diff options
| author | Danilo M. <danix@danix.xyz> | 2026-09-28 21:21:36 +0200 |
|---|---|---|
| committer | Danilo M. <danix@danix.xyz> | 2026-09-28 21:21:36 +0200 |
| commit | 5c08d1f69d6d110172f393ae0f9765e0fa46e5a3 (patch) | |
| tree | e02555e09c746c274a4c9ba878524e560afbf6a9 /stable-diffusion.cpp-vulkan | |
| parent | 293b629d7db2f1c24fac76c0c89102bdf500b7f6 (diff) | |
| download | my-slackbuilds-5c08d1f69d6d110172f393ae0f9765e0fa46e5a3.tar.gz my-slackbuilds-5c08d1f69d6d110172f393ae0f9765e0fa46e5a3.zip | |
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>
Diffstat (limited to 'stable-diffusion.cpp-vulkan')
| -rw-r--r-- | stable-diffusion.cpp-vulkan/patches/lokr-nd-input.patch | 27 | ||||
| -rw-r--r-- | stable-diffusion.cpp-vulkan/stable-diffusion.cpp-vulkan.SlackBuild | 4 |
2 files changed, 31 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]; diff --git a/stable-diffusion.cpp-vulkan/stable-diffusion.cpp-vulkan.SlackBuild b/stable-diffusion.cpp-vulkan/stable-diffusion.cpp-vulkan.SlackBuild index 450e3a8..13c16fd 100644 --- a/stable-diffusion.cpp-vulkan/stable-diffusion.cpp-vulkan.SlackBuild +++ b/stable-diffusion.cpp-vulkan/stable-diffusion.cpp-vulkan.SlackBuild @@ -84,6 +84,10 @@ find -L . \ \( -perm 666 -o -perm 664 -o -perm 640 -o -perm 600 -o -perm 444 \ -o -perm 440 -o -perm 400 \) -exec chmod 644 {} \; +# Runtime LoKr assumed 2D linear inputs; Krea2 feeds [q, tokens, batch] +# and every LoKr LoRA aborted on a reshape assert. Not upstream yet. +patch -p1 < $CWD/patches/lokr-nd-input.patch + # Build the server web UI offline. Every npm tarball listed in the .info # (the linux-x64 closure of the pnpm lockfile minus typescript/vue-tsc, # which only type-check) is unpacked into a flat node_modules; the lockfile |