From: bssrdf <redacted>
Date: Fri, 7 Nov 2025 22:41:58 +0000 (-0500)
Subject: CUDA: properly handle nb00=nb02 case for cpy (llama/17081)
X-Git-Tag: upstream/1.8.3~355
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=547724b0a5d43540255647788b59ca2e78bba27d;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp

CUDA: properly handle nb00=nb02 case for cpy (llama/17081)
---

diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
index 1dba60eb..50612237 100644
--- a/ggml/src/ggml-cuda/cpy.cu
+++ b/ggml/src/ggml-cuda/cpy.cu
@@ -198,7 +198,7 @@ static void ggml_cpy_flt_cuda(
     if (transposed) {
         GGML_ASSERT(ne == ne00*ne01*ne02);  // ne[3] is 1 assumed
         int ne00n, ne01n, ne02n;
-        if (nb00 < nb02) {
+        if (nb00 <= nb02) { // most likely safe to handle nb00 = nb02 case here
             ne00n = ne00;
             ne01n = ne01;
             ne02n = ne02;
@@ -206,8 +206,6 @@ static void ggml_cpy_flt_cuda(
             ne00n = ne00;
             ne01n = ne01*ne02;
             ne02n = 1;
-        } else {
-            GGML_ASSERT(false);
         }
 
         dim3 dimGrid( (ne01n + CUDA_CPY_TILE_DIM_2D - 1) / CUDA_CPY_TILE_DIM_2D,