DiT W4A8 torchao sidecar (int8_act)

torchao/w4a8_int8_act/diffusion_w4a8_int8_act_torchao_info.json

@@ -0,0 +1,1724 @@
+{
+  "format": "torchao_int8_dynamic_activation_int4_weight",
+  "variant_id": "int8_act",
+  "out_stem": "diffusion_w4a8_int8_act_torchao",
+  "torchao_config": "Int8DynamicActivationIntxWeightConfig",
+  "activation_quantization": "dynamic_at_runtime",
+  "activation_note": "Activations: dynamic int8 per torchao (scales computed during forward; not stored as a static tensor).",
+  "weight_quantization": "offline_in_this_script",
+  "source_hf_repo": "Soul-AILab/SoulX-FlashTalk-14B",
+  "tensor_count": 1633,
+  "tensor_names_sample": [
+    "patch_embedding.weight",
+    "patch_embedding.bias",
+    "text_embedding.0.weight",
+    "text_embedding.0.bias",
+    "text_embedding.2.weight",
+    "text_embedding.2.bias",
+    "time_embedding.0.weight",
+    "time_embedding.0.bias",
+    "time_embedding.2.weight",
+    "time_embedding.2.bias",
+    "time_projection.1.weight",
+    "time_projection.1.bias",
+    "blocks.0.modulation",
+    "blocks.0.self_attn.q.weight",
+    "blocks.0.self_attn.q.bias",
+    "blocks.0.self_attn.k.weight",
+    "blocks.0.self_attn.k.bias",
+    "blocks.0.self_attn.v.weight",
+    "blocks.0.self_attn.v.bias",
+    "blocks.0.self_attn.o.weight",
+    "blocks.0.self_attn.o.bias",
+    "blocks.0.self_attn.norm_q.weight",
+    "blocks.0.self_attn.norm_k.weight",
+    "blocks.0.norm3.weight",
+    "blocks.0.norm3.bias",
+    "blocks.0.cross_attn.q.weight",
+    "blocks.0.cross_attn.q.bias",
+    "blocks.0.cross_attn.k.weight",
+    "blocks.0.cross_attn.k.bias",
+    "blocks.0.cross_attn.v.weight",
+    "blocks.0.cross_attn.v.bias",
+    "blocks.0.cross_attn.o.weight",
+    "blocks.0.cross_attn.o.bias",
+    "blocks.0.cross_attn.norm_q.weight",
+    "blocks.0.cross_attn.norm_k.weight",
+    "blocks.0.cross_attn.k_img.weight",
+    "blocks.0.cross_attn.k_img.bias",
+    "blocks.0.cross_attn.v_img.weight",
+    "blocks.0.cross_attn.v_img.bias",
+    "blocks.0.cross_attn.norm_k_img.weight",
+    "blocks.0.ffn.0.weight",
+    "blocks.0.ffn.0.bias",
+    "blocks.0.ffn.2.weight",
+    "blocks.0.ffn.2.bias",
+    "blocks.0.audio_cross_attn.q_linear.weight",
+    "blocks.0.audio_cross_attn.q_linear.bias",
+    "blocks.0.audio_cross_attn.proj.weight",
+    "blocks.0.audio_cross_attn.proj.bias"
+  ],
+  "safetensors_metadata_keys": [
+    "audio_proj.norm.bias",
+    "audio_proj.norm.weight",
+    "audio_proj.proj1.bias",
+    "audio_proj.proj1.weight",
+    "audio_proj.proj1_vf.bias",
+    "audio_proj.proj1_vf.weight",
+    "audio_proj.proj2.bias",
+    "audio_proj.proj2.weight",
+    "audio_proj.proj3.bias",
+    "audio_proj.proj3.weight",
+    "blocks.0.audio_cross_attn.kv_linear.bias",
+    "blocks.0.audio_cross_attn.kv_linear.weight",
+    "blocks.0.audio_cross_attn.proj.bias",
+    "blocks.0.audio_cross_attn.proj.weight",
+    "blocks.0.audio_cross_attn.q_linear.bias",
+    "blocks.0.audio_cross_attn.q_linear.weight",
+    "blocks.0.cross_attn.k.bias",
+    "blocks.0.cross_attn.k.weight",
+    "blocks.0.cross_attn.k_img.bias",
+    "blocks.0.cross_attn.k_img.weight",
+    "blocks.0.cross_attn.norm_k.weight",
+    "blocks.0.cross_attn.norm_k_img.weight",
+    "blocks.0.cross_attn.norm_q.weight",
+    "blocks.0.cross_attn.o.bias",
+    "blocks.0.cross_attn.o.weight",
+    "blocks.0.cross_attn.q.bias",
+    "blocks.0.cross_attn.q.weight",
+    "blocks.0.cross_attn.v.bias",
+    "blocks.0.cross_attn.v.weight",
+    "blocks.0.cross_attn.v_img.bias",
+    "blocks.0.cross_attn.v_img.weight",
+    "blocks.0.ffn.0.bias",
+    "blocks.0.ffn.0.weight",
+    "blocks.0.ffn.2.bias",
+    "blocks.0.ffn.2.weight",
+    "blocks.0.modulation",
+    "blocks.0.norm3.bias",
+    "blocks.0.norm3.weight",
+    "blocks.0.norm_x.bias",
+    "blocks.0.norm_x.weight",
+    "blocks.0.self_attn.k.bias",
+    "blocks.0.self_attn.k.weight",
+    "blocks.0.self_attn.norm_k.weight",
+    "blocks.0.self_attn.norm_q.weight",
+    "blocks.0.self_attn.o.bias",
+    "blocks.0.self_attn.o.weight",
+    "blocks.0.self_attn.q.bias",
+    "blocks.0.self_attn.q.weight",
+    "blocks.0.self_attn.v.bias",
+    "blocks.0.self_attn.v.weight",
+    "blocks.1.audio_cross_attn.kv_linear.bias",
+    "blocks.1.audio_cross_attn.kv_linear.weight",
+    "blocks.1.audio_cross_attn.proj.bias",
+    "blocks.1.audio_cross_attn.proj.weight",
+    "blocks.1.audio_cross_attn.q_linear.bias",
+    "blocks.1.audio_cross_attn.q_linear.weight",
+    "blocks.1.cross_attn.k.bias",
+    "blocks.1.cross_attn.k.weight",
+    "blocks.1.cross_attn.k_img.bias",
+    "blocks.1.cross_attn.k_img.weight",
+    "blocks.1.cross_attn.norm_k.weight",
+    "blocks.1.cross_attn.norm_k_img.weight",
+    "blocks.1.cross_attn.norm_q.weight",
+    "blocks.1.cross_attn.o.bias",
+    "blocks.1.cross_attn.o.weight",
+    "blocks.1.cross_attn.q.bias",
+    "blocks.1.cross_attn.q.weight",
+    "blocks.1.cross_attn.v.bias",
+    "blocks.1.cross_attn.v.weight",
+    "blocks.1.cross_attn.v_img.bias",
+    "blocks.1.cross_attn.v_img.weight",
+    "blocks.1.ffn.0.bias",
+    "blocks.1.ffn.0.weight",
+    "blocks.1.ffn.2.bias",
+    "blocks.1.ffn.2.weight",
+    "blocks.1.modulation",
+    "blocks.1.norm3.bias",
+    "blocks.1.norm3.weight",
+    "blocks.1.norm_x.bias",
+    "blocks.1.norm_x.weight",
+    "blocks.1.self_attn.k.bias",
+    "blocks.1.self_attn.k.weight",
+    "blocks.1.self_attn.norm_k.weight",
+    "blocks.1.self_attn.norm_q.weight",
+    "blocks.1.self_attn.o.bias",
+    "blocks.1.self_attn.o.weight",
+    "blocks.1.self_attn.q.bias",
+    "blocks.1.self_attn.q.weight",
+    "blocks.1.self_attn.v.bias",
+    "blocks.1.self_attn.v.weight",
+    "blocks.10.audio_cross_attn.kv_linear.bias",
+    "blocks.10.audio_cross_attn.kv_linear.weight",
+    "blocks.10.audio_cross_attn.proj.bias",
+    "blocks.10.audio_cross_attn.proj.weight",
+    "blocks.10.audio_cross_attn.q_linear.bias",
+    "blocks.10.audio_cross_attn.q_linear.weight",
+    "blocks.10.cross_attn.k.bias",
+    "blocks.10.cross_attn.k.weight",
+    "blocks.10.cross_attn.k_img.bias",
+    "blocks.10.cross_attn.k_img.weight",
+    "blocks.10.cross_attn.norm_k.weight",
+    "blocks.10.cross_attn.norm_k_img.weight",
+    "blocks.10.cross_attn.norm_q.weight",
+    "blocks.10.cross_attn.o.bias",
+    "blocks.10.cross_attn.o.weight",
+    "blocks.10.cross_attn.q.bias",
+    "blocks.10.cross_attn.q.weight",
+    "blocks.10.cross_attn.v.bias",
+    "blocks.10.cross_attn.v.weight",
+    "blocks.10.cross_attn.v_img.bias",
+    "blocks.10.cross_attn.v_img.weight",
+    "blocks.10.ffn.0.bias",
+    "blocks.10.ffn.0.weight",
+    "blocks.10.ffn.2.bias",
+    "blocks.10.ffn.2.weight",
+    "blocks.10.modulation",
+    "blocks.10.norm3.bias",
+    "blocks.10.norm3.weight",
+    "blocks.10.norm_x.bias",
+    "blocks.10.norm_x.weight",
+    "blocks.10.self_attn.k.bias",
+    "blocks.10.self_attn.k.weight",
+    "blocks.10.self_attn.norm_k.weight",
+    "blocks.10.self_attn.norm_q.weight",
+    "blocks.10.self_attn.o.bias",
+    "blocks.10.self_attn.o.weight",
+    "blocks.10.self_attn.q.bias",
+    "blocks.10.self_attn.q.weight",
+    "blocks.10.self_attn.v.bias",
+    "blocks.10.self_attn.v.weight",
+    "blocks.11.audio_cross_attn.kv_linear.bias",
+    "blocks.11.audio_cross_attn.kv_linear.weight",
+    "blocks.11.audio_cross_attn.proj.bias",
+    "blocks.11.audio_cross_attn.proj.weight",
+    "blocks.11.audio_cross_attn.q_linear.bias",
+    "blocks.11.audio_cross_attn.q_linear.weight",
+    "blocks.11.cross_attn.k.bias",
+    "blocks.11.cross_attn.k.weight",
+    "blocks.11.cross_attn.k_img.bias",
+    "blocks.11.cross_attn.k_img.weight",
+    "blocks.11.cross_attn.norm_k.weight",
+    "blocks.11.cross_attn.norm_k_img.weight",
+    "blocks.11.cross_attn.norm_q.weight",
+    "blocks.11.cross_attn.o.bias",
+    "blocks.11.cross_attn.o.weight",
+    "blocks.11.cross_attn.q.bias",
+    "blocks.11.cross_attn.q.weight",
+    "blocks.11.cross_attn.v.bias",
+    "blocks.11.cross_attn.v.weight",
+    "blocks.11.cross_attn.v_img.bias",
+    "blocks.11.cross_attn.v_img.weight",
+    "blocks.11.ffn.0.bias",
+    "blocks.11.ffn.0.weight",
+    "blocks.11.ffn.2.bias",
+    "blocks.11.ffn.2.weight",
+    "blocks.11.modulation",
+    "blocks.11.norm3.bias",
+    "blocks.11.norm3.weight",
+    "blocks.11.norm_x.bias",
+    "blocks.11.norm_x.weight",
+    "blocks.11.self_attn.k.bias",
+    "blocks.11.self_attn.k.weight",
+    "blocks.11.self_attn.norm_k.weight",
+    "blocks.11.self_attn.norm_q.weight",
+    "blocks.11.self_attn.o.bias",
+    "blocks.11.self_attn.o.weight",
+    "blocks.11.self_attn.q.bias",
+    "blocks.11.self_attn.q.weight",
+    "blocks.11.self_attn.v.bias",
+    "blocks.11.self_attn.v.weight",
+    "blocks.12.audio_cross_attn.kv_linear.bias",
+    "blocks.12.audio_cross_attn.kv_linear.weight",
+    "blocks.12.audio_cross_attn.proj.bias",
+    "blocks.12.audio_cross_attn.proj.weight",
+    "blocks.12.audio_cross_attn.q_linear.bias",
+    "blocks.12.audio_cross_attn.q_linear.weight",
+    "blocks.12.cross_attn.k.bias",
+    "blocks.12.cross_attn.k.weight",
+    "blocks.12.cross_attn.k_img.bias",
+    "blocks.12.cross_attn.k_img.weight",
+    "blocks.12.cross_attn.norm_k.weight",
+    "blocks.12.cross_attn.norm_k_img.weight",
+    "blocks.12.cross_attn.norm_q.weight",
+    "blocks.12.cross_attn.o.bias",
+    "blocks.12.cross_attn.o.weight",
+    "blocks.12.cross_attn.q.bias",
+    "blocks.12.cross_attn.q.weight",
+    "blocks.12.cross_attn.v.bias",
+    "blocks.12.cross_attn.v.weight",
+    "blocks.12.cross_attn.v_img.bias",
+    "blocks.12.cross_attn.v_img.weight",
+    "blocks.12.ffn.0.bias",
+    "blocks.12.ffn.0.weight",
+    "blocks.12.ffn.2.bias",
+    "blocks.12.ffn.2.weight",
+    "blocks.12.modulation",
+    "blocks.12.norm3.bias",
+    "blocks.12.norm3.weight",
+    "blocks.12.norm_x.bias",
+    "blocks.12.norm_x.weight",
+    "blocks.12.self_attn.k.bias",
+    "blocks.12.self_attn.k.weight",
+    "blocks.12.self_attn.norm_k.weight",
+    "blocks.12.self_attn.norm_q.weight",
+    "blocks.12.self_attn.o.bias",
+    "blocks.12.self_attn.o.weight",
+    "blocks.12.self_attn.q.bias",
+    "blocks.12.self_attn.q.weight",
+    "blocks.12.self_attn.v.bias",
+    "blocks.12.self_attn.v.weight",
+    "blocks.13.audio_cross_attn.kv_linear.bias",
+    "blocks.13.audio_cross_attn.kv_linear.weight",
+    "blocks.13.audio_cross_attn.proj.bias",
+    "blocks.13.audio_cross_attn.proj.weight",
+    "blocks.13.audio_cross_attn.q_linear.bias",
+    "blocks.13.audio_cross_attn.q_linear.weight",
+    "blocks.13.cross_attn.k.bias",
+    "blocks.13.cross_attn.k.weight",
+    "blocks.13.cross_attn.k_img.bias",
+    "blocks.13.cross_attn.k_img.weight",
+    "blocks.13.cross_attn.norm_k.weight",
+    "blocks.13.cross_attn.norm_k_img.weight",
+    "blocks.13.cross_attn.norm_q.weight",
+    "blocks.13.cross_attn.o.bias",
+    "blocks.13.cross_attn.o.weight",
+    "blocks.13.cross_attn.q.bias",
+    "blocks.13.cross_attn.q.weight",
+    "blocks.13.cross_attn.v.bias",
+    "blocks.13.cross_attn.v.weight",
+    "blocks.13.cross_attn.v_img.bias",
+    "blocks.13.cross_attn.v_img.weight",
+    "blocks.13.ffn.0.bias",
+    "blocks.13.ffn.0.weight",
+    "blocks.13.ffn.2.bias",
+    "blocks.13.ffn.2.weight",
+    "blocks.13.modulation",
+    "blocks.13.norm3.bias",
+    "blocks.13.norm3.weight",
+    "blocks.13.norm_x.bias",
+    "blocks.13.norm_x.weight",
+    "blocks.13.self_attn.k.bias",
+    "blocks.13.self_attn.k.weight",
+    "blocks.13.self_attn.norm_k.weight",
+    "blocks.13.self_attn.norm_q.weight",
+    "blocks.13.self_attn.o.bias",
+    "blocks.13.self_attn.o.weight",
+    "blocks.13.self_attn.q.bias",
+    "blocks.13.self_attn.q.weight",
+    "blocks.13.self_attn.v.bias",
+    "blocks.13.self_attn.v.weight",
+    "blocks.14.audio_cross_attn.kv_linear.bias",
+    "blocks.14.audio_cross_attn.kv_linear.weight",
+    "blocks.14.audio_cross_attn.proj.bias",
+    "blocks.14.audio_cross_attn.proj.weight",
+    "blocks.14.audio_cross_attn.q_linear.bias",
+    "blocks.14.audio_cross_attn.q_linear.weight",
+    "blocks.14.cross_attn.k.bias",
+    "blocks.14.cross_attn.k.weight",
+    "blocks.14.cross_attn.k_img.bias",
+    "blocks.14.cross_attn.k_img.weight",
+    "blocks.14.cross_attn.norm_k.weight",
+    "blocks.14.cross_attn.norm_k_img.weight",
+    "blocks.14.cross_attn.norm_q.weight",
+    "blocks.14.cross_attn.o.bias",
+    "blocks.14.cross_attn.o.weight",
+    "blocks.14.cross_attn.q.bias",
+    "blocks.14.cross_attn.q.weight",
+    "blocks.14.cross_attn.v.bias",
+    "blocks.14.cross_attn.v.weight",
+    "blocks.14.cross_attn.v_img.bias",
+    "blocks.14.cross_attn.v_img.weight",
+    "blocks.14.ffn.0.bias",
+    "blocks.14.ffn.0.weight",
+    "blocks.14.ffn.2.bias",
+    "blocks.14.ffn.2.weight",
+    "blocks.14.modulation",
+    "blocks.14.norm3.bias",
+    "blocks.14.norm3.weight",
+    "blocks.14.norm_x.bias",
+    "blocks.14.norm_x.weight",
+    "blocks.14.self_attn.k.bias",
+    "blocks.14.self_attn.k.weight",
+    "blocks.14.self_attn.norm_k.weight",
+    "blocks.14.self_attn.norm_q.weight",
+    "blocks.14.self_attn.o.bias",
+    "blocks.14.self_attn.o.weight",
+    "blocks.14.self_attn.q.bias",
+    "blocks.14.self_attn.q.weight",
+    "blocks.14.self_attn.v.bias",
+    "blocks.14.self_attn.v.weight",
+    "blocks.15.audio_cross_attn.kv_linear.bias",
+    "blocks.15.audio_cross_attn.kv_linear.weight",
+    "blocks.15.audio_cross_attn.proj.bias",
+    "blocks.15.audio_cross_attn.proj.weight",
+    "blocks.15.audio_cross_attn.q_linear.bias",
+    "blocks.15.audio_cross_attn.q_linear.weight",
+    "blocks.15.cross_attn.k.bias",
+    "blocks.15.cross_attn.k.weight",
+    "blocks.15.cross_attn.k_img.bias",
+    "blocks.15.cross_attn.k_img.weight",
+    "blocks.15.cross_attn.norm_k.weight",
+    "blocks.15.cross_attn.norm_k_img.weight",
+    "blocks.15.cross_attn.norm_q.weight",
+    "blocks.15.cross_attn.o.bias",
+    "blocks.15.cross_attn.o.weight",
+    "blocks.15.cross_attn.q.bias",
+    "blocks.15.cross_attn.q.weight",
+    "blocks.15.cross_attn.v.bias",
+    "blocks.15.cross_attn.v.weight",
+    "blocks.15.cross_attn.v_img.bias",
+    "blocks.15.cross_attn.v_img.weight",
+    "blocks.15.ffn.0.bias",
+    "blocks.15.ffn.0.weight",
+    "blocks.15.ffn.2.bias",
+    "blocks.15.ffn.2.weight",
+    "blocks.15.modulation",
+    "blocks.15.norm3.bias",
+    "blocks.15.norm3.weight",
+    "blocks.15.norm_x.bias",
+    "blocks.15.norm_x.weight",
+    "blocks.15.self_attn.k.bias",
+    "blocks.15.self_attn.k.weight",
+    "blocks.15.self_attn.norm_k.weight",
+    "blocks.15.self_attn.norm_q.weight",
+    "blocks.15.self_attn.o.bias",
+    "blocks.15.self_attn.o.weight",
+    "blocks.15.self_attn.q.bias",
+    "blocks.15.self_attn.q.weight",
+    "blocks.15.self_attn.v.bias",
+    "blocks.15.self_attn.v.weight",
+    "blocks.16.audio_cross_attn.kv_linear.bias",
+    "blocks.16.audio_cross_attn.kv_linear.weight",
+    "blocks.16.audio_cross_attn.proj.bias",
+    "blocks.16.audio_cross_attn.proj.weight",
+    "blocks.16.audio_cross_attn.q_linear.bias",
+    "blocks.16.audio_cross_attn.q_linear.weight",
+    "blocks.16.cross_attn.k.bias",
+    "blocks.16.cross_attn.k.weight",
+    "blocks.16.cross_attn.k_img.bias",
+    "blocks.16.cross_attn.k_img.weight",
+    "blocks.16.cross_attn.norm_k.weight",
+    "blocks.16.cross_attn.norm_k_img.weight",
+    "blocks.16.cross_attn.norm_q.weight",
+    "blocks.16.cross_attn.o.bias",
+    "blocks.16.cross_attn.o.weight",
+    "blocks.16.cross_attn.q.bias",
+    "blocks.16.cross_attn.q.weight",
+    "blocks.16.cross_attn.v.bias",
+    "blocks.16.cross_attn.v.weight",
+    "blocks.16.cross_attn.v_img.bias",
+    "blocks.16.cross_attn.v_img.weight",
+    "blocks.16.ffn.0.bias",
+    "blocks.16.ffn.0.weight",
+    "blocks.16.ffn.2.bias",
+    "blocks.16.ffn.2.weight",
+    "blocks.16.modulation",
+    "blocks.16.norm3.bias",
+    "blocks.16.norm3.weight",
+    "blocks.16.norm_x.bias",
+    "blocks.16.norm_x.weight",
+    "blocks.16.self_attn.k.bias",
+    "blocks.16.self_attn.k.weight",
+    "blocks.16.self_attn.norm_k.weight",
+    "blocks.16.self_attn.norm_q.weight",
+    "blocks.16.self_attn.o.bias",
+    "blocks.16.self_attn.o.weight",
+    "blocks.16.self_attn.q.bias",
+    "blocks.16.self_attn.q.weight",
+    "blocks.16.self_attn.v.bias",
+    "blocks.16.self_attn.v.weight",
+    "blocks.17.audio_cross_attn.kv_linear.bias",
+    "blocks.17.audio_cross_attn.kv_linear.weight",
+    "blocks.17.audio_cross_attn.proj.bias",
+    "blocks.17.audio_cross_attn.proj.weight",
+    "blocks.17.audio_cross_attn.q_linear.bias",
+    "blocks.17.audio_cross_attn.q_linear.weight",
+    "blocks.17.cross_attn.k.bias",
+    "blocks.17.cross_attn.k.weight",
+    "blocks.17.cross_attn.k_img.bias",
+    "blocks.17.cross_attn.k_img.weight",
+    "blocks.17.cross_attn.norm_k.weight",
+    "blocks.17.cross_attn.norm_k_img.weight",
+    "blocks.17.cross_attn.norm_q.weight",
+    "blocks.17.cross_attn.o.bias",
+    "blocks.17.cross_attn.o.weight",
+    "blocks.17.cross_attn.q.bias",
+    "blocks.17.cross_attn.q.weight",
+    "blocks.17.cross_attn.v.bias",
+    "blocks.17.cross_attn.v.weight",
+    "blocks.17.cross_attn.v_img.bias",
+    "blocks.17.cross_attn.v_img.weight",
+    "blocks.17.ffn.0.bias",
+    "blocks.17.ffn.0.weight",
+    "blocks.17.ffn.2.bias",
+    "blocks.17.ffn.2.weight",
+    "blocks.17.modulation",
+    "blocks.17.norm3.bias",
+    "blocks.17.norm3.weight",
+    "blocks.17.norm_x.bias",
+    "blocks.17.norm_x.weight",
+    "blocks.17.self_attn.k.bias",
+    "blocks.17.self_attn.k.weight",
+    "blocks.17.self_attn.norm_k.weight",
+    "blocks.17.self_attn.norm_q.weight",
+    "blocks.17.self_attn.o.bias",
+    "blocks.17.self_attn.o.weight",
+    "blocks.17.self_attn.q.bias",
+    "blocks.17.self_attn.q.weight",
+    "blocks.17.self_attn.v.bias",
+    "blocks.17.self_attn.v.weight",
+    "blocks.18.audio_cross_attn.kv_linear.bias",
+    "blocks.18.audio_cross_attn.kv_linear.weight",
+    "blocks.18.audio_cross_attn.proj.bias",
+    "blocks.18.audio_cross_attn.proj.weight",
+    "blocks.18.audio_cross_attn.q_linear.bias",
+    "blocks.18.audio_cross_attn.q_linear.weight",
+    "blocks.18.cross_attn.k.bias",
+    "blocks.18.cross_attn.k.weight",
+    "blocks.18.cross_attn.k_img.bias",
+    "blocks.18.cross_attn.k_img.weight",
+    "blocks.18.cross_attn.norm_k.weight",
+    "blocks.18.cross_attn.norm_k_img.weight",
+    "blocks.18.cross_attn.norm_q.weight",
+    "blocks.18.cross_attn.o.bias",
+    "blocks.18.cross_attn.o.weight",
+    "blocks.18.cross_attn.q.bias",
+    "blocks.18.cross_attn.q.weight",
+    "blocks.18.cross_attn.v.bias",
+    "blocks.18.cross_attn.v.weight",
+    "blocks.18.cross_attn.v_img.bias",
+    "blocks.18.cross_attn.v_img.weight",
+    "blocks.18.ffn.0.bias",
+    "blocks.18.ffn.0.weight",
+    "blocks.18.ffn.2.bias",
+    "blocks.18.ffn.2.weight",
+    "blocks.18.modulation",
+    "blocks.18.norm3.bias",
+    "blocks.18.norm3.weight",
+    "blocks.18.norm_x.bias",
+    "blocks.18.norm_x.weight",
+    "blocks.18.self_attn.k.bias",
+    "blocks.18.self_attn.k.weight",
+    "blocks.18.self_attn.norm_k.weight",
+    "blocks.18.self_attn.norm_q.weight",
+    "blocks.18.self_attn.o.bias",
+    "blocks.18.self_attn.o.weight",
+    "blocks.18.self_attn.q.bias",
+    "blocks.18.self_attn.q.weight",
+    "blocks.18.self_attn.v.bias",
+    "blocks.18.self_attn.v.weight",
+    "blocks.19.audio_cross_attn.kv_linear.bias",
+    "blocks.19.audio_cross_attn.kv_linear.weight",
+    "blocks.19.audio_cross_attn.proj.bias",
+    "blocks.19.audio_cross_attn.proj.weight",
+    "blocks.19.audio_cross_attn.q_linear.bias",
+    "blocks.19.audio_cross_attn.q_linear.weight",
+    "blocks.19.cross_attn.k.bias",
+    "blocks.19.cross_attn.k.weight",
+    "blocks.19.cross_attn.k_img.bias",
+    "blocks.19.cross_attn.k_img.weight",
+    "blocks.19.cross_attn.norm_k.weight",
+    "blocks.19.cross_attn.norm_k_img.weight",
+    "blocks.19.cross_attn.norm_q.weight",
+    "blocks.19.cross_attn.o.bias",
+    "blocks.19.cross_attn.o.weight",
+    "blocks.19.cross_attn.q.bias",
+    "blocks.19.cross_attn.q.weight",
+    "blocks.19.cross_attn.v.bias",
+    "blocks.19.cross_attn.v.weight",
+    "blocks.19.cross_attn.v_img.bias",
+    "blocks.19.cross_attn.v_img.weight",
+    "blocks.19.ffn.0.bias",
+    "blocks.19.ffn.0.weight",
+    "blocks.19.ffn.2.bias",
+    "blocks.19.ffn.2.weight",
+    "blocks.19.modulation",
+    "blocks.19.norm3.bias",
+    "blocks.19.norm3.weight",
+    "blocks.19.norm_x.bias",
+    "blocks.19.norm_x.weight",
+    "blocks.19.self_attn.k.bias",
+    "blocks.19.self_attn.k.weight",
+    "blocks.19.self_attn.norm_k.weight",
+    "blocks.19.self_attn.norm_q.weight",
+    "blocks.19.self_attn.o.bias",
+    "blocks.19.self_attn.o.weight",
+    "blocks.19.self_attn.q.bias",
+    "blocks.19.self_attn.q.weight",
+    "blocks.19.self_attn.v.bias",
+    "blocks.19.self_attn.v.weight",
+    "blocks.2.audio_cross_attn.kv_linear.bias",
+    "blocks.2.audio_cross_attn.kv_linear.weight",
+    "blocks.2.audio_cross_attn.proj.bias",
+    "blocks.2.audio_cross_attn.proj.weight",
+    "blocks.2.audio_cross_attn.q_linear.bias",
+    "blocks.2.audio_cross_attn.q_linear.weight",
+    "blocks.2.cross_attn.k.bias",
+    "blocks.2.cross_attn.k.weight",
+    "blocks.2.cross_attn.k_img.bias",
+    "blocks.2.cross_attn.k_img.weight",
+    "blocks.2.cross_attn.norm_k.weight",
+    "blocks.2.cross_attn.norm_k_img.weight",
+    "blocks.2.cross_attn.norm_q.weight",
+    "blocks.2.cross_attn.o.bias",
+    "blocks.2.cross_attn.o.weight",
+    "blocks.2.cross_attn.q.bias",
+    "blocks.2.cross_attn.q.weight",
+    "blocks.2.cross_attn.v.bias",
+    "blocks.2.cross_attn.v.weight",
+    "blocks.2.cross_attn.v_img.bias",
+    "blocks.2.cross_attn.v_img.weight",
+    "blocks.2.ffn.0.bias",
+    "blocks.2.ffn.0.weight",
+    "blocks.2.ffn.2.bias",
+    "blocks.2.ffn.2.weight",
+    "blocks.2.modulation",
+    "blocks.2.norm3.bias",
+    "blocks.2.norm3.weight",
+    "blocks.2.norm_x.bias",
+    "blocks.2.norm_x.weight",
+    "blocks.2.self_attn.k.bias",
+    "blocks.2.self_attn.k.weight",
+    "blocks.2.self_attn.norm_k.weight",
+    "blocks.2.self_attn.norm_q.weight",
+    "blocks.2.self_attn.o.bias",
+    "blocks.2.self_attn.o.weight",
+    "blocks.2.self_attn.q.bias",
+    "blocks.2.self_attn.q.weight",
+    "blocks.2.self_attn.v.bias",
+    "blocks.2.self_attn.v.weight",
+    "blocks.20.audio_cross_attn.kv_linear.bias",
+    "blocks.20.audio_cross_attn.kv_linear.weight",
+    "blocks.20.audio_cross_attn.proj.bias",
+    "blocks.20.audio_cross_attn.proj.weight",
+    "blocks.20.audio_cross_attn.q_linear.bias",
+    "blocks.20.audio_cross_attn.q_linear.weight",
+    "blocks.20.cross_attn.k.bias",
+    "blocks.20.cross_attn.k.weight",
+    "blocks.20.cross_attn.k_img.bias",
+    "blocks.20.cross_attn.k_img.weight",
+    "blocks.20.cross_attn.norm_k.weight",
+    "blocks.20.cross_attn.norm_k_img.weight",
+    "blocks.20.cross_attn.norm_q.weight",
+    "blocks.20.cross_attn.o.bias",
+    "blocks.20.cross_attn.o.weight",
+    "blocks.20.cross_attn.q.bias",
+    "blocks.20.cross_attn.q.weight",
+    "blocks.20.cross_attn.v.bias",
+    "blocks.20.cross_attn.v.weight",
+    "blocks.20.cross_attn.v_img.bias",
+    "blocks.20.cross_attn.v_img.weight",
+    "blocks.20.ffn.0.bias",
+    "blocks.20.ffn.0.weight",
+    "blocks.20.ffn.2.bias",
+    "blocks.20.ffn.2.weight",
+    "blocks.20.modulation",
+    "blocks.20.norm3.bias",
+    "blocks.20.norm3.weight",
+    "blocks.20.norm_x.bias",
+    "blocks.20.norm_x.weight",
+    "blocks.20.self_attn.k.bias",
+    "blocks.20.self_attn.k.weight",
+    "blocks.20.self_attn.norm_k.weight",
+    "blocks.20.self_attn.norm_q.weight",
+    "blocks.20.self_attn.o.bias",
+    "blocks.20.self_attn.o.weight",
+    "blocks.20.self_attn.q.bias",
+    "blocks.20.self_attn.q.weight",
+    "blocks.20.self_attn.v.bias",
+    "blocks.20.self_attn.v.weight",
+    "blocks.21.audio_cross_attn.kv_linear.bias",
+    "blocks.21.audio_cross_attn.kv_linear.weight",
+    "blocks.21.audio_cross_attn.proj.bias",
+    "blocks.21.audio_cross_attn.proj.weight",
+    "blocks.21.audio_cross_attn.q_linear.bias",
+    "blocks.21.audio_cross_attn.q_linear.weight",
+    "blocks.21.cross_attn.k.bias",
+    "blocks.21.cross_attn.k.weight",
+    "blocks.21.cross_attn.k_img.bias",
+    "blocks.21.cross_attn.k_img.weight",
+    "blocks.21.cross_attn.norm_k.weight",
+    "blocks.21.cross_attn.norm_k_img.weight",
+    "blocks.21.cross_attn.norm_q.weight",
+    "blocks.21.cross_attn.o.bias",
+    "blocks.21.cross_attn.o.weight",
+    "blocks.21.cross_attn.q.bias",
+    "blocks.21.cross_attn.q.weight",
+    "blocks.21.cross_attn.v.bias",
+    "blocks.21.cross_attn.v.weight",
+    "blocks.21.cross_attn.v_img.bias",
+    "blocks.21.cross_attn.v_img.weight",
+    "blocks.21.ffn.0.bias",
+    "blocks.21.ffn.0.weight",
+    "blocks.21.ffn.2.bias",
+    "blocks.21.ffn.2.weight",
+    "blocks.21.modulation",
+    "blocks.21.norm3.bias",
+    "blocks.21.norm3.weight",
+    "blocks.21.norm_x.bias",
+    "blocks.21.norm_x.weight",
+    "blocks.21.self_attn.k.bias",
+    "blocks.21.self_attn.k.weight",
+    "blocks.21.self_attn.norm_k.weight",
+    "blocks.21.self_attn.norm_q.weight",
+    "blocks.21.self_attn.o.bias",
+    "blocks.21.self_attn.o.weight",
+    "blocks.21.self_attn.q.bias",
+    "blocks.21.self_attn.q.weight",
+    "blocks.21.self_attn.v.bias",
+    "blocks.21.self_attn.v.weight",
+    "blocks.22.audio_cross_attn.kv_linear.bias",
+    "blocks.22.audio_cross_attn.kv_linear.weight",
+    "blocks.22.audio_cross_attn.proj.bias",
+    "blocks.22.audio_cross_attn.proj.weight",
+    "blocks.22.audio_cross_attn.q_linear.bias",
+    "blocks.22.audio_cross_attn.q_linear.weight",
+    "blocks.22.cross_attn.k.bias",
+    "blocks.22.cross_attn.k.weight",
+    "blocks.22.cross_attn.k_img.bias",
+    "blocks.22.cross_attn.k_img.weight",
+    "blocks.22.cross_attn.norm_k.weight",
+    "blocks.22.cross_attn.norm_k_img.weight",
+    "blocks.22.cross_attn.norm_q.weight",
+    "blocks.22.cross_attn.o.bias",
+    "blocks.22.cross_attn.o.weight",
+    "blocks.22.cross_attn.q.bias",
+    "blocks.22.cross_attn.q.weight",
+    "blocks.22.cross_attn.v.bias",
+    "blocks.22.cross_attn.v.weight",
+    "blocks.22.cross_attn.v_img.bias",
+    "blocks.22.cross_attn.v_img.weight",
+    "blocks.22.ffn.0.bias",
+    "blocks.22.ffn.0.weight",
+    "blocks.22.ffn.2.bias",
+    "blocks.22.ffn.2.weight",
+    "blocks.22.modulation",
+    "blocks.22.norm3.bias",
+    "blocks.22.norm3.weight",
+    "blocks.22.norm_x.bias",
+    "blocks.22.norm_x.weight",
+    "blocks.22.self_attn.k.bias",
+    "blocks.22.self_attn.k.weight",
+    "blocks.22.self_attn.norm_k.weight",
+    "blocks.22.self_attn.norm_q.weight",
+    "blocks.22.self_attn.o.bias",
+    "blocks.22.self_attn.o.weight",
+    "blocks.22.self_attn.q.bias",
+    "blocks.22.self_attn.q.weight",
+    "blocks.22.self_attn.v.bias",
+    "blocks.22.self_attn.v.weight",
+    "blocks.23.audio_cross_attn.kv_linear.bias",
+    "blocks.23.audio_cross_attn.kv_linear.weight",
+    "blocks.23.audio_cross_attn.proj.bias",
+    "blocks.23.audio_cross_attn.proj.weight",
+    "blocks.23.audio_cross_attn.q_linear.bias",
+    "blocks.23.audio_cross_attn.q_linear.weight",
+    "blocks.23.cross_attn.k.bias",
+    "blocks.23.cross_attn.k.weight",
+    "blocks.23.cross_attn.k_img.bias",
+    "blocks.23.cross_attn.k_img.weight",
+    "blocks.23.cross_attn.norm_k.weight",
+    "blocks.23.cross_attn.norm_k_img.weight",
+    "blocks.23.cross_attn.norm_q.weight",
+    "blocks.23.cross_attn.o.bias",
+    "blocks.23.cross_attn.o.weight",
+    "blocks.23.cross_attn.q.bias",
+    "blocks.23.cross_attn.q.weight",
+    "blocks.23.cross_attn.v.bias",
+    "blocks.23.cross_attn.v.weight",
+    "blocks.23.cross_attn.v_img.bias",
+    "blocks.23.cross_attn.v_img.weight",
+    "blocks.23.ffn.0.bias",
+    "blocks.23.ffn.0.weight",
+    "blocks.23.ffn.2.bias",
+    "blocks.23.ffn.2.weight",
+    "blocks.23.modulation",
+    "blocks.23.norm3.bias",
+    "blocks.23.norm3.weight",
+    "blocks.23.norm_x.bias",
+    "blocks.23.norm_x.weight",
+    "blocks.23.self_attn.k.bias",
+    "blocks.23.self_attn.k.weight",
+    "blocks.23.self_attn.norm_k.weight",
+    "blocks.23.self_attn.norm_q.weight",
+    "blocks.23.self_attn.o.bias",
+    "blocks.23.self_attn.o.weight",
+    "blocks.23.self_attn.q.bias",
+    "blocks.23.self_attn.q.weight",
+    "blocks.23.self_attn.v.bias",
+    "blocks.23.self_attn.v.weight",
+    "blocks.24.audio_cross_attn.kv_linear.bias",
+    "blocks.24.audio_cross_attn.kv_linear.weight",
+    "blocks.24.audio_cross_attn.proj.bias",
+    "blocks.24.audio_cross_attn.proj.weight",
+    "blocks.24.audio_cross_attn.q_linear.bias",
+    "blocks.24.audio_cross_attn.q_linear.weight",
+    "blocks.24.cross_attn.k.bias",
+    "blocks.24.cross_attn.k.weight",
+    "blocks.24.cross_attn.k_img.bias",
+    "blocks.24.cross_attn.k_img.weight",
+    "blocks.24.cross_attn.norm_k.weight",
+    "blocks.24.cross_attn.norm_k_img.weight",
+    "blocks.24.cross_attn.norm_q.weight",
+    "blocks.24.cross_attn.o.bias",
+    "blocks.24.cross_attn.o.weight",
+    "blocks.24.cross_attn.q.bias",
+    "blocks.24.cross_attn.q.weight",
+    "blocks.24.cross_attn.v.bias",
+    "blocks.24.cross_attn.v.weight",
+    "blocks.24.cross_attn.v_img.bias",
+    "blocks.24.cross_attn.v_img.weight",
+    "blocks.24.ffn.0.bias",
+    "blocks.24.ffn.0.weight",
+    "blocks.24.ffn.2.bias",
+    "blocks.24.ffn.2.weight",
+    "blocks.24.modulation",
+    "blocks.24.norm3.bias",
+    "blocks.24.norm3.weight",
+    "blocks.24.norm_x.bias",
+    "blocks.24.norm_x.weight",
+    "blocks.24.self_attn.k.bias",
+    "blocks.24.self_attn.k.weight",
+    "blocks.24.self_attn.norm_k.weight",
+    "blocks.24.self_attn.norm_q.weight",
+    "blocks.24.self_attn.o.bias",
+    "blocks.24.self_attn.o.weight",
+    "blocks.24.self_attn.q.bias",
+    "blocks.24.self_attn.q.weight",
+    "blocks.24.self_attn.v.bias",
+    "blocks.24.self_attn.v.weight",
+    "blocks.25.audio_cross_attn.kv_linear.bias",
+    "blocks.25.audio_cross_attn.kv_linear.weight",
+    "blocks.25.audio_cross_attn.proj.bias",
+    "blocks.25.audio_cross_attn.proj.weight",
+    "blocks.25.audio_cross_attn.q_linear.bias",
+    "blocks.25.audio_cross_attn.q_linear.weight",
+    "blocks.25.cross_attn.k.bias",
+    "blocks.25.cross_attn.k.weight",
+    "blocks.25.cross_attn.k_img.bias",
+    "blocks.25.cross_attn.k_img.weight",
+    "blocks.25.cross_attn.norm_k.weight",
+    "blocks.25.cross_attn.norm_k_img.weight",
+    "blocks.25.cross_attn.norm_q.weight",
+    "blocks.25.cross_attn.o.bias",
+    "blocks.25.cross_attn.o.weight",
+    "blocks.25.cross_attn.q.bias",
+    "blocks.25.cross_attn.q.weight",
+    "blocks.25.cross_attn.v.bias",
+    "blocks.25.cross_attn.v.weight",
+    "blocks.25.cross_attn.v_img.bias",
+    "blocks.25.cross_attn.v_img.weight",
+    "blocks.25.ffn.0.bias",
+    "blocks.25.ffn.0.weight",
+    "blocks.25.ffn.2.bias",
+    "blocks.25.ffn.2.weight",
+    "blocks.25.modulation",
+    "blocks.25.norm3.bias",
+    "blocks.25.norm3.weight",
+    "blocks.25.norm_x.bias",
+    "blocks.25.norm_x.weight",
+    "blocks.25.self_attn.k.bias",
+    "blocks.25.self_attn.k.weight",
+    "blocks.25.self_attn.norm_k.weight",
+    "blocks.25.self_attn.norm_q.weight",
+    "blocks.25.self_attn.o.bias",
+    "blocks.25.self_attn.o.weight",
+    "blocks.25.self_attn.q.bias",
+    "blocks.25.self_attn.q.weight",
+    "blocks.25.self_attn.v.bias",
+    "blocks.25.self_attn.v.weight",
+    "blocks.26.audio_cross_attn.kv_linear.bias",
+    "blocks.26.audio_cross_attn.kv_linear.weight",
+    "blocks.26.audio_cross_attn.proj.bias",
+    "blocks.26.audio_cross_attn.proj.weight",
+    "blocks.26.audio_cross_attn.q_linear.bias",
+    "blocks.26.audio_cross_attn.q_linear.weight",
+    "blocks.26.cross_attn.k.bias",
+    "blocks.26.cross_attn.k.weight",
+    "blocks.26.cross_attn.k_img.bias",
+    "blocks.26.cross_attn.k_img.weight",
+    "blocks.26.cross_attn.norm_k.weight",
+    "blocks.26.cross_attn.norm_k_img.weight",
+    "blocks.26.cross_attn.norm_q.weight",
+    "blocks.26.cross_attn.o.bias",
+    "blocks.26.cross_attn.o.weight",
+    "blocks.26.cross_attn.q.bias",
+    "blocks.26.cross_attn.q.weight",
+    "blocks.26.cross_attn.v.bias",
+    "blocks.26.cross_attn.v.weight",
+    "blocks.26.cross_attn.v_img.bias",
+    "blocks.26.cross_attn.v_img.weight",
+    "blocks.26.ffn.0.bias",
+    "blocks.26.ffn.0.weight",
+    "blocks.26.ffn.2.bias",
+    "blocks.26.ffn.2.weight",
+    "blocks.26.modulation",
+    "blocks.26.norm3.bias",
+    "blocks.26.norm3.weight",
+    "blocks.26.norm_x.bias",
+    "blocks.26.norm_x.weight",
+    "blocks.26.self_attn.k.bias",
+    "blocks.26.self_attn.k.weight",
+    "blocks.26.self_attn.norm_k.weight",
+    "blocks.26.self_attn.norm_q.weight",
+    "blocks.26.self_attn.o.bias",
+    "blocks.26.self_attn.o.weight",
+    "blocks.26.self_attn.q.bias",
+    "blocks.26.self_attn.q.weight",
+    "blocks.26.self_attn.v.bias",
+    "blocks.26.self_attn.v.weight",
+    "blocks.27.audio_cross_attn.kv_linear.bias",
+    "blocks.27.audio_cross_attn.kv_linear.weight",
+    "blocks.27.audio_cross_attn.proj.bias",
+    "blocks.27.audio_cross_attn.proj.weight",
+    "blocks.27.audio_cross_attn.q_linear.bias",
+    "blocks.27.audio_cross_attn.q_linear.weight",
+    "blocks.27.cross_attn.k.bias",
+    "blocks.27.cross_attn.k.weight",
+    "blocks.27.cross_attn.k_img.bias",
+    "blocks.27.cross_attn.k_img.weight",
+    "blocks.27.cross_attn.norm_k.weight",
+    "blocks.27.cross_attn.norm_k_img.weight",
+    "blocks.27.cross_attn.norm_q.weight",
+    "blocks.27.cross_attn.o.bias",
+    "blocks.27.cross_attn.o.weight",
+    "blocks.27.cross_attn.q.bias",
+    "blocks.27.cross_attn.q.weight",
+    "blocks.27.cross_attn.v.bias",
+    "blocks.27.cross_attn.v.weight",
+    "blocks.27.cross_attn.v_img.bias",
+    "blocks.27.cross_attn.v_img.weight",
+    "blocks.27.ffn.0.bias",
+    "blocks.27.ffn.0.weight",
+    "blocks.27.ffn.2.bias",
+    "blocks.27.ffn.2.weight",
+    "blocks.27.modulation",
+    "blocks.27.norm3.bias",
+    "blocks.27.norm3.weight",
+    "blocks.27.norm_x.bias",
+    "blocks.27.norm_x.weight",
+    "blocks.27.self_attn.k.bias",
+    "blocks.27.self_attn.k.weight",
+    "blocks.27.self_attn.norm_k.weight",
+    "blocks.27.self_attn.norm_q.weight",
+    "blocks.27.self_attn.o.bias",
+    "blocks.27.self_attn.o.weight",
+    "blocks.27.self_attn.q.bias",
+    "blocks.27.self_attn.q.weight",
+    "blocks.27.self_attn.v.bias",
+    "blocks.27.self_attn.v.weight",
+    "blocks.28.audio_cross_attn.kv_linear.bias",
+    "blocks.28.audio_cross_attn.kv_linear.weight",
+    "blocks.28.audio_cross_attn.proj.bias",
+    "blocks.28.audio_cross_attn.proj.weight",
+    "blocks.28.audio_cross_attn.q_linear.bias",
+    "blocks.28.audio_cross_attn.q_linear.weight",
+    "blocks.28.cross_attn.k.bias",
+    "blocks.28.cross_attn.k.weight",
+    "blocks.28.cross_attn.k_img.bias",
+    "blocks.28.cross_attn.k_img.weight",
+    "blocks.28.cross_attn.norm_k.weight",
+    "blocks.28.cross_attn.norm_k_img.weight",
+    "blocks.28.cross_attn.norm_q.weight",
+    "blocks.28.cross_attn.o.bias",
+    "blocks.28.cross_attn.o.weight",
+    "blocks.28.cross_attn.q.bias",
+    "blocks.28.cross_attn.q.weight",
+    "blocks.28.cross_attn.v.bias",
+    "blocks.28.cross_attn.v.weight",
+    "blocks.28.cross_attn.v_img.bias",
+    "blocks.28.cross_attn.v_img.weight",
+    "blocks.28.ffn.0.bias",
+    "blocks.28.ffn.0.weight",
+    "blocks.28.ffn.2.bias",
+    "blocks.28.ffn.2.weight",
+    "blocks.28.modulation",
+    "blocks.28.norm3.bias",
+    "blocks.28.norm3.weight",
+    "blocks.28.norm_x.bias",
+    "blocks.28.norm_x.weight",
+    "blocks.28.self_attn.k.bias",
+    "blocks.28.self_attn.k.weight",
+    "blocks.28.self_attn.norm_k.weight",
+    "blocks.28.self_attn.norm_q.weight",
+    "blocks.28.self_attn.o.bias",
+    "blocks.28.self_attn.o.weight",
+    "blocks.28.self_attn.q.bias",
+    "blocks.28.self_attn.q.weight",
+    "blocks.28.self_attn.v.bias",
+    "blocks.28.self_attn.v.weight",
+    "blocks.29.audio_cross_attn.kv_linear.bias",
+    "blocks.29.audio_cross_attn.kv_linear.weight",
+    "blocks.29.audio_cross_attn.proj.bias",
+    "blocks.29.audio_cross_attn.proj.weight",
+    "blocks.29.audio_cross_attn.q_linear.bias",
+    "blocks.29.audio_cross_attn.q_linear.weight",
+    "blocks.29.cross_attn.k.bias",
+    "blocks.29.cross_attn.k.weight",
+    "blocks.29.cross_attn.k_img.bias",
+    "blocks.29.cross_attn.k_img.weight",
+    "blocks.29.cross_attn.norm_k.weight",
+    "blocks.29.cross_attn.norm_k_img.weight",
+    "blocks.29.cross_attn.norm_q.weight",
+    "blocks.29.cross_attn.o.bias",
+    "blocks.29.cross_attn.o.weight",
+    "blocks.29.cross_attn.q.bias",
+    "blocks.29.cross_attn.q.weight",
+    "blocks.29.cross_attn.v.bias",
+    "blocks.29.cross_attn.v.weight",
+    "blocks.29.cross_attn.v_img.bias",
+    "blocks.29.cross_attn.v_img.weight",
+    "blocks.29.ffn.0.bias",
+    "blocks.29.ffn.0.weight",
+    "blocks.29.ffn.2.bias",
+    "blocks.29.ffn.2.weight",
+    "blocks.29.modulation",
+    "blocks.29.norm3.bias",
+    "blocks.29.norm3.weight",
+    "blocks.29.norm_x.bias",
+    "blocks.29.norm_x.weight",
+    "blocks.29.self_attn.k.bias",
+    "blocks.29.self_attn.k.weight",
+    "blocks.29.self_attn.norm_k.weight",
+    "blocks.29.self_attn.norm_q.weight",
+    "blocks.29.self_attn.o.bias",
+    "blocks.29.self_attn.o.weight",
+    "blocks.29.self_attn.q.bias",
+    "blocks.29.self_attn.q.weight",
+    "blocks.29.self_attn.v.bias",
+    "blocks.29.self_attn.v.weight",
+    "blocks.3.audio_cross_attn.kv_linear.bias",
+    "blocks.3.audio_cross_attn.kv_linear.weight",
+    "blocks.3.audio_cross_attn.proj.bias",
+    "blocks.3.audio_cross_attn.proj.weight",
+    "blocks.3.audio_cross_attn.q_linear.bias",
+    "blocks.3.audio_cross_attn.q_linear.weight",
+    "blocks.3.cross_attn.k.bias",
+    "blocks.3.cross_attn.k.weight",
+    "blocks.3.cross_attn.k_img.bias",
+    "blocks.3.cross_attn.k_img.weight",
+    "blocks.3.cross_attn.norm_k.weight",
+    "blocks.3.cross_attn.norm_k_img.weight",
+    "blocks.3.cross_attn.norm_q.weight",
+    "blocks.3.cross_attn.o.bias",
+    "blocks.3.cross_attn.o.weight",
+    "blocks.3.cross_attn.q.bias",
+    "blocks.3.cross_attn.q.weight",
+    "blocks.3.cross_attn.v.bias",
+    "blocks.3.cross_attn.v.weight",
+    "blocks.3.cross_attn.v_img.bias",
+    "blocks.3.cross_attn.v_img.weight",
+    "blocks.3.ffn.0.bias",
+    "blocks.3.ffn.0.weight",
+    "blocks.3.ffn.2.bias",
+    "blocks.3.ffn.2.weight",
+    "blocks.3.modulation",
+    "blocks.3.norm3.bias",
+    "blocks.3.norm3.weight",
+    "blocks.3.norm_x.bias",
+    "blocks.3.norm_x.weight",
+    "blocks.3.self_attn.k.bias",
+    "blocks.3.self_attn.k.weight",
+    "blocks.3.self_attn.norm_k.weight",
+    "blocks.3.self_attn.norm_q.weight",
+    "blocks.3.self_attn.o.bias",
+    "blocks.3.self_attn.o.weight",
+    "blocks.3.self_attn.q.bias",
+    "blocks.3.self_attn.q.weight",
+    "blocks.3.self_attn.v.bias",
+    "blocks.3.self_attn.v.weight",
+    "blocks.30.audio_cross_attn.kv_linear.bias",
+    "blocks.30.audio_cross_attn.kv_linear.weight",
+    "blocks.30.audio_cross_attn.proj.bias",
+    "blocks.30.audio_cross_attn.proj.weight",
+    "blocks.30.audio_cross_attn.q_linear.bias",
+    "blocks.30.audio_cross_attn.q_linear.weight",
+    "blocks.30.cross_attn.k.bias",
+    "blocks.30.cross_attn.k.weight",
+    "blocks.30.cross_attn.k_img.bias",
+    "blocks.30.cross_attn.k_img.weight",
+    "blocks.30.cross_attn.norm_k.weight",
+    "blocks.30.cross_attn.norm_k_img.weight",
+    "blocks.30.cross_attn.norm_q.weight",
+    "blocks.30.cross_attn.o.bias",
+    "blocks.30.cross_attn.o.weight",
+    "blocks.30.cross_attn.q.bias",
+    "blocks.30.cross_attn.q.weight",
+    "blocks.30.cross_attn.v.bias",
+    "blocks.30.cross_attn.v.weight",
+    "blocks.30.cross_attn.v_img.bias",
+    "blocks.30.cross_attn.v_img.weight",
+    "blocks.30.ffn.0.bias",
+    "blocks.30.ffn.0.weight",
+    "blocks.30.ffn.2.bias",
+    "blocks.30.ffn.2.weight",
+    "blocks.30.modulation",
+    "blocks.30.norm3.bias",
+    "blocks.30.norm3.weight",
+    "blocks.30.norm_x.bias",
+    "blocks.30.norm_x.weight",
+    "blocks.30.self_attn.k.bias",
+    "blocks.30.self_attn.k.weight",
+    "blocks.30.self_attn.norm_k.weight",
+    "blocks.30.self_attn.norm_q.weight",
+    "blocks.30.self_attn.o.bias",
+    "blocks.30.self_attn.o.weight",
+    "blocks.30.self_attn.q.bias",
+    "blocks.30.self_attn.q.weight",
+    "blocks.30.self_attn.v.bias",
+    "blocks.30.self_attn.v.weight",
+    "blocks.31.audio_cross_attn.kv_linear.bias",
+    "blocks.31.audio_cross_attn.kv_linear.weight",
+    "blocks.31.audio_cross_attn.proj.bias",
+    "blocks.31.audio_cross_attn.proj.weight",
+    "blocks.31.audio_cross_attn.q_linear.bias",
+    "blocks.31.audio_cross_attn.q_linear.weight",
+    "blocks.31.cross_attn.k.bias",
+    "blocks.31.cross_attn.k.weight",
+    "blocks.31.cross_attn.k_img.bias",
+    "blocks.31.cross_attn.k_img.weight",
+    "blocks.31.cross_attn.norm_k.weight",
+    "blocks.31.cross_attn.norm_k_img.weight",
+    "blocks.31.cross_attn.norm_q.weight",
+    "blocks.31.cross_attn.o.bias",
+    "blocks.31.cross_attn.o.weight",
+    "blocks.31.cross_attn.q.bias",
+    "blocks.31.cross_attn.q.weight",
+    "blocks.31.cross_attn.v.bias",
+    "blocks.31.cross_attn.v.weight",
+    "blocks.31.cross_attn.v_img.bias",
+    "blocks.31.cross_attn.v_img.weight",
+    "blocks.31.ffn.0.bias",
+    "blocks.31.ffn.0.weight",
+    "blocks.31.ffn.2.bias",
+    "blocks.31.ffn.2.weight",
+    "blocks.31.modulation",
+    "blocks.31.norm3.bias",
+    "blocks.31.norm3.weight",
+    "blocks.31.norm_x.bias",
+    "blocks.31.norm_x.weight",
+    "blocks.31.self_attn.k.bias",
+    "blocks.31.self_attn.k.weight",
+    "blocks.31.self_attn.norm_k.weight",
+    "blocks.31.self_attn.norm_q.weight",
+    "blocks.31.self_attn.o.bias",
+    "blocks.31.self_attn.o.weight",
+    "blocks.31.self_attn.q.bias",
+    "blocks.31.self_attn.q.weight",
+    "blocks.31.self_attn.v.bias",
+    "blocks.31.self_attn.v.weight",
+    "blocks.32.audio_cross_attn.kv_linear.bias",
+    "blocks.32.audio_cross_attn.kv_linear.weight",
+    "blocks.32.audio_cross_attn.proj.bias",
+    "blocks.32.audio_cross_attn.proj.weight",
+    "blocks.32.audio_cross_attn.q_linear.bias",
+    "blocks.32.audio_cross_attn.q_linear.weight",
+    "blocks.32.cross_attn.k.bias",
+    "blocks.32.cross_attn.k.weight",
+    "blocks.32.cross_attn.k_img.bias",
+    "blocks.32.cross_attn.k_img.weight",
+    "blocks.32.cross_attn.norm_k.weight",
+    "blocks.32.cross_attn.norm_k_img.weight",
+    "blocks.32.cross_attn.norm_q.weight",
+    "blocks.32.cross_attn.o.bias",
+    "blocks.32.cross_attn.o.weight",
+    "blocks.32.cross_attn.q.bias",
+    "blocks.32.cross_attn.q.weight",
+    "blocks.32.cross_attn.v.bias",
+    "blocks.32.cross_attn.v.weight",
+    "blocks.32.cross_attn.v_img.bias",
+    "blocks.32.cross_attn.v_img.weight",
+    "blocks.32.ffn.0.bias",
+    "blocks.32.ffn.0.weight",
+    "blocks.32.ffn.2.bias",
+    "blocks.32.ffn.2.weight",
+    "blocks.32.modulation",
+    "blocks.32.norm3.bias",
+    "blocks.32.norm3.weight",
+    "blocks.32.norm_x.bias",
+    "blocks.32.norm_x.weight",
+    "blocks.32.self_attn.k.bias",
+    "blocks.32.self_attn.k.weight",
+    "blocks.32.self_attn.norm_k.weight",
+    "blocks.32.self_attn.norm_q.weight",
+    "blocks.32.self_attn.o.bias",
+    "blocks.32.self_attn.o.weight",
+    "blocks.32.self_attn.q.bias",
+    "blocks.32.self_attn.q.weight",
+    "blocks.32.self_attn.v.bias",
+    "blocks.32.self_attn.v.weight",
+    "blocks.33.audio_cross_attn.kv_linear.bias",
+    "blocks.33.audio_cross_attn.kv_linear.weight",
+    "blocks.33.audio_cross_attn.proj.bias",
+    "blocks.33.audio_cross_attn.proj.weight",
+    "blocks.33.audio_cross_attn.q_linear.bias",
+    "blocks.33.audio_cross_attn.q_linear.weight",
+    "blocks.33.cross_attn.k.bias",
+    "blocks.33.cross_attn.k.weight",
+    "blocks.33.cross_attn.k_img.bias",
+    "blocks.33.cross_attn.k_img.weight",
+    "blocks.33.cross_attn.norm_k.weight",
+    "blocks.33.cross_attn.norm_k_img.weight",
+    "blocks.33.cross_attn.norm_q.weight",
+    "blocks.33.cross_attn.o.bias",
+    "blocks.33.cross_attn.o.weight",
+    "blocks.33.cross_attn.q.bias",
+    "blocks.33.cross_attn.q.weight",
+    "blocks.33.cross_attn.v.bias",
+    "blocks.33.cross_attn.v.weight",
+    "blocks.33.cross_attn.v_img.bias",
+    "blocks.33.cross_attn.v_img.weight",
+    "blocks.33.ffn.0.bias",
+    "blocks.33.ffn.0.weight",
+    "blocks.33.ffn.2.bias",
+    "blocks.33.ffn.2.weight",
+    "blocks.33.modulation",
+    "blocks.33.norm3.bias",
+    "blocks.33.norm3.weight",
+    "blocks.33.norm_x.bias",
+    "blocks.33.norm_x.weight",
+    "blocks.33.self_attn.k.bias",
+    "blocks.33.self_attn.k.weight",
+    "blocks.33.self_attn.norm_k.weight",
+    "blocks.33.self_attn.norm_q.weight",
+    "blocks.33.self_attn.o.bias",
+    "blocks.33.self_attn.o.weight",
+    "blocks.33.self_attn.q.bias",
+    "blocks.33.self_attn.q.weight",
+    "blocks.33.self_attn.v.bias",
+    "blocks.33.self_attn.v.weight",
+    "blocks.34.audio_cross_attn.kv_linear.bias",
+    "blocks.34.audio_cross_attn.kv_linear.weight",
+    "blocks.34.audio_cross_attn.proj.bias",
+    "blocks.34.audio_cross_attn.proj.weight",
+    "blocks.34.audio_cross_attn.q_linear.bias",
+    "blocks.34.audio_cross_attn.q_linear.weight",
+    "blocks.34.cross_attn.k.bias",
+    "blocks.34.cross_attn.k.weight",
+    "blocks.34.cross_attn.k_img.bias",
+    "blocks.34.cross_attn.k_img.weight",
+    "blocks.34.cross_attn.norm_k.weight",
+    "blocks.34.cross_attn.norm_k_img.weight",
+    "blocks.34.cross_attn.norm_q.weight",
+    "blocks.34.cross_attn.o.bias",
+    "blocks.34.cross_attn.o.weight",
+    "blocks.34.cross_attn.q.bias",
+    "blocks.34.cross_attn.q.weight",
+    "blocks.34.cross_attn.v.bias",
+    "blocks.34.cross_attn.v.weight",
+    "blocks.34.cross_attn.v_img.bias",
+    "blocks.34.cross_attn.v_img.weight",
+    "blocks.34.ffn.0.bias",
+    "blocks.34.ffn.0.weight",
+    "blocks.34.ffn.2.bias",
+    "blocks.34.ffn.2.weight",
+    "blocks.34.modulation",
+    "blocks.34.norm3.bias",
+    "blocks.34.norm3.weight",
+    "blocks.34.norm_x.bias",
+    "blocks.34.norm_x.weight",
+    "blocks.34.self_attn.k.bias",
+    "blocks.34.self_attn.k.weight",
+    "blocks.34.self_attn.norm_k.weight",
+    "blocks.34.self_attn.norm_q.weight",
+    "blocks.34.self_attn.o.bias",
+    "blocks.34.self_attn.o.weight",
+    "blocks.34.self_attn.q.bias",
+    "blocks.34.self_attn.q.weight",
+    "blocks.34.self_attn.v.bias",
+    "blocks.34.self_attn.v.weight",
+    "blocks.35.audio_cross_attn.kv_linear.bias",
+    "blocks.35.audio_cross_attn.kv_linear.weight",
+    "blocks.35.audio_cross_attn.proj.bias",
+    "blocks.35.audio_cross_attn.proj.weight",
+    "blocks.35.audio_cross_attn.q_linear.bias",
+    "blocks.35.audio_cross_attn.q_linear.weight",
+    "blocks.35.cross_attn.k.bias",
+    "blocks.35.cross_attn.k.weight",
+    "blocks.35.cross_attn.k_img.bias",
+    "blocks.35.cross_attn.k_img.weight",
+    "blocks.35.cross_attn.norm_k.weight",
+    "blocks.35.cross_attn.norm_k_img.weight",
+    "blocks.35.cross_attn.norm_q.weight",
+    "blocks.35.cross_attn.o.bias",
+    "blocks.35.cross_attn.o.weight",
+    "blocks.35.cross_attn.q.bias",
+    "blocks.35.cross_attn.q.weight",
+    "blocks.35.cross_attn.v.bias",
+    "blocks.35.cross_attn.v.weight",
+    "blocks.35.cross_attn.v_img.bias",
+    "blocks.35.cross_attn.v_img.weight",
+    "blocks.35.ffn.0.bias",
+    "blocks.35.ffn.0.weight",
+    "blocks.35.ffn.2.bias",
+    "blocks.35.ffn.2.weight",
+    "blocks.35.modulation",
+    "blocks.35.norm3.bias",
+    "blocks.35.norm3.weight",
+    "blocks.35.norm_x.bias",
+    "blocks.35.norm_x.weight",
+    "blocks.35.self_attn.k.bias",
+    "blocks.35.self_attn.k.weight",
+    "blocks.35.self_attn.norm_k.weight",
+    "blocks.35.self_attn.norm_q.weight",
+    "blocks.35.self_attn.o.bias",
+    "blocks.35.self_attn.o.weight",
+    "blocks.35.self_attn.q.bias",
+    "blocks.35.self_attn.q.weight",
+    "blocks.35.self_attn.v.bias",
+    "blocks.35.self_attn.v.weight",
+    "blocks.36.audio_cross_attn.kv_linear.bias",
+    "blocks.36.audio_cross_attn.kv_linear.weight",
+    "blocks.36.audio_cross_attn.proj.bias",
+    "blocks.36.audio_cross_attn.proj.weight",
+    "blocks.36.audio_cross_attn.q_linear.bias",
+    "blocks.36.audio_cross_attn.q_linear.weight",
+    "blocks.36.cross_attn.k.bias",
+    "blocks.36.cross_attn.k.weight",
+    "blocks.36.cross_attn.k_img.bias",
+    "blocks.36.cross_attn.k_img.weight",
+    "blocks.36.cross_attn.norm_k.weight",
+    "blocks.36.cross_attn.norm_k_img.weight",
+    "blocks.36.cross_attn.norm_q.weight",
+    "blocks.36.cross_attn.o.bias",
+    "blocks.36.cross_attn.o.weight",
+    "blocks.36.cross_attn.q.bias",
+    "blocks.36.cross_attn.q.weight",
+    "blocks.36.cross_attn.v.bias",
+    "blocks.36.cross_attn.v.weight",
+    "blocks.36.cross_attn.v_img.bias",
+    "blocks.36.cross_attn.v_img.weight",
+    "blocks.36.ffn.0.bias",
+    "blocks.36.ffn.0.weight",
+    "blocks.36.ffn.2.bias",
+    "blocks.36.ffn.2.weight",
+    "blocks.36.modulation",
+    "blocks.36.norm3.bias",
+    "blocks.36.norm3.weight",
+    "blocks.36.norm_x.bias",
+    "blocks.36.norm_x.weight",
+    "blocks.36.self_attn.k.bias",
+    "blocks.36.self_attn.k.weight",
+    "blocks.36.self_attn.norm_k.weight",
+    "blocks.36.self_attn.norm_q.weight",
+    "blocks.36.self_attn.o.bias",
+    "blocks.36.self_attn.o.weight",
+    "blocks.36.self_attn.q.bias",
+    "blocks.36.self_attn.q.weight",
+    "blocks.36.self_attn.v.bias",
+    "blocks.36.self_attn.v.weight",
+    "blocks.37.audio_cross_attn.kv_linear.bias",
+    "blocks.37.audio_cross_attn.kv_linear.weight",
+    "blocks.37.audio_cross_attn.proj.bias",
+    "blocks.37.audio_cross_attn.proj.weight",
+    "blocks.37.audio_cross_attn.q_linear.bias",
+    "blocks.37.audio_cross_attn.q_linear.weight",
+    "blocks.37.cross_attn.k.bias",
+    "blocks.37.cross_attn.k.weight",
+    "blocks.37.cross_attn.k_img.bias",
+    "blocks.37.cross_attn.k_img.weight",
+    "blocks.37.cross_attn.norm_k.weight",
+    "blocks.37.cross_attn.norm_k_img.weight",
+    "blocks.37.cross_attn.norm_q.weight",
+    "blocks.37.cross_attn.o.bias",
+    "blocks.37.cross_attn.o.weight",
+    "blocks.37.cross_attn.q.bias",
+    "blocks.37.cross_attn.q.weight",
+    "blocks.37.cross_attn.v.bias",
+    "blocks.37.cross_attn.v.weight",
+    "blocks.37.cross_attn.v_img.bias",
+    "blocks.37.cross_attn.v_img.weight",
+    "blocks.37.ffn.0.bias",
+    "blocks.37.ffn.0.weight",
+    "blocks.37.ffn.2.bias",
+    "blocks.37.ffn.2.weight",
+    "blocks.37.modulation",
+    "blocks.37.norm3.bias",
+    "blocks.37.norm3.weight",
+    "blocks.37.norm_x.bias",
+    "blocks.37.norm_x.weight",
+    "blocks.37.self_attn.k.bias",
+    "blocks.37.self_attn.k.weight",
+    "blocks.37.self_attn.norm_k.weight",
+    "blocks.37.self_attn.norm_q.weight",
+    "blocks.37.self_attn.o.bias",
+    "blocks.37.self_attn.o.weight",
+    "blocks.37.self_attn.q.bias",
+    "blocks.37.self_attn.q.weight",
+    "blocks.37.self_attn.v.bias",
+    "blocks.37.self_attn.v.weight",
+    "blocks.38.audio_cross_attn.kv_linear.bias",
+    "blocks.38.audio_cross_attn.kv_linear.weight",
+    "blocks.38.audio_cross_attn.proj.bias",
+    "blocks.38.audio_cross_attn.proj.weight",
+    "blocks.38.audio_cross_attn.q_linear.bias",
+    "blocks.38.audio_cross_attn.q_linear.weight",
+    "blocks.38.cross_attn.k.bias",
+    "blocks.38.cross_attn.k.weight",
+    "blocks.38.cross_attn.k_img.bias",
+    "blocks.38.cross_attn.k_img.weight",
+    "blocks.38.cross_attn.norm_k.weight",
+    "blocks.38.cross_attn.norm_k_img.weight",
+    "blocks.38.cross_attn.norm_q.weight",
+    "blocks.38.cross_attn.o.bias",
+    "blocks.38.cross_attn.o.weight",
+    "blocks.38.cross_attn.q.bias",
+    "blocks.38.cross_attn.q.weight",
+    "blocks.38.cross_attn.v.bias",
+    "blocks.38.cross_attn.v.weight",
+    "blocks.38.cross_attn.v_img.bias",
+    "blocks.38.cross_attn.v_img.weight",
+    "blocks.38.ffn.0.bias",
+    "blocks.38.ffn.0.weight",
+    "blocks.38.ffn.2.bias",
+    "blocks.38.ffn.2.weight",
+    "blocks.38.modulation",
+    "blocks.38.norm3.bias",
+    "blocks.38.norm3.weight",
+    "blocks.38.norm_x.bias",
+    "blocks.38.norm_x.weight",
+    "blocks.38.self_attn.k.bias",
+    "blocks.38.self_attn.k.weight",
+    "blocks.38.self_attn.norm_k.weight",
+    "blocks.38.self_attn.norm_q.weight",
+    "blocks.38.self_attn.o.bias",
+    "blocks.38.self_attn.o.weight",
+    "blocks.38.self_attn.q.bias",
+    "blocks.38.self_attn.q.weight",
+    "blocks.38.self_attn.v.bias",
+    "blocks.38.self_attn.v.weight",
+    "blocks.39.audio_cross_attn.kv_linear.bias",
+    "blocks.39.audio_cross_attn.kv_linear.weight",
+    "blocks.39.audio_cross_attn.proj.bias",
+    "blocks.39.audio_cross_attn.proj.weight",
+    "blocks.39.audio_cross_attn.q_linear.bias",
+    "blocks.39.audio_cross_attn.q_linear.weight",
+    "blocks.39.cross_attn.k.bias",
+    "blocks.39.cross_attn.k.weight",
+    "blocks.39.cross_attn.k_img.bias",
+    "blocks.39.cross_attn.k_img.weight",
+    "blocks.39.cross_attn.norm_k.weight",
+    "blocks.39.cross_attn.norm_k_img.weight",
+    "blocks.39.cross_attn.norm_q.weight",
+    "blocks.39.cross_attn.o.bias",
+    "blocks.39.cross_attn.o.weight",
+    "blocks.39.cross_attn.q.bias",
+    "blocks.39.cross_attn.q.weight",
+    "blocks.39.cross_attn.v.bias",
+    "blocks.39.cross_attn.v.weight",
+    "blocks.39.cross_attn.v_img.bias",
+    "blocks.39.cross_attn.v_img.weight",
+    "blocks.39.ffn.0.bias",
+    "blocks.39.ffn.0.weight",
+    "blocks.39.ffn.2.bias",
+    "blocks.39.ffn.2.weight",
+    "blocks.39.modulation",
+    "blocks.39.norm3.bias",
+    "blocks.39.norm3.weight",
+    "blocks.39.norm_x.bias",
+    "blocks.39.norm_x.weight",
+    "blocks.39.self_attn.k.bias",
+    "blocks.39.self_attn.k.weight",
+    "blocks.39.self_attn.norm_k.weight",
+    "blocks.39.self_attn.norm_q.weight",
+    "blocks.39.self_attn.o.bias",
+    "blocks.39.self_attn.o.weight",
+    "blocks.39.self_attn.q.bias",
+    "blocks.39.self_attn.q.weight",
+    "blocks.39.self_attn.v.bias",
+    "blocks.39.self_attn.v.weight",
+    "blocks.4.audio_cross_attn.kv_linear.bias",
+    "blocks.4.audio_cross_attn.kv_linear.weight",
+    "blocks.4.audio_cross_attn.proj.bias",
+    "blocks.4.audio_cross_attn.proj.weight",
+    "blocks.4.audio_cross_attn.q_linear.bias",
+    "blocks.4.audio_cross_attn.q_linear.weight",
+    "blocks.4.cross_attn.k.bias",
+    "blocks.4.cross_attn.k.weight",
+    "blocks.4.cross_attn.k_img.bias",
+    "blocks.4.cross_attn.k_img.weight",
+    "blocks.4.cross_attn.norm_k.weight",
+    "blocks.4.cross_attn.norm_k_img.weight",
+    "blocks.4.cross_attn.norm_q.weight",
+    "blocks.4.cross_attn.o.bias",
+    "blocks.4.cross_attn.o.weight",
+    "blocks.4.cross_attn.q.bias",
+    "blocks.4.cross_attn.q.weight",
+    "blocks.4.cross_attn.v.bias",
+    "blocks.4.cross_attn.v.weight",
+    "blocks.4.cross_attn.v_img.bias",
+    "blocks.4.cross_attn.v_img.weight",
+    "blocks.4.ffn.0.bias",
+    "blocks.4.ffn.0.weight",
+    "blocks.4.ffn.2.bias",
+    "blocks.4.ffn.2.weight",
+    "blocks.4.modulation",
+    "blocks.4.norm3.bias",
+    "blocks.4.norm3.weight",
+    "blocks.4.norm_x.bias",
+    "blocks.4.norm_x.weight",
+    "blocks.4.self_attn.k.bias",
+    "blocks.4.self_attn.k.weight",
+    "blocks.4.self_attn.norm_k.weight",
+    "blocks.4.self_attn.norm_q.weight",
+    "blocks.4.self_attn.o.bias",
+    "blocks.4.self_attn.o.weight",
+    "blocks.4.self_attn.q.bias",
+    "blocks.4.self_attn.q.weight",
+    "blocks.4.self_attn.v.bias",
+    "blocks.4.self_attn.v.weight",
+    "blocks.5.audio_cross_attn.kv_linear.bias",
+    "blocks.5.audio_cross_attn.kv_linear.weight",
+    "blocks.5.audio_cross_attn.proj.bias",
+    "blocks.5.audio_cross_attn.proj.weight",
+    "blocks.5.audio_cross_attn.q_linear.bias",
+    "blocks.5.audio_cross_attn.q_linear.weight",
+    "blocks.5.cross_attn.k.bias",
+    "blocks.5.cross_attn.k.weight",
+    "blocks.5.cross_attn.k_img.bias",
+    "blocks.5.cross_attn.k_img.weight",
+    "blocks.5.cross_attn.norm_k.weight",
+    "blocks.5.cross_attn.norm_k_img.weight",
+    "blocks.5.cross_attn.norm_q.weight",
+    "blocks.5.cross_attn.o.bias",
+    "blocks.5.cross_attn.o.weight",
+    "blocks.5.cross_attn.q.bias",
+    "blocks.5.cross_attn.q.weight",
+    "blocks.5.cross_attn.v.bias",
+    "blocks.5.cross_attn.v.weight",
+    "blocks.5.cross_attn.v_img.bias",
+    "blocks.5.cross_attn.v_img.weight",
+    "blocks.5.ffn.0.bias",
+    "blocks.5.ffn.0.weight",
+    "blocks.5.ffn.2.bias",
+    "blocks.5.ffn.2.weight",
+    "blocks.5.modulation",
+    "blocks.5.norm3.bias",
+    "blocks.5.norm3.weight",
+    "blocks.5.norm_x.bias",
+    "blocks.5.norm_x.weight",
+    "blocks.5.self_attn.k.bias",
+    "blocks.5.self_attn.k.weight",
+    "blocks.5.self_attn.norm_k.weight",
+    "blocks.5.self_attn.norm_q.weight",
+    "blocks.5.self_attn.o.bias",
+    "blocks.5.self_attn.o.weight",
+    "blocks.5.self_attn.q.bias",
+    "blocks.5.self_attn.q.weight",
+    "blocks.5.self_attn.v.bias",
+    "blocks.5.self_attn.v.weight",
+    "blocks.6.audio_cross_attn.kv_linear.bias",
+    "blocks.6.audio_cross_attn.kv_linear.weight",
+    "blocks.6.audio_cross_attn.proj.bias",
+    "blocks.6.audio_cross_attn.proj.weight",
+    "blocks.6.audio_cross_attn.q_linear.bias",
+    "blocks.6.audio_cross_attn.q_linear.weight",
+    "blocks.6.cross_attn.k.bias",
+    "blocks.6.cross_attn.k.weight",
+    "blocks.6.cross_attn.k_img.bias",
+    "blocks.6.cross_attn.k_img.weight",
+    "blocks.6.cross_attn.norm_k.weight",
+    "blocks.6.cross_attn.norm_k_img.weight",
+    "blocks.6.cross_attn.norm_q.weight",
+    "blocks.6.cross_attn.o.bias",
+    "blocks.6.cross_attn.o.weight",
+    "blocks.6.cross_attn.q.bias",
+    "blocks.6.cross_attn.q.weight",
+    "blocks.6.cross_attn.v.bias",
+    "blocks.6.cross_attn.v.weight",
+    "blocks.6.cross_attn.v_img.bias",
+    "blocks.6.cross_attn.v_img.weight",
+    "blocks.6.ffn.0.bias",
+    "blocks.6.ffn.0.weight",
+    "blocks.6.ffn.2.bias",
+    "blocks.6.ffn.2.weight",
+    "blocks.6.modulation",
+    "blocks.6.norm3.bias",
+    "blocks.6.norm3.weight",
+    "blocks.6.norm_x.bias",
+    "blocks.6.norm_x.weight",
+    "blocks.6.self_attn.k.bias",
+    "blocks.6.self_attn.k.weight",
+    "blocks.6.self_attn.norm_k.weight",
+    "blocks.6.self_attn.norm_q.weight",
+    "blocks.6.self_attn.o.bias",
+    "blocks.6.self_attn.o.weight",
+    "blocks.6.self_attn.q.bias",
+    "blocks.6.self_attn.q.weight",
+    "blocks.6.self_attn.v.bias",
+    "blocks.6.self_attn.v.weight",
+    "blocks.7.audio_cross_attn.kv_linear.bias",
+    "blocks.7.audio_cross_attn.kv_linear.weight",
+    "blocks.7.audio_cross_attn.proj.bias",
+    "blocks.7.audio_cross_attn.proj.weight",
+    "blocks.7.audio_cross_attn.q_linear.bias",
+    "blocks.7.audio_cross_attn.q_linear.weight",
+    "blocks.7.cross_attn.k.bias",
+    "blocks.7.cross_attn.k.weight",
+    "blocks.7.cross_attn.k_img.bias",
+    "blocks.7.cross_attn.k_img.weight",
+    "blocks.7.cross_attn.norm_k.weight",
+    "blocks.7.cross_attn.norm_k_img.weight",
+    "blocks.7.cross_attn.norm_q.weight",
+    "blocks.7.cross_attn.o.bias",
+    "blocks.7.cross_attn.o.weight",
+    "blocks.7.cross_attn.q.bias",
+    "blocks.7.cross_attn.q.weight",
+    "blocks.7.cross_attn.v.bias",
+    "blocks.7.cross_attn.v.weight",
+    "blocks.7.cross_attn.v_img.bias",
+    "blocks.7.cross_attn.v_img.weight",
+    "blocks.7.ffn.0.bias",
+    "blocks.7.ffn.0.weight",
+    "blocks.7.ffn.2.bias",
+    "blocks.7.ffn.2.weight",
+    "blocks.7.modulation",
+    "blocks.7.norm3.bias",
+    "blocks.7.norm3.weight",
+    "blocks.7.norm_x.bias",
+    "blocks.7.norm_x.weight",
+    "blocks.7.self_attn.k.bias",
+    "blocks.7.self_attn.k.weight",
+    "blocks.7.self_attn.norm_k.weight",
+    "blocks.7.self_attn.norm_q.weight",
+    "blocks.7.self_attn.o.bias",
+    "blocks.7.self_attn.o.weight",
+    "blocks.7.self_attn.q.bias",
+    "blocks.7.self_attn.q.weight",
+    "blocks.7.self_attn.v.bias",
+    "blocks.7.self_attn.v.weight",
+    "blocks.8.audio_cross_attn.kv_linear.bias",
+    "blocks.8.audio_cross_attn.kv_linear.weight",
+    "blocks.8.audio_cross_attn.proj.bias",
+    "blocks.8.audio_cross_attn.proj.weight",
+    "blocks.8.audio_cross_attn.q_linear.bias",
+    "blocks.8.audio_cross_attn.q_linear.weight",
+    "blocks.8.cross_attn.k.bias",
+    "blocks.8.cross_attn.k.weight",
+    "blocks.8.cross_attn.k_img.bias",
+    "blocks.8.cross_attn.k_img.weight",
+    "blocks.8.cross_attn.norm_k.weight",
+    "blocks.8.cross_attn.norm_k_img.weight",
+    "blocks.8.cross_attn.norm_q.weight",
+    "blocks.8.cross_attn.o.bias",
+    "blocks.8.cross_attn.o.weight",
+    "blocks.8.cross_attn.q.bias",
+    "blocks.8.cross_attn.q.weight",
+    "blocks.8.cross_attn.v.bias",
+    "blocks.8.cross_attn.v.weight",
+    "blocks.8.cross_attn.v_img.bias",
+    "blocks.8.cross_attn.v_img.weight",
+    "blocks.8.ffn.0.bias",
+    "blocks.8.ffn.0.weight",
+    "blocks.8.ffn.2.bias",
+    "blocks.8.ffn.2.weight",
+    "blocks.8.modulation",
+    "blocks.8.norm3.bias",
+    "blocks.8.norm3.weight",
+    "blocks.8.norm_x.bias",
+    "blocks.8.norm_x.weight",
+    "blocks.8.self_attn.k.bias",
+    "blocks.8.self_attn.k.weight",
+    "blocks.8.self_attn.norm_k.weight",
+    "blocks.8.self_attn.norm_q.weight",
+    "blocks.8.self_attn.o.bias",
+    "blocks.8.self_attn.o.weight",
+    "blocks.8.self_attn.q.bias",
+    "blocks.8.self_attn.q.weight",
+    "blocks.8.self_attn.v.bias",
+    "blocks.8.self_attn.v.weight",
+    "blocks.9.audio_cross_attn.kv_linear.bias",
+    "blocks.9.audio_cross_attn.kv_linear.weight",
+    "blocks.9.audio_cross_attn.proj.bias",
+    "blocks.9.audio_cross_attn.proj.weight",
+    "blocks.9.audio_cross_attn.q_linear.bias",
+    "blocks.9.audio_cross_attn.q_linear.weight",
+    "blocks.9.cross_attn.k.bias",
+    "blocks.9.cross_attn.k.weight",
+    "blocks.9.cross_attn.k_img.bias",
+    "blocks.9.cross_attn.k_img.weight",
+    "blocks.9.cross_attn.norm_k.weight",
+    "blocks.9.cross_attn.norm_k_img.weight",
+    "blocks.9.cross_attn.norm_q.weight",
+    "blocks.9.cross_attn.o.bias",
+    "blocks.9.cross_attn.o.weight",
+    "blocks.9.cross_attn.q.bias",
+    "blocks.9.cross_attn.q.weight",
+    "blocks.9.cross_attn.v.bias",
+    "blocks.9.cross_attn.v.weight",
+    "blocks.9.cross_attn.v_img.bias",
+    "blocks.9.cross_attn.v_img.weight",
+    "blocks.9.ffn.0.bias",
+    "blocks.9.ffn.0.weight",
+    "blocks.9.ffn.2.bias",
+    "blocks.9.ffn.2.weight",
+    "blocks.9.modulation",
+    "blocks.9.norm3.bias",
+    "blocks.9.norm3.weight",
+    "blocks.9.norm_x.bias",
+    "blocks.9.norm_x.weight",
+    "blocks.9.self_attn.k.bias",
+    "blocks.9.self_attn.k.weight",
+    "blocks.9.self_attn.norm_k.weight",
+    "blocks.9.self_attn.norm_q.weight",
+    "blocks.9.self_attn.o.bias",
+    "blocks.9.self_attn.o.weight",
+    "blocks.9.self_attn.q.bias",
+    "blocks.9.self_attn.q.weight",
+    "blocks.9.self_attn.v.bias",
+    "blocks.9.self_attn.v.weight",
+    "head.head.bias",
+    "head.head.weight",
+    "head.modulation",
+    "img_emb.proj.0.bias",
+    "img_emb.proj.0.weight",
+    "img_emb.proj.1.bias",
+    "img_emb.proj.1.weight",
+    "img_emb.proj.3.bias",
+    "img_emb.proj.3.weight",
+    "img_emb.proj.4.bias",
+    "img_emb.proj.4.weight",
+    "patch_embedding.bias",
+    "patch_embedding.weight",
+    "tensor_names",
+    "text_embedding.0.bias",
+    "text_embedding.0.weight",
+    "text_embedding.2.bias",
+    "text_embedding.2.weight",
+    "time_embedding.0.bias",
+    "time_embedding.0.weight",
+    "time_embedding.2.bias",
+    "time_embedding.2.weight",
+    "time_projection.1.bias",
+    "time_projection.1.weight"
+  ],
+  "wan_config_json": {
+    "_class_name": "WanModel",
+    "_diffusers_version": "0.34.0",
+    "audio_window": 5,
+    "context_tokens": 32,
+    "dim": 5120,
+    "eps": 1e-06,
+    "ffn_dim": 13824,
+    "freq_dim": 256,
+    "in_dim": 36,
+    "intermediate_dim": 512,
+    "model_type": "i2v",
+    "norm_input_visual": true,
+    "norm_output_audio": true,
+    "num_heads": 40,
+    "num_layers": 40,
+    "out_dim": 16,
+    "output_dim": 768,
+    "text_len": 512,
+    "vae_scale": 4,
+    "weight_init": false
+  },
+  "extra": {
+    "int8_weight_group_size": 32,
+    "config_version": 2
+  },
+  "reload_note": "Reload: instantiate WanModel, apply same torchao quant config, unflatten safetensors via torchao.prototype.safetensors, load_state_dict."
+}