llama.cpp/src/llama-context.cpp

#include "llama-context.h"

#include "llama-impl.h"
#include "llama-io.h"
#include "llama-mmap.h"
#include "llama-model.h"
#include "llama-kv-cache.h"

#include <cassert>
#include <cstring>
#include <stdexcept>
#include <cinttypes>

//
// llama_context
//

llama_context::llama_context(
        const llama_model & model,
              llama_context_params params) :
    model(model) {
    LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__);

    t_start_us = model.t_start_us;
    t_load_us  = model.t_load_us;

    const auto & hparams = model.hparams;

    cparams.n_seq_max        = std::max(1u, params.n_seq_max);
    cparams.n_threads        = params.n_threads;
    cparams.n_threads_batch  = params.n_threads_batch;
    cparams.yarn_ext_factor  = params.yarn_ext_factor;
    cparams.yarn_attn_factor = params.yarn_attn_factor;
    cparams.yarn_beta_fast   = params.yarn_beta_fast;
    cparams.yarn_beta_slow   = params.yarn_beta_slow;
    cparams.defrag_thold     = params.defrag_thold;
    cparams.embeddings       = params.embeddings;
    cparams.offload_kqv      = params.offload_kqv;
    cparams.flash_attn       = params.flash_attn;
    cparams.no_perf          = params.no_perf;
    cparams.pooling_type     = params.pooling_type;
    cparams.warmup           = false;

    cparams.n_ctx            = params.n_ctx           == 0    ? hparams.n_ctx_train           : params.n_ctx;
    cparams.rope_freq_base   = params.rope_freq_base  == 0.0f ? hparams.rope_freq_base_train  : params.rope_freq_base;
    cparams.rope_freq_scale  = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale;

    cparams.n_ctx_orig_yarn  = params.yarn_orig_ctx    != 0 ? params.yarn_orig_ctx    :
                               hparams.n_ctx_orig_yarn != 0 ? hparams.n_ctx_orig_yarn :
                                                              hparams.n_ctx_train;

    cparams.cb_eval           = params.cb_eval;
    cparams.cb_eval_user_data = params.cb_eval_user_data;

    auto rope_scaling_type = params.rope_scaling_type;
    if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED) {
        rope_scaling_type = hparams.rope_scaling_type_train;
    }

    if (rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_NONE) {
        cparams.rope_freq_scale = 1.0f; // never scale if scaling type is none
    }

    if (cparams.yarn_ext_factor < 0.0f) { // negative indicates 'not set'
        cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_TYPE_YARN ? 1.0f : 0.0f;
    }

    cparams.yarn_attn_factor *= hparams.rope_attn_factor;

    if (cparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
        if (hparams.pooling_type == LLAMA_POOLING_TYPE_UNSPECIFIED) {
            cparams.pooling_type = LLAMA_POOLING_TYPE_NONE;
        } else {
            cparams.pooling_type = hparams.pooling_type;
        }
    }

    if (params.attention_type == LLAMA_ATTENTION_TYPE_UNSPECIFIED) {
        cparams.causal_attn = hparams.causal_attn;
    } else {
        cparams.causal_attn = params.attention_type == LLAMA_ATTENTION_TYPE_CAUSAL;
    }

    // with causal attention, the batch size is limited by the context size
    cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;

    // the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
    // this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
    // ref: https://github.com/ggerganov/llama.cpp/pull/5021
    // TODO: this padding is not needed for the cache-less context so we should probably move it to llama_context_kv_self
    if (cparams.n_batch < GGML_KQ_MASK_PAD) {
        LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
        cparams.n_batch = GGML_KQ_MASK_PAD;
    }

    cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);

    const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;

    LLAMA_LOG_INFO("%s: n_seq_max     = %u\n",   __func__, cparams.n_seq_max);
    LLAMA_LOG_INFO("%s: n_ctx         = %u\n",   __func__, cparams.n_ctx);
    LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n",   __func__, n_ctx_per_seq);
    LLAMA_LOG_INFO("%s: n_batch       = %u\n",   __func__, cparams.n_batch);
    LLAMA_LOG_INFO("%s: n_ubatch      = %u\n",   __func__, cparams.n_ubatch);
    LLAMA_LOG_INFO("%s: causal_attn   = %d\n",   __func__, cparams.causal_attn);
    LLAMA_LOG_INFO("%s: flash_attn    = %d\n",   __func__, cparams.flash_attn);
    LLAMA_LOG_INFO("%s: freq_base     = %.1f\n", __func__, cparams.rope_freq_base);
    LLAMA_LOG_INFO("%s: freq_scale    = %g\n",   __func__, cparams.rope_freq_scale);

    if (n_ctx_per_seq < hparams.n_ctx_train) {
        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n",
                __func__, n_ctx_per_seq, hparams.n_ctx_train);
    }

    if (n_ctx_per_seq > hparams.n_ctx_train) {
        LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
                __func__, n_ctx_per_seq, hparams.n_ctx_train);
    }

    logits_all = params.logits_all;

    if (!hparams.vocab_only) {
        // GPU backends
        for (auto * dev : model.devices) {
            ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr);
            if (backend == nullptr) {
                throw std::runtime_error(format("failed to initialize %s backend", ggml_backend_dev_name(dev)));
            }
            backends.emplace_back(backend);
        }

        // add ACCEL backends (such as BLAS)
        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
            if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_ACCEL) {
                ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr);
                if (backend == nullptr) {
                    throw std::runtime_error(format("failed to initialize %s backend", ggml_backend_dev_name(dev)));
                }
                backends.emplace_back(backend);
            }
        }

        // add CPU backend
        backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
        if (backend_cpu == nullptr) {
            throw std::runtime_error("failed to initialize CPU backend");
        }
        backends.emplace_back(backend_cpu);

        // create a list of the set_n_threads functions in the backends
        for (auto & backend : backends) {
            ggml_backend_dev_t dev = ggml_backend_get_device(backend.get());
            ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
            if (reg) {
                auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
                if (ggml_backend_set_n_threads_fn) {
                    set_n_threads_fns.emplace_back(backend.get(), ggml_backend_set_n_threads_fn);
                }
            }
        }

        llama_set_abort_callback(this, params.abort_callback, params.abort_callback_data);

        // graph outputs buffer
        {
            // resized during inference when a batch uses more outputs
            if ((uint32_t) output_reserve(params.n_seq_max) < params.n_seq_max) {
                throw std::runtime_error("failed to reserve initial output buffer");
            }

            LLAMA_LOG_INFO("%s: %10s  output buffer size = %8.2f MiB\n", __func__,
                    ggml_backend_buffer_name    (buf_output.get()),
                    ggml_backend_buffer_get_size(buf_output.get()) / 1024.0 / 1024.0);
        }
    }

    // init the memory module
    // TODO: for now, always create a unified KV cache
    if (!hparams.vocab_only) {
        kv_self.reset(static_cast<llama_kv_cache_unified *>(model.create_memory()));

        LLAMA_LOG_DEBUG("%s: n_ctx = %u\n", __func__, cparams.n_ctx);

        cparams.n_ctx = GGML_PAD(cparams.n_ctx, kv_self->get_padding(cparams));

        LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx);

        uint32_t kv_size = cparams.n_ctx;
        ggml_type type_k = params.type_k;
        ggml_type type_v = params.type_v;

        if (llama_model_is_recurrent(&model)) {
            // Mamba needs at least as many KV cells as there are sequences kept at any time
            kv_size = std::max((uint32_t) 1, params.n_seq_max);
            // it's probably best to keep as much precision as possible for the states
            type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states
            type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states
        }

        GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0);
        GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0);

        if (!kv_self->init(model, cparams, type_k, type_v, kv_size, cparams.offload_kqv)) {
            throw std::runtime_error("failed to initialize self-attention cache");
        }

        {
            const size_t memory_size_k = kv_self->size_k_bytes();
            const size_t memory_size_v = kv_self->size_v_bytes();

            LLAMA_LOG_INFO("%s: KV self size  = %7.2f MiB, K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
                    (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f),
                    ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
                    ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
        }
    }

    // init backends
    if (!hparams.vocab_only) {
        LLAMA_LOG_DEBUG("%s: enumerating backends\n", __func__);

        backend_buft.clear();
        backend_ptrs.clear();

        for (auto & backend : backends) {
            auto * buft = ggml_backend_get_default_buffer_type(backend.get());
            auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));

            if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model.devices.empty()) {
                // use the host buffer of the first device CPU for faster transfer of the intermediate state
                auto * dev = model.devices[0];
                auto * host_buft = ggml_backend_dev_host_buffer_type(dev);
                if (host_buft) {
                    buft = host_buft;
                }
            }

            backend_buft.push_back(buft);
            backend_ptrs.push_back(backend.get());
        }

        LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());

        const size_t max_nodes = this->graph_max_nodes();

        LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);

        // buffer used to store the computation graph and the tensor meta data
        buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));

        // TODO: move these checks to ggml_backend_sched
        // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
        bool pipeline_parallel =
            model.n_devices() > 1 &&
            model.params.n_gpu_layers > (int) model.hparams.n_layer &&
            model.params.split_mode == LLAMA_SPLIT_MODE_LAYER &&
            cparams.offload_kqv;

        // pipeline parallelism requires support for async compute and events in all devices
        if (pipeline_parallel) {
            for (auto & backend : backends) {
                auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
                if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) {
                    // ignore CPU backend
                    continue;
                }
                auto * dev = ggml_backend_get_device(backend.get());
                ggml_backend_dev_props props;
                ggml_backend_dev_get_props(dev, &props);
                if (!props.caps.async || !props.caps.events) {
                    // device does not support async compute or events
                    pipeline_parallel = false;
                    break;
                }
            }
        }

        sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel));

        if (pipeline_parallel) {
            LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
        }
    }

    // reserve worst-case graph
    if (!hparams.vocab_only) {
        const uint32_t n_seqs = 1; // TODO: worst-case number of sequences
        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);

        llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph

        // restore later
        // TODO: something cleaner
        const auto n_outputs_save = n_outputs;

        LLAMA_LOG_DEBUG("%s: worst-case: n_tokens = %d, n_seqs = %d, n_outputs = %d\n", __func__, n_tokens, n_seqs, n_outputs);

        int n_splits_pp = -1;
        int n_nodes_pp  = -1;

        int n_splits_tg = -1;
        int n_nodes_tg  = -1;

        // simulate full KV cache
        kv_self->n = kv_self->size;

        cross.v_embd.clear();

        // reserve pp graph first so that buffers are only allocated once
        {
            llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};

            // max number of outputs
            n_outputs = ubatch_pp.n_tokens;

            LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_pp.n_tokens, ubatch_pp.n_seqs);

            auto * gf = graph_init();
            graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT);

            if (!ggml_backend_sched_reserve(sched.get(), gf)) {
                throw std::runtime_error("failed to allocate compute pp buffers");
            }

            n_splits_pp = ggml_backend_sched_get_n_splits(sched.get());
            n_nodes_pp  = ggml_graph_n_nodes(gf);
        }

        // reserve with tg graph to get the number of splits and nodes
        {
            llama_ubatch ubatch_tg = { true, 1, 1, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};

            n_outputs = ubatch_tg.n_tokens;

            LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_tg.n_tokens, ubatch_tg.n_seqs);

            auto * gf = graph_init();
            graph_build(ctx_compute.get(), gf, ubatch_tg, LLM_GRAPH_TYPE_DEFAULT);

            if (!ggml_backend_sched_reserve(sched.get(), gf)) {
                throw std::runtime_error("failed to allocate compute tg buffers");
            }

            n_splits_tg = ggml_backend_sched_get_n_splits(sched.get());
            n_nodes_tg  = ggml_graph_n_nodes(gf);
        }

        // reserve again with pp graph to avoid ggml-alloc reallocations during inference
        {
            llama_ubatch ubatch_pp = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};

            n_outputs = ubatch_pp.n_tokens;

            LLAMA_LOG_DEBUG("%s: reserving graph for n_tokens = %d, n_seqs = %d\n", __func__, ubatch_pp.n_tokens, ubatch_pp.n_seqs);

            auto * gf = graph_init();
            graph_build(ctx_compute.get(), gf, ubatch_pp, LLM_GRAPH_TYPE_DEFAULT);

            if (!ggml_backend_sched_reserve(sched.get(), gf)) {
                throw std::runtime_error("failed to allocate compute pp buffers");
            }
        }

        n_outputs = n_outputs_save;

        for (size_t i = 0; i < backend_ptrs.size(); ++i) {
            ggml_backend_t             backend = backend_ptrs[i];
            ggml_backend_buffer_type_t buft    = backend_buft[i];
            size_t size = ggml_backend_sched_get_buffer_size(sched.get(), backend);
            if (size > 1) {
                LLAMA_LOG_INFO("%s: %10s compute buffer size = %8.2f MiB\n", __func__,
                        ggml_backend_buft_name(buft),
                        size / 1024.0 / 1024.0);
            }
        }

        if (n_nodes_pp == n_nodes_tg) {
            LLAMA_LOG_INFO("%s: graph nodes  = %d\n", __func__, n_nodes_pp);
        } else {
            LLAMA_LOG_INFO("%s: graph nodes  = %d (with bs=%d), %d (with bs=1)\n", __func__, n_nodes_pp, n_tokens, n_nodes_tg);
        }

        if (n_splits_pp == n_splits_tg) {
            LLAMA_LOG_INFO("%s: graph splits = %d\n", __func__, n_splits_pp);
        } else {
            LLAMA_LOG_INFO("%s: graph splits = %d (with bs=%d), %d (with bs=1)\n", __func__, n_splits_pp, n_tokens, n_splits_tg);
        }
    }
}

llama_context::~llama_context() = default;

void llama_context::synchronize() {
    ggml_backend_sched_synchronize(sched.get());

    // FIXME: if multiple single tokens are evaluated without a synchronization,
    // the stats will be added to the prompt evaluation stats
    // this should only happen when using batch size 1 to evaluate a batch

    // add the evaluation to the stats
    if (n_queued_tokens == 1) {
        if (!cparams.no_perf) {
            t_eval_us += ggml_time_us() - t_compute_start_us;
        }
        n_eval++;
    } else if (n_queued_tokens > 1) {
        if (!cparams.no_perf) {
            t_p_eval_us += ggml_time_us() - t_compute_start_us;
        }
        n_p_eval += n_queued_tokens;
    }

    // get a more accurate load time, upon first eval
    if (n_queued_tokens > 0 && !has_evaluated_once) {
        t_load_us = ggml_time_us() - t_start_us;
        has_evaluated_once = true;
    }

    n_queued_tokens = 0;
    t_compute_start_us = 0;
}

const llama_model & llama_context::get_model() const {
    return model;
}

uint32_t llama_context::n_ctx() const {
    return cparams.n_ctx;
}

uint32_t llama_context::n_ctx_per_seq() const {
    return cparams.n_ctx / cparams.n_seq_max;
}

uint32_t llama_context::n_batch() const {
    return cparams.n_batch;
}

uint32_t llama_context::n_ubatch() const {
    return cparams.n_ubatch;
}

uint32_t llama_context::n_seq_max() const {
    return cparams.n_seq_max;
}

uint32_t llama_context::n_threads() const {
    return cparams.n_threads;
}

uint32_t llama_context::n_threads_batch() const {
    return cparams.n_threads_batch;
}

llama_kv_cache * llama_context::get_kv_self() {
    return kv_self.get();
}

const llama_kv_cache * llama_context::get_kv_self() const {
    return kv_self.get();
}

ggml_tensor * llama_context::build_rope_shift(
        ggml_context * ctx0,
        ggml_tensor * cur,
        ggml_tensor * shift,
        ggml_tensor * factors,
              float   freq_base,
              float   freq_scale,
        ggml_backend_buffer * bbuf) const {
    const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;

    const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
    const auto & yarn_attn_factor = cparams.yarn_attn_factor;
    const auto & yarn_beta_fast   = cparams.yarn_beta_fast;
    const auto & yarn_beta_slow   = cparams.yarn_beta_slow;

    const auto & hparams = model.hparams;

    const auto & n_rot     = hparams.n_rot;
    const auto & rope_type = hparams.rope_type;

    ggml_tensor * tmp;

    if (ggml_is_quantized(cur->type)) {
        // dequantize to f32 -> RoPE -> quantize back
        tmp = ggml_cast(ctx0, cur, GGML_TYPE_F32);

        if (bbuf) {
            for (const auto & backend : backends) {
                // Figure out which backend KV cache belongs to
                if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(bbuf))) {
                    ggml_backend_sched_set_tensor_backend(sched.get(), tmp, backend.get());
                    break;
                }
            }
        }

        tmp = ggml_rope_ext_inplace(ctx0, tmp,
                shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);

        tmp = ggml_cpy(ctx0, tmp, cur);
    } else {
        // we rotate only the first n_rot dimensions
        tmp = ggml_rope_ext_inplace(ctx0, cur,
                shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);
    }

    return tmp;
}

class llm_graph_input_k_shift : public llm_graph_input_i {
public:
    llm_graph_input_k_shift(const llama_kv_cache_unified * kv_self) : kv_self(kv_self) {}
    virtual ~llm_graph_input_k_shift() = default;

    void set_input(const llama_ubatch * ubatch) override;

    ggml_tensor * k_shift; // I32 [kv_size]

    const llama_kv_cache_unified * kv_self;
};

void llm_graph_input_k_shift::set_input(const llama_ubatch * ubatch) {
    GGML_UNUSED(ubatch);

    if (k_shift) {
        assert(ggml_backend_buffer_is_host(k_shift->buffer));

        int32_t * data = (int32_t *) k_shift->data;

        for (uint32_t i = 0; i < kv_self->size; ++i) {
            data[i] = kv_self->cells[i].delta;
        }
    }
}

llm_graph_result_ptr llama_context::build_kv_self_shift(
        ggml_context * ctx0,
        ggml_cgraph * gf) const {
    auto res = std::make_unique<llm_graph_result>();

    const auto & hparams = model.hparams;

    const auto & n_layer = hparams.n_layer;

    const auto & n_embd_head_k = hparams.n_embd_head_k;
  //const auto & n_embd_head_v = hparams.n_embd_head_v;

    //GGML_ASSERT(kv_self->size == n_ctx);

    auto inp = std::make_unique<llm_graph_input_k_shift>(kv_self.get());

    inp->k_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cparams.n_ctx);
    ggml_set_input(inp->k_shift);

    for (uint32_t il = 0; il < n_layer; ++il) {
        const int64_t n_head_kv    = hparams.n_head_kv(il);
        const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);

        const bool is_swa = hparams.is_swa(il);

        // note: the swa rope params could become part of the cparams in the future
        //       if we decide to make them configurable, like the non-sliding ones
        const float freq_base_l  = is_swa ? hparams.rope_freq_base_train_swa  : cparams.rope_freq_base;
        const float freq_scale_l = is_swa ? hparams.rope_freq_scale_train_swa : cparams.rope_freq_scale;

        ggml_tensor * rope_factors = kv_self->cbs.get_rope_factors(n_ctx_per_seq(), il);

        ggml_tensor * k =
            ggml_view_3d(ctx0, kv_self->k_l[il],
                n_embd_head_k, n_head_kv, kv_self->size,
                ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k),
                ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
                0);

        ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, kv_self->k_l[il]->buffer);

        ggml_build_forward_expand(gf, cur);
    }

    res->add_input(std::move(inp));

    return res;
}

llm_graph_result_ptr llama_context::build_kv_self_defrag(
        ggml_context * ctx0,
        ggml_cgraph * gf) const {
    auto res = std::make_unique<llm_graph_result>();

    const auto & hparams = model.hparams;

    const auto & ids = kv_self->defrag_info.ids;

#if 0
    // CPU defrag
    //
    // TODO: optimizations are possible:
    //       - multiple threads
    //       - avoid copying to the host memory when already there
    //
    // likely not worth the effort, as we have ggml_graph based defrag
    //

    const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa();
    const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa();

    const uint32_t kv_size = size;

    std::vector<uint8_t> buf_k;
    std::vector<uint8_t> buf_v;

    for (uint32_t il = 0; il < n_layer; ++il) {
        const size_t k_size_row = ggml_row_size(k_l[il]->type, n_embd_k_gqa);
        const size_t k_size     = ggml_row_size(k_l[il]->type, n_embd_k_gqa*kv_size);

        const size_t v_size_el = ggml_type_size(v_l[il]->type);
        const size_t v_size    = ggml_row_size (v_l[il]->type, n_embd_v_gqa*kv_size);

        buf_k.resize(k_size);
        buf_v.resize(v_size);

        ggml_backend_tensor_get(k_l[il], buf_k.data(), 0, buf_k.size());
        ggml_backend_tensor_get(v_l[il], buf_v.data(), 0, buf_v.size());

        // batch move [i, i+nm) to [id, id+nm)
        // note: cells can move only to a lower index
        for (uint32_t i = 0; i < n_kv; ++i) {
            const uint32_t id = ids[i];

            if (i == id || id == n_kv) {
                continue;
            }

            uint32_t nm = 1;

            while (i + nm < n_kv && ids[i + nm] == id + nm) {
                nm++;
            }

            // move keys
            {
                const int64_t os =  i*k_size_row;
                const int64_t od = id*k_size_row;

                memcpy(buf_k.data() + od, buf_k.data() + os, nm*k_size_row);
            }

            // move values (note: they are transposed)
            {
                const int64_t os =  i;
                const int64_t od = id;

                for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
                    memcpy(buf_v.data() + (od + j*kv_size)*v_size_el, buf_v.data() + (os + j*kv_size)*v_size_el, nm*v_size_el);
                }
            }

            i += nm - 1;
        }

        ggml_backend_tensor_set(k_l[il], buf_k.data(), 0, buf_k.size());
        ggml_backend_tensor_set(v_l[il], buf_v.data(), 0, buf_v.size());
    }
#else
    for (uint32_t i = 0; i < ids.size(); ++i) {
        const uint32_t id = ids[i];

        if (i == id || id == ids.size()) {
            continue;
        }

        uint32_t nm = 1;

        while (i + nm < ids.size() && ids[i + nm] == id + nm) {
            nm++;
        }

        for (uint32_t il = 0; il < hparams.n_layer; ++il) { // NOLINT
            const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il);
            const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(il);

            ggml_tensor * view_k_src = ggml_view_2d(ctx0, kv_self->k_l[il],
                    n_embd_k_gqa, nm,
                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*i));

            ggml_tensor * view_k_dst = ggml_view_2d(ctx0, kv_self->k_l[il],
                    n_embd_k_gqa, nm,
                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
                    ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa*id));

            ggml_tensor * view_v_src;
            ggml_tensor * view_v_dst;

            if (cparams.flash_attn) {
                // NOTE: the V cache is not transposed when using flash attention
                view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il],
                        n_embd_v_gqa, nm,
                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*i));

                view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il],
                        n_embd_v_gqa, nm,
                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
                        ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa*id));
            } else {
                view_v_src = ggml_view_2d(ctx0, kv_self->v_l[il],
                        nm, n_embd_v_gqa,
                        ggml_row_size(kv_self->v_l[il]->type, kv_self->size),
                        ggml_row_size(kv_self->v_l[il]->type, i));

                view_v_dst = ggml_view_2d(ctx0, kv_self->v_l[il],
                        nm, n_embd_v_gqa,
                        ggml_row_size(kv_self->v_l[il]->type, kv_self->size),
                        ggml_row_size(kv_self->v_l[il]->type, id));
            }

            ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_k_src, view_k_dst));
            ggml_build_forward_expand(gf, ggml_cpy(ctx0, view_v_src, view_v_dst));
        }

        i += nm - 1;
    }

    //LLAMA_LOG_INFO("gf->n_nodes = %d\n", gf->n_nodes);
#endif

    return res;
}

void llama_context::kv_self_update() {
    auto & kv = kv_self;

    bool need_reserve = false;

    if (kv->has_shift) {
        if (!kv->get_can_shift()) {
            GGML_ABORT("The current context does not support K-shift");
        }

        LLAMA_LOG_DEBUG("%s: applying K-shift\n", __func__);

        // apply K-shift if needed
        if (model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) {
            ggml_backend_sched_reset(sched.get());

            auto * gf = graph_init();

            auto res = build_kv_self_shift(ctx_compute.get(), gf);

            ggml_backend_sched_alloc_graph(sched.get(), gf);

            res->set_inputs(nullptr);

            graph_compute(gf, false);

            need_reserve = true;
        }

        {
            kv->has_shift = false;

            for (uint32_t i = 0; i < kv->size; ++i) {
                kv->cells[i].delta = 0;
            }
        }
    }

    // defragment the KV cache if needed
    if (kv->do_defrag) {
        LLAMA_LOG_DEBUG("%s: defragmenting KV cache\n", __func__);

        if (kv->defrag_prepare(graph_max_nodes())) {
            ggml_backend_sched_reset(sched.get());

            auto * gf = graph_init();

            auto res = build_kv_self_defrag(ctx_compute.get(), gf);

            ggml_backend_sched_alloc_graph(sched.get(), gf);

            res->set_inputs(nullptr);

            graph_compute(gf, false);

            need_reserve = true;
        }

        kv->do_defrag = false;
    }

    // reserve a worst case graph if needed
    if (need_reserve) {
        LLAMA_LOG_DEBUG("%s: reserving a worst case graph\n", __func__);

        // build worst-case graph
        uint32_t n_seqs = 1; // TODO: worst-case number of sequences
        uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);

        // simulate full KV cache
        kv_self->n = kv_self->size;

        llama_token token = model.vocab.token_bos(); // not actually used by llama_build_graph, but required to choose between token and embedding inputs graph
        llama_ubatch ubatch = { true, n_tokens, n_tokens / n_seqs, n_seqs, &token, nullptr, nullptr, nullptr, nullptr, nullptr};

        auto * gf = graph_init();
        graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT);

        // initialize scheduler with the worst-case graph
        ggml_backend_sched_reset(sched.get());
        if (!ggml_backend_sched_reserve(sched.get(), gf)) {
            LLAMA_LOG_ERROR("%s: failed to allocate compute buffers\n", __func__);
        }
    }
}

enum llama_pooling_type llama_context::pooling_type() const {
    return cparams.pooling_type;
}

float * llama_context::get_logits() {
    // reorder logits for backward compatibility
    output_reorder();

    return logits;
}

float * llama_context::get_logits_ith(int32_t i) {
    int32_t j = -1;

    try {
        if (logits == nullptr) {
            throw std::runtime_error("no logits");
        }

        if (i < 0) {
            j = n_outputs + i;
            if (j < 0) {
                throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs));
            }
        } else if ((size_t) i >= output_ids.size()) {
            throw std::runtime_error(format("out of range [0, %zu)", output_ids.size()));
        } else {
            j = output_ids[i];
        }

        if (j < 0) {
            throw std::runtime_error(format("batch.logits[%d] != true", i));
        }
        if (j >= n_outputs) {
            // This should not happen
            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
        }

        return logits + j*model.vocab.n_tokens();
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
#ifndef NDEBUG
        GGML_ABORT("fatal error");
#else
        return nullptr;
#endif
    }
}

float * llama_context::get_embeddings() {
    // reorder embeddings for backward compatibility
    output_reorder();

    return embd;
}

float * llama_context::get_embeddings_ith(int32_t i) {
    int32_t j = -1;

    try {
        if (embd == nullptr) {
            throw std::runtime_error("no embeddings");
        }

        if (i < 0) {
            j = n_outputs + i;
            if (j < 0) {
                throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs));
            }
        } else if ((size_t) i >= output_ids.size()) {
            throw std::runtime_error(format("out of range [0, %zu)", output_ids.size()));
        } else {
            j = output_ids[i];
        }

        if (j < 0) {
            throw std::runtime_error(format("batch.logits[%d] != true", i));
        }
        if (j >= n_outputs) {
            // This should not happen
            throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
        }

        return embd + j*model.hparams.n_embd;
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what());
#ifndef NDEBUG
        GGML_ABORT("fatal error");
#else
        return nullptr;
#endif
    }
}

float * llama_context::get_embeddings_seq(llama_seq_id seq_id) {
    auto it = embd_seq.find(seq_id);
    if (it == embd_seq.end()) {
        return nullptr;
    }

    return it->second.data();
}

void llama_context::attach_threadpool(
           ggml_threadpool_t threadpool,
           ggml_threadpool_t threadpool_batch) {
    LLAMA_LOG_DEBUG("%s: call\n", __func__);

    this->threadpool       = threadpool;
    this->threadpool_batch = threadpool_batch ? threadpool_batch : threadpool;
}

void llama_context::detach_threadpool() {
    LLAMA_LOG_DEBUG("%s: call\n", __func__);

    this->threadpool       = nullptr;
    this->threadpool_batch = nullptr;
}

void llama_context::set_n_threads(int32_t n_threads, int32_t n_threads_batch) {
    LLAMA_LOG_DEBUG("%s: n_threads = %d, n_threads_batch = %d\n", __func__, n_threads, n_threads_batch);

    cparams.n_threads       = n_threads;
    cparams.n_threads_batch = n_threads_batch;
}

void llama_context::set_abort_callback(bool (*abort_callback)(void * data), void * abort_callback_data) {
    LLAMA_LOG_DEBUG("%s: call\n", __func__);

    this->abort_callback      = abort_callback;
    this->abort_callback_data = abort_callback_data;

    for (auto & backend : backends) {
        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get()));
        auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback");
        if (set_abort_callback_fn) {
            set_abort_callback_fn(backend.get(), this->abort_callback, this->abort_callback_data);
        }
    }
}

void llama_context::set_embeddings(bool value) {
    LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value);

    cparams.embeddings = value;
}

void llama_context::set_causal_attn(bool value) {
    LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value);

    cparams.causal_attn = value;
}

void llama_context::set_warmup(bool value) {
    LLAMA_LOG_DEBUG("%s: value = %d\n", __func__, value);

    cparams.warmup = value;
}

void llama_context::set_adapter_lora(
            llama_adapter_lora * adapter,
            float scale) {
    LLAMA_LOG_DEBUG("%s: adapter = %p, scale = %f\n", __func__, (void *) adapter, scale);

    loras[adapter] = scale;
}

bool llama_context::rm_adapter_lora(
            llama_adapter_lora * adapter) {
    LLAMA_LOG_DEBUG("%s: adapter = %p\n", __func__, (void *) adapter);

    auto pos = loras.find(adapter);
    if (pos != loras.end()) {
        loras.erase(pos);
        return true;
    }

    return false;
}

void llama_context::clear_adapter_lora() {
    LLAMA_LOG_DEBUG("%s: call\n", __func__);

    loras.clear();
}

bool llama_context::apply_adapter_cvec(
            const float * data,
                 size_t   len,
                int32_t   n_embd,
                int32_t   il_start,
                int32_t   il_end) {
    LLAMA_LOG_DEBUG("%s: il_start = %d, il_end = %d\n", __func__, il_start, il_end);

    return cvec.apply(model, data, len, n_embd, il_start, il_end);
}

int llama_context::encode(llama_batch & inp_batch) {
    if (inp_batch.n_tokens == 0) {
        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
        return -1;
    }

    // temporary allocate memory for the input batch if needed
    // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences
    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1);

    const llama_batch & batch = batch_allocr.batch;
    const int32_t n_tokens = batch.n_tokens;

    const auto & hparams = model.hparams;

    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT

    if (batch.token) {
        for (int32_t i = 0; i < n_tokens; ++i) {
            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
                LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
                return -1;
            }
        }
    }

    // micro-batching is not possible for non-causal encoding, so we process the batch in a single shot
    GGML_ASSERT(cparams.n_ubatch >= (uint32_t) n_tokens && "encoder requires n_ubatch >= n_tokens");

    if (t_compute_start_us == 0) {
        t_compute_start_us = ggml_time_us();
    }

    n_queued_tokens += n_tokens;

    const int64_t n_embd = hparams.n_embd;

    sbatch.from_batch(batch, n_embd, /* simple_split */ true, /* logits_all */ true);

    const llama_ubatch ubatch = sbatch.split_simple(n_tokens);

    // reserve output buffer
    if (output_reserve(n_tokens) < n_tokens) {
        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %u outputs\n", __func__, n_tokens);
        return -2;
    };

    for (int32_t i = 0; i < n_tokens; ++i) {
        output_ids[i] = i;
    }

    n_outputs = n_tokens;

    //batch_manager->prepare(ubatch);

    ggml_backend_sched_reset(sched.get());
    ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);

    const auto causal_attn_org = cparams.causal_attn;

    // always use non-causal attention for encoder graphs
    // TODO: this is a tmp solution until we have a proper way to support enc-dec models
    //       ref: https://github.com/ggml-org/llama.cpp/pull/12181#issuecomment-2730451223
    cparams.causal_attn = false;

    auto * gf = graph_init();
    auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_ENCODER);

    ggml_backend_sched_alloc_graph(sched.get(), gf);

    res->set_inputs(&ubatch);

    cparams.causal_attn = causal_attn_org;

    const auto compute_status = graph_compute(gf, n_tokens > 1);
    switch (compute_status) {
        case GGML_STATUS_SUCCESS:
            break;
        case GGML_STATUS_ABORTED:
            return 2;
        case GGML_STATUS_ALLOC_FAILED:
            return -2;
        case GGML_STATUS_FAILED:
        default:
            return -3;
    }

    auto * t_embd = res->get_embd_pooled() ? res->get_embd_pooled() : res->get_embd();

    // extract embeddings
    if (t_embd) {
        ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd);
        GGML_ASSERT(backend_embd != nullptr);

        GGML_ASSERT(embd != nullptr);

        switch (cparams.pooling_type) {
            case LLAMA_POOLING_TYPE_NONE:
                {
                    // extract token embeddings
                    GGML_ASSERT(n_tokens*n_embd <= (int64_t) embd_size);
                    ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd*sizeof(float));
                } break;
            case LLAMA_POOLING_TYPE_MEAN:
            case LLAMA_POOLING_TYPE_CLS:
            case LLAMA_POOLING_TYPE_LAST:
                {
                    // extract sequence embeddings
                    auto & embd_seq_out = embd_seq;
                    embd_seq_out.clear();

                    GGML_ASSERT(!ubatch.equal_seqs); // TODO: handle equal splits

                    for (int32_t i = 0; i < n_tokens; i++) {
                        const llama_seq_id seq_id = ubatch.seq_id[i][0];
                        if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
                            continue;
                        }
                        embd_seq_out[seq_id].resize(n_embd);
                        ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
                    }
                } break;
            case LLAMA_POOLING_TYPE_RANK:
                {
                    // TODO: this likely should be the same logic as in llama_decoder_internal, but better to
                    //       wait for an encoder model that requires this pooling type in order to test it
                    //       https://github.com/ggerganov/llama.cpp/pull/9510
                    GGML_ABORT("RANK pooling not implemented yet");
                }
            case LLAMA_POOLING_TYPE_UNSPECIFIED:
                {
                    GGML_ABORT("unknown pooling type");
                }
        }
    }

    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
    // overlap with device computation.
    ggml_backend_sched_reset(sched.get());

    // TODO: hacky solution
    if (model.arch == LLM_ARCH_T5 && t_embd) {
        //cross.t_embd = t_embd;

        synchronize();

        cross.n_embd = t_embd->ne[0];
        cross.n_enc  = t_embd->ne[1];
        cross.v_embd.resize(cross.n_embd*cross.n_enc);
        memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd));

        // remember the sequence ids used during the encoding - needed for cross attention later
        cross.seq_ids_enc.resize(n_tokens);
        for (int32_t i = 0; i < n_tokens; i++) {
            cross.seq_ids_enc[i].clear();
            for (int s = 0; s < ubatch.n_seq_id[i]; s++) {
                llama_seq_id seq_id = ubatch.seq_id[i][s];
                cross.seq_ids_enc[i].insert(seq_id);
            }
        }
    }

    return 0;
}

int llama_context::decode(llama_batch & inp_batch) {
    if (inp_batch.n_tokens == 0) {
        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
        return -1;
    }

    // temporary allocate memory for the input batch if needed
    // TODO: this is incorrect for multiple sequences because pos_max() is the maximum across all sequences
    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : kv_self->pos_max() + 1);

    const llama_batch & batch = batch_allocr.batch;

    const auto & vocab   = model.vocab;
    const auto & hparams = model.hparams;

    const int32_t n_vocab = vocab.n_tokens();

    const int64_t n_tokens_all = batch.n_tokens;
    const int64_t n_embd       = hparams.n_embd;

    // TODO: remove this stuff
    class batch_guard {
    public:
        batch_guard(llama_kv_cache_unified & kv_self) : kv_slot_restorer(kv_self) {
        }

        ~batch_guard() {
            if (!is_done) {
                kv_slot_restorer.restore();
            }
        }

        void done() {
            is_done = true;
        }

        void save(const llama_kv_cache_slot_info & slot_info) {
            kv_slot_restorer.save(slot_info);
        }

    private:
        bool is_done = false;

        llama_kv_slot_restorer kv_slot_restorer;
    };

    batch_guard bg(*kv_self);

    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT

    if (batch.token) {
        for (int64_t i = 0; i < n_tokens_all; ++i) {
            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
                LLAMA_LOG_ERROR("%s: invalid token[%" PRId64 "] = %d\n", __func__, i, batch.token[i]);
                throw std::runtime_error("invalid token");
            }
        }
    }

    GGML_ASSERT(n_tokens_all <= cparams.n_batch);

    GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens");

    if (t_compute_start_us == 0) {
        t_compute_start_us = ggml_time_us();
    }
    n_queued_tokens += n_tokens_all;

    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;

    embd_seq.clear();

    int64_t n_outputs_all = 0;

    // count outputs
    if (batch.logits && !embd_pooled) {
        for (uint32_t i = 0; i < n_tokens_all; ++i) {
            n_outputs_all += batch.logits[i] != 0;
        }
    } else if (logits_all || embd_pooled) {
        n_outputs_all = n_tokens_all;
    } else {
        // keep last output only
        n_outputs_all = 1;
    }

    const bool logits_all = n_outputs_all == n_tokens_all;

    sbatch.from_batch(batch, n_embd,
            /* simple_split */ !kv_self->recurrent,
            /* logits_all   */ logits_all);

    // reserve output buffer
    if (output_reserve(n_outputs_all) < n_outputs_all) {
        LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
        return -2;
    };

    int64_t n_outputs_prev = 0;

    while (sbatch.n_tokens > 0) {
        llama_ubatch ubatch = llama_ubatch();

        const auto & n_ubatch = cparams.n_ubatch;

        if (kv_self->recurrent) {
            if (embd_pooled) {
                // Pooled embeddings cannot be split across ubatches (yet)
                ubatch = sbatch.split_seq(cparams.n_ubatch);
            } else {
                // recurrent model architectures are easier to implement
                // with equal-length sequences
                ubatch = sbatch.split_equal(cparams.n_ubatch);
            }
        } else {
            ubatch = sbatch.split_simple(n_ubatch);
        }

        // count the outputs in this u_batch
        {
            int32_t n_outputs_new = 0;

            if (n_outputs_all == n_tokens_all) {
                n_outputs_new = ubatch.n_tokens;
            } else {
                GGML_ASSERT(ubatch.output);
                for (uint32_t i = 0; i < ubatch.n_tokens; i++) {
                    n_outputs_new += (int32_t) (ubatch.output[i] != 0);
                }
            }

            // needs to happen before the graph is built
            n_outputs = n_outputs_new;
        }

        // non-causal masks do not use the KV cache
        if (hparams.causal_attn) {
            kv_self_update();

            // if we have enough unused cells before the current head ->
            //   better to start searching from the beginning of the cache, hoping to fill it
            if (kv_self->head > kv_self->used + 2*ubatch.n_tokens) {
                kv_self->head = 0;
            }

            const auto slot_info = kv_self->find_slot(ubatch);
            if (!slot_info) {
                LLAMA_LOG_ERROR("%s: failed to prepare ubatch\n", __func__);
                return -3;
            }

            bg.save(slot_info);

            if (!kv_self->recurrent) {
                // a heuristic, to avoid attending the full cache if it is not yet utilized
                // after enough generations, the benefit from this heuristic disappears
                // if we start defragmenting the cache, the benefit from this will be more important
                const uint32_t pad = kv_self->get_padding(cparams);
                kv_self->n = std::min(kv_self->size, std::max(pad, GGML_PAD(kv_self->cell_max(), pad)));
            }
        }

        //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self->n, kv_self->used, kv_self->head);

        ggml_backend_sched_reset(sched.get());
        ggml_backend_sched_set_eval_callback(sched.get(), cparams.cb_eval, cparams.cb_eval_user_data);

        auto * gf = graph_init();
        auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DECODER);

        // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);

        ggml_backend_sched_alloc_graph(sched.get(), gf);

        res->set_inputs(&ubatch);

        const auto compute_status = graph_compute(gf, ubatch.n_tokens > 1);
        if (compute_status != GGML_STATUS_SUCCESS) {
            switch (compute_status) {
                case GGML_STATUS_ABORTED:
                    return 2;
                case GGML_STATUS_ALLOC_FAILED:
                    return -2;
                case GGML_STATUS_FAILED:
                default:
                    return -3;
            }
        }

        // update the kv ring buffer
        {
            kv_self->head += ubatch.n_tokens;

            // Ensure kv cache head points to a valid index.
            if (kv_self->head >= kv_self->size) {
                kv_self->head = 0;
            }
        }

        // plot the computation graph in dot format (for debugging purposes)
        //if (n_past%100 == 0) {
        //    ggml_graph_dump_dot(gf, NULL, "llama.dot");
        //}

        auto * t_logits = cparams.embeddings ? nullptr         : res->get_logits();
        auto * t_embd   = cparams.embeddings ? res->get_embd() : nullptr;

        if (t_embd && res->get_embd_pooled()) {
            t_embd = res->get_embd_pooled();
        }

        // extract logits
        if (t_logits && n_outputs > 0) {
            ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits);
            GGML_ASSERT(backend_res != nullptr);
            GGML_ASSERT(logits != nullptr);

            float * logits_out = logits + n_outputs_prev*n_vocab;

            if (n_outputs) {
                GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all);
                GGML_ASSERT((n_outputs_prev + n_outputs)*n_vocab <= (int64_t) logits_size);
                ggml_backend_tensor_get_async(backend_res, t_logits, logits_out, 0, n_outputs*n_vocab*sizeof(float));
            }
        }

        // extract embeddings
        if (t_embd && n_outputs > 0) {
            ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd);
            GGML_ASSERT(backend_embd != nullptr);

            switch (cparams.pooling_type) {
                case LLAMA_POOLING_TYPE_NONE:
                    {
                        // extract token embeddings
                        GGML_ASSERT(embd != nullptr);
                        float * embd_out = embd + n_outputs_prev*n_embd;

                        if (n_outputs) {
                            GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all);
                            GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd <= (int64_t) embd_size);
                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd*sizeof(float));
                        }
                    } break;
                case LLAMA_POOLING_TYPE_MEAN:
                case LLAMA_POOLING_TYPE_CLS:
                case LLAMA_POOLING_TYPE_LAST:
                    {
                        // extract sequence embeddings (cleared before processing each batch)
                        auto & embd_seq_out = embd_seq;

                        for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
                            const llama_seq_id seq_id = ubatch.seq_id[s][0];
                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
                                continue;
                            }
                            embd_seq_out[seq_id].resize(n_embd);
                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (n_embd*seq_id)*sizeof(float), n_embd*sizeof(float));
                        }
                    } break;
                case LLAMA_POOLING_TYPE_RANK:
                    {
                        // extract the rerank score - a single float per sequence
                        auto & embd_seq_out = embd_seq;

                        for (uint32_t s = 0; s < ubatch.n_seqs; ++s) {
                            const llama_seq_id seq_id = ubatch.seq_id[s][0];
                            if (embd_seq_out.find(seq_id) != embd_seq_out.end()) {
                                continue;
                            }
                            embd_seq_out[seq_id].resize(1);
                            ggml_backend_tensor_get_async(backend_embd, t_embd, embd_seq_out[seq_id].data(), (seq_id)*sizeof(float), sizeof(float));
                        }
                    } break;
                case LLAMA_POOLING_TYPE_UNSPECIFIED:
                    {
                        GGML_ABORT("unknown pooling type");
                    }
            }
        }

        n_outputs_prev += n_outputs;
    }

    // finalize the batch processing
    bg.done();

    // set output mappings
    {
        bool sorted_output = true;

        GGML_ASSERT(sbatch.out_ids.size() == (size_t) n_outputs_all);

        for (int64_t i = 0; i < n_outputs_all; ++i) {
            int64_t out_id = sbatch.out_ids[i];
            output_ids[out_id] = i;
            if (out_id != i) {
                sorted_output = false;
            }
        }

        if (sorted_output) {
            sbatch.out_ids.clear();
        }
    }

    // set to total number of outputs in the batch, for use in llama_get_logits_ith
    n_outputs = n_outputs_all;

    // wait for the computation to finish (automatically done when obtaining the model output)
    //synchronize();

    // decide if we need to defrag the kv cache
    if (cparams.causal_attn && cparams.defrag_thold > 0.0f) {
        // - do not defrag small contexts (i.e. < 2048 tokens)
        // - count the padding towards the number of used tokens
        const float fragmentation = kv_self->n >= 2048 ? std::max(0.0f, 1.0f - float(kv_self->used + kv_self->get_padding(cparams))/float(kv_self->n)) : 0.0f;

        // queue defragmentation for next llama_kv_cache_update
        if (fragmentation > cparams.defrag_thold) {
            LLAMA_LOG_DEBUG("%s: fragmentation: %.2f - requesting defrag\n", __func__, fragmentation);

            kv_self->defrag();
        }
    }

    // Reset state for the next token before backend sync, to allow the CPU activities in the reset to
    // overlap with device computation.
    ggml_backend_sched_reset(sched.get());

    return 0;
}

//
// output
//

int32_t llama_context::output_reserve(int32_t n_outputs) {
    const auto & hparams = model.hparams;
    const auto & vocab   = model.vocab;

    const int64_t n_outputs_max = std::max<int64_t>(n_outputs, n_seq_max());

    const auto n_batch = cparams.n_batch;
    const auto n_vocab = vocab.n_tokens();
    const auto n_embd  = hparams.n_embd;

    // TODO: use a per-batch flag for logits presence instead
    bool has_logits = !cparams.embeddings;
    bool has_embd   =  cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE);

    // TODO: hacky enc-dec support
    if (model.arch == LLM_ARCH_T5) {
        has_logits = true;
        has_embd   = true;
    }

    logits_size = has_logits ? n_vocab*n_outputs_max : 0;
    embd_size   = has_embd   ?  n_embd*n_outputs_max : 0;

    if (output_ids.empty()) {
        // init, never resized afterwards
        output_ids.resize(n_batch);
    }

    const size_t prev_size = buf_output ? ggml_backend_buffer_get_size(buf_output.get()) : 0;
    const size_t new_size  = (logits_size + embd_size) * sizeof(float);

    // alloc only when more than the current capacity is required
    // TODO: also consider shrinking the buffer
    if (!buf_output || prev_size < new_size) {
        if (buf_output) {
#ifndef NDEBUG
            // This doesn't happen often, but may be annoying in some cases (like the HellaSwag benchmark)
            LLAMA_LOG_INFO("%s: reallocating output buffer from size %.02f MiB to %.02f MiB\n", __func__, prev_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
#endif
            buf_output = nullptr;
            logits = nullptr;
            embd = nullptr;
        }

        auto * buft = ggml_backend_cpu_buffer_type();
        // try to use the host buffer of the device where the output tensor is allocated for faster transfer to system memory
        auto * output_dev = model.dev_output();
        auto * output_dev_host_buft = output_dev ? ggml_backend_dev_host_buffer_type(output_dev) : nullptr;
        if (output_dev_host_buft) {
            buft = output_dev_host_buft;
        }
        buf_output.reset(ggml_backend_buft_alloc_buffer(buft, new_size));
        if (buf_output == nullptr) {
            LLAMA_LOG_ERROR("%s: failed to allocate output buffer of size %.2f MiB\n", __func__, new_size / (1024.0 * 1024.0));
            return 0;
        }
    }

    float * output_base = (float *) ggml_backend_buffer_get_base(buf_output.get());

    logits = has_logits ? output_base               : nullptr;
    embd   = has_embd   ? output_base + logits_size : nullptr;

    // set all ids as invalid (negative)
    std::fill(output_ids.begin(), output_ids.end(), -1);

    ggml_backend_buffer_clear(buf_output.get(), 0);

    this->n_outputs     = 0;
    this->n_outputs_max = n_outputs_max;

    return n_outputs_max;
}

void llama_context::output_reorder() {
    auto & out_ids = sbatch.out_ids;
    if (!out_ids.empty()) {
        const uint32_t n_vocab = model.vocab.n_tokens();
        const uint32_t n_embd  = model.hparams.n_embd;

        GGML_ASSERT((size_t) n_outputs == out_ids.size());

        // TODO: is there something more efficient which also minimizes swaps?
        // selection sort, to minimize swaps (from https://en.wikipedia.org/wiki/Selection_sort)
        for (int32_t i = 0; i < n_outputs - 1; ++i) {
            int32_t j_min = i;
            for (int32_t j = i + 1; j < n_outputs; ++j) {
                if (out_ids[j] < out_ids[j_min]) {
                    j_min = j;
                }
            }
            if (j_min == i) { continue; }
            std::swap(out_ids[i], out_ids[j_min]);
            if (logits_size > 0) {
                for (uint32_t k = 0; k < n_vocab; k++) {
                    std::swap(logits[i*n_vocab + k], logits[j_min*n_vocab + k]);
                }
            }
            if (embd_size > 0) {
                for (uint32_t k = 0; k < n_embd; k++) {
                    std::swap(embd[i*n_embd + k], embd[j_min*n_embd + k]);
                }
            }
        }
        std::fill(output_ids.begin(), output_ids.end(), -1);
        for (int32_t i = 0; i < n_outputs; ++i) {
            output_ids[out_ids[i]] = i;
        }
        out_ids.clear();
    }
}

//
// graph
//

int32_t llama_context::graph_max_nodes() const {
    return std::max<int32_t>(65536, 5*model.n_tensors());
}

ggml_cgraph * llama_context::graph_init() {
    ggml_init_params params = {
        /*.mem_size   =*/ buf_compute_meta.size(),
        /*.mem_buffer =*/ buf_compute_meta.data(),
        /*.no_alloc   =*/ true,
    };

    ctx_compute.reset(ggml_init(params));

    return ggml_new_graph_custom(ctx_compute.get(), graph_max_nodes(), false);
}

llm_graph_result_ptr llama_context::graph_build(
            ggml_context * ctx,
             ggml_cgraph * gf,
      const llama_ubatch & ubatch,
            llm_graph_type gtype) {
    return model.build_graph(
            {
                /*.ctx         =*/ ctx,
                /*.arch        =*/ model.arch,
                /*.hparams     =*/ model.hparams,
                /*.cparams     =*/ cparams,
                /*.ubatch      =*/ ubatch,
                /*.sched       =*/ sched.get(),
                /*.backend_cpu =*/ backend_cpu,
                /*.cvec        =*/ &cvec,
                /*.loras       =*/ &loras,
                /*.memory      =*/ kv_self.get(),
                /*.cross       =*/ &cross,
                /*.n_outputs   =*/ n_outputs,
                /*.cb          =*/ graph_get_cb(),
            }, gf, gtype);
}

ggml_status llama_context::graph_compute(
            ggml_cgraph * gf,
                   bool   batched) {
    int n_threads        = batched ? cparams.n_threads_batch : cparams.n_threads;
    ggml_threadpool_t tp = batched ? threadpool_batch        : threadpool;

    if (backend_cpu != nullptr) {
        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_cpu));
        auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
        set_threadpool_fn(backend_cpu, tp);
    }

    // set the number of threads for all the backends
    for (const auto & set_n_threads_fn : set_n_threads_fns) {
        set_n_threads_fn.second(set_n_threads_fn.first, n_threads);
    }

    auto status = ggml_backend_sched_graph_compute_async(sched.get(), gf);
    if (status != GGML_STATUS_SUCCESS) {
        LLAMA_LOG_ERROR("%s: ggml_backend_sched_graph_compute_async failed with error %d\n", __func__, status);
    }

    // fprintf(stderr, "splits: %d\n", ggml_backend_sched_get_n_splits(sched));

    return status;
}

llm_graph_cb llama_context::graph_get_cb() const {
    return [&](const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il) {
        if (il >= 0) {
            ggml_format_name(cur, "%s-%d", name, il);
        } else {
            ggml_set_name(cur, name);
        }

        if (!cparams.offload_kqv) {
            if (strcmp(name, "kqv_merged_cont") == 0) {
                // all nodes between the KV store and the attention output are run on the CPU
                ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend_cpu);
            }
        }

        // norm may be automatically assigned to the backend of the previous layer, increasing data transfer between backends
        // FIXME: fix in ggml_backend_sched
        const bool full_offload = model.params.n_gpu_layers > (int) model.hparams.n_layer;
        if (ubatch.n_tokens < 32 || full_offload) {
            if (il != -1 && strcmp(name, "norm") == 0) {
                const auto & dev_layer = model.dev_layer(il);
                for (const auto & backend : backends) {
                    if (ggml_backend_get_device(backend.get()) == dev_layer) {
                        if (ggml_backend_supports_op(backend.get(), cur)) {
                            ggml_backend_sched_set_tensor_backend(sched.get(), cur, backend.get());
                        }
                    }
                }
            }
        }
    };
}

//
// state save/load
//

class llama_io_write_dummy : public llama_io_write_i {
public:
    llama_io_write_dummy() = default;

    void write(const void * /* src */, size_t size) override {
        size_written += size;
    }

    void write_tensor(const ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override {
        size_written += size;
    }

    size_t n_bytes() override {
        return size_written;
    }

private:
    size_t size_written = 0;
};

class llama_io_write_buffer : public llama_io_write_i {
public:
    llama_io_write_buffer(
            uint8_t * p, size_t len) : ptr(p), buf_size(len) {}

    void write(const void * src, size_t size) override {
        if (size > buf_size) {
            throw std::runtime_error("unexpectedly reached end of buffer");
        }
        memcpy(ptr, src, size);
        ptr += size;
        size_written += size;
        buf_size -= size;
    }

    void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) override {
        if (size > buf_size) {
            throw std::runtime_error("unexpectedly reached end of buffer");
        }
        ggml_backend_tensor_get(tensor, ptr, offset, size);
        ptr += size;
        size_written += size;
        buf_size -= size;
    }

    size_t n_bytes() override {
        return size_written;
    }

private:
    uint8_t * ptr;
    size_t buf_size = 0;
    size_t size_written = 0;
};

class llama_io_read_buffer : public llama_io_read_i {
public:
    llama_io_read_buffer(const uint8_t * p, size_t len) : ptr(p), buf_size(len) {}

    const uint8_t * read(size_t size) override {
        const uint8_t * base_ptr = ptr;
        if (size > buf_size) {
            throw std::runtime_error("unexpectedly reached end of buffer");
        }
        ptr += size;
        size_read += size;
        buf_size -= size;
        return base_ptr;
    }

    void read_to(void * dst, size_t size) override {
        memcpy(dst, read(size), size);
    }

    size_t n_bytes() override {
        return size_read;
    }

private:
    const uint8_t * ptr;
    size_t buf_size = 0;
    size_t size_read = 0;
};

class llama_io_write_file : public llama_io_write_i {
public:
    llama_io_write_file(llama_file * f) : file(f) {}

    void write(const void * src, size_t size) override {
        file->write_raw(src, size);
        size_written += size;
    }

    void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) override {
        temp_buffer.resize(size);
        ggml_backend_tensor_get(tensor, temp_buffer.data(), offset, size);
        write(temp_buffer.data(), temp_buffer.size());
    }

    size_t n_bytes() override {
        return size_written;
    }

private:
    llama_file * file;
    size_t size_written = 0;
    std::vector<uint8_t> temp_buffer;
};

class llama_io_read_file : public llama_io_read_i {
public:
    llama_io_read_file(llama_file * f) : file(f) {}

    void read_to(void * dst, size_t size) override {
        file->read_raw(dst, size);
        size_read += size;
    }

    const uint8_t * read(size_t size) override {
        temp_buffer.resize(size);
        read_to(temp_buffer.data(), size);
        return temp_buffer.data();
    }

    size_t n_bytes() override {
        return size_read;
    }

private:
    llama_file * file;
    size_t size_read = 0;
    std::vector<uint8_t> temp_buffer;
};

size_t llama_context::state_get_size() {
    llama_io_write_dummy io;
    try {
        return state_write_data(io);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
        return 0;
    }
}

size_t llama_context::state_get_data(uint8_t * dst, size_t size) {
    llama_io_write_buffer io(dst, size);
    try {
        return state_write_data(io);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
        return 0;
    }
}

size_t llama_context::state_set_data(const uint8_t * src, size_t size) {
    llama_io_read_buffer io(src, size);
    try {
        return state_read_data(io);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
        return 0;
    }
}

size_t llama_context::state_seq_get_size(llama_seq_id seq_id) {
    llama_io_write_dummy io;
    try {
        return state_seq_write_data(io, seq_id);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
        return 0;
    }
}

size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size) {
    llama_io_write_buffer io(dst, size);
    try {
        return state_seq_write_data(io, seq_id);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
        return 0;
    }
}

size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size) {
    llama_io_read_buffer io(src, size);
    try {
        return state_seq_read_data(io, seq_id);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
        return 0;
    }
}

bool llama_context::state_load_file(const char * filepath, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
    llama_file file(filepath, "rb");

    // sanity checks
    {
        const uint32_t magic   = file.read_u32();
        const uint32_t version = file.read_u32();

        if (magic != LLAMA_SESSION_MAGIC || version != LLAMA_SESSION_VERSION) {
            LLAMA_LOG_ERROR("%s: unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
            return false;
        }
    }

    // load the prompt
    {
        const uint32_t n_token_count = file.read_u32();

        if (n_token_count > n_token_capacity) {
            LLAMA_LOG_ERROR("%s: token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
            return false;
        }

        file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
        *n_token_count_out = n_token_count;
    }

    // restore the context state
    {
        const size_t n_state_size_cur = file.size() - file.tell();

        llama_io_read_file io( &file);
        const size_t n_read = state_read_data(io);

        if (n_read != n_state_size_cur) {
            LLAMA_LOG_ERROR("%s: did not read all of the session file data! size %zu, got %zu\n", __func__, n_state_size_cur, n_read);
            return false;
        }
    }

    return true;
}

bool llama_context::state_save_file(const char * filepath, const llama_token * tokens, size_t n_token_count) {
    llama_file file(filepath, "wb");

    file.write_u32(LLAMA_SESSION_MAGIC);
    file.write_u32(LLAMA_SESSION_VERSION);

    // save the prompt
    file.write_u32((uint32_t) n_token_count);
    file.write_raw(tokens, sizeof(llama_token) * n_token_count);

    // save the context state using stream saving
    llama_io_write_file io(&file);
    state_write_data(io);

    return true;
}

size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * filepath, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
    llama_file file(filepath, "rb");

    // version checks
    {
        const uint32_t magic   = file.read_u32();
        const uint32_t version = file.read_u32();

        if (magic != LLAMA_STATE_SEQ_MAGIC || version != LLAMA_STATE_SEQ_VERSION) {
            LLAMA_LOG_ERROR("%s: unknown (magic, version) for sequence state file: %08x, %08x\n", __func__, magic, version);
            return 0;
        }
    }

    // load the prompt
    {
        const uint32_t n_token_count = file.read_u32();

        if (n_token_count > n_token_capacity) {
            LLAMA_LOG_ERROR("%s: token count in sequence state file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
            return 0;
        }

        file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
        *n_token_count_out = n_token_count;
    }

    // restore the context state
    {
        const size_t state_size = file.size() - file.tell();
        llama_io_read_file io(&file);
        const size_t nread = state_seq_read_data(io, seq_id);
        if (!nread) {
            LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
            return 0;
        }
        GGML_ASSERT(nread <= state_size);
        GGML_ASSERT(nread + sizeof(uint32_t) * 3 + sizeof(llama_token) * *n_token_count_out == file.tell());
    }

    return file.tell();
}

size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * filepath, const llama_token * tokens, size_t n_token_count) {
    llama_file file(filepath, "wb");

    file.write_u32(LLAMA_STATE_SEQ_MAGIC);
    file.write_u32(LLAMA_STATE_SEQ_VERSION);

    // save the prompt
    file.write_u32((uint32_t) n_token_count);
    file.write_raw(tokens, sizeof(llama_token) * n_token_count);

    // save the context state using stream saving
    llama_io_write_file io(&file);
    state_seq_write_data(io, seq_id);

    const size_t res = file.tell();
    GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + io.n_bytes());

    return res;
}

size_t llama_context::state_write_data(llama_io_write_i & io) {
    LLAMA_LOG_DEBUG("%s: writing state\n", __func__);

    // write model info
    {
        LLAMA_LOG_DEBUG("%s: - writing model info\n", __func__);

        const std::string arch_str = llm_arch_name(model.arch);
        io.write_string(arch_str);
        // TODO: add more model-specific info which should prevent loading the session file if not identical
    }

    // write output ids
    {
        LLAMA_LOG_DEBUG("%s: - writing output ids\n", __func__);

        output_reorder();

        const auto n_outputs    = this->n_outputs;
        const auto & output_ids = this->output_ids;

        std::vector<int32_t> w_output_pos;

        GGML_ASSERT(n_outputs <= n_outputs_max);

        w_output_pos.resize(n_outputs);

        // build a more compact representation of the output ids
        for (size_t i = 0; i < n_batch(); ++i) {
            // map an output id to a position in the batch
            int32_t pos = output_ids[i];
            if (pos >= 0) {
                GGML_ASSERT(pos < n_outputs);
                w_output_pos[pos] = i;
            }
        }

        io.write(&n_outputs, sizeof(n_outputs));

        if (n_outputs) {
            io.write(w_output_pos.data(), n_outputs * sizeof(int32_t));
        }
    }

    // write logits
    {
        LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);

        const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.vocab.n_tokens());

        io.write(&logits_size, sizeof(logits_size));

        if (logits_size) {
            io.write(logits, logits_size * sizeof(float));
        }
    }

    // write embeddings
    {
        LLAMA_LOG_DEBUG("%s: - writing embeddings\n", __func__);

        const uint64_t embd_size = std::min((uint64_t) this->embd_size, (uint64_t) n_outputs * model.hparams.n_embd);

        io.write(&embd_size, sizeof(embd_size));

        if (embd_size) {
            io.write(embd, embd_size * sizeof(float));
        }
    }

    LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__);
    kv_self->state_write(io);

    return io.n_bytes();
}

size_t llama_context::state_read_data(llama_io_read_i & io) {
    LLAMA_LOG_DEBUG("%s: reading state\n", __func__);

    // read model info
    {
        LLAMA_LOG_DEBUG("%s: - reading model info\n", __func__);

        const std::string cur_arch_str = llm_arch_name(model.arch);

        std::string arch_str;
        io.read_string(arch_str);
        if (cur_arch_str != arch_str) {
            throw std::runtime_error(format("wrong model arch: '%s' instead of '%s'", arch_str.c_str(), cur_arch_str.c_str()));
        }
        // TODO: add more info which needs to be identical but which is not verified otherwise
    }

    // read output ids
    {
        LLAMA_LOG_DEBUG("%s: - reading output ids\n", __func__);

        auto n_outputs = this->n_outputs;
        io.read_to(&n_outputs, sizeof(n_outputs));

        if (n_outputs > output_reserve(n_outputs)) {
            throw std::runtime_error("could not reserve outputs");
        }

        std::vector<int32_t> output_pos;

        if (n_outputs) {
            output_pos.resize(n_outputs);
            io.read_to(output_pos.data(), n_outputs * sizeof(int32_t));

            for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) {
                int32_t id = output_pos[i];
                if ((uint32_t) id >= n_batch()) {
                    throw std::runtime_error(format("invalid output id, %d does not fit in batch size of %u", id, n_batch()));
                }
                this->output_ids[id] = i;
            }

            this->n_outputs = n_outputs;
        }
    }

    // read logits
    {
        LLAMA_LOG_DEBUG("%s: - reading logits\n", __func__);

        uint64_t logits_size;
        io.read_to(&logits_size, sizeof(logits_size));

        if (this->logits_size < logits_size) {
            throw std::runtime_error("logits buffer too small");
        }

        if (logits_size) {
            io.read_to(this->logits, logits_size * sizeof(float));
        }
    }

    // read embeddings
    {
        LLAMA_LOG_DEBUG("%s: - reading embeddings\n", __func__);

        uint64_t embd_size;
        io.read_to(&embd_size, sizeof(embd_size));

        if (this->embd_size < embd_size) {
            throw std::runtime_error("embeddings buffer too small");
        }

        if (embd_size) {
            io.read_to(this->embd, embd_size * sizeof(float));
        }
    }

    LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
    kv_self->state_read(io);

    return io.n_bytes();
}

size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
    GGML_UNUSED(seq_id);

    kv_self->state_write(io, seq_id);

    return io.n_bytes();
}

size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
    GGML_UNUSED(seq_id);

    kv_self->state_read(io, seq_id);

    return io.n_bytes();
}

//
// perf
//

llama_perf_context_data llama_context::perf_get_data() const {
    llama_perf_context_data data = {};

    data.t_start_ms  = 1e-3 * t_start_us;
    data.t_load_ms   = 1e-3 * t_load_us;
    data.t_p_eval_ms = 1e-3 * t_p_eval_us;
    data.t_eval_ms   = 1e-3 * t_eval_us;
    data.n_p_eval    = std::max(1, n_p_eval);
    data.n_eval      = std::max(1, n_eval);

    return data;
}

void llama_context::perf_reset() {
    t_start_us  = ggml_time_us();
    t_eval_us   = n_eval = 0;
    t_p_eval_us = n_p_eval = 0;
}

//
// interface implementation
//

llama_context_params llama_context_default_params() {
    llama_context_params result = {
        /*.n_ctx                       =*/ 512,
        /*.n_batch                     =*/ 2048,
        /*.n_ubatch                    =*/ 512,
        /*.n_seq_max                   =*/ 1,
        /*.n_threads                   =*/ GGML_DEFAULT_N_THREADS, // TODO: better default
        /*.n_threads_batch             =*/ GGML_DEFAULT_N_THREADS,
        /*.rope_scaling_type           =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED,
        /*.pooling_type                =*/ LLAMA_POOLING_TYPE_UNSPECIFIED,
        /*.attention_type              =*/ LLAMA_ATTENTION_TYPE_UNSPECIFIED,
        /*.rope_freq_base              =*/ 0.0f,
        /*.rope_freq_scale             =*/ 0.0f,
        /*.yarn_ext_factor             =*/ -1.0f,
        /*.yarn_attn_factor            =*/ 1.0f,
        /*.yarn_beta_fast              =*/ 32.0f,
        /*.yarn_beta_slow              =*/ 1.0f,
        /*.yarn_orig_ctx               =*/ 0,
        /*.defrag_thold                =*/ -1.0f,
        /*.cb_eval                     =*/ nullptr,
        /*.cb_eval_user_data           =*/ nullptr,
        /*.type_k                      =*/ GGML_TYPE_F16,
        /*.type_v                      =*/ GGML_TYPE_F16,
        /*.logits_all                  =*/ false,
        /*.embeddings                  =*/ false,
        /*.offload_kqv                 =*/ true,
        /*.flash_attn                  =*/ false,
        /*.no_perf                     =*/ true,
        /*.abort_callback              =*/ nullptr,
        /*.abort_callback_data         =*/ nullptr,
    };

    return result;
}

llama_context * llama_init_from_model(
                 llama_model * model,
        llama_context_params   params) {
    if (!model) {
        LLAMA_LOG_ERROR("%s: model cannot be NULL\n", __func__);
        return nullptr;
    }

    if (params.n_batch == 0 && params.n_ubatch == 0) {
        LLAMA_LOG_ERROR("%s: n_batch and n_ubatch cannot both be zero\n", __func__);
        return nullptr;
    }

    if (params.n_ctx == 0 && model->hparams.n_ctx_train == 0) {
        LLAMA_LOG_ERROR("%s: n_ctx and model->hparams.n_ctx_train cannot both be zero\n", __func__);
        return nullptr;
    }

    if (params.flash_attn && model->arch == LLM_ARCH_GROK) {
        LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
        params.flash_attn = false;
    }

    if (params.flash_attn && model->hparams.n_embd_head_k != model->hparams.n_embd_head_v) {
        LLAMA_LOG_WARN("%s: flash_attn requires n_embd_head_k == n_embd_head_v - forcing off\n", __func__);
        params.flash_attn = false;
    }

    if (ggml_is_quantized(params.type_v) && !params.flash_attn) {
        LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__);
        return nullptr;
    }

    try {
        auto * ctx = new llama_context(*model, params);
        return ctx;
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: failed to initialize the context: %s\n", __func__, err.what());
    }

    return nullptr;
}

// deprecated
llama_context * llama_new_context_with_model(
                 llama_model * model,
        llama_context_params   params) {
    return llama_init_from_model(model, params);
}

void llama_free(llama_context * ctx) {
    delete ctx;
}

uint32_t llama_n_ctx(const llama_context * ctx) {
    return ctx->n_ctx();
}

uint32_t llama_n_batch(const llama_context * ctx) {
    return ctx->n_batch();
}

uint32_t llama_n_ubatch(const llama_context * ctx) {
    return ctx->n_ubatch();
}

uint32_t llama_n_seq_max(const llama_context * ctx) {
    return ctx->n_seq_max();
}

const llama_model * llama_get_model(const llama_context * ctx) {
    return &ctx->get_model();
}

llama_kv_cache * llama_get_kv_self(llama_context * ctx) {
    return ctx->get_kv_self();
}

void llama_kv_self_update(llama_context * ctx) {
    ctx->kv_self_update();
}

enum llama_pooling_type llama_pooling_type(const llama_context * ctx) {
    return ctx->pooling_type();
}

void llama_attach_threadpool(
            llama_context * ctx,
        ggml_threadpool_t   threadpool,
        ggml_threadpool_t   threadpool_batch) {
    ctx->attach_threadpool(threadpool, threadpool_batch);
}

void llama_detach_threadpool(llama_context * ctx) {
    ctx->detach_threadpool();
}

void llama_set_n_threads(llama_context * ctx, int32_t n_threads, int32_t n_threads_batch) {
    ctx->set_n_threads(n_threads, n_threads_batch);
}

int32_t llama_n_threads(llama_context * ctx) {
    return ctx->n_threads();
}

int32_t llama_n_threads_batch(llama_context * ctx) {
    return ctx->n_threads_batch();
}

void llama_set_abort_callback(llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
    ctx->set_abort_callback(abort_callback, abort_callback_data);
}

void llama_set_embeddings(llama_context * ctx, bool embeddings) {
    ctx->set_embeddings(embeddings);
}

void llama_set_causal_attn(llama_context * ctx, bool causal_attn) {
    ctx->set_causal_attn(causal_attn);
}

void llama_set_warmup(llama_context * ctx, bool warmup) {
    ctx->set_warmup(warmup);
}

void llama_synchronize(llama_context * ctx) {
    ctx->synchronize();
}

float * llama_get_logits(llama_context * ctx) {
    ctx->synchronize();

    return ctx->get_logits();
}

float * llama_get_logits_ith(llama_context * ctx, int32_t i) {
    ctx->synchronize();

    return ctx->get_logits_ith(i);
}

float * llama_get_embeddings(llama_context * ctx) {
    ctx->synchronize();

    return ctx->get_embeddings();
}

float * llama_get_embeddings_ith(llama_context * ctx, int32_t i) {
    ctx->synchronize();

    return ctx->get_embeddings_ith(i);
}

float * llama_get_embeddings_seq(llama_context * ctx, llama_seq_id seq_id) {
    ctx->synchronize();

    return ctx->get_embeddings_seq(seq_id);
}

// llama adapter API

int32_t llama_set_adapter_lora(
            llama_context * ctx,
            llama_adapter_lora * adapter,
            float scale) {
    ctx->set_adapter_lora(adapter, scale);

    return 0;
}

int32_t llama_rm_adapter_lora(
            llama_context * ctx,
            llama_adapter_lora * adapter) {
    bool res = ctx->rm_adapter_lora(adapter);

    return res ? 0 : -1;
}

void llama_clear_adapter_lora(llama_context * ctx) {
    ctx->clear_adapter_lora();
}

int32_t llama_apply_adapter_cvec(
        llama_context * ctx,
                 const float * data,
                      size_t   len,
                     int32_t   n_embd,
                     int32_t   il_start,
                     int32_t   il_end) {
    bool res = ctx->apply_adapter_cvec(data, len, n_embd, il_start, il_end);

    return res ? 0 : -1;
}

//
// kv cache view
//

llama_kv_cache_view llama_kv_cache_view_init(const llama_context * ctx, int32_t n_seq_max) {
    const auto * kv = ctx->get_kv_self();
    if (kv == nullptr) {
        LLAMA_LOG_WARN("%s: the context does not have a KV cache\n", __func__);
        return {};
    }

    return llama_kv_cache_view_init(*kv, n_seq_max);
}

void llama_kv_cache_view_update(const llama_context * ctx, llama_kv_cache_view * view) {
    const auto * kv = ctx->get_kv_self();
    if (kv == nullptr) {
        LLAMA_LOG_WARN("%s: the context does not have a KV cache\n", __func__);
        return;
    }

    llama_kv_cache_view_update(view, kv);
}

//
// kv cache
//

// deprecated
int32_t llama_get_kv_cache_token_count(const llama_context * ctx) {
    return llama_kv_self_n_tokens(ctx);
}

int32_t llama_kv_self_n_tokens(const llama_context * ctx) {
    return llama_kv_cache_n_tokens(ctx->get_kv_self());
}

// deprecated
int32_t llama_get_kv_cache_used_cells(const llama_context * ctx) {
    return llama_kv_self_used_cells(ctx);
}

int32_t llama_kv_self_used_cells(const llama_context * ctx) {
    return llama_kv_cache_used_cells(ctx->get_kv_self());
}

// deprecated
void llama_kv_cache_clear(llama_context * ctx) {
    llama_kv_self_clear(ctx);
}

void llama_kv_self_clear(llama_context * ctx) {
    llama_kv_cache_clear(ctx->get_kv_self());
}

// deprecated
bool llama_kv_cache_seq_rm(
        llama_context * ctx,
         llama_seq_id   seq_id,
            llama_pos   p0,
            llama_pos   p1) {
    return llama_kv_self_seq_rm(ctx, seq_id, p0, p1);
}

bool llama_kv_self_seq_rm(
        llama_context * ctx,
         llama_seq_id   seq_id,
            llama_pos   p0,
            llama_pos   p1) {
    return llama_kv_cache_seq_rm(ctx->get_kv_self(), seq_id, p0, p1);
}

// deprecated
void llama_kv_cache_seq_cp(
        llama_context * ctx,
         llama_seq_id   seq_id_src,
         llama_seq_id   seq_id_dst,
            llama_pos   p0,
            llama_pos   p1) {
    return llama_kv_self_seq_cp(ctx, seq_id_src, seq_id_dst, p0, p1);
}

void llama_kv_self_seq_cp(
        llama_context * ctx,
         llama_seq_id   seq_id_src,
         llama_seq_id   seq_id_dst,
            llama_pos   p0,
            llama_pos   p1) {
    return llama_kv_cache_seq_cp(ctx->get_kv_self(), seq_id_src, seq_id_dst, p0, p1);
}

// deprecated
void llama_kv_cache_seq_keep(
        llama_context * ctx,
         llama_seq_id   seq_id) {
    return llama_kv_self_seq_keep(ctx, seq_id);
}

void llama_kv_self_seq_keep(llama_context * ctx, llama_seq_id seq_id) {
    return llama_kv_cache_seq_keep(ctx->get_kv_self(), seq_id);
}

// deprecated
void llama_kv_cache_seq_add(
        llama_context * ctx,
         llama_seq_id   seq_id,
            llama_pos   p0,
            llama_pos   p1,
            llama_pos   delta) {
    return llama_kv_self_seq_add(ctx, seq_id, p0, p1, delta);
}

void llama_kv_self_seq_add(
        llama_context * ctx,
         llama_seq_id   seq_id,
            llama_pos   p0,
            llama_pos   p1,
            llama_pos   delta) {
    return llama_kv_cache_seq_add(ctx->get_kv_self(), seq_id, p0, p1, delta);
}

// deprecated
void llama_kv_cache_seq_div(
        llama_context * ctx,
         llama_seq_id   seq_id,
            llama_pos   p0,
            llama_pos   p1,
                  int   d) {
    return llama_kv_self_seq_div(ctx, seq_id, p0, p1, d);
}

void llama_kv_self_seq_div(
        llama_context * ctx,
         llama_seq_id   seq_id,
            llama_pos   p0,
            llama_pos   p1,
                  int   d) {
    return llama_kv_cache_seq_div(ctx->get_kv_self(), seq_id, p0, p1, d);
}

// deprecated
llama_pos llama_kv_cache_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
    return llama_kv_self_seq_pos_max(ctx, seq_id);
}

llama_pos llama_kv_self_seq_pos_max(llama_context * ctx, llama_seq_id seq_id) {
    return llama_kv_cache_seq_pos_max(ctx->get_kv_self(), seq_id);
}

// deprecated
void llama_kv_cache_defrag(llama_context * ctx) {
    return llama_kv_self_defrag(ctx);
}

void llama_kv_self_defrag(llama_context * ctx) {
    llama_kv_cache_defrag(ctx->get_kv_self());
}

// deprecated
bool llama_kv_cache_can_shift(const llama_context * ctx) {
    return llama_kv_self_can_shift(ctx);
}

bool llama_kv_self_can_shift(const llama_context * ctx) {
    return llama_kv_cache_can_shift(ctx->get_kv_self());
}

// deprecated
void llama_kv_cache_update(llama_context * ctx) {
    llama_kv_self_update(ctx);
}

// llama state API

// deprecated
size_t llama_get_state_size(llama_context * ctx) {
    return llama_state_get_size(ctx);
}

// deprecated
size_t llama_copy_state_data(llama_context * ctx, uint8_t * dst) {
    return llama_state_get_data(ctx, dst, -1);
}

// deprecated
size_t llama_set_state_data(llama_context * ctx, const uint8_t * src) {
    return llama_state_set_data(ctx, src, -1);
}

// deprecated
bool llama_load_session_file(llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
    return llama_state_load_file(ctx, path_session, tokens_out, n_token_capacity, n_token_count_out);
}

// deprecated
bool llama_save_session_file(llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
    return llama_state_save_file(ctx, path_session, tokens, n_token_count);
}

// Returns the *actual* size of the state.
// Intended to be used when saving to state to a buffer.
size_t llama_state_get_size(llama_context * ctx) {
    return ctx->state_get_size();
}

size_t llama_state_get_data(llama_context * ctx, uint8_t * dst, size_t size) {
    ctx->synchronize();

    return ctx->state_get_data(dst, size);
}

// Sets the state reading from the specified source address
size_t llama_state_set_data(llama_context * ctx, const uint8_t * src, size_t size) {
    ctx->synchronize();

    return ctx->state_set_data(src, size);
}

bool llama_state_load_file(llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
    ctx->synchronize();

    try {
        return ctx->state_load_file(path_session, tokens_out, n_token_capacity, n_token_count_out);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error loading session file: %s\n", __func__, err.what());
        return false;
    }
}

bool llama_state_save_file(llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
    ctx->synchronize();

    try {
        return ctx->state_save_file(path_session, tokens, n_token_count);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error saving session file: %s\n", __func__, err.what());
        return false;
    }
}

size_t llama_state_seq_get_size(llama_context * ctx, llama_seq_id seq_id) {
    return ctx->state_seq_get_size(seq_id);
}

size_t llama_state_seq_get_data(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) {
    ctx->synchronize();

    return ctx->state_seq_get_data(seq_id, dst, size);
}

size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) {
    ctx->synchronize();

    return ctx->state_seq_set_data(seq_id, src, size);
}

size_t llama_state_seq_save_file(llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {
    ctx->synchronize();

    try {
        return ctx->state_seq_save_file(seq_id, filepath, tokens, n_token_count);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error saving sequence state file: %s\n", __func__, err.what());
        return 0;
    }
}

size_t llama_state_seq_load_file(llama_context * ctx, const char * filepath, llama_seq_id dest_seq_id, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out) {
    ctx->synchronize();

    try {
        return ctx->state_seq_load_file(dest_seq_id, filepath, tokens_out, n_token_capacity, n_token_count_out);
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: error loading sequence state file: %s\n", __func__, err.what());
        return 0;
    }
}

///

int32_t llama_encode(
        llama_context * ctx,
          llama_batch   batch) {
    const int ret = ctx->encode(batch);
    if (ret != 0) {
        LLAMA_LOG_ERROR("%s: failed to encode, ret = %d\n", __func__, ret);
    }

    return ret;
}

int32_t llama_decode(
        llama_context * ctx,
          llama_batch   batch) {
    const int ret = ctx->decode(batch);
    if (ret != 0) {
        LLAMA_LOG_ERROR("%s: failed to decode, ret = %d\n", __func__, ret);
    }

    return ret;
}

//
// perf
//

llama_perf_context_data llama_perf_context(const llama_context * ctx) {
    llama_perf_context_data data = {};

    if (ctx == nullptr) {
        return data;
    }

    data = ctx->perf_get_data();

    return data;
}

void llama_perf_context_print(const llama_context * ctx) {
    const auto data = llama_perf_context(ctx);

    const double t_end_ms = 1e-3 * ggml_time_us();

    LLAMA_LOG_INFO("%s:        load time = %10.2f ms\n", __func__, data.t_load_ms);
    LLAMA_LOG_INFO("%s: prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)\n",
            __func__, data.t_p_eval_ms, data.n_p_eval, data.t_p_eval_ms / data.n_p_eval, 1e3 / data.t_p_eval_ms * data.n_p_eval);
    LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
            __func__, data.t_eval_ms, data.n_eval, data.t_eval_ms / data.n_eval, 1e3 / data.t_eval_ms * data.n_eval);
    LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (t_end_ms - data.t_start_ms), (data.n_p_eval + data.n_eval));
}

void llama_perf_context_reset(llama_context * ctx) {
    ctx->perf_reset();
}