diff options
| -rw-r--r-- | llama.cpp | 15 |
1 files changed, 12 insertions, 3 deletions
@@ -1156,6 +1156,7 @@ static void llama_model_load_internal( } } #endif // GGML_USE_CUBLAS + #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer)); @@ -1164,6 +1165,10 @@ static void llama_model_load_internal( fprintf(stderr, "%s: offloading non-repeating layers to GPU\n", __func__); } size_t vram_kv_cache = 0; + +#ifdef GGML_USE_CUBLAS + const int max_backend_supported_layers = hparams.n_layer + 3; + const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3; if (n_gpu_layers > (int) hparams.n_layer + 1) { if (low_vram) { fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__); @@ -1180,14 +1185,18 @@ static void llama_model_load_internal( vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2; } } - const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3; +#elif defined(GGML_USE_CLBLAST) + const int max_backend_supported_layers = hparams.n_layer + 1; + const int max_offloadable_layers = hparams.n_layer + 1; +#endif // GGML_USE_CUBLAS + fprintf(stderr, "%s: offloaded %d/%d layers to GPU\n", - __func__, std::min(n_gpu_layers, max_offloadable_layers), hparams.n_layer + 3); + __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers); fprintf(stderr, "%s: total VRAM used: %zu MB\n", __func__, (vram_weights + vram_scratch + vram_kv_cache + MB - 1) / MB); // round up #else (void) n_gpu_layers; -#endif +#endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) } // populate `tensors_by_name` |