vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Prevent offloading of more than 33 layers

Browse source
This commit is contained in:
Galunid 2023-10-24 15:05:58 +02:00
parent 27d0c11897
commit 51b3b56c08
1 changed files with 8 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

8
llama.cpp
View file

@ -3020,6 +3020,14 @@ static void llm_load_tensors(
ggml_backend_type backend_norm; ggml_backend_type backend_norm;
ggml_backend_type backend_output; ggml_backend_type backend_output;
// Don't allow for offloading of more than 33 layers.
// Offloading 34 layers causes model to respond with letter 'E'
// Offloading 35 layers doesn't work because of missing cuda implementation for rope:
// GGML_ASSERT: ggml-cuda.cu:6402: ne00 == n_dims && "ne00 != n_dims is not implemented for CUDA yet"
if (n_gpu_layers > 33) {
n_gpu_layers = 33;
}
if (n_gpu_layers > int(n_layer)) { if (n_gpu_layers > int(n_layer)) {
// norm is not performance relevant on its own but keeping it in VRAM reduces data copying // norm is not performance relevant on its own but keeping it in VRAM reduces data copying
// on Windows however this is detrimental unless everything is on the GPU // on Windows however this is detrimental unless everything is on the GPU