llama_cpp 0.16.1 → 0.17.0

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

@@ -1,221 +0,0 @@
-#include "norm.cuh"
-
-template <int block_size>
-static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) {
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    const int tid = threadIdx.x;
-
-    float2 mean_var = make_float2(0.f, 0.f);
-
-    for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row*ncols + col];
-        mean_var.x += xi;
-        mean_var.y += xi * xi;
-    }
-
-    // sum up partial sums
-    mean_var = warp_reduce_sum(mean_var);
-    if (block_size > WARP_SIZE) {
-        __shared__ float2 s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = mean_var;
-        }
-        __syncthreads();
-        mean_var = s_sum[lane_id];
-        mean_var = warp_reduce_sum(mean_var);
-    }
-
-    const float mean = mean_var.x / ncols;
-    const float var = mean_var.y / ncols - mean * mean;
-    const float inv_std = rsqrtf(var + eps);
-
-    for (int col = tid; col < ncols; col += block_size) {
-        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
-    }
-}
-
-template <int block_size>
-static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
-    // blockIdx.x: num_groups idx
-    // threadIdx.x: block_size idx
-    int start = blockIdx.x * group_size;
-    int end = start + group_size;
-
-    start += threadIdx.x;
-
-    if (end >= ne_elements) {
-        end = ne_elements;
-    }
-
-    float tmp = 0.0f; // partial sum for thread in warp
-
-    for (int j = start; j < end; j += block_size) {
-        tmp += x[j];
-    }
-
-    tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
-        __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
-        }
-        __syncthreads();
-        tmp = s_sum[lane_id];
-        tmp = warp_reduce_sum(tmp);
-    }
-
-    float mean = tmp / group_size;
-    tmp = 0.0f;
-
-    for (int j = start; j < end; j += block_size) {
-        float xi = x[j] - mean;
-        dst[j] = xi;
-        tmp += xi * xi;
-    }
-
-    tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
-        __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
-        }
-        __syncthreads();
-        tmp = s_sum[lane_id];
-        tmp = warp_reduce_sum(tmp);
-    }
-
-    float variance = tmp / group_size;
-    float scale = rsqrtf(variance + eps);
-    for (int j = start; j < end; j += block_size) {
-        dst[j] *= scale;
-    }
-}
-
-template <int block_size>
-static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    const int tid = threadIdx.x;
-
-    float tmp = 0.0f; // partial sum for thread in warp
-
-    for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row*ncols + col];
-        tmp += xi * xi;
-    }
-
-    // sum up partial sums
-    tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
-        __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
-        }
-        __syncthreads();
-        tmp = s_sum[lane_id];
-        tmp = warp_reduce_sum(tmp);
-    }
-
-    const float mean = tmp / ncols;
-    const float scale = rsqrtf(mean + eps);
-
-    for (int col = tid; col < ncols; col += block_size) {
-        dst[row*ncols + col] = scale * x[row*ncols + col];
-    }
-}
-
-static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
-    if (ncols < 1024) {
-        const dim3 block_dims(WARP_SIZE, 1, 1);
-        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
-    } else {
-        const dim3 block_dims(1024, 1, 1);
-        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
-    }
-}
-
-static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
-    static const float eps = 1e-6f;
-    if (group_size < 1024) {
-        const dim3 block_dims(WARP_SIZE, 1, 1);
-        group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
-    } else {
-        const dim3 block_dims(1024, 1, 1);
-        group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
-    }
-}
-
-static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
-    if (ncols < 1024) {
-        const dim3 block_dims(WARP_SIZE, 1, 1);
-        rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
-    } else {
-        const dim3 block_dims(1024, 1, 1);
-        rms_norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
-    }
-}
-
-void ggml_cuda_op_norm(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(ggml_is_contiguous(src0));
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
-
-    float eps;
-    memcpy(&eps, dst->op_params, sizeof(float));
-
-    norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
-}
-
-void ggml_cuda_op_group_norm(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(ggml_is_contiguous(src0));
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    int num_groups = dst->op_params[0];
-    int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
-    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], group_size, ggml_nelements(src0), stream);
-}
-
-void ggml_cuda_op_rms_norm(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(ggml_is_contiguous(src0));
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
-
-    float eps;
-    memcpy(&eps, dst->op_params, sizeof(float));
-
-    rms_norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
-}

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

@@ -1,49 +0,0 @@
-#include "pad.cuh"
-
-static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
-    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
-    // blockIdx.y: idx of ne1
-    // blockIDx.x: idx of ne0 / BLOCK_SIZE
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-
-    // operation
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne00 +
-            blockIdx.z * ne00 * ne01;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        dst[offset_dst] = 0.0f;
-    }
-}
-
-static void pad_f32_cuda(const float * x, float * dst,
-    const int ne00, const int ne01, const int ne02, const int ne03,
-    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2*ne3);
-    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
-}
-
-void ggml_cuda_op_pad(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_F32);
-    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
-
-    pad_f32_cuda(src0_d, dst_d,
-        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
-        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
-}

data/vendor/tmp/llama.cpp/ggml-cuda/pool2d.cu

@@ -1,94 +0,0 @@
-#include "pool2d.cuh"
-
-template <typename Ti, typename To>
-static  __global__ void pool2d_nchw_kernel(
-        const int ih, const int iw, const int oh, const int ow,
-        const int kh, const int kw, const int sh, const int sw,
-        const int ph, const int pw, const int parallel_elements,
-        const Ti* src, To* dst, const enum ggml_op_pool op) {
-    int idx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (idx >= parallel_elements) {
-        return;
-    }
-
-    const int I_HW = ih * iw;
-    const int O_HW = oh * ow;
-    const int nc = idx / O_HW;
-    const int cur_oh = idx % O_HW / ow;
-    const int cur_ow = idx % O_HW % ow;
-    const Ti* i_ptr = src + nc * I_HW;
-    To* o_ptr = dst + nc * O_HW;
-    const int start_h = cur_oh * sh - ph;
-    const int bh = max(0, start_h);
-    const int eh = min(ih, start_h + kh);
-    const int start_w = cur_ow * sw - pw;
-    const int bw = max(0, start_w);
-    const int ew = min(iw, start_w + kw);
-    const To scale = 1. / (kh * kw);
-    To res = 0;
-
-    switch (op) {
-        case GGML_OP_POOL_AVG: res = 0; break;
-        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
-        default: assert(false);
-    }
-
-    for (int i = bh; i < eh; i += 1) {
-        for (int j = bw; j < ew; j += 1) {
-#if __CUDA_ARCH__ >= 350
-            Ti cur = __ldg(i_ptr + i * iw + j);
-#else
-            Ti cur = i_ptr[i * iw + j];
-#endif
-            switch (op) {
-                case GGML_OP_POOL_AVG: res += cur * scale; break;
-                case GGML_OP_POOL_MAX: res = max(res, (To)cur); break;
-                default: assert(false);
-            }
-        }
-    }
-    o_ptr[cur_oh * ow + cur_ow] = res;
-}
-
-static void pool2d_nchw_kernel_f32_f32_cuda(
-        const int ih, const int iw, const int oh, const int ow,
-        const int kh, const int kw, const int sh, const int sw,
-        const int ph, const int pw, const int parallel_elements,
-        const float * src, float * dst, const enum ggml_op_pool op,
-        cudaStream_t stream) {
-
-    const int num_blocks = (parallel_elements + CUDA_POOL2D_BLOCK_SIZE - 1) / CUDA_POOL2D_BLOCK_SIZE;
-    dim3 block_nums(num_blocks);
-    pool2d_nchw_kernel<<<block_nums, CUDA_POOL2D_BLOCK_SIZE, 0, stream>>>(ih, iw, oh, ow, kh, kw, sh, sw, ph, pw, parallel_elements, src, dst, op);
-}
-
-void ggml_cuda_op_pool2d(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_F32);
-
-    const int32_t * opts = (const int32_t *)dst->op_params;
-    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
-    const int k0 = opts[1];
-    const int k1 = opts[2];
-    const int s0 = opts[3];
-    const int s1 = opts[4];
-    const int p0 = opts[5];
-    const int p1 = opts[6];
-
-    const int64_t IH = src0->ne[1];
-    const int64_t IW = src0->ne[0];
-
-    const int64_t N = dst->ne[3];
-    const int64_t OC = dst->ne[2];
-    const int64_t OH = dst->ne[1];
-    const int64_t OW = dst->ne[0];
-
-    const int parallel_elements = N * OC * OH * OW;
-
-    pool2d_nchw_kernel_f32_f32_cuda(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0, parallel_elements, src0_d, dst_d, op, stream);
-}

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

@@ -1,112 +0,0 @@
-#include "quantize.cuh"
-#include <cstdint>
-
-static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx0_padded) {
-    const int64_t ix0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (ix0 >= kx0_padded) {
-        return;
-    }
-
-    const int64_t ix1 = blockIdx.y;
-
-    const int64_t i_padded = ix1*kx0_padded + ix0;
-
-    block_q8_1 * y = (block_q8_1 *) vy;
-
-    const int64_t ib = i_padded / QK8_1; // block index
-    const int64_t iqs = i_padded % QK8_1; // quant index
-
-    const float xi = ix0 < kx ? x[ix1*kx + ix0] : 0.0f;
-    float amax = fabsf(xi);
-    float sum = xi;
-
-    amax = warp_reduce_max(amax);
-    sum = warp_reduce_sum(sum);
-
-    const float d = amax / 127;
-    const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
-
-    y[ib].qs[iqs] = q;
-
-    if (iqs > 0) {
-        return;
-    }
-
-    reinterpret_cast<half&>(y[ib].ds.x) = d;
-    reinterpret_cast<half&>(y[ib].ds.y) = sum;
-}
-
-template <bool need_sum>
-static __global__ void quantize_mmq_q8_1(
-    const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
-
-    const int64_t ix0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (ix0 >= kx0_padded) {
-        return;
-    }
-
-    const int64_t ix1 = kx1*blockIdx.z + blockIdx.y;
-
-    block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
-
-    const int64_t ib0 = blockIdx.z*(gridDim.y*gridDim.x*blockDim.x/(4*QK8_1)); // first block of channel
-    const int64_t ib  = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y;              // block index in channel
-    const int64_t iqs = ix0 % (4*QK8_1);                                       // quant index in block
-
-    const float xi = ix0 < kx0 ? x[ix1*kx0 + ix0] : 0.0f;
-    float amax = fabsf(xi);
-
-    amax = warp_reduce_max(amax);
-
-    float sum;
-    if (need_sum) {
-        sum = warp_reduce_sum(xi);
-    }
-
-    const float d = amax / 127;
-    const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
-
-    y[ib].qs[iqs] = q;
-
-    if (iqs % QK8_1 != 0) {
-        return;
-    }
-
-    if (need_sum) {
-        y[ib].ds[iqs/QK8_1] = make_half2(d, sum);
-    } else {
-        ((float *) y[ib].ds)[iqs/QK8_1] = d;
-    }
-}
-
-void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
-    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
-
-    GGML_ASSERT(kx0_padded % QK8_1 == 0);
-
-    const int64_t block_num_x = (kx0_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
-    const dim3 num_blocks(block_num_x, kx1*channels, 1);
-    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx0_padded);
-
-    GGML_UNUSED(type_x);
-}
-
-void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
-    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
-
-    GGML_ASSERT(kx0_padded % (4*QK8_1) == 0);
-
-    const int64_t block_num_x = (kx0_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
-    const dim3 num_blocks(block_num_x, kx1, channels);
-    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    if (mmq_need_sum(type_x)) {
-        quantize_mmq_q8_1<true><<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
-    } else {
-        quantize_mmq_q8_1<false><<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
-    }
-}