fix permute ctx for videeo decoding

Co-authored-by: Ollin Boer Bohan <[email protected]>

Author: stduhpf

stable-diffusion.cpp

@@ -2312,10 +2312,6 @@ class StableDiffusionGGML {
             first_stage_model->free_compute_buffer();
             process_vae_output_tensor(result);
         } else {
-            if (sd_version_is_wan(version)) {
-                x = ggml_permute(work_ctx, x, 0, 1, 3, 2);
-            }
-
             if (vae_tiling_params.enabled && !decode_video) {
                 // split latent in 64x64 tiles and compute in several steps
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
@@ -2326,7 +2322,6 @@ class StableDiffusionGGML {
                 tae_first_stage->compute(n_threads, x, true, &result);
             }
             tae_first_stage->free_compute_buffer();
-
         }
 
         int64_t t1 = ggml_time_ms();

tae.hpp

@@ -372,6 +372,10 @@ class TAEHV : public GGMLBlock {
 
     struct ggml_tensor* decode(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
         auto decoder = std::dynamic_pointer_cast<TinyVideoDecoder>(blocks["decoder"]);
+        if (sd_version_is_wan(version)) {
+            // (W, H, C, T) -> (W, H, T, C)
+            z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 0, 1, 3, 2));
+        }
         auto result  = decoder->forward(ctx, z);
         if (sd_version_is_wan(version)) {
             // (W, H, C, T) -> (W, H, T, C)