diff --git a/common/speculative.cpp b/common/speculative.cpp index 854a97b85a..2db381d580 100644 --- a/common/speculative.cpp +++ b/common/speculative.cpp @@ -2467,7 +2467,10 @@ common_params common_base_params_to_speculative(const common_params & params) { result.pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; if (has_draft) { - result.devices = params_spec.devices; + // default to global devices value + if (!params_spec.devices.empty()) { + result.devices = params_spec.devices; + } result.model = params_spec.mparams; result.n_gpu_layers = params_spec.n_gpu_layers; result.tensor_buft_overrides = params_spec.tensor_buft_overrides;