root
/
ollama
mirror of https://github.com/ollama/ollama.git
1
0
Fork 0
Code Issues Actions Packages Projects Releases Wiki Activity

ggml: Use assigned layers when reporting loading stats 

Reporting params.NumGPULayers can be misleading because it is the
requested number of layers, not the actual number that is loaded.
While they are often the same, there are cases where they might mismatch,
such as if the GPU backend is missing.
This commit is contained in:
Jesse Gross 2025-07-07 13:10:14 -07:00 committed by Jesse Gross
parent 9a43994c45
commit acef9b4c1b
1 changed files with 13 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

20
ml/backend/ggml/ggml.go
View File

@ -356,23 +356,25 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
}
// Mimic llama runner logs summarizing layers and memory
slog.Info(fmt.Sprintf("offloading %d repeating layers to GPU", max(0, params.NumGPULayers-1)))
gpuLayers := 0
for _, layer := range layers {
if C.ggml_backend_dev_type(layer.d) == C.GGML_BACKEND_DEVICE_TYPE_GPU {
gpuLayers++
}
}
slog.Info(fmt.Sprintf("offloading %d repeating layers to GPU", gpuLayers))
switch C.ggml_backend_dev_type(output.d) {
case 0: // CPU
case C.GGML_BACKEND_DEVICE_TYPE_CPU:
slog.Info("offloading output layer to CPU")
case 1: // GPU
case C.GGML_BACKEND_DEVICE_TYPE_GPU:
slog.Info("offloading output layer to GPU")
gpuLayers++
case 2: // ACCEL
case C.GGML_BACKEND_DEVICE_TYPE_ACCEL:
slog.Info("offloading output layer to ACCEL")
}
for _, layer := range layers {
if C.ggml_backend_dev_type(layer.d) == 1 {
gpuLayers++
}
}
slog.Info(fmt.Sprintf("offloaded %d/%d layers to GPU", gpuLayers, len(layers)+1))
for bs := range maps.Values(bbs) {
slog.Info("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)), "size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
}