From de7fa0a3c6a2e1b4cd9f22eb8d6bf5b12dbdb63b Mon Sep 17 00:00:00 2001
From: bri-prism <288398250+bri-prism@users.noreply.github.com>
Date: Thu, 8 Oct 2026 12:56:54 -0700
Subject: [PATCH] Musa FWHT fix (#30167)

---
 ggml/src/ggml-cuda/fwht.cu | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/ggml/src/ggml-cuda/fwht.cu b/ggml/src/ggml-cuda/fwht.cu
index d69e2ed81..78768eace 100644
--- a/ggml/src/ggml-cuda/fwht.cu
+++ b/ggml/src/ggml-cuda/fwht.cu
@@ -200,9 +200,12 @@ static bool ggml_cuda_op_fwht_impl(ggml_backend_cuda_context & ctx, const ggml_t
             case 4096:
                 ggml_cuda_kernel_launch(fwht_cuda_block<4096, nt, T>, launch_params_w, src_d, dst_d, rows, scale);
                 return true;
+#if !defined(GGML_USE_MUSA)
+            // 32 KB of shared memory, above the MUSA limit; falls back there
             case 8192:
                 ggml_cuda_kernel_launch(fwht_cuda_block<8192, nt, T>, launch_params_w, src_d, dst_d, rows, scale);
                 return true;
+#endif // !defined(GGML_USE_MUSA)
             default:
                 return false;
         }