensorInfo { name: "transformer.blocks.9.ffn.0.bias", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/97", storage_size: 960 }
+transformer.blocks.9.ffn.0.weight: TensorInfo { name: "transformer.blocks.9.ffn.0.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/96", storage_size: 960 }
+transformer.blocks.9.ffn.1.weight: TensorInfo { name: "transformer.blocks.9.ffn.1.weight", dtype: F32, layout: Layout { shape: [5120, 960], stride: [960, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/98", storage_size: 4915200 }
+transformer.blocks.9.ffn.3.weight: TensorInfo { name: "transformer.blocks.9.ffn.3.weight", dtype: F32, layout: Layout { shape: [960, 2560], stride: [2560, 1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/99", storage_size: 2457600 }
+
+transformer.norm.weight: TensorInfo { name: "transformer.norm.weight", dtype: F32, layout: Layout { shape: [960], stride: [1], start_offset: 0 }, path: "esmc_300m_v0_fp32/data/300", storage_size: 960 }
+```
diff --git a/ferritin-esm/examples/esmc/main.rs b/ferritin-esm/examples/esmc/main.rs
index 04afd948..9b4f128f 100644
--- a/ferritin-esm/examples/esmc/main.rs
+++ b/ferritin-esm/examples/esmc/main.rs
@@ -4,7 +4,6 @@ use candle_core::{DType, Device, D};
 use candle_hf_hub::{api::sync::Api, Repo, RepoType};
 use candle_nn::VarBuilder;
 use ferritin_esm::{ESMCConfig, ESMC};
-use safetensors::SafeTensors;
 
 // pub fn esmc_300m_202412(device: &Device) -> Result<Box<dyn Model>> {
 //     let tokenizer = get_model_tokenizers(ESM3_OPEN_SMALL)?.sequence;
@@ -38,8 +37,15 @@ fn main() -> Result<()> {
 
     let vb = VarBuilder::from_backend(Box::new(pth), DType::F32, Device::Cpu);
     let config = ESMCConfig::esmc_300m();
-    let esmc = ESMC::load(vb, config);
+    let esmc = ESMC::load(vb.clone(), config)?;
+    // println!("ESMC Loaded: {}", esmc);
+
+    // Error: cannot find tensor transformer.layer.attention.layer_norm.weight
+
+    println!(
+        "VB: {}",
+        vb.contains_tensor("transformer.blocks.6.attn.layernorm_qkv.1.weight")
+    );
 
-    println!("ESMC Loaded");
     Ok(())
 }
diff --git a/ferritin-esm/src/esm/layers/attention.rs b/ferritin-esm/src/esm/layers/attention.rs
index c2651f7b..206edf1f 100644
--- a/ferritin-esm/src/esm/layers/attention.rs
+++ b/ferritin-esm/src/esm/layers/attention.rs
@@ -55,16 +55,21 @@ impl MultiHeadAttention {
         } = config;
 
         let d_head = d_model / n_heads;
-        let ln_conf = LayerNormConfig::from(1e-5);
-        let layernorm = nn::layer_norm(*d_model, ln_conf, vb.pp("layer_norm"))?;
-        let linear = nn::linear(*d_model, d_model * 3, vb.pp("linear1"))?;
+        // let ln_conf = LayerNormConfig::from(1e-5);
+        let ln_conf = LayerNormConfig {
+            eps: 1e-5,
+            remove_mean: true,
+            affine: false,
+        };
+        let layernorm = nn::layer_norm(*d_model, ln_conf, vb.pp("layernorm_qkv.0"))?;
+        let linear = nn::linear_no_bias(*d_model, d_model * 3, vb.pp("layernorm_qkv.1"))?;
         let layernorm_qkv = nn::seq().add(layernorm).add(linear);
-        let out_proj = nn::linear(*d_model, *d_model, vb.pp("out_proj"))?;
-
+        let out_proj = nn::linear_no_bias(*d_model, *d_model, vb.pp("out_proj"))?;
         // note: only handling the True case for the moment
         // let  qk_layernorm = true
         let q_ln = Box::new(nn::layer_norm(*d_model, ln_conf, vb.pp("q_ln"))?);
         let k_ln = Box::new(nn::layer_norm(*d_model, ln_conf, vb.pp("k_ln"))?);
+
         let rotary = RotaryEmbedding::load(vb.pp("rotary"), config)?;
 
         Ok(Self {
diff --git a/ferritin-esm/src/esm/layers/blocks.rs b/ferritin-esm/src/esm/layers/blocks.rs
index c5659de6..33c66802 100644
--- a/ferritin-esm/src/esm/layers/blocks.rs
+++ b/ferritin-esm/src/esm/layers/blocks.rs
@@ -1,7 +1,7 @@
 use super::attention::MultiHeadAttention;
 use super::geom_attention::GeometricReasoningOriginalImpl;
 use crate::esm::models::esmc::{ESMCConfig, Ffn_Type};
-use crate::esm::utils::structure::affine3d::Affine3D;
+// use crate::esm::utils::structure::affine3d::Affine3D;
 use candle_core::{Module, Result, Tensor, D};
 use candle_nn::ops::silu;
 use candle_nn::{self as nn, VarBuilder};
@@ -29,9 +29,9 @@ impl SwiGLU {
         let hidden_dim = Self::swiglu_correction_fn(*expansion_ratio, *d_model);
 
         Ok(Self {
-            layer_norm: nn::layer_norm(*d_model, 1e-5, vb.pp("layer_norm"))?,
-            linear1: nn::linear(*d_model, hidden_dim * 2, vb.pp("linear1"))?,
-            linear2: nn::linear(hidden_dim, *d_model, vb.pp("linear2"))?,
+            layer_norm: nn::layer_norm(*d_model, 1e-5, vb.pp("0"))?,
+            linear1: nn::linear_no_bias(*d_model, hidden_dim * 2, vb.pp("1"))?,
+            linear2: nn::linear_no_bias(hidden_dim, *d_model, vb.pp("3"))?,
         })
     }
 }
@@ -131,29 +131,33 @@ impl UnifiedTransformerBlock {
     pub fn load(vb: VarBuilder, config: &ESMCConfig, layer: usize) -> Result<Self> {
         let ESMCConfig {
             ffn_type,
+            v_head_transformer,
             use_plain_attn,
             n_layers_geom,
             residue_scaling_factor,
             ..
         } = config;
 
-        let use_geom_attn: bool = layer < *n_layers_geom;
-
         let attn = match use_plain_attn {
             false => None,
-            true => Some(MultiHeadAttention::load(vb.pp("attention"), config)?),
+            true => Some(MultiHeadAttention::load(vb.pp("attn"), config)?),
         };
 
-        let geom_attn = match use_geom_attn {
-            false => None,
-            true => Some(GeometricReasoningOriginalImpl::load(
-                vb.pp("geometric"),
-                config,
-            )?),
-        };
+        // println!("LAYER; GEOM: {}, {}", layer, n_layers_geom);
+        let use_geom_attn: bool = layer < *n_layers_geom;
+        // println!("Geom ATTN {}", use_geom_attn);
+        // let geom_attn = match use_geom_attn {
+        //     false => None,
+        //     true => Some(GeometricReasoningOriginalImpl::load(
+        //         vb.pp("geometric"),
+        //         config,
+        //     )?),
+        // };
+
+        let geom_attn = None;
 
         let ffn = match ffn_type {
-            Ffn_Type::SWIGLU => SwiGLU::load(vb.pp("swiglue"), config)?,
+            Ffn_Type::SWIGLU => SwiGLU::load(vb.pp("ffn"), config)?,
             _ => unimplemented!(), // Ffn_Type::GLU => unimplemented!(),
         };
 
diff --git a/ferritin-esm/src/esm/layers/geom_attention.rs b/ferritin-esm/src/esm/layers/geom_attention.rs
index 2e953c94..6dc30c23 100644
--- a/ferritin-esm/src/esm/layers/geom_attention.rs
+++ b/ferritin-esm/src/esm/layers/geom_attention.rs
@@ -48,7 +48,9 @@ impl GeometricReasoningOriginalImpl {
         } = config;
 
         let num_vector_messages = 1usize;
-        let v_heads = v_head_transformer.unwrap();
+
+        // todo: this is a hidden param. Needs to be fixed
+        let v_heads = v_head_transformer.unwrap_or(128);
 
         let dim_proj = 4 * v_heads * 3 + v_heads * 3 * num_vector_messages;
         let channels_out = v_heads * 3 * num_vector_messages;
diff --git a/ferritin-esm/src/esm/layers/regression_head.rs b/ferritin-esm/src/esm/layers/regression_head.rs
index 114146f5..41e1b8b9 100644
--- a/ferritin-esm/src/esm/layers/regression_head.rs
+++ b/ferritin-esm/src/esm/layers/regression_head.rs
@@ -27,18 +27,14 @@ impl RegressionHead {
             ..
         } = config;
 
-        let linear1 = nn::linear(
-            *d_model,
-            *regression_head_hidden_dim,
-            vb.pp("regression_linear"),
-        )?;
+        let linear1 = nn::linear(*d_model, *regression_head_hidden_dim, vb.pp("0"))?;
         let gelu = candle_nn::Activation::Gelu;
         let ln_conf = LayerNormConfig::from(1e-5);
-        let norm = nn::layer_norm(*regression_head_hidden_dim, ln_conf, vb.pp("layer_norm"))?;
+        let norm = nn::layer_norm(*regression_head_hidden_dim, ln_conf, vb.pp("2"))?;
         let linear2 = nn::linear(
             *regression_head_hidden_dim,
             *regression_head_output_dim,
-            vb.pp("linear2"),
+            vb.pp("3"),
         )?;
 
         let model = nn::seq().add(linear1).add(gelu).add(norm).add(linear2);
diff --git a/ferritin-esm/src/esm/layers/transformer_stack.rs b/ferritin-esm/src/esm/layers/transformer_stack.rs
index 04a83912..d4375b16 100644
--- a/ferritin-esm/src/esm/layers/transformer_stack.rs
+++ b/ferritin-esm/src/esm/layers/transformer_stack.rs
@@ -1,6 +1,6 @@
 use crate::esm::layers::blocks::UnifiedTransformerBlock;
 use crate::esm::models::esmc::ESMCConfig;
-use crate::esm::utils::structure::affine3d::Affine3D;
+// use crate::esm::utils::structure::affine3d::Affine3D;
 use candle_core::{Module, Result, Tensor, D};
 use candle_nn::{self as nn, LayerNorm, LayerNormConfig};
 
@@ -31,11 +31,20 @@ impl TransformerStack {
 
         let mut blocks = Vec::with_capacity(*n_layers as usize);
         for i in 0..*n_layers {
-            blocks.push(UnifiedTransformerBlock::load(vb.pp("layer"), &config, i)?);
+            blocks.push(UnifiedTransformerBlock::load(
+                vb.pp(format!("blocks.{}", i)),
+                &config,
+                i,
+            )?);
         }
 
-        let ln_conf = LayerNormConfig::from(1e-5);
-        let norm = nn::layer_norm(*d_model, ln_conf, vb.pp("layer_norm"))?;
+        // let ln_conf = LayerNormConfig::from(1e-5);
+        let ln_conf = LayerNormConfig {
+            eps: 1e-5,
+            remove_mean: true,
+            affine: false,
+        };
+        let norm = nn::layer_norm(*d_model, ln_conf, vb.pp("norm"))?;
 
         Ok(Self { blocks, norm })
     }
diff --git a/ferritin-esm/src/esm/models/esmc.rs b/ferritin-esm/src/esm/models/esmc.rs
index 1b3eba0c..ff29e486 100644
--- a/ferritin-esm/src/esm/models/esmc.rs
+++ b/ferritin-esm/src/esm/models/esmc.rs
@@ -122,30 +122,17 @@ impl ESMC {
     pub fn load(vb: VarBuilder, config: ESMCConfig) -> Result<Self> {
         let ESMCConfig {
             d_model,
-            n_heads,
-            n_layers,
-            v_head_transformer,
-            ffn_type,
             tokenizer,
-            use_plain_attn,
-            n_layers_geom,
-            scale_residue,
-            residue_scaling_factor,
-            mask_and_zero_frameless,
-            bias,
-            qk_layernorm,
-            expansion_ratio,
-            regression_head_output_dim
-            regression_head_hidden_dim,
             embedding_dim,
+            ..
         } = config;
 
         let tokenizer_collection = tokenizer.get_model_tokenizers();
 
         Ok(Self {
-            embed: nn::embedding(embedding_dim, d_model as usize, vb.pp("embedding"))?,
+            embed: nn::embedding(embedding_dim, d_model as usize, vb.pp("embed"))?,
             transformer: TransformerStack::load(vb.pp("transformer"), &config)?,
-            sequence_head: RegressionHead::load(vb.pp("regression"), &config)?,
+            sequence_head: RegressionHead::load(vb.pp("sequence_head"), &config)?,
             tokenizer: tokenizer_collection.sequence,
         })
     }
diff --git a/ferritin-esm/src/esm/utils/structure/affine3D.rs b/ferritin-esm/src/esm/utils/structure/affine3D.rs
index c3429fe4..d79783e6 100644
--- a/ferritin-esm/src/esm/utils/structure/affine3D.rs
+++ b/ferritin-esm/src/esm/utils/structure/affine3D.rs
@@ -24,7 +24,8 @@ pub trait Rotation: Sized + Clone {
     }
 
     fn requires_grad(&self) -> bool {
-        self.tensor().requires_grad()
+        // self.tensor().requires_grad()
+        unimplemented!()
     }
 
     fn to_dtype(&self, dtype: DType) -> Result<Self>;
diff --git a/ferritin-esm/src/esm/utils/structure/mod.rs b/ferritin-esm/src/esm/utils/structure/mod.rs
index efd3860b..8cd74c91 100644
--- a/ferritin-esm/src/esm/utils/structure/mod.rs
+++ b/ferritin-esm/src/esm/utils/structure/mod.rs
@@ -1,4 +1,4 @@
 pub mod affine3d;
-mod protein_chain;
-mod protein_complex;
-mod protein_structure;
+// mod protein_chain;
+// mod protein_complex;
+// mod protein_structure;
diff --git a/justfile b/justfile
index 9c90ca76..c022372d 100644
--- a/justfile
+++ b/justfile
@@ -45,4 +45,5 @@ test-ligandmpnn:
     cargo test --features metal -p ferritin-ligandmpnn test_cli_command_run_example_06 -- --nocapture
 
 esmc:
+    #RUST_BACKTRACE=1 cargo run --example esmc
     cargo run --example esmc