ggml : add GGML_CUDA_USE_GRAPHS option, restore GGML_CUDA_FORCE_CUBLAS (cmake) (...

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 18297834e9225d386ea4a70f2ca59d1939161213..7a7197282f9467a594bd7cca539c3f01db68a347 100644 (file)
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -80,6 +80,7 @@ set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED})
 set(GGML_ALL_WARNINGS       ${LLAMA_ALL_WARNINGS})
 set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})
 set(GGML_LLAMAFILE          ON)
+set(GGML_CUDA_USE_GRAPHS    ON)
 
 # transition helpers
 function (llama_option_depr TYPE OLD NEW)

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index f3763f7eb9fef117888eefac70c49877ec330d7e..0d0d52d57597100250e2caceb106865ed62522a2 100644 (file)
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -109,6 +109,7 @@ option(GGML_LLAMAFILE                       "ggml: use ggml SGEMM"
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
+option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
 set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
 set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
 option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
@@ -119,6 +120,7 @@ set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
 option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
+option(GGML_CUDA_USE_GRAPHS                 "ggml: use CUDA graphs (llama.cpp only)"          OFF)
 
 option(GGML_CURL                            "ggml: use libcurl to download model from an URL" OFF)
 option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)

diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index ba341d3749050bab42c782777ea87a40e9e7c782..d0f4097d8cd0c84935f878a27826fe00e222e844 100644 (file)
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -295,12 +295,15 @@ if (GGML_CUDA)
 
         list(APPEND GGML_CDEF_PUBLIC GGML_USE_CUDA)
 
-        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
         add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
         add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
         add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
+        if (GGML_CUDA_USE_GRAPHS)
+            add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+        endif()
+
         if (GGML_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()