Allow compiling ggml-cuda without mmq or flash attention

Author: milot-mirdita

ggml/CMakeLists.txt

@@ -149,6 +149,7 @@ set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 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                         "ggml: compile with FlashAttention"               ON)
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 

ggml/src/ggml-cuda/CMakeLists.txt

@@ -28,24 +28,35 @@ if (CUDAToolkit_FOUND)
     list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
 
     file(GLOB   GGML_SOURCES_CUDA "*.cu")
-    file(GLOB   SRCS "template-instances/fattn-wmma*.cu")
-    list(APPEND GGML_SOURCES_CUDA ${SRCS})
-    file(GLOB   SRCS "template-instances/mmq*.cu")
-    list(APPEND GGML_SOURCES_CUDA ${SRCS})
-
-    if (GGML_CUDA_FA_ALL_QUANTS)
-        file(GLOB   SRCS "template-instances/fattn-vec*.cu")
+    if (GGML_CUDA_FA)
+        file(GLOB SRCS "template-instances/fattn-wmma*.cu")
         list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
     else()
-        file(GLOB   SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        file(GLOB   SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        file(GLOB   SRCS "template-instances/fattn-vec*f16-f16.cu")
+        list(FILTER GGML_SOURCES_CUDA EXCLUDE REGEX ".*fattn.*")
+        list(FILTER GGML_HEADERS_CUDA EXCLUDE REGEX ".*fattn.*")
+        # message(FATAL_ERROR ${GGML_SOURCES_CUDA})
+    endif()
+    if (NOT GGML_CUDA_FORCE_CUBLAS)
+        file(GLOB   SRCS "template-instances/mmq*.cu")
         list(APPEND GGML_SOURCES_CUDA ${SRCS})
     endif()
 
+    if (GGML_CUDA_FA)
+        add_compile_definitions(GGML_CUDA_FA)
+        if (GGML_CUDA_FA_ALL_QUANTS)
+            file(GLOB   SRCS "template-instances/fattn-vec*.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+        else()
+            file(GLOB   SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            file(GLOB   SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            file(GLOB   SRCS "template-instances/fattn-vec*f16-f16.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        endif()
+    endif()
+
     ggml_add_backend_library(ggml-cuda
                              ${GGML_HEADERS_CUDA}
                              ${GGML_SOURCES_CUDA}

ggml/src/ggml-cuda/common.cuh

@@ -151,6 +151,10 @@ typedef float2 dfloat2;
 #define FLASH_ATTN_AVAILABLE
 #endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
 
+#if !defined(GGML_CUDA_FA)
+#undef FLASH_ATTN_AVAILABLE
+#endif
+
 static constexpr bool fast_fp16_available(const int cc) {
     return cc >= GGML_CUDA_CC_PASCAL && cc != 610;
 }

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -16,7 +16,9 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/cross-entropy-loss.cuh"
 #include "ggml-cuda/diagmask.cuh"
+#ifdef FLASH_ATTN_AVAILABLE
 #include "ggml-cuda/fattn.cuh"
+#endif
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
 #include "ggml-cuda/mmq.cuh"
@@ -2160,8 +2162,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
             ggml_cuda_op_argsort(ctx, dst);
             break;
         case GGML_OP_FLASH_ATTN_EXT:
+#ifdef FLASH_ATTN_AVAILABLE
             ggml_cuda_flash_attn_ext(ctx, dst);
             break;
+#else
+            return false;
+#endif
         case GGML_OP_CROSS_ENTROPY_LOSS:
             ggml_cuda_cross_entropy_loss(ctx, dst);
             break;

ggml/src/ggml-cuda/mmq.cu

@@ -1,5 +1,12 @@
 #include "mmq.cuh"
 
+#ifdef GGML_CUDA_FORCE_CUBLAS
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context &,
+    const ggml_tensor *, const ggml_tensor *, ggml_tensor *, const char *, const float *,
+    const char *, float *, const int64_t, const int64_t, const int64_t,
+    const int64_t, cudaStream_t) {}
+#else
 void ggml_cuda_op_mul_mat_q(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
@@ -94,6 +101,7 @@ void ggml_cuda_op_mul_mat_q(
     GGML_UNUSED(dst);
     GGML_UNUSED(src1_ddf_i);
 }
+#endif // GGML_CUDA_FORCE_CUBLAS
 
 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
 #ifdef GGML_CUDA_FORCE_CUBLAS

ggml/src/ggml-cuda/mmq.cuh

@@ -2906,6 +2906,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
 #define DECL_MMQ_CASE(type)                                                        \
     template void mul_mat_q_case<type>(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) \
 
+#if !defined(GGML_CUDA_FORCE_CUBLAS)
 extern DECL_MMQ_CASE(GGML_TYPE_Q4_0);
 extern DECL_MMQ_CASE(GGML_TYPE_Q4_1);
 extern DECL_MMQ_CASE(GGML_TYPE_Q5_0);
@@ -2924,6 +2925,7 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ3_S);
 extern DECL_MMQ_CASE(GGML_TYPE_IQ1_S);
 extern DECL_MMQ_CASE(GGML_TYPE_IQ4_NL);
 extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS);
+#endif
 
 // -------------------------------------------------------------------------------------------------------------------------