Test on GGML_OP_FUSED_RMS_NORM

Nexesenex · Nexesenex · commit 8bf8519638c4 · 2024-11-02T19:37:23.000+01:00
Similar to LCPP ggml-org#10111
diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
@@ -3133,6 +3133,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_FUSED_RMS_NORM:
+            return ggml_is_contiguous(op->src[0]) && op->ne[0] % WARP_SIZE == 0;
+            break;
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
