ggml-blas: add note

Signed-off-by: Aaron Teo <[email protected]>

Author: taronaeo

ggml/src/ggml-blas/ggml-blas.cpp

@@ -513,13 +513,14 @@ static bool ggml_backend_blas_device_supports_op(ggml_backend_dev_t dev, const g
             // TODO: find the optimal value
             const int64_t min_batch = 32;
 
-            return ggml_is_contiguous(src0) &&
-                   ggml_is_contiguous(src1) &&
-                   src0->view_src == nullptr &&
-                   src1->view_src == nullptr &&
-                   src1->type == GGML_TYPE_F32 &&
-                   (ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch) &&
-                   (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
+            return ggml_is_contiguous(src0)
+                   && ggml_is_contiguous(src1)
+                   && src1->type == GGML_TYPE_F32
+                   // NOTE: llama-bench creates views that somehow does not go through init_tensor
+                   //       this prevents the uninitialized views from being used in BLAS
+                   && src0->view_src == nullptr && src1->view_src == nullptr
+                   && (ne0 >= min_batch && ne1 >= min_batch && ne10 >= min_batch)
+                   && (src0->type == GGML_TYPE_F32 || ggml_get_type_traits(src0->type)->to_float != NULL);
         }
 
         default: