jiuge.cpp

#include "jiuge_impl.hpp"
#include "jiuge_weight.hpp"

#include "../../tensor.hpp"
#include "../../utils.hpp"
#include "../inference_context.hpp"
#include "infinicore_infer.h"

#include <random>
#include <thread>
#include <vector>

void createDeviceResource(DeviceResource *rsrc, const JiugeMeta *meta,
                          const JiugeWeights *weights,
                          infiniDevice_t device, int idev,
                          int ndev, int dev_id,
                          infinicclComm_t comm) {
    RUN_INFINI(infinirtSetDevice(device, dev_id));
    infiniopHandle_t handle;
    infiniopCreateHandle(&handle);
    infinirtStream_t stream;
    infinirtStreamCreate(&stream);

    std::vector<std::shared_ptr<Tensor>> w_attn_norm, w_attn_qkv, b_attn_qkv, w_attn_out,
        w_ffn_norm, w_ffn_gate_up, w_ffn_down;
    for (size_t layer = 0; layer < meta->nlayer; layer++) {
        w_attn_norm.push_back(
            getAttnNorm(meta, weights, layer));
        w_attn_qkv.push_back(
            getAttnQKV(meta, weights, layer, idev, ndev));
        if (weights->attn_qkv_b != nullptr) {
            b_attn_qkv.push_back(
                getAttnQKVBias(meta, weights, layer, idev, ndev));
        }
        w_attn_out.push_back(
            getAttnO(meta, weights, layer, idev, ndev));
        w_ffn_norm.push_back(
            getFFNNorm(meta, weights, layer));
        w_ffn_gate_up.push_back(
            getFFNGateUp(meta, weights, layer, idev, ndev));
        w_ffn_down.push_back(
            getFFNDown(meta, weights, layer, idev, ndev));
    }

    auto memory_pool = std::make_shared<MemoryPool>(128 * 1024 * 1024);

    *rsrc = DeviceResource{
        device,
        dev_id,
        handle,
        getInEmbd(meta, weights),
        getOutNorm(meta, weights),
        getOutEmbd(meta, weights),
        getSinTable(meta),
        getCosTable(meta),
        w_attn_norm,
        w_attn_qkv,
        b_attn_qkv,
        w_attn_out,
        w_ffn_norm,
        w_ffn_gate_up,
        w_ffn_down,
        stream,
        comm,
        memory_pool,
    };
    RUN_INFINI(infinirtDeviceSynchronize());
}

void releaseDeviceResource(DeviceResource &res) {
    infinirtDeviceSynchronize();
    // Release individual Tensors
    res.w_in_embd.reset();
    res.w_out_norm.reset();
    res.w_out_embd.reset();
    res.sin_table.reset();
    res.cos_table.reset();
    for (auto &t : res.w_attn_norm) {
        t.reset();
    }
    res.w_attn_norm.clear();
    for (auto &t : res.w_attn_qkv) {
        t.reset();
    }
    res.w_attn_qkv.clear();
    for (auto &t : res.b_attn_qkv) {
        t.reset();
    }
    res.b_attn_qkv.clear();
    for (auto &t : res.w_attn_out) {
        t.reset();
    }
    res.w_attn_out.clear();
    for (auto &t : res.w_ffn_norm) {
        t.reset();
    }
    res.w_ffn_norm.clear();
    for (auto &t : res.w_ffn_gate_up) {
        t.reset();
    }
    res.w_ffn_gate_up.clear();
    for (auto &t : res.w_ffn_down) {
        t.reset();
    }
    res.w_ffn_down.clear();
    infiniopDestroyHandle(res.handle);
    res.handle = nullptr;
    infinirtStreamDestroy(res.stream);
    res.stream = nullptr;
    infinicclCommDestroy(res.comm);
    res.comm = nullptr;
}

void inferDeviceBatch(const JiugeMeta &meta, DeviceResource &rsrc,
                      uint32_t idev, uint32_t ndev,
                      const uint32_t *tokens, uint32_t ntok,
                      const uint32_t *req_lens, uint32_t nreq, const uint32_t *req_pos,
                      struct KVCache **kv_caches,
                      const float *temperature, const uint32_t *topk, const float *topp,
                      uint32_t *output, InferenceContext &ctx) {
    auto nlayer = meta.nlayer;
    auto nkvh = meta.nkvh / ndev;
    auto nh = meta.nh / ndev;
    auto ngroup = nh / nkvh;
    // auto dctx = meta.dctx;
    auto dh = meta.dh;
    auto d = meta.d;
    auto dt_logits = meta.dt_logits;
    auto di = meta.di / ndev;
    auto dvoc = meta.dvoc;
    auto stream = rsrc.stream;
    bool has_qkv_bias = rsrc.b_attn_qkv.size() > 0;

    // Allocate buffers
    auto logits_in = Tensor::buffer(dt_logits, {ntok, d}, rsrc.memory_pool);
    auto logits_out = Tensor::buffer(dt_logits, {ntok, d}, rsrc.memory_pool);
    auto qkv_buf = Tensor::buffer(dt_logits, {ntok, (nh + nkvh * 2) * dh}, rsrc.memory_pool);
    auto gate_up_buf = Tensor::buffer(dt_logits, {ntok, 2 * di}, rsrc.memory_pool);
    auto o_buf = Tensor::buffer(dt_logits, {ntok, nh * dh}, rsrc.memory_pool);
    auto prob_buf = Tensor::buffer(dt_logits, {nreq, dvoc}, rsrc.memory_pool);
    auto result_buf = Tensor::buffer(INFINI_DTYPE_I64, {nreq}, rsrc.memory_pool);
    auto result_cpu = std::vector<int64_t>(nreq);

    // Prepare inputs
    auto batch_pos_ids = std::vector<uint32_t>(ntok);
    size_t req_start = 0;
    for (uint32_t req = 0; req < nreq; req++) {
        for (uint32_t i = 0; i < req_lens[req]; i++) {
            batch_pos_ids[req_start + i] = req_pos[req] + i;
        }
        req_start += req_lens[req];
    }

    std::shared_ptr<Tensor> pos_ids_buf;
    if (rsrc.device == INFINI_DEVICE_CPU) {
        pos_ids_buf = Tensor::weight(batch_pos_ids.data(), INFINI_DTYPE_U32, {ntok});
    } else {
        pos_ids_buf = Tensor::buffer(INFINI_DTYPE_U32, {ntok}, rsrc.memory_pool);
        RUN_INFINI(infinirtMemcpyAsync(pos_ids_buf->data(), batch_pos_ids.data(), sizeof(uint32_t) * ntok,
                                       INFINIRT_MEMCPY_H2D, stream));
    }
    for (uint32_t i = 0; i < ntok; i++) {
        RUN_INFINI(infinirtMemcpyAsync(logits_in->data(i * d),
                                       rsrc.w_in_embd->data(tokens[i] * d),
                                       dsize(dt_logits) * d, INFINIRT_MEMCPY_D2D, stream));
    }

    // Attention
    // attention inner
    size_t max_qk_size = 0;
    size_t max_seq_len = 0;

    for (uint32_t req = 0; req < nreq; req++) {
        auto past_len = req_pos[req];
        auto seq_len = req_lens[req];
        auto total_len = past_len + seq_len;

        max_qk_size = std::max(max_qk_size, size_t(seq_len * total_len));
        max_seq_len = std::max(max_seq_len, size_t(seq_len));
    }

    auto qk_buf = Tensor::buffer(dt_logits, {nh, max_qk_size}, rsrc.memory_pool);
    auto rearrange_q_buf = Tensor::buffer(dt_logits, {nkvh, ngroup * max_seq_len, dh}, rsrc.memory_pool);
    auto attn_val_buf = Tensor::buffer(dt_logits, {nh, max_seq_len, dh}, rsrc.memory_pool);

    // MLP buffers
    auto gate_buf = gate_up_buf->slice(1, 0, di);
    auto up_buf = gate_up_buf->slice(1, di, di);

    // Compute
    for (uint32_t layer = 0; layer < nlayer; layer++) {
        // 1. Attention
        // rms norm
        ctx.rmsnorm(logits_out, logits_in, rsrc.w_attn_norm[layer], meta.epsilon);
        // qkv_proj
        if (has_qkv_bias) {
            ctx.rearrange(qkv_buf, rsrc.b_attn_qkv[layer]->view({ntok, (nh + nkvh * 2) * dh}, {0, 1}));
        }
        ctx.linear(qkv_buf, logits_out, rsrc.w_attn_qkv[layer], 1.0, 0.0, has_qkv_bias ? qkv_buf : nullptr);
        // rope
        auto qkv_rope = qkv_buf->viewReshaped({ntok, nh + nkvh * 2, dh});
        ctx.rope(qkv_rope->slice(1, 0, nh), qkv_rope->slice(1, 0, nh), pos_ids_buf, rsrc.sin_table, rsrc.cos_table);
        ctx.rope(qkv_rope->slice(1, nh, nkvh), qkv_rope->slice(1, nh, nkvh), pos_ids_buf, rsrc.sin_table, rsrc.cos_table);

        size_t token_offset = 0;
        for (uint32_t req = 0; req < nreq; req++) {
            auto past_len = req_pos[req];
            auto seq_len = req_lens[req];
            auto total_len = past_len + seq_len;
            auto o = o_buf->viewReshaped({ntok, nh, dh})->slice({{0, token_offset, seq_len}})->dimSplit(1, {nkvh, ngroup})->permute({1, 2, 0, 3});
            auto q = qkv_rope->slice({{0, token_offset, seq_len}, {1, 0, nh}})->dimSplit(1, {nkvh, ngroup})->permute({1, 2, 0, 3});
            auto k = qkv_rope->slice({{0, token_offset, seq_len}, {1, nh, nkvh}});
            auto v = qkv_rope->slice({{0, token_offset, seq_len}, {1, nh + nkvh, nkvh}});

            // self attention
            // concat
            ctx.rearrange(kv_caches[req]->k[idev][layer]->slice(0, past_len, seq_len), k);
            ctx.rearrange(kv_caches[req]->v[idev][layer]->slice(0, past_len, seq_len), v);
            // qk
            auto q_rearrange = rearrange_q_buf->viewReshaped({nkvh, ngroup, seq_len, dh});
            ctx.rearrange(q_rearrange, q);
            auto qk_gemm = qk_buf->viewReshaped({nkvh, ngroup * seq_len, total_len});
            auto k_gemm = kv_caches[req]->k[idev][layer]->slice(0, 0, total_len)->permute({1, 2, 0});
            ctx.linear(qk_gemm, rearrange_q_buf, k_gemm, 1. / sqrt(dh), 0.0, nullptr);
            // softmax
            auto qk_softmax = qk_buf->viewReshaped({nh, seq_len, total_len});
            ctx.causalSoftmax(qk_softmax, qk_softmax);
            auto v_gemm = kv_caches[req]->v[idev][layer]->slice(0, 0, total_len)->permute({1, 0, 2});
            ctx.linear(attn_val_buf, qk_gemm, v_gemm, 1.0, 0.0, nullptr);
            // rearrange attn val
            auto attn_val_gemm = attn_val_buf->viewReshaped({nkvh, ngroup, max_seq_len, dh});
            ctx.rearrange(o, attn_val_gemm);

            token_offset += seq_len;
        }

        // o_proj
        ctx.linear(logits_in, o_buf, rsrc.w_attn_out[layer], 1.0, 0.0, idev == 0 ? logits_in : nullptr); // only rank 0 adds residual

        // All_reduce if distributed
        if (rsrc.comm != nullptr) {
            RUN_INFINI(infinicclAllReduce(
                logits_in->data(), logits_in->data(), ntok * d, dt_logits,
                INFINICCL_SUM, rsrc.comm, stream));
            RUN_INFINI(infinirtStreamSynchronize(stream));
        }
        // 2. FFN
        ctx.rmsnorm(logits_out, logits_in, rsrc.w_ffn_norm[layer], meta.epsilon);
        ctx.linear(gate_up_buf, logits_out, rsrc.w_ffn_gate_up[layer], 1.0, 0.0, nullptr);
        ctx.swiglu(gate_buf, up_buf, gate_buf);
        ctx.linear(logits_in, gate_buf, rsrc.w_ffn_down[layer], 1.0, 0.0, idev == 0 ? logits_in : nullptr); // only rank 0 adds residual

        // All_reduce if distributed
        if (rsrc.comm != nullptr) {
            RUN_INFINI(infinicclAllReduce(
                logits_in->data(), logits_in->data(), ntok * d, dt_logits,
                INFINICCL_SUM, rsrc.comm, stream));
            RUN_INFINI(infinirtStreamSynchronize(stream));
        }
    }
    // Sample and Output
    if (idev == 0) {
        size_t token_offset = 0;
        for (uint32_t req = 0; req < nreq; req++) {
            auto seq_len = req_lens[req];
            token_offset += seq_len;
            ctx.rmsnorm(logits_out->slice(0, req, 1),
                        logits_in->slice(0, token_offset - 1, 1),
                        rsrc.w_out_norm,
                        meta.epsilon);
        }
        ctx.linear(prob_buf, logits_out->slice(0, 0, nreq), rsrc.w_out_embd, 1.0, 0.0, nullptr);
        std::random_device _rd;
        std::mt19937 gen(_rd());
        token_offset = 0;
        for (uint32_t req = 0; req < nreq; req++) {
            auto seq_len = req_lens[req];
            float random_val = std::uniform_real_distribution<float>(0, 1)(gen);
            ctx.randomSample(result_buf->view({}, {}),
                             prob_buf->view({dvoc}, {1}),
                             random_val, topp[req], topk[req], temperature[req]);
            token_offset += seq_len;
        }
        RUN_INFINI(infinirtStreamSynchronize(stream));
        RUN_INFINI(infinirtMemcpy(result_cpu.data(), result_buf->data(),
                                  sizeof(int64_t) * nreq, INFINIRT_MEMCPY_D2H));
        for (uint32_t req = 0; req < nreq; req++) {
            output[req] = result_cpu[req];
        }
    }
}

__C void
inferBatch(struct JiugeModel *model,
           const uint32_t *tokens, uint32_t ntok,
           const uint32_t *req_lens, uint32_t nreq, const uint32_t *req_pos,
           struct KVCache **kv_caches,
           const float *temperature, const uint32_t *topk, const float *topp,
           uint32_t *output) {
    model->req.tokens = tokens;
    model->req.ntok = ntok;
    model->req.req_lens = req_lens;
    model->req.nreq = nreq;
    model->req.req_pos = req_pos;
    model->req.kv_caches = kv_caches;
    model->req.output = output;
    model->req.temperature = temperature;
    model->req.topk = topk;
    model->req.topp = topp;

    for (size_t idev = 0; idev < model->dev_ids.size(); idev++) {
        std::unique_lock<std::mutex> lock(model->states[idev].mtx);
        model->states[idev].proceed = true;
        lock.unlock();
        model->states[idev].cv_start.notify_one();
    }
    for (size_t i = model->dev_ids.size(); i > 0; i--) {
        auto idev = i - 1;
        std::unique_lock<std::mutex> lock(model->states[idev].mtx);
        model->states[idev].cv_done.wait(lock, [&] { return !(model->states[idev].proceed); });
        lock.unlock();
    }
}

void launchDevice(const JiugeMeta &meta, const JiugeWeights *weights, DeviceResource *rsrc, InferState &state, InferRequest &req,
                  infiniDevice_t device, int idev, int ndev, int dev_id, infinicclComm_t comm) {
    CacheManager cache_manager(256);
    InferenceContext ctx(rsrc, &cache_manager, rsrc->stream);

    // Create Device Resource
    createDeviceResource(rsrc, &meta, weights, device, idev, ndev, dev_id, comm);
    {
        std::unique_lock<std::mutex> lock(state.mtx);
        state.loaded = true;
        lock.unlock();
        state.cv_load.notify_one();
    }

    // Infer Loop
    while (true) {
        std::unique_lock<std::mutex> lock(state.mtx);
        state.cv_start.wait(lock, [&] { return state.proceed || state.exit_flag; });
        // quit if exit_flag is set
        if (state.exit_flag) {
            break;
        }

        inferDeviceBatch(meta, *rsrc, idev, ndev, req.tokens, req.ntok,
                         req.req_lens, req.nreq, req.req_pos, req.kv_caches,
                         req.temperature, req.topk, req.topp, req.output,
                         ctx);

        state.proceed = false;
        lock.unlock();
        state.cv_done.notify_one();
    }

    // Clean-Up
    releaseDeviceResource(*rsrc);
}

JiugeModel::JiugeModel(const JiugeMeta *_meta, const JiugeWeights *weights, infiniDevice_t device_, std::vector<int> device_ids) : meta(*_meta) {
    int ndev = int(device_ids.size());
    device = device_;
    dev_ids = device_ids;
    dev_resources = std::vector<DeviceResource>(ndev);
    states = std::vector<InferState>(ndev);
    threads.resize(ndev);
    RUN_INFINI(infinirtInit());
    auto comms = std::vector<infinicclComm_t>(ndev, nullptr);
    if (ndev > 1) {
        RUN_INFINI(infinicclCommInitAll(device, comms.data(), ndev, dev_ids.data()));
    }

    for (int i = 0; i < ndev; i++) {
        threads[i] = std::thread(launchDevice, std::cref(meta), weights, &dev_resources[i], std::ref(states[i]), std::ref(req), device, i, ndev, dev_ids[i], comms[i]);
    }
    for (int i = 0; i < ndev; i++) {
        std::unique_lock<std::mutex> lock(states[i].mtx);
        states[i].cv_load.wait(lock, [&] { return states[i].loaded; });
        lock.unlock();
    }
}

__C struct JiugeModel *
createJiugeModel(const JiugeMeta *meta,
                 const JiugeWeights *weights,
                 infiniDevice_t device,
                 int ndev,
                 const int *dev_ids) {
    std::vector<int> device_ids(ndev);
    std::copy(dev_ids, dev_ids + ndev, device_ids.begin());
    JiugeModel *model = new JiugeModel(meta, weights, device, device_ids);
    return model;
}

__C void destroyJiugeModel(struct JiugeModel *model) {
    auto ndev = model->dev_resources.size();

    for (size_t idev = 0; idev < ndev; idev++) {
        std::unique_lock<std::mutex> lock(model->states[idev].mtx);
        model->states[idev].exit_flag = true;
        lock.unlock();
        model->states[idev].cv_start.notify_one();
    }

    for (size_t idev = 0; idev < ndev; idev++) {
        model->threads[idev].join();
    }

    delete model;
}