From: Piotr Wilkin (ilintar) <redacted>
Date: Wed, 19 Nov 2025 09:36:33 +0000 (+0100)
Subject: Fix too relaxed check on CUDA "fast copy" (can_be_transposed) condition (llama/17332)
X-Git-Tag: upstream/1.8.3~271
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=746cbed20adadcc199e9e1c50173411276e6a360;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp

Fix too relaxed check on CUDA "fast copy" (can_be_transposed) condition (llama/17332)

* Fix too relaxed check on CUDA "fast copy" (can_be_transposed) condition

* Argh.

* Making CISC happy ;)

* Integrate CONT tests

* Use loopy loop

* Skip new tests for (B)F16 for now.
---

diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu
index 50612237..c1afde96 100644
--- a/ggml/src/ggml-cuda/cpy.cu
+++ b/ggml/src/ggml-cuda/cpy.cu
@@ -384,7 +384,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     char * src1_ddc = (char *) src1->data;
 
     const bool contiguous_srcs = ggml_is_contiguous(src0) && ggml_is_contiguous(src1);
-    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) && src0->ne[3] == 1;
+    const bool can_be_transposed = nb01 == (int64_t)ggml_element_size(src0) &&
+        src0->ne[3] == 1 && nb02 == ne00 * ne01 * (int64_t)ggml_element_size(src0);
 
     if (src0->type == src1->type && contiguous_srcs) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));