From: Georgi Gerganov <redacted>
Date: Wed, 28 Jan 2026 07:15:27 +0000 (+0200)
Subject: cuda : fix "V is K view" check for non-unified KV cache (llama/19145)
X-Git-Tag: upstream/1.8.3+155~101
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=7fb0f823de760f1be8e6d951516d3adc95b8bb61;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp

cuda : fix "V is K view" check for non-unified KV cache (llama/19145)
---

diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index 3d7daccf..b6a7460d 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -789,7 +789,7 @@ void launch_fattn(
     const ggml_tensor * K = dst->src[1];
     const ggml_tensor * V = dst->src[2];
 
-    const bool V_is_K_view = V->view_src && V->view_offs == 0 && (V->view_src == K || V->view_src == K->view_src);
+    const bool V_is_K_view = V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs));
 
     const ggml_tensor * mask  = dst->src[3];
     const ggml_tensor * sinks = dst->src[4];
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index fe18ff6c..195904ee 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -310,7 +310,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         }
     }
 
-    const bool V_is_K_view = V->view_src && V->view_offs == 0 && (V->view_src == K || V->view_src == K->view_src);
+    const bool V_is_K_view = V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs));
 
     const int cc = ggml_cuda_info().devices[device].cc;