llama_cpp 0.16.0 → 0.16.1

data/vendor/tmp/llama.cpp/ggml-blas.cpp

@@ -0,0 +1,363 @@
+#include "ggml-blas.h"
+#include "ggml-backend-impl.h"
+
+#include <future>
+#include <vector>
+
+#if defined(GGML_USE_ACCELERATE)
+#   include <Accelerate/Accelerate.h>
+#elif defined(GGML_BLAS_USE_MKL)
+#   include <mkl.h>
+#else
+#   include <cblas.h>
+#   ifdef BLIS_ENABLE_CBLAS
+#       include <blis.h>
+#   endif
+#endif
+
+struct ggml_backend_blas_context {
+    int n_threads = GGML_DEFAULT_N_THREADS;
+    std::unique_ptr<char[]> work_data;
+    size_t work_size = 0;
+#ifndef GGML_USE_OPENMP
+    std::vector<std::future<void>> tasks;
+#endif
+};
+
+// helper function to determine if it is better to use BLAS or not
+// for large matrices, BLAS is faster
+static bool ggml_backend_blas_use_blas(const struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    const int64_t ne10 = src1->ne[0];
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+
+    // TODO: find the optimal values for these
+    if (ggml_is_contiguous(src0) &&
+        ggml_is_contiguous(src1) &&
+        src1->type == GGML_TYPE_F32 &&
+        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
+
+        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
+        return true;
+    }
+
+    return false;
+}
+
+static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    const int64_t ne_plane      = ne01*ne00;
+    const size_t  desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);
+
+    if (ctx->work_size < desired_wsize) {
+        ctx->work_data.reset(new char[desired_wsize]);
+        ctx->work_size = desired_wsize;
+    }
+    void * wdata = ctx->work_data.get();
+
+    // convert src0 to float
+    if (type != GGML_TYPE_F32) {
+        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
+        ggml_to_float_t const to_float = type_traits.to_float;
+
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                const void  *       x      = (char *)  src0->data + i02*nb02          + i03*nb03;
+                      float * const wplane = (float *) wdata      + i02*ne_plane      + i03*ne02*ne_plane;
+
+                const int min_cols_per_thread = 4096;
+                const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
+                const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);
+
+#ifdef GGML_USE_OPENMP
+                #pragma omp parallel for num_threads(n_threads)
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                }
+#else
+                for (int i = 1; i < n_threads; i++) {
+                    const int64_t start =       i*ne01/n_threads;
+                    const int64_t end   = (i + 1)*ne01/n_threads;
+                    if (start < end) {
+                        ctx->tasks.push_back(std::async(std::launch::async, [=]() {
+                            for (int64_t i01 = start; i01 < end; i01++) {
+                                to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                            }
+                        }));
+                    }
+                }
+                {
+                    // reuse the current thread for the first task
+                    const int64_t start = 0;
+                    const int64_t end   = ne01/n_threads;
+                    for (int64_t i01 = start; i01 < end; i01++) {
+                        to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                    }
+                }
+#endif
+            }
+        }
+
+#ifndef GGML_USE_OPENMP
+        // wait for all tasks to finish
+        for (auto & task : ctx->tasks) {
+            task.get();
+        }
+        ctx->tasks.clear();
+#endif
+    }
+
+#if defined(OPENBLAS_VERSION)
+    openblas_set_num_threads(ctx->n_threads);
+#endif
+
+#if defined(BLIS_ENABLE_CBLAS)
+    bli_thread_set_num_threads(ctx->n_threads);
+#endif
+
+    for (int64_t i13 = 0; i13 < ne13; i13++) {
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            const int64_t i03 = i13/r3;
+            const int64_t i02 = i12/r2;
+
+            const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
+            const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
+                  float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
+
+            if (type != GGML_TYPE_F32) {
+                x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
+            }
+
+            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
+                        ne1, ne01, ne10,
+                        1.0f,   y, ne10,
+                                x, ne00,
+                        0.0f,   d, ne01);
+        }
+    }
+}
+
+static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(ne0  == ne00);
+    GGML_ASSERT(ne1  == ne10);
+    GGML_ASSERT(ne2  == ne02);
+    GGML_ASSERT(ne02 == ne12);
+    GGML_ASSERT(ne3  == ne13);
+    GGML_ASSERT(ne03 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    // GGML_ASSERT(nb0 <= nb1);
+    // GGML_ASSERT(nb1 <= nb2);
+    // GGML_ASSERT(nb2 <= nb3);
+
+    // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
+    // src0: (k,n)
+    // src1: (k,m)
+    // dst:  (m,n)
+    //
+    // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
+    // Also expressed as (major,minor)
+    // a: (m,k): so src1 transposed
+    // b: (k,n): so src0
+    // c: (m,n)
+    //
+    // However, if ggml_is_transposed(src1) is true, then
+    // src1->data already contains a transposed version, so sgemm mustn't
+    // transpose it further.
+
+    int n = src0->ne[0];
+    int k = src0->ne[1];
+    int m = src1->ne[0];
+
+    CBLAS_TRANSPOSE transposeA;
+    int lda;
+
+    if (!ggml_is_transposed(src1)) {
+        transposeA = CblasTrans;
+        lda = m;
+    } else {
+        transposeA = CblasNoTrans;
+        lda = k;
+    }
+
+    float * a = (float *) ((char *) src1->data);
+    float * b = (float *) ((char *) src0->data);
+    float * c = (float *) ((char *) dst->data);
+
+    cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
+
+    GGML_UNUSED(ctx);
+}
+
+// backend interface
+
+GGML_CALL static const char * ggml_backend_blas_name(ggml_backend_t backend) {
+    return "BLAS";
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static void ggml_backend_blas_free(ggml_backend_t backend) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+    delete ctx;
+    delete backend;
+}
+
+GGML_CALL static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        switch (node->op) {
+            case GGML_OP_MUL_MAT:
+                ggml_backend_blas_mul_mat(ctx, node);
+                break;
+
+            case GGML_OP_OUT_PROD:
+                ggml_backend_blas_out_prod(ctx, node);
+                break;
+
+            case GGML_OP_NONE:
+            case GGML_OP_RESHAPE:
+            case GGML_OP_VIEW:
+            case GGML_OP_PERMUTE:
+            case GGML_OP_TRANSPOSE:
+                break;
+
+            default:
+                fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
+                GGML_ASSERT(false);
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    return (op->op == GGML_OP_MUL_MAT  && ggml_backend_blas_use_blas(op)) ||
+           (op->op == GGML_OP_OUT_PROD && op->src[0]->type == GGML_TYPE_F32 &&
+                                          op->src[1]->type == GGML_TYPE_F32 &&
+                                          ggml_is_matrix(src0) &&
+                                          ggml_is_matrix(src1) &&
+                                          ggml_is_contiguous(src0) &&
+                                          (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
+
+    GGML_UNUSED(backend);
+}
+
+GGML_CALL static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i blas_backend_i = {
+    /* .get_name                = */ ggml_backend_blas_name,
+    /* .free                    = */ ggml_backend_blas_free,
+    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
+    /* .supports_op             = */ ggml_backend_blas_supports_op,
+    /* .supports_buft           = */ ggml_backend_blas_supports_buft,
+    /* .offload_op              = */ NULL,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_blas_guid(void) {
+    static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_blas_init(void) {
+    ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
+
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_blas_guid(),
+        /* .interface = */ blas_backend_i,
+        /* .context   = */ ctx,
+    };
+
+#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
+    if (openblas_get_parallel() != OPENBLAS_OPENMP) {
+        fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
+    }
+#endif
+
+#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
+    fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
+#endif
+
+    return backend;
+}
+
+GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
+}
+
+void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_blas(backend_blas));
+
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
+    ctx->n_threads = n_threads;
+}

data/vendor/tmp/llama.cpp/ggml-blas.h

@@ -0,0 +1,23 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_API GGML_CALL ggml_backend_t ggml_backend_blas_init(void);
+
+GGML_API GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend);
+
+// number of threads used for conversion to float
+// for openblas and blis, this will also set the number of threads used for blas operations
+GGML_API GGML_CALL void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+
+
+#ifdef  __cplusplus
+}
+#endif

data/vendor/tmp/llama.cpp/ggml-common.h

@@ -123,12 +123,18 @@ typedef sycl::half2 ggml_half2;
 #define QI1_S (QK_K / (4*QR1_S))
 #define QR1_S 8
 
+#define QI1_M (QK_K / (4*QR1_M))
+#define QR1_M 8
+
 #define QI4_NL (QK4_NL / (4*QR4_NL))
 #define QR4_NL 2
 
 #define QI4_XS (QK_K / (4*QR4_XS))
 #define QR4_XS 8
 
+#define QI3_S (QK_K / (4*QR3_S))
+#define QR3_S 8
+
 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
 
 #define QK4_0 32

data/vendor/tmp/llama.cpp/ggml-cuda/argsort.cu

@@ -73,6 +73,7 @@ static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, co
     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) {

data/vendor/tmp/llama.cpp/ggml-cuda/dmmv.cu

@@ -422,10 +422,22 @@ static __device__ void convert_f16(const void * vx, const int64_t ib, const int
     v.y = x[ib + iqs + 1];
 }
 
-template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
+static constexpr __device__ dequantize_kernel_t get_dequantize_kernel(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? dequantize_q4_0 :
+        type == GGML_TYPE_Q4_1 ? dequantize_q4_1 :
+        type == GGML_TYPE_Q5_0 ? dequantize_q5_0 :
+        type == GGML_TYPE_Q5_1 ? dequantize_q5_1 :
+        type == GGML_TYPE_Q8_0 ? dequantize_q8_0 :
+        type == GGML_TYPE_F16 ? convert_f16 :
+        nullptr;
+}
+
+template <ggml_type type>
 static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
-    // qk = quantized weights per x block
-    // qr = number of quantized weights per data value in x block
+    constexpr int qk = ggml_cuda_type_traits<type>::qk; // quantized weights per x block
+    constexpr int qr = ggml_cuda_type_traits<type>::qr; // number of quantized weights per data value in x block
+    constexpr dequantize_kernel_t dequantize_kernel = get_dequantize_kernel(type);
+
     const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
 
     if (row >= nrows) {
@@ -493,7 +505,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y,
     // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
+    dequantize_mul_mat_vec<GGML_TYPE_Q4_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -502,7 +514,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y,
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
+    dequantize_mul_mat_vec<GGML_TYPE_Q4_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -511,7 +523,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y,
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
+    dequantize_mul_mat_vec<GGML_TYPE_Q5_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -520,7 +532,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y,
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
+    dequantize_mul_mat_vec<GGML_TYPE_Q5_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -529,7 +541,7 @@ static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y,
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
+    dequantize_mul_mat_vec<GGML_TYPE_Q8_0>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 
@@ -580,7 +592,7 @@ static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, floa
     const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
     const dim3 block_nums(block_num_y, 1, 1);
     const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<1, 1, convert_f16>
+    dequantize_mul_mat_vec<GGML_TYPE_F16>
         <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
 }
 

data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f16.cu

@@ -43,7 +43,7 @@ static __global__ void flash_attn_tile_ext_f16(
         const int ne1,
         const int ne2,
         const int ne3) {
-#if FP16_AVAILABLE
+#ifdef FP16_AVAILABLE
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.