mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 29 Packages Projects Releases Wiki Activity

llama: offload output layer to GPU first (#18148)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-12-18 08:12:18 +01:00
committed by GitHub
parent 9cff4cc554
commit 57c1e05643
2 changed files with 27 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model.cpp
+7 -5
View File
@@ -2378,10 +2378,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
if (cpu_dev == nullptr) {
throw std::runtime_error(format("%s: no CPU backend found", __func__));
}
const int i_gpu_start = std::max((int) hparams.n_layer - n_gpu_layers, (int) 0);
const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, (int)n_layer + 1);
const int i_gpu_start = std::max(int(hparams.n_layer) + 1 - n_gpu_layers, 0);
const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, int(n_layer) + 1);
auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
const bool is_swa = il < (int) hparams.n_layer && hparams.is_swa(il);
const bool is_swa = il < int(hparams.n_layer) && hparams.is_swa(il);
if (il < i_gpu_start || (il - i_gpu_start) >= act_gpu_layers) {
LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s, is_swa = %d\n", il, ggml_backend_dev_name(cpu_dev), is_swa);
return {cpu_dev, &pimpl->cpu_buft_list};
@@ -6693,10 +6693,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
if (llama_supports_gpu_offload()) {
const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_gpu);
if (n_gpu_layers > (int) hparams.n_layer) {
int n_repeating = n_gpu;
if (n_repeating > 0) {
LLAMA_LOG_INFO("%s: offloading output layer to GPU\n", __func__);
n_repeating--;
}
LLAMA_LOG_INFO("%s: offloading %d repeating layers to GPU\n", __func__, n_repeating);
const int max_backend_supported_layers = hparams.n_layer + 1;
const int max_offloadable_layers = hparams.n_layer + 1;