@novastera-oss/llamarn 0.0.1-alpha.4

package/cpp/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt

@@ -0,0 +1,184 @@
+cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+find_package(CUDAToolkit)
+
+if (CUDAToolkit_FOUND)
+    message(STATUS "CUDA Toolkit found")
+
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # native == GPUs available at build time
+        # 50     == Maxwell, lowest CUDA 12 standard
+        # 60     == P100, FP16 CUDA intrinsics
+        # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
+        # 70     == V100, FP16 tensor cores
+        # 75     == Turing, int8 tensor cores
+        # 80     == Ampere, asynchronous data loading, faster tensor core instructions
+        # 86     == RTX 3000, needs CUDA v11.1
+        # 89     == RTX 4000, needs CUDA v11.8
+        #
+        # XX-virtual == compile CUDA code as PTX, do JIT compilation to binary code on first run
+        # XX-real    == compile CUDA code as device code for this specific architecture
+        # no suffix  == compile as both PTX and device code
+        #
+        # The default behavior for a non-native is to build virtual architectures as needed to cover all features needed
+        #     for best performance and to also build real architectures for the most commonly used GPUs.
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+            set(CMAKE_CUDA_ARCHITECTURES "native")
+        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
+                set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
+            else()
+                set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real")
+            endif()
+        else()
+            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
+                set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
+            else()
+                set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real")
+            endif()
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+    enable_language(CUDA)
+
+    file(GLOB   GGML_HEADERS_CUDA "*.cuh")
+    list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_CUDA "*.cu")
+    file(GLOB   SRCS "template-instances/fattn-mma*.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")
+        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()
+
+    ggml_add_backend_library(ggml-cuda
+                             ${GGML_HEADERS_CUDA}
+                             ${GGML_SOURCES_CUDA}
+                            )
+
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_CUDA_GRAPHS)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (NOT GGML_CUDA_FA)
+        add_compile_definitions(GGML_CUDA_NO_FA)
+    endif()
+
+    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        add_compile_definitions(GGML_CUDA_F16)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        if (WIN32)
+            # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+        else ()
+            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+        endif()
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
+    endif()
+
+    set(CUDA_CXX_FLAGS "")
+
+    set(CUDA_FLAGS -use_fast_math -extended-lambda)
+
+    if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
+        # Options are:
+        # - none (not recommended)
+        # - speed (nvcc's default)
+        # - balance
+        # - size
+        list(APPEND CUDA_FLAGS -compress-mode=${GGML_CUDA_COMPRESSION_MODE})
+    endif()
+
+    if (GGML_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    endif()
+
+    if (GGML_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
+
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+                OUTPUT_STRIP_TRAILING_WHITESPACE
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message(STATUS "CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+    endif()
+
+    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+
+    target_compile_options(ggml-cuda PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+else()
+    message(FATAL_ERROR "CUDA Toolkit not found")
+endif()

package/cpp/llama.cpp/ggml/src/ggml-cuda/acc.cu

@@ -0,0 +1,47 @@
+#include "acc.cuh"
+
+static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, int offset) {
+    const int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i >= ne) {
+        return;
+    }
+    int src1_idx = i - offset;
+    int oz = src1_idx / nb2;
+    int oy = (src1_idx - (oz * nb2)) / nb1;
+    int ox = src1_idx % nb1;
+    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
+        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
+    } else {
+        dst[i] = x[i];
+    }
+}
+
+static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
+    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
+    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
+}
+
+void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
+
+    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
+    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
+    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
+    int offset = dst->op_params[3] / 4; // offset in bytes
+
+    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
+}

package/cpp/llama.cpp/ggml/src/ggml-cuda/acc.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ACC_BLOCK_SIZE 256
+
+void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

package/cpp/llama.cpp/ggml/src/ggml-cuda/arange.cu

@@ -0,0 +1,34 @@
+#include "arange.cuh"
+
+static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) {
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    dst[nidx] = start + step * nidx;
+}
+
+static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
+    arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start,  step);
+}
+
+void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    float start;
+    float stop;
+    float step;
+    memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
+    memcpy(&stop,  (float *)dst->op_params + 1, sizeof(float));
+    memcpy(&step,  (float *)dst->op_params + 2, sizeof(float));
+
+    int64_t steps = (int64_t)ceil((stop - start) / step);
+    GGML_ASSERT(ggml_nelements(dst) == steps);
+
+    arange_f32_cuda(dst_d, dst->ne[0], start, step, stream);
+}

package/cpp/llama.cpp/ggml/src/ggml-cuda/arange.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ARANGE_BLOCK_SIZE 256
+
+void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

package/cpp/llama.cpp/ggml/src/ggml-cuda/argmax.cu

@@ -0,0 +1,91 @@
+#include <algorithm>
+#include <cstdint>
+
+#include "argmax.cuh"
+#include "common.cuh"
+#include "sum.cuh"
+
+static __global__ void argmax_f32(const float * __restrict__ x, int32_t * __restrict__ dst, const int64_t ncols) {
+    const int64_t row = blockIdx.x;
+
+    float maxval = -FLT_MAX;
+    int   argmax = -1;
+    const float * rowx = x + row * ncols;
+
+    for (int32_t col = threadIdx.x; col < ncols; col += blockDim.x) {
+        const float val = rowx[col];
+        if (val > maxval) {
+            maxval = val;
+            argmax = col;
+        }
+    }
+
+#pragma unroll
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
+        const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
+        if (val > maxval) {
+            maxval = val;
+            argmax = col;
+        }
+    }
+
+    const int n_warps = blockDim.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    if (n_warps > 1) {
+        constexpr int    max_warps = 1024 / WARP_SIZE;
+        __shared__ float shared_maxval[max_warps];
+        __shared__ int   shared_argmax[max_warps];
+        if (lane_id == 0) {
+            shared_maxval[warp_id] = maxval;
+            shared_argmax[warp_id] = argmax;
+        }
+
+        __syncthreads();
+
+        if (warp_id == 0) {
+            if (lane_id < n_warps) {
+                maxval = shared_maxval[lane_id];
+                argmax = shared_argmax[lane_id];
+            }
+#pragma unroll
+            for (int offset = 16; offset > 0; offset >>= 1) {
+                const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
+                const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
+                if (val > maxval) {
+                    maxval = val;
+                    argmax = col;
+                }
+            }
+        }
+    }
+
+    if (warp_id == 0 && lane_id == 0) {
+        dst[row] = argmax;
+    }
+}
+
+void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_I32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ne00  = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const float * src0_d = (const float *) src0->data;
+    int32_t     * dst_d  = (int32_t     *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t num_blocks = nrows;
+    const int64_t num_threads = std::min<int64_t>(1024, (ne00 + WARP_SIZE - 1) / WARP_SIZE * WARP_SIZE);
+    const dim3 blocks_dim(num_threads, 1, 1);
+    const dim3 blocks_num(num_blocks, 1, 1);
+
+    argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00);
+}

package/cpp/llama.cpp/ggml/src/ggml-cuda/argmax.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

package/cpp/llama.cpp/ggml/src/ggml-cuda/argsort.cu

@@ -0,0 +1,104 @@
+#include "argsort.cuh"
+
+template<typename T>
+static inline __device__ void ggml_cuda_swap(T & a, T & b) {
+    T tmp = a;
+    a = b;
+    b = tmp;
+}
+
+template<ggml_sort_order order>
+static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
+    // bitonic sort
+    int col = threadIdx.x;
+    int row = blockIdx.y;
+
+    if (col >= ncols_pad) {
+        return;
+    }
+
+    const float * x_row = x + row * ncols;
+    extern __shared__ int dst_row[];
+
+    // initialize indices
+    dst_row[col] = col;
+
+    __syncthreads();
+
+    for (int k = 2; k <= ncols_pad; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= ncols ||
+                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    ) {
+                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
+                    }
+                } else {
+                    if (dst_row[ixj] >= ncols ||
+                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    ) {
+                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
+                    }
+                }
+            }
+            __syncthreads();
+        }
+    }
+
+    // copy the result to dst without the padding
+    if (col < ncols) {
+        dst[row * ncols + col] = dst_row[col];
+    }
+}
+
+static int next_power_of_2(int x) {
+    int n = 1;
+    while (n < x) {
+        n *= 2;
+    }
+    return n;
+}
+
+static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
+    // bitonic sort requires ncols to be power of 2
+    const int ncols_pad = next_power_of_2(ncols);
+
+    const dim3 block_dims(ncols_pad, 1, 1);
+    const dim3 block_nums(1, nrows, 1);
+    const size_t shared_mem = ncols_pad * sizeof(int);
+
+    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
+    GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
+
+    if (order == GGML_SORT_ORDER_ASC) {
+        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+    } else if (order == GGML_SORT_ORDER_DESC) {
+        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_I32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
+
+    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
+}

package/cpp/llama.cpp/ggml/src/ggml-cuda/argsort.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst);