CUDA: mmq CLI option, fixed mmq build issues (#2453)

1 files changed, 8 insertions, 2 deletions

diff --git a/llama.cpp b/llama.cpp
index 50da427..d427054 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -901,6 +901,7 @@ struct llama_context_params llama_context_default_params() {
         /*.progress_callback           =*/ nullptr,
         /*.progress_callback_user_data =*/ nullptr,
         /*.low_vram                    =*/ false,
+        /*.mul_mat_q                   =*/ false,
         /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
         /*.vocab_only                  =*/ false,
@@ -1028,6 +1029,7 @@ static void llama_model_load_internal(
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
+        const bool mul_mat_q,
         float rope_freq_base,
         float rope_freq_scale,
         bool low_vram,
@@ -1156,9 +1158,11 @@ static void llama_model_load_internal(
     }
 
     (void) main_gpu;
+    (void) mul_mat_q;
 #if defined(GGML_USE_CUBLAS)
     fprintf(stderr, "%s: using CUDA for GPU acceleration\n", __func__);
     ggml_cuda_set_main_device(main_gpu);
+    ggml_cuda_set_mul_mat_q(mul_mat_q);
 #define LLAMA_BACKEND_OFFLOAD       GGML_BACKEND_GPU
 #define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT
 #elif defined(GGML_USE_CLBLAST)
@@ -1367,6 +1371,7 @@ static bool llama_model_load(
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
+        const bool mul_mat_q,
         float rope_freq_base,
         float rope_freq_scale,
         bool low_vram,
@@ -1377,7 +1382,8 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gqa, rms_norm_eps, n_gpu_layers, main_gpu, tensor_split, rope_freq_base, rope_freq_scale, low_vram, memory_type,
+        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gqa, rms_norm_eps, n_gpu_layers,
+                                  main_gpu, tensor_split, mul_mat_q, rope_freq_base, rope_freq_scale, low_vram, memory_type,
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
@@ -3192,7 +3198,7 @@ struct llama_model * llama_load_model_from_file(
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
     if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gqa, params.rms_norm_eps, params.n_gpu_layers,
-                params.main_gpu, params.tensor_split, params.rope_freq_base, params.rope_freq_scale,params.low_vram,
+                params.main_gpu, params.tensor_split, params.mul_mat_q, params.rope_freq_base, params.rope_freq_scale,params.low_vram,
                 memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback,
                 params.progress_callback_user_data)) {
         delete model;