@@ -128,6 +128,11 @@ func New(model string, opts api.Options) (*LLM, error) {
C.llama_backend_init(C.bool(llm.UseNUMA))
+ // TODO: GQA == 8 suggests 70B model which doesn't support metal
+ if llm.NumGQA == 8 {
+ llm.NumGPU = 0
+ }
+
params := C.llama_context_default_params()
params.seed = C.uint(llm.Seed)
params.n_ctx = C.int(llm.NumCtx)