Gg/flash attn

CMakeLists.txt

@@ -43,6 +43,18 @@ else()
     set(LLAMA_METAL_DEFAULT OFF)
 endif()
 
+# TODO: fix this for Android CI
+#       https://github.com/ggerganov/llama.cpp/pull/6716#issuecomment-2061509191
+#if (CMAKE_SYSTEM_NAME MATCHES "ANDROID")
+#    set(LLAMA_LLAMAFILE_DEFAULT OFF)
+#else()
+#    set(LLAMA_LLAMAFILE_DEFAULT ON)
+#endif()
+
+# TODO: temporary disable until MoE is fixed
+#       https://github.com/ggerganov/llama.cpp/pull/6716
+set(LLAMA_LLAMAFILE_DEFAULT OFF)
+
 # general
 option(BUILD_SHARED_LIBS                "build shared libraries"                                OFF)
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
@@ -88,6 +100,7 @@ endif()
 # 3rd party libs
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
+option(LLAMA_LLAMAFILE                       "llama: use llamafile SGEMM"                       ${LLAMA_LLAMAFILE_DEFAULT})
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
 option(LLAMA_CUDA                            "llama: use CUDA"                                  OFF)
 option(LLAMA_CUBLAS                          "llama: use CUDA (deprecated, use LLAMA_CUDA)"     OFF)
@@ -286,6 +299,7 @@ if (LLAMA_METAL)
         ${METALKIT_FRAMEWORK}
         )
 endif()
+
 if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
@@ -368,6 +382,13 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
+if (LLAMA_LLAMAFILE)
+    add_compile_definitions(GGML_USE_LLAMAFILE)
+
+    set(GGML_HEADERS_LLAMAFILE sgemm.h)
+    set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
+endif()
+
 if (LLAMA_QKK_64)
     add_compile_definitions(GGML_QKK_64)
 endif()
@@ -1151,15 +1172,16 @@ add_library(ggml OBJECT
             ggml-backend.h
             ggml-quants.c
             ggml-quants.h
-            ${GGML_SOURCES_CUDA}    ${GGML_HEADERS_CUDA}
-            ${GGML_SOURCES_OPENCL}  ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_METAL}   ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI}     ${GGML_HEADERS_MPI}
-            ${GGML_SOURCES_EXTRA}   ${GGML_HEADERS_EXTRA}
-            ${GGML_SOURCES_SYCL}    ${GGML_HEADERS_SYCL}
-            ${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
-            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
-            ${GGML_SOURCES_ROCM}    ${GGML_HEADERS_ROCM}
+            ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
+            ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI}       ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
+            ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
+            ${GGML_SOURCES_VULKAN}    ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
+            ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})

Makefile

@@ -384,6 +384,15 @@ ifdef LLAMA_OPENBLAS
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
+# TODO: temporary disable until MoE is fixed
+#       https://github.com/ggerganov/llama.cpp/pull/6716
+LLAMA_NO_LLAMAFILE := 1
+
+ifndef LLAMA_NO_LLAMAFILE
+	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
+	OBJS        += sgemm.o
+endif
+
 ifdef LLAMA_BLIS
 	MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
 	MK_LDFLAGS  += -lblis -L/usr/local/lib
@@ -480,11 +489,9 @@ ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/com
 
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
 	$(NVCC_COMPILE)
-
 endif # LLAMA_CUDA
 
 ifdef LLAMA_CLBLAST
-
 	MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other clblast OpenCL)
 	MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
@@ -603,6 +610,11 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
+ifndef LLAMA_NO_LLAMAFILE
+sgemm.o: sgemm.cpp sgemm.h ggml.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif
+
 GF_CC := $(CC)
 include scripts/get-flags.mk
 

Package.swift

@@ -4,6 +4,7 @@ import PackageDescription
 
 var sources = [
     "ggml.c",
+    "sgemm.cpp",
     "llama.cpp",
     "unicode.cpp",
     "unicode-data.cpp",

README.md

@@ -189,6 +189,8 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [MindMac](https://mindmac.app) (proprietary)
 - [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
+- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
+
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 ---

build.zig

@@ -112,6 +112,7 @@ pub fn build(b: *std.build.Builder) !void {
     make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
 
     const ggml = make.obj("ggml", "ggml.c");
+    const sgemm = make.obj("sgemm", "sgemm.cpp");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
     const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
@@ -128,14 +129,14 @@ pub fn build(b: *std.build.Builder) !void {
     const clip = make.obj("clip", "examples/llava/clip.cpp");
     const llava = make.obj("llava", "examples/llava/llava.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

common/common.cpp

@@ -108,6 +108,79 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
+#if defined(__x86_64__) && defined(__linux__)
+#include <pthread.h>
+
+static void cpuid(unsigned leaf, unsigned subleaf,
+                  unsigned *eax, unsigned *ebx, unsigned *ecx, unsigned *edx) {
+    __asm__("movq\t%%rbx,%%rsi\n\t"
+            "cpuid\n\t"
+            "xchgq\t%%rbx,%%rsi"
+            : "=a"(*eax), "=S"(*ebx), "=c"(*ecx), "=d"(*edx)
+            : "0"(leaf), "2"(subleaf));
+}
+
+static int pin_cpu(int cpu) {
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+    CPU_SET(cpu, &mask);
+    return pthread_setaffinity_np(pthread_self(), sizeof(mask), &mask);
+}
+
+static bool is_hybrid_cpu(void) {
+    unsigned eax, ebx, ecx, edx;
+    cpuid(7, 0, &eax, &ebx, &ecx, &edx);
+    return !!(edx & (1u << 15));
+}
+
+static bool is_running_on_efficiency_core(void) {
+    unsigned eax, ebx, ecx, edx;
+    cpuid(0x1a, 0, &eax, &ebx, &ecx, &edx);
+    int intel_atom = 0x20;
+    int core_type = (eax & 0xff000000u) >> 24;
+    return core_type == intel_atom;
+}
+
+static int count_math_cpus(int cpu_count) {
+    int result = 0;
+    for (int cpu = 0; cpu < cpu_count; ++cpu) {
+        if (pin_cpu(cpu)) {
+            return -1;
+        }
+        if (is_running_on_efficiency_core()) {
+            continue; // efficiency cores harm lockstep threading
+        }
+        ++cpu; // hyperthreading isn't useful for linear algebra
+        ++result;
+    }
+    return result;
+}
+
+#endif // __x86_64__ && __linux__
+
+/**
+ * Returns number of CPUs on system that are useful for math.
+ */
+int get_math_cpu_count() {
+#if defined(__x86_64__) && defined(__linux__)
+    int cpu_count = sysconf(_SC_NPROCESSORS_ONLN);
+    if (cpu_count < 1) {
+        return get_num_physical_cores();
+    }
+    if (is_hybrid_cpu()) {
+        cpu_set_t affinity;
+        if (!pthread_getaffinity_np(pthread_self(), sizeof(affinity), &affinity)) {
+            int result = count_math_cpus(cpu_count);
+            pthread_setaffinity_np(pthread_self(), sizeof(affinity), &affinity);
+            if (result > 0) {
+                return result;
+            }
+        }
+    }
+#endif
+    return get_num_physical_cores();
+}
+
 void process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -827,6 +900,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.cont_batching = true;
         return true;
     }
+    if (arg == "-fa" || arg == "--flash-attn") {
+        params.flash_attn = true;
+        return true;
+    }
     if (arg == "--color") {
         params.use_color = true;
         return true;
@@ -1763,6 +1840,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.cb_eval           = params.cb_eval;
     cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
+    cparams.flash_attn        = params.flash_attn;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
     cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
@@ -2600,6 +2678,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "seed: %u # default: -1 (random seed)\n", params.seed);
     fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
     fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
+    fprintf(stream, "flash_attn: %s # default: false\n", params.flash_attn ? "true" : "false");
     fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
 
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());

common/common.h

@@ -39,6 +39,7 @@ extern char const *LLAMA_BUILD_TARGET;
 
 struct llama_control_vector_load_info;
 
+int get_math_cpu_count();
 int32_t get_num_physical_cores();
 
 //
@@ -48,7 +49,7 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = get_num_physical_cores();
+    int32_t n_threads             = get_math_cpu_count();
     int32_t n_threads_draft       = -1;
     int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_threads_batch_draft = -1;
@@ -147,6 +148,7 @@ struct gpt_params {
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
+    bool flash_attn        = false; // flash attention
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens