CUDA: fix MMQ for non-contiguous src0, add tests (ggml-org#10021)

* CUDA: fix MMQ for non-contiguous src0, add tests

* revise test code

Author: JohannesGaessler

ggml/src/ggml-cuda.cu

@@ -1538,20 +1538,13 @@ static void ggml_cuda_op_mul_mat(
             const size_t nbytes_data    = ggml_nbytes(src0);
             const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding);
-            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream));
-        }
-
-        // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
-        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
-            const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
-            const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
         }
 
-        // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
+        // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared:
         if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
-            const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
-            const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
+            const size_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
+            const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
             CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
         }
 

ggml/src/ggml.c

@@ -4132,7 +4132,7 @@ GGML_CALL int64_t ggml_blck_size(enum ggml_type type) {
 
 GGML_CALL size_t ggml_nbytes(const struct ggml_tensor * tensor) {
     size_t nbytes;
-    size_t blck_size = ggml_blck_size(tensor->type);
+    const size_t blck_size = ggml_blck_size(tensor->type);
     if (blck_size == 1) {
         nbytes = ggml_type_size(tensor->type);
         for (int i = 0; i < GGML_MAX_DIMS; ++i) {