ollama/llama/ggml-cuda/mmq.cu

/**
 * llama.cpp - commit 46e3556e01b824e52395fb050b29804b6cff2a7c - do not edit this file
 *
 * MIT License
 *
 * Copyright (c) 2023-2024 The ggml authors
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to deal
 * in the Software without restriction, including without limitation the rights
 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#include "mmq.cuh"

void ggml_cuda_op_mul_mat_q(
    ggml_backend_cuda_context & ctx,
    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
    const int64_t src1_padded_row_size, cudaStream_t stream) {

    const int64_t ne00 = src0->ne[0];

    const int64_t ne10 = src1->ne[0];
    const int64_t ne11 = src1->ne[1];
    GGML_ASSERT(ne10 % QK8_1 == 0);

    const int64_t ne0 = dst->ne[0];

    const int64_t row_diff = row_high - row_low;
    const int64_t stride00 = ne00 / ggml_blck_size(src0->type);

    int id = ggml_cuda_get_device();
    const int compute_capability = ggml_cuda_info().devices[id].cc;

    // the main device has a larger memory buffer to hold the results from all GPUs
    // nrows_dst == nrows of the matrix that the kernel writes into
    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;

    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
    // Also its fixup needs to allocate a temporary buffer in the memory pool.
    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
    const bool use_stream_k = compute_capability >= GGML_CUDA_CC_VOLTA && compute_capability < GGML_CUDA_CC_OFFSET_AMD && src1_ncols == ne11;
    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};

    switch (src0->type) {
        case GGML_TYPE_Q4_0:
            mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
            break;
        case GGML_TYPE_Q4_1:
            mul_mat_q_case<GGML_TYPE_Q4_1>(ctx, args, stream);
            break;
        case GGML_TYPE_Q5_0:
            mul_mat_q_case<GGML_TYPE_Q5_0>(ctx, args, stream);
            break;
        case GGML_TYPE_Q5_1:
            mul_mat_q_case<GGML_TYPE_Q5_1>(ctx, args, stream);
            break;
        case GGML_TYPE_Q8_0:
            mul_mat_q_case<GGML_TYPE_Q8_0>(ctx, args, stream);
            break;
        case GGML_TYPE_Q2_K:
            mul_mat_q_case<GGML_TYPE_Q2_K>(ctx, args, stream);
            break;
        case GGML_TYPE_Q3_K:
            mul_mat_q_case<GGML_TYPE_Q3_K>(ctx, args, stream);
            break;
        case GGML_TYPE_Q4_K:
            mul_mat_q_case<GGML_TYPE_Q4_K>(ctx, args, stream);
            break;
        case GGML_TYPE_Q5_K:
            mul_mat_q_case<GGML_TYPE_Q5_K>(ctx, args, stream);
            break;
        case GGML_TYPE_Q6_K:
            mul_mat_q_case<GGML_TYPE_Q6_K>(ctx, args, stream);
            break;
        case GGML_TYPE_IQ2_XXS:
            mul_mat_q_case<GGML_TYPE_IQ2_XXS>(ctx, args, stream);
            break;
        case GGML_TYPE_IQ2_XS:
            mul_mat_q_case<GGML_TYPE_IQ2_XS>(ctx, args, stream);
            break;
        case GGML_TYPE_IQ2_S:
            mul_mat_q_case<GGML_TYPE_IQ2_S>(ctx, args, stream);
            break;
        case GGML_TYPE_IQ3_XXS:
            mul_mat_q_case<GGML_TYPE_IQ3_XXS>(ctx, args, stream);
            break;
        case GGML_TYPE_IQ3_S:
            mul_mat_q_case<GGML_TYPE_IQ3_S>(ctx, args, stream);
            break;
        case GGML_TYPE_IQ1_S:
            mul_mat_q_case<GGML_TYPE_IQ1_S>(ctx, args, stream);
            break;
        case GGML_TYPE_IQ4_XS:
            mul_mat_q_case<GGML_TYPE_IQ4_XS>(ctx, args, stream);
            break;
        case GGML_TYPE_IQ4_NL:
            mul_mat_q_case<GGML_TYPE_IQ4_NL>(ctx, args, stream);
            break;
        default:
            GGML_ABORT("fatal error");
            break;
    }

    GGML_UNUSED(src1);
    GGML_UNUSED(dst);
    GGML_UNUSED(src1_ddf_i);
}

bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
#ifdef GGML_CUDA_FORCE_CUBLAS
    return false;
#endif // GGML_CUDA_FORCE_CUBLAS

    bool mmq_supported;

    switch (type) {
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
        case GGML_TYPE_Q5_1:
        case GGML_TYPE_Q8_0:
        case GGML_TYPE_Q2_K:
        case GGML_TYPE_Q3_K:
        case GGML_TYPE_Q4_K:
        case GGML_TYPE_Q5_K:
        case GGML_TYPE_Q6_K:
        case GGML_TYPE_IQ2_XXS:
        case GGML_TYPE_IQ2_XS:
        case GGML_TYPE_IQ2_S:
        case GGML_TYPE_IQ3_XXS:
        case GGML_TYPE_IQ3_S:
        case GGML_TYPE_IQ1_S:
        case GGML_TYPE_IQ4_XS:
        case GGML_TYPE_IQ4_NL:
            mmq_supported = true;
            break;
        default:
            mmq_supported = false;
            break;
    }

    if (!mmq_supported) {
        return false;
    }

    if (int8_mma_available(cc)) {
        return true;
    }

    if (cc < GGML_CUDA_CC_DP4A) {
        return false;
    }

#ifdef GGML_CUDA_FORCE_MMQ
    return true;
#endif //GGML_CUDA_FORCE_MMQ

    if (cc < GGML_CUDA_CC_OFFSET_AMD) {
        return cc < GGML_CUDA_CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
    }

    return (cc < GGML_CUDA_CC_RDNA3 && cc != GGML_CUDA_CC_CDNA && cc != GGML_CUDA_CC_VEGA20) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
}