Error(s) in loading state_dict for NextDiT

[makisekurisu-jp]

(venv) PS D:\ComfyUI> cd custom_nodes\ComfyUI-GGUF\tools
(venv) PS D:\ComfyUI\custom_nodes\ComfyUI-GGUF\tools> python convert.py --src D:\ComfyUI\models\diffusion_models\Beyond_Reality_Super-Z-Image_3.0-BF16.safetensors
cap_embedder.0.weight                         torch.bfloat16 --> F32, shape = {2560}
cap_embedder.1.bias                           torch.bfloat16 --> F32, shape = {3840}
cap_embedder.1.weight                         torch.bfloat16 --> BF16, shape = {5120, 3840}
cap_pad_token                                 torch.bfloat16 --> BF16, shape = {7680, 1}
context_refiner.0.attention.k_norm.weight     torch.bfloat16 --> F32, shape = {128}
context_refiner.0.attention.out.weight        torch.bfloat16 --> BF16, shape = {7680, 3840}
context_refiner.0.attention.q_norm.weight     torch.bfloat16 --> F32, shape = {128}
context_refiner.0.attention.qkv.weight        torch.bfloat16 --> BF16, shape = {7680, 11520}
context_refiner.0.attention_norm1.weight      torch.bfloat16 --> F32, shape = {3840}
context_refiner.0.attention_norm2.weight      torch.bfloat16 --> F32, shape = {3840}
context_refiner.0.feed_forward.w1.weight      torch.bfloat16 --> BF16, shape = {7680, 10240}
context_refiner.0.feed_forward.w2.weight      torch.bfloat16 --> BF16, shape = {20480, 3840}
context_refiner.0.feed_forward.w3.weight      torch.bfloat16 --> BF16, shape = {7680, 10240}
context_refiner.0.ffn_norm1.weight            torch.bfloat16 --> F32, shape = {3840}
context_refiner.0.ffn_norm2.weight            torch.bfloat16 --> F32, shape = {3840}
context_refiner.1.attention.k_norm.weight     torch.bfloat16 --> F32, shape = {128}
context_refiner.1.attention.out.weight        torch.bfloat16 --> BF16, shape = {7680, 3840}
context_refiner.1.attention.q_norm.weight     torch.bfloat16 --> F32, shape = {128}
context_refiner.1.attention.qkv.weight        torch.bfloat16 --> BF16, shape = {7680, 11520}
context_refiner.1.attention_norm1.weight      torch.bfloat16 --> F32, shape = {3840}
context_refiner.1.attention_norm2.weight      torch.bfloat16 --> F32, shape = {3840}
context_refiner.1.feed_forward.w1.weight      torch.bfloat16 --> BF16, shape = {7680, 10240}
context_refiner.1.feed_forward.w2.weight      torch.bfloat16 --> BF16, shape = {20480, 3840}
context_refiner.1.feed_forward.w3.weight      torch.bfloat16 --> BF16, shape = {7680, 10240}
context_refiner.1.ffn_norm1.weight            torch.bfloat16 --> F32, shape = {3840}
context_refiner.1.ffn_norm2.weight            torch.bfloat16 --> F32, shape = {3840}
final_layer.adaLN_modulation.1.bias           torch.bfloat16 --> F32, shape = {3840}
final_layer.adaLN_modulation.1.weight         torch.bfloat16 --> BF16, shape = {512, 3840}
final_layer.linear.bias                       torch.bfloat16 --> F32, shape = {64}
final_layer.linear.weight                     torch.bfloat16 --> BF16, shape = {7680, 64}
layers.0.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.0.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.0.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.0.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.0.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.0.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.0.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.0.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.0.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.0.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.0.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.0.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.0.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.1.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.1.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.1.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.1.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.1.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.1.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.1.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.1.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.1.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.1.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.1.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.1.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.1.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.10.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.10.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.10.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.10.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.10.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.10.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.10.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.10.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.10.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.10.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.10.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.10.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.10.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.11.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.11.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.11.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.11.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.11.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.11.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.11.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.11.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.11.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.11.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.11.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.11.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.11.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.12.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.12.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.12.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.12.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.12.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.12.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.12.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.12.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.12.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.12.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.12.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.12.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.12.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.13.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.13.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.13.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.13.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.13.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.13.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.13.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.13.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.13.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.13.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.13.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.13.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.13.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.14.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.14.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.14.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.14.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.14.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.14.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.14.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.14.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.14.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.14.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.14.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.14.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.14.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.15.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.15.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.15.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.15.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.15.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.15.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.15.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.15.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.15.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.15.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.15.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.15.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.15.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.16.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.16.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.16.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.16.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.16.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.16.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.16.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.16.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.16.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.16.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.16.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.16.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.16.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.17.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.17.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.17.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.17.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.17.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.17.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.17.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.17.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.17.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.17.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.17.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.17.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.17.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.18.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.18.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.18.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.18.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.18.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.18.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.18.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.18.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.18.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.18.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.18.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.18.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.18.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.19.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.19.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.19.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.19.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.19.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.19.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.19.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.19.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.19.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.19.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.19.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.19.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.19.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.2.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.2.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.2.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.2.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.2.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.2.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.2.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.2.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.2.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.2.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.2.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.2.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.2.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.20.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.20.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.20.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.20.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.20.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.20.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.20.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.20.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.20.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.20.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.20.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.20.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.20.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.21.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.21.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.21.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.21.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.21.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.21.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.21.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.21.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.21.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.21.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.21.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.21.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.21.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.22.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.22.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.22.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.22.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.22.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.22.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.22.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.22.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.22.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.22.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.22.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.22.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.22.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.23.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.23.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.23.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.23.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.23.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.23.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.23.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.23.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.23.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.23.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.23.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.23.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.23.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.24.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.24.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.24.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.24.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.24.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.24.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.24.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.24.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.24.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.24.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.24.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.24.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.24.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.25.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.25.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.25.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.25.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.25.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.25.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.25.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.25.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.25.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.25.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.25.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.25.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.25.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.26.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.26.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.26.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.26.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.26.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.26.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.26.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.26.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.26.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.26.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.26.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.26.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.26.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.27.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.27.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.27.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.27.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.27.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.27.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.27.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.27.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.27.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.27.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.27.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.27.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.27.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.28.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.28.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.28.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.28.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.28.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.28.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.28.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.28.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.28.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.28.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.28.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.28.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.28.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.29.adaLN_modulation.0.bias             torch.bfloat16 --> F32, shape = {15360}
layers.29.adaLN_modulation.0.weight           torch.bfloat16 --> BF16, shape = {512, 15360}
layers.29.attention.k_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.29.attention.out.weight                torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.29.attention.q_norm.weight             torch.bfloat16 --> F32, shape = {128}
layers.29.attention.qkv.weight                torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.29.attention_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
layers.29.attention_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
layers.29.feed_forward.w1.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.29.feed_forward.w2.weight              torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.29.feed_forward.w3.weight              torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.29.ffn_norm1.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.29.ffn_norm2.weight                    torch.bfloat16 --> F32, shape = {3840}
layers.3.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.3.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.3.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.3.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.3.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.3.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.3.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.3.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.3.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.3.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.3.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.3.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.3.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.4.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.4.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.4.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.4.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.4.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.4.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.4.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.4.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.4.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.4.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.4.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.4.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.4.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.5.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.5.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.5.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.5.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.5.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.5.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.5.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.5.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.5.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.5.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.5.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.5.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.5.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.6.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.6.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.6.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.6.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.6.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.6.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.6.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.6.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.6.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.6.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.6.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.6.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.6.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.7.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.7.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.7.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.7.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.7.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.7.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.7.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.7.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.7.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.7.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.7.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.7.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.7.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.8.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.8.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.8.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.8.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.8.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.8.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.8.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.8.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.8.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.8.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.8.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.8.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.8.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.9.adaLN_modulation.0.bias              torch.bfloat16 --> F32, shape = {15360}
layers.9.adaLN_modulation.0.weight            torch.bfloat16 --> BF16, shape = {512, 15360}
layers.9.attention.k_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.9.attention.out.weight                 torch.bfloat16 --> BF16, shape = {7680, 3840}
layers.9.attention.q_norm.weight              torch.bfloat16 --> F32, shape = {128}
layers.9.attention.qkv.weight                 torch.bfloat16 --> BF16, shape = {7680, 11520}
layers.9.attention_norm1.weight               torch.bfloat16 --> F32, shape = {3840}
layers.9.attention_norm2.weight               torch.bfloat16 --> F32, shape = {3840}
layers.9.feed_forward.w1.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.9.feed_forward.w2.weight               torch.bfloat16 --> BF16, shape = {20480, 3840}
layers.9.feed_forward.w3.weight               torch.bfloat16 --> BF16, shape = {7680, 10240}
layers.9.ffn_norm1.weight                     torch.bfloat16 --> F32, shape = {3840}
layers.9.ffn_norm2.weight                     torch.bfloat16 --> F32, shape = {3840}
noise_refiner.0.adaLN_modulation.0.bias       torch.bfloat16 --> F32, shape = {15360}
noise_refiner.0.adaLN_modulation.0.weight     torch.bfloat16 --> BF16, shape = {512, 15360}
noise_refiner.0.attention.k_norm.weight       torch.bfloat16 --> F32, shape = {128}
noise_refiner.0.attention.out.weight          torch.bfloat16 --> BF16, shape = {7680, 3840}
noise_refiner.0.attention.q_norm.weight       torch.bfloat16 --> F32, shape = {128}
noise_refiner.0.attention.qkv.weight          torch.bfloat16 --> BF16, shape = {7680, 11520}
noise_refiner.0.attention_norm1.weight        torch.bfloat16 --> F32, shape = {3840}
noise_refiner.0.attention_norm2.weight        torch.bfloat16 --> F32, shape = {3840}
noise_refiner.0.feed_forward.w1.weight        torch.bfloat16 --> BF16, shape = {7680, 10240}
noise_refiner.0.feed_forward.w2.weight        torch.bfloat16 --> BF16, shape = {20480, 3840}
noise_refiner.0.feed_forward.w3.weight        torch.bfloat16 --> BF16, shape = {7680, 10240}
noise_refiner.0.ffn_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
noise_refiner.0.ffn_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
noise_refiner.1.adaLN_modulation.0.bias       torch.bfloat16 --> F32, shape = {15360}
noise_refiner.1.adaLN_modulation.0.weight     torch.bfloat16 --> BF16, shape = {512, 15360}
noise_refiner.1.attention.k_norm.weight       torch.bfloat16 --> F32, shape = {128}
noise_refiner.1.attention.out.weight          torch.bfloat16 --> BF16, shape = {7680, 3840}
noise_refiner.1.attention.q_norm.weight       torch.bfloat16 --> F32, shape = {128}
noise_refiner.1.attention.qkv.weight          torch.bfloat16 --> BF16, shape = {7680, 11520}
noise_refiner.1.attention_norm1.weight        torch.bfloat16 --> F32, shape = {3840}
noise_refiner.1.attention_norm2.weight        torch.bfloat16 --> F32, shape = {3840}
noise_refiner.1.feed_forward.w1.weight        torch.bfloat16 --> BF16, shape = {7680, 10240}
noise_refiner.1.feed_forward.w2.weight        torch.bfloat16 --> BF16, shape = {20480, 3840}
noise_refiner.1.feed_forward.w3.weight        torch.bfloat16 --> BF16, shape = {7680, 10240}
noise_refiner.1.ffn_norm1.weight              torch.bfloat16 --> F32, shape = {3840}
noise_refiner.1.ffn_norm2.weight              torch.bfloat16 --> F32, shape = {3840}
t_embedder.mlp.0.bias                         torch.bfloat16 --> F32, shape = {1024}
t_embedder.mlp.0.weight                       torch.bfloat16 --> BF16, shape = {512, 1024}
t_embedder.mlp.2.bias                         torch.bfloat16 --> F32, shape = {256}
t_embedder.mlp.2.weight                       torch.bfloat16 --> BF16, shape = {2048, 256}
x_embedder.bias                               torch.bfloat16 --> F32, shape = {3840}
x_embedder.weight                             torch.bfloat16 --> BF16, shape = {128, 3840}
x_pad_token                                   torch.bfloat16 --> BF16, shape = {7680, 1}
100%|████████████████████████████████████████████████████████████████████████████████| 453/453 [00:45<00:00, 10.03it/s]
Writing: 100%|██████████████████████████████████████████████████████████████████| 12.3G/12.3G [00:34<00:00, 357Mbyte/s]
(venv) PS D:\ComfyUI\custom_nodes\ComfyUI-GGUF\tools>

[makisekurisu-jp]

@city96 I need your help, please🙏

[qskousen]

I just discussed this in #419

x_pad_token and cap_pad_token can't be BF16, though I am not sure of the reason. If you downcast them to FP16 (with some precision loss) or upcast them to FP32, then it should work fine.

[makisekurisu-jp]

#392

[82sound-jpg]

我在使用zimage base q4km 和q80时候也出现这个错误

[pizzadelivery01]

Adding a data point to this issue — I hit the same thing converting a Z-Image finetune (NextDiT) and traced it two layers deep.

Setup

• ComfyUI on current main, ComfyUI-GGUF on current main (6ea2651)
• ComfyUI-GGUF/tools at current main, llama.cpp @ b3962 + lcpp.patch
• Source: Z-Image finetune from Civitai, diffusion model extracted via ModelSave, BF16 safetensors
• Quantized to Q6_K (11741 MB -> 5636 MB)
• Text encoder: qwen_4b_fp8_scaled.safetensors (not a GGUF encoder)
• RTX 5080, torch 2.11.0+cu128, Linux/WSL2

Problem 1 — load fails on pad tokens

UnetLoaderGGUF:

RuntimeError: Error(s) in loading state_dict for NextDiT:
  size mismatch for x_pad_token: copying a param with shape torch.Size([3840]) from checkpoint, the shape in current model is torch.Size([1, 3840]).
  size mismatch for cap_pad_token: copying a param with shape torch.Size([3840]) from checkpoint, the shape in current model is torch.Size([1, 3840]).

Both tensors are [1, 3840] in the source. GGUF drops the trailing dim-1, and convert.py does not emit comfy.gguf.orig_shape.* for them, so the loader reconstructs [3840].

Problem 2 — patching orig_shape moves the failure downstream

I rewrote the quantized GGUF adding comfy.gguf.orig_shape.x_pad_token = [1, 3840] and the same for cap_pad_token, leaving all tensor data byte-identical (verified: 0 shape mismatches against the original across all 453 tensors).

That clears the load. Sampling then fails:

KSampler
RuntimeError: Sizes of tensors must match except in dimension 1.
Expected size 3840 but got size 7680 for tensor number 1 in the list.

7680 is 2x 3840. Stored state is consistent as far as I can inspect it — both tensors report shape [3840], type 30 (BF16), data.shape == (7680,), which is the correct byte count for 3840 BF16 elements.

Control

Same workflow, same fp8 text encoder, Load Diffusion Model with the original BF16 safetensors instead of the GGUF loader: renders fine. The only variable is the unet loader.

So the metadata patch is not the fix — declaring [1, 3840] satisfies load_state_dict but something downstream still reads 7680, and declaring [3840] is equivalent to no metadata at all. Both states fail, just in different places.

Happy to share the conversion log, the patched file, or test a fix.