stable-diffusion.cpp-vulkan: fix LoKr LoRAs on Krea2main

Runtime LoKr assumed 2D linear inputs ([q, batch]). Krea2 feeds
[q, tokens, batch], so every LoKr LoRA aborted on a reshape assert
at the first sampling step. Flatten the extra dims before the
Kronecker split and restore the shape on output. Not upstream yet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 27 insertions, 0 deletions

diff --git a/stable-diffusion.cpp-vulkan/patches/lokr-nd-input.patch b/stable-diffusion.cpp-vulkan/patches/lokr-nd-input.patch
new file mode 100644
index 0000000..6dd2d01
--- /dev/null
+++ b/stable-diffusion.cpp-vulkan/patches/lokr-nd-input.patch
@@ -0,0 +1,27 @@
+diff --git a/src/model/adapter/lora_ops.cpp b/src/model/adapter/lora_ops.cpp
+index 7dacfd1..686b8d3 100644
+--- a/src/model/adapter/lora_ops.cpp
++++ b/src/model/adapter/lora_ops.cpp
+@@ -65,6 +65,13 @@ ggml_tensor* ggml_ext_lokr_forward(
+     ggml_tensor* hb;
+ 
+     if (!is_conv) {
++        // Linear inputs may carry extra dims ([q, tokens, batch, ...], e.g. Krea2):
++        // flatten them into one batch dim and restore the shape on output.
++        ggml_tensor* h_in = h;
++        if (!ggml_is_contiguous(h)) {
++            h = ggml_cont(ctx, h);
++        }
++        h                  = ggml_reshape_2d(ctx, h, h->ne[0], ggml_nelements(h) / h->ne[0]);
+         int batch          = (int)h->ne[1];
+         int merge_batch_uq = batch;
+         int merge_batch_vp = batch;
+@@ -118,7 +125,7 @@ ggml_tensor* ggml_ext_lokr_forward(
+         }
+ 
+         ggml_tensor* hc  = ggml_transpose(ctx, hc_t);
+-        ggml_tensor* out = ggml_reshape_2d(ctx, ggml_cont(ctx, hc), up * vp, batch);
++        ggml_tensor* out = ggml_reshape_4d(ctx, ggml_cont(ctx, hc), up * vp, h_in->ne[1], h_in->ne[2], h_in->ne[3]);
+         return ggml_ext_scale(ctx, out, scale);
+     } else {
+         int batch = (int)h->ne[3];