diff --git a/src/models/limite.cpp b/src/models/limite.cpp index 2bd45561a..887bfef63 100644 --- a/src/models/limite.cpp +++ b/src/models/limite.cpp @@ -18,6 +18,12 @@ void llama_model_limite::load_arch_hparams(llama_model_loader & ml) { std::fill(hparams.mudd_tap_idx.begin(), hparams.mudd_tap_idx.end(), -1); ml.get_key(LLM_KV_MUDD_TAP_COUNT, hparams.n_mudd_taps, false); if (hparams.n_mudd_taps > 0) { + // reject metadata where n_layer * n_mudd_taps would index past mudd_tap_idx + if (hparams.n_mudd_taps > LLAMA_MAX_LAYERS || (uint64_t) hparams.n_layer() * hparams.n_mudd_taps > LLAMA_MAX_LAYERS) { + throw std::runtime_error(format( + "invalid limite mudd config: n_layer (%u) * n_mudd_taps (%u) exceeds LLAMA_MAX_LAYERS (%d)", + hparams.n_layer(), hparams.n_mudd_taps, LLAMA_MAX_LAYERS)); + } ml.get_key(LLM_KV_MUDD_FEED_FORWARD_LENGTH, hparams.n_mudd_ff); ml.get_arr(LLM_KV_MUDD_TAP_INDICES, hparams.mudd_tap_idx); }