minor tweak

Author: bssrdf

ggml/src/ggml-cuda/cpy.cu

@@ -394,18 +394,18 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
                 ne01n = ne01*ne02;
                 ne02n = 1;
             }
-            ne00 = ne00n ;
+            ne00 = ne00n;
             ne01 = ne01n;
             ne02 = ne02n;
             can_be_transposed = true;
         }
-        if ((nb02 == (int64_t)ggml_element_size(src0) &&
+        if ((nb02 == (int64_t)ggml_element_size(src0) && nb00 <= nb01 &&
             nb01 == ne02 * ne00 * (int64_t)ggml_element_size(src0))) {
-            GGML_ASSERT(nb00 <= nb01);
+            // GGML_ASSERT(nb00 <= nb01);
             ne00n = ne00*ne01;
             ne01n = ne02;
             ne02n = 1; // not used
-            ne00 = ne00n ;
+            ne00 = ne00n;
             ne01 = ne01n;
             ne02 = ne02n;
             can_be_transposed = true;

tests/test-backend-ops.cpp

@@ -7928,6 +7928,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768, 1024, 256, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {768, 1024, 256, 1}, {0, 0, 0, 0}, {0, 0, 0, 0}, true));
 
+    // sd.cpp cases
     test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {4352, 1, 9216, 1}, {1, 2, 0, 3}, {0, 0, 0, 0}));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {4352, 1, 9216, 1}, {1, 2, 0, 3}, {0, 0, 0, 0}));
     test_cases.emplace_back(new test_cpy(GGML_TYPE_BF16, GGML_TYPE_BF16, {21504, 4352, 1, 1}, {2, 0, 1, 3}, {0, 0, 0, 0}));