CUDA: use mul_mat_q kernels by default (#2683)

diff --git a/common/common.cpp b/common/common.cpp
index 1623ba21f461ae4939e64bfb9ac1fe876a33d6ff..2a83b379ec4f56a117a950a34518f8a277f63ccb 100644 (file)
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -387,11 +387,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 #else
             fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
 #endif // GGML_USE_CUBLAS
-        } else if (arg == "--mul-mat-q" || arg == "-mmq") {
+        } else if (arg == "--no-mul-mat-q" || arg == "-nommq") {
 #ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = true;
+            params.mul_mat_q = false;
 #else
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n");
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
 #endif // GGML_USE_CUBLAS
         } else if (arg == "--low-vram" || arg == "-lv") {
 #ifdef GGML_USE_CUBLAS
@@ -599,11 +599,11 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stdout, "                        number of layers to store in VRAM\n");
     fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
-    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
-    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
-    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
-    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
+    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n");
+    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
+    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif
     fprintf(stdout, "  --mtest               compute maximum memory usage\n");
     fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");

diff --git a/common/common.h b/common/common.h
index c50a6edfc4124728e7facde01ef91221aba05dec..18fd951ead9df6581ce4a3fd993b4ed52eff9cf1 100644 (file)
--- a/common/common.h
+++ b/common/common.h
@@ -68,7 +68,7 @@ struct gpt_params {
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
     bool low_vram          = false; // if true, reduce VRAM usage at the cost of performance
-    bool mul_mat_q         = false; // if true, use experimental mul_mat_q kernels
+    bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs

diff --git a/examples/server/server.cpp b/examples/server/server.cpp
index 39fdf3307c93d8bd8ee97f407ffb04f26e9a71f4..e5bc52cd00624b12f55092f1972b1b3c45472e41 100644 (file)
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -671,12 +671,11 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     fprintf(stdout, "                        number of layers to store in VRAM\n");
     fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
     fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
     fprintf(stdout, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
-    fprintf(stdout, "  -mmq, --mul-mat-q     use experimental mul_mat_q CUDA kernels instead of cuBLAS. TEMP!!!\n" );
-    fprintf(stdout, "                        Reduces VRAM usage by 700/970/1430 MiB for 7b/13b/33b but prompt processing speed\n" );
-    fprintf(stdout, "                        is still suboptimal, especially q2_K, q3_K, q5_K, and q6_K.\n" );
+    fprintf(stdout, "  -nommq, --no-mul-mat-q\n");
+    fprintf(stdout, "                        use cuBLAS instead of custom mul_mat_q CUDA kernels.\n");
+    fprintf(stdout, "                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif
     fprintf(stdout, "  -m FNAME, --model FNAME\n");
     fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
@@ -867,12 +866,12 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n", {});
 #endif // GGML_USE_CUBLAS
         }
-        else if (arg == "--mul-mat-q" || arg == "-mmq")
+        else if (arg == "--no-mul-mat-q" || arg == "-nommq")
         {
 #ifdef GGML_USE_CUBLAS
-            params.mul_mat_q = true;
+            params.mul_mat_q = false;
 #else
-            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. It is not possible to use mul_mat_q kernels.\n", {});
+            LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n", {});
 #endif // GGML_USE_CUBLAS
         }
         else if (arg == "--main-gpu" || arg == "-mg")

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 4fe378c210030fcbbcd8834777a6ac05e9bcbbdb..70a950bb58b9b0b09356d8e9a226a9f339b91067 100644 (file)
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -287,7 +287,7 @@ static int g_device_count = -1;
 static int g_main_device = 0;
 static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES];
 static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
-static bool g_mul_mat_q = false;
+static bool g_mul_mat_q = true;
 
 static void * g_scratch_buffer = nullptr;
 static size_t g_scratch_size = 1024*1024*1024; // 1 GB by default