Merge branch 'master' into HEAD

Author: ggerganov

CMakeLists.txt

@@ -75,6 +75,7 @@ set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 OFF)
 option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
+option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
 option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
@@ -225,6 +226,14 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
+if (LLAMA_K_QUANTS)
+    set(GGML_SOURCES_EXTRA ${GGML_SOURCES_EXTRA} k_quants.c k_quants.h)
+    add_compile_definitions(GGML_USE_K_QUANTS)
+    if (LLAMA_QKK_64)
+        add_compile_definitions(GGML_QKK_64)
+    endif()
+endif()
+
 if (LLAMA_CUBLAS)
     cmake_minimum_required(VERSION 3.17)
 
@@ -250,6 +259,15 @@ if (LLAMA_CUBLAS)
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
 
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        if (LLAMA_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "61") # needed for f16 CUDA intrinsics
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "52") # lowest CUDA 12 standard
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
     else()
         message(WARNING "cuBLAS not found")
     endif()
@@ -280,11 +298,6 @@ if (LLAMA_METAL)
         )
 endif()
 
-if (LLAMA_K_QUANTS)
-    set(GGML_SOURCES_EXTRA ${GGML_SOURCES_EXTRA} k_quants.c k_quants.h)
-    add_compile_definitions(GGML_USE_K_QUANTS)
-endif()
-
 if (LLAMA_CLBLAST)
     find_package(CLBlast)
     if (CLBlast_FOUND)
@@ -493,22 +506,6 @@ if (BUILD_SHARED_LIBS)
     endif()
 endif()
 
-if (GGML_SOURCES_CUDA)
-    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
-    set_property(TARGET ggml  PROPERTY CUDA_ARCHITECTURES "native")
-    set_property(TARGET ggml  PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
-
-    set_property(TARGET ggml_static PROPERTY CUDA_ARCHITECTURES "native")
-    set_property(TARGET ggml_static PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
-
-    if (BUILD_SHARED_LIBS)
-        set_property(TARGET ggml_shared PROPERTY CUDA_ARCHITECTURES "native")
-        set_property(TARGET ggml_shared PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
-    endif()
-
-    set_property(TARGET llama PROPERTY CUDA_ARCHITECTURES "native")
-endif()
-
 
 #
 # programs, examples and tests

Makefile

@@ -43,8 +43,11 @@ endif
 
 # keep standard at C11 and C++11
 # -Ofast tends to produce faster code, but may not be available for some compilers.
-#OPT = -Ofast
+ifdef LLAMA_FAST
+OPT = -Ofast
+else
 OPT = -O3
+endif
 CFLAGS   = -I.              $(OPT) -std=c11   -fPIC
 CXXFLAGS = -I. -I./examples $(OPT) -std=c++11 -fPIC
 LDFLAGS  =
@@ -131,6 +134,10 @@ ifndef LLAMA_NO_K_QUANTS
 	CFLAGS   += -DGGML_USE_K_QUANTS
 	CXXFLAGS += -DGGML_USE_K_QUANTS
 	OBJS     += k_quants.o
+ifdef LLAMA_QKK_64
+	CFLAGS   += -DGGML_QKK_64
+	CXXFLAGS += -DGGML_QKK_64
+endif
 endif
 
 ifndef LLAMA_NO_ACCELERATE

README.md

@@ -5,16 +5,16 @@
 [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
+
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
-- Roadmap June 2023: https://github.com/ggerganov/llama.cpp/discussions/1729
-- GPU support with Metal (Apple Silicon): https://github.com/ggerganov/llama.cpp/pull/1642
-- High-quality 2,3,4,5,6-bit quantization: https://github.com/ggerganov/llama.cpp/pull/1684
-- Multi-GPU support: https://github.com/ggerganov/llama.cpp/pull/1607
-- Training LLaMA models from scratch: https://github.com/ggerganov/llama.cpp/pull/1652
-- CPU threading improvements: https://github.com/ggerganov/llama.cpp/pull/1632
+- k-quants now support super-block size of 64: https://github.com/ggerganov/llama.cpp/pull/2001
+- New roadmap: https://github.com/users/ggerganov/projects/7
+- Azure CI brainstorming: https://github.com/ggerganov/llama.cpp/discussions/1985
+- p1 : LLM-based code completion engine at the edge : https://github.com/ggml-org/p1/discussions/1
 
 <details>
   <summary>Table of Contents</summary>
@@ -33,6 +33,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
         <li><a href="#quantization">Quantization</a></li>
         <li><a href="#interactive-mode">Interactive mode</a></li>
         <li><a href="#instruction-mode-with-alpaca">Instruction mode with Alpaca</a></li>
+        <li><a href="#using-openllama">Using OpenLLaMA</a></li>
         <li><a href="#using-gpt4all">Using GPT4All</a></li>
         <li><a href="#using-pygmalion-7b--metharme-7b">Using Pygmalion 7B & Metharme 7B</a></li>
         <li><a href="#obtaining-the-facebook-llama-original-model-and-stanford-alpaca-model-data">Obtaining the Facebook LLaMA original model and Stanford Alpaca model data</a></li>
@@ -344,7 +345,7 @@ Building the program with BLAS support may lead to some performance improvements
   | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA dequantization + mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_F16     | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels. Can improve performance on relatively recent GPUs. |
-  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value 2 1 can improve performance for slow GPUs. |
+  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
 
 - #### CLBlast
 
@@ -378,7 +379,7 @@ Building the program with BLAS support may lead to some performance improvements
       ```sh
       git clone https://github.com/CNugteren/CLBlast.git
       mkdir CLBlast/build
-      cd CLBLast/build
+      cd CLBlast/build
       cmake .. -DBUILD_SHARED_LIBS=OFF -DTUNERS=OFF
       cmake --build . --config Release
       cmake --install . --prefix /some/path
@@ -547,6 +548,13 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 >
 ```
 
+### Using [OpenLLaMA](https://github.com/openlm-research/open_llama)
+
+OpenLLaMA is an openly licensed reproduction of Meta's original LLaMA model. It uses the same architecture and is a drop-in replacement for the original LLaMA weights.
+
+- Download the [3B](https://huggingface.co/openlm-research/open_llama_3b), [7B](https://huggingface.co/openlm-research/open_llama_7b), or [13B](https://huggingface.co/openlm-research/open_llama_13b) model from Hugging Face.
+- Convert the model to ggml FP16 format using `python convert.py <path to OpenLLaMA directory>`
+
 ### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
 
 - Obtain the `tokenizer.model` file from LLaMA model and put it to `models`
@@ -676,12 +684,13 @@ Upon completion of the aforementioned steps, you will have successfully compiled
 ```
 GGML_OPENCL_PLATFORM=0
 GGML_OPENCL_DEVICE=0
-export LD_LIBRARY_PATH=/system/vendor/lib64:$LD_LIBRARY_PATH
-./main (...)
+export LD_LIBRARY_PATH=/vendor/lib64:$LD_LIBRARY_PATH
 ```
 
 For easy and swift re-execution, consider documenting this final part in a .sh script file. This will enable you to rerun the process with minimal hassle.
 
+Place your desired model into the `/llama.cpp/models/` directory and execute the `./main (...)` script.
+
 ### Docker
 
 #### Prerequisites

build.zig

@@ -1,61 +1,58 @@
 const std = @import("std");
 
+// Zig Version: 0.11.0-dev.3379+629f0d23b
 pub fn build(b: *std.build.Builder) void {
     const target = b.standardTargetOptions(.{});
-    const optimize = b.standardReleaseOptions();
-    const want_lto = b.option(bool, "lto", "Want -fLTO");
-
-    const lib = b.addStaticLibrary("llama", null);
-    lib.want_lto = want_lto;
-    lib.setTarget(target);
-    lib.setBuildMode(optimize);
+    const optimize = b.standardOptimizeOption(.{});
+    const lib = b.addStaticLibrary(.{
+        .name = "llama",
+        .target = target,
+        .optimize = optimize,
+    });
+    lib.linkLibC();
     lib.linkLibCpp();
     lib.addIncludePath(".");
-    lib.addIncludePath("examples");
+    lib.addIncludePath("./examples");
     lib.addCSourceFiles(&.{
         "ggml.c",
     }, &.{"-std=c11"});
     lib.addCSourceFiles(&.{
         "llama.cpp",
     }, &.{"-std=c++11"});
-    lib.install();
-
-    const build_args = .{ .b = b, .lib = lib, .target = target, .optimize = optimize, .want_lto = want_lto };
-
-    const exe = build_example("main", build_args);
-    _ = build_example("quantize", build_args);
-    _ = build_example("perplexity", build_args);
-    _ = build_example("embedding", build_args);
-
-    // create "zig build run" command for ./main
-
-    const run_cmd = exe.run();
-    run_cmd.step.dependOn(b.getInstallStep());
-    if (b.args) |args| {
-        run_cmd.addArgs(args);
+    b.installArtifact(lib);
+
+    const examples = .{
+        "main",
+        "baby-llama",
+        "embedding",
+        // "metal",
+        "perplexity",
+        "quantize",
+        "quantize-stats",
+        "save-load-state",
+        // "server",
+        "simple",
+        "train-text-from-scratch",
+    };
+
+    inline for (examples) |example_name| {
+        const exe = b.addExecutable(.{
+            .name = example_name,
+            .target = target,
+            .optimize = optimize,
+        });
+        exe.addIncludePath(".");
+        exe.addIncludePath("./examples");
+        exe.addCSourceFiles(&.{
+            std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{example_name, example_name}),
+            "examples/common.cpp",
+        }, &.{"-std=c++11"});
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_" ++ example_name, "Run the app");
+        run_step.dependOn(&run_cmd.step);
     }
-
-    const run_step = b.step("run", "Run the app");
-    run_step.dependOn(&run_cmd.step);
-}
-
-fn build_example(comptime name: []const u8, args: anytype) *std.build.LibExeObjStep {
-    const b = args.b;
-    const lib = args.lib;
-    const want_lto = args.want_lto;
-
-    const exe = b.addExecutable(name, null);
-    exe.want_lto = want_lto;
-    lib.setTarget(args.target);
-    lib.setBuildMode(args.optimize);
-    exe.addIncludePath(".");
-    exe.addIncludePath("examples");
-    exe.addCSourceFiles(&.{
-        std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{name, name}),
-        "examples/common.cpp",
-    }, &.{"-std=c++11"});
-    exe.linkLibrary(lib);
-    exe.install();
-
-    return exe;
 }