llama.cpp/src/llama-graph.cpp

#include "llama-graph.h"

#include "llama-impl.h"
#include "llama-batch.h"
#include "llama-cparams.h"

#include "llama-kv-cache.h"
#include "llama-kv-cache-iswa.h"
#include "llama-memory-hybrid.h"
#include "llama-memory-recurrent.h"

#include <cassert>
#include <cmath>
#include <cstring>

void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
    if (ubatch->token) {
        const int64_t n_tokens = ubatch->n_tokens;

        ggml_backend_tensor_set(tokens, ubatch->token, 0, n_tokens*ggml_element_size(tokens));
    }

    if (ubatch->embd) {
        const int64_t n_embd   = embd->ne[0];
        const int64_t n_tokens = ubatch->n_tokens;

        ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*ggml_element_size(embd));
    }
}

bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) {
    bool res = true;

    res &= (!tokens && !params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens);
    res &= (!embd   && !params.ubatch.embd)  || (embd   &&   embd->ne[0] == params.ubatch.n_tokens);

    return res;
}

void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
    if (ubatch->pos && pos) {
        const int64_t n_tokens = ubatch->n_tokens;

        if (ubatch->token && n_pos_per_embd == 4) {
            // in case we're using M-RoPE with text tokens, convert the 1D positions to 4D
            // the 3 first dims are the same, and 4th dim is all 0
            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd);
            // copy the first dimension
            for (int i = 0; i < n_tokens; ++i) {
                pos_data[               i] = ubatch->pos[i];
                pos_data[    n_tokens + i] = ubatch->pos[i];
                pos_data[2 * n_tokens + i] = ubatch->pos[i];
                pos_data[3 * n_tokens + i] = 0; // 4th dim is 0
            }
            ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
        } else {
            ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_embd*ggml_element_size(pos));
        }
    }
}

bool llm_graph_input_pos::can_reuse(const llm_graph_params & params) {
    bool res = true;

    res &= pos->ne[0] == params.ubatch.n_tokens;

    return res;
}

void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
    if (ubatch->pos && attn_scale) {
        const int64_t n_tokens = ubatch->n_tokens;

        std::vector<float> attn_scale_data(n_tokens, 0.0f);
        for (int i = 0; i < n_tokens; ++i) {
            const float pos = ubatch->pos[i];
            attn_scale_data[i] = std::log(
                std::floor((pos + 1.0f) / n_attn_temp_floor_scale) + 1.0
            ) * f_attn_temp_scale + 1.0;
        }

        ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*ggml_element_size(attn_scale));
    }
}

void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {
    if (pos_bucket) {
        const int64_t n_tokens = ubatch->n_tokens;

        GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer));
        GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing

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

        for (int h = 0; h < 1; ++h) {
            for (int j = 0; j < n_tokens; ++j) {
                for (int i = 0; i < n_tokens; ++i) {
                    data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch->pos[i], ubatch->pos[j], hparams.n_rel_attn_bkts, true);
                }
            }
        }
    }
}

void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) {
    if (pos_bucket) {
        mctx->set_input_pos_bucket(pos_bucket, ubatch);
    }
}

void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) {
    GGML_ASSERT(out_ids);

    const int64_t n_tokens = ubatch->n_tokens;

    GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer));
    int32_t * data = (int32_t *) out_ids->data;

    if (n_outputs == n_tokens) {
        for (int i = 0; i < n_tokens; ++i) {
            data[i] = i;
        }

        return;
    }

    GGML_ASSERT(ubatch->output);

    int n_outputs = 0;

    for (int i = 0; i < n_tokens; ++i) {
        if (ubatch->output[i]) {
            data[n_outputs++] = i;
        }
    }
}

bool llm_graph_input_out_ids::can_reuse(const llm_graph_params & params) {
    bool res = true;

    res &= n_outputs == params.n_outputs;

    return res;
}

void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
        const int64_t n_tokens     = ubatch->n_tokens;
        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
        const int64_t n_seqs_unq   = ubatch->n_seqs_unq;

        GGML_ASSERT(mean);
        GGML_ASSERT(ggml_backend_buffer_is_host(mean->buffer));

        float * data = (float *) mean->data;
        memset(mean->data, 0, n_tokens*n_seqs_unq*ggml_element_size(mean));

        std::vector<uint64_t> sums(n_seqs_unq, 0);
        for (int i = 0; i < n_tokens; i += n_seq_tokens) {
            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
                const int32_t      seq_idx = ubatch->seq_idx[seq_id];

                sums[seq_idx] += ubatch->n_seq_tokens;
            }
        }

        std::vector<float> div(n_seqs_unq, 0.0f);
        for (int s = 0; s < n_seqs_unq; ++s) {
            const uint64_t sum = sums[s];
            if (sum > 0) {
                div[s] = 1.0f/float(sum);
            }
        }

        for (int i = 0; i < n_tokens; i += n_seq_tokens) {
            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
                const int32_t      seq_idx = ubatch->seq_idx[seq_id];

                for (int j = 0; j < n_seq_tokens; ++j) {
                    data[seq_idx*n_tokens + i + j] = div[seq_idx];
                }
            }
        }
    }
}

void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
    const int64_t n_tokens     = ubatch->n_tokens;
    const int64_t n_seqs_unq   = ubatch->n_seqs_unq;

    if (cparams.embeddings && (
        cparams.pooling_type == LLAMA_POOLING_TYPE_CLS  ||
        cparams.pooling_type == LLAMA_POOLING_TYPE_RANK ||
        cparams.pooling_type == LLAMA_POOLING_TYPE_LAST
    )) {
        GGML_ASSERT(cls);
        GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer));

        uint32_t * data = (uint32_t *) cls->data;
        memset(cls->data, 0, n_seqs_unq*ggml_element_size(cls));

        std::vector<int> target_pos(n_seqs_unq, -1);
        std::vector<int> target_row(n_seqs_unq, -1);

        bool last = cparams.pooling_type == LLAMA_POOLING_TYPE_LAST;

        for (int i = 0; i < n_tokens; ++i) {
            const llama_pos pos = ubatch->pos[i];

            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
                const llama_seq_id seq_id  = ubatch->seq_id[i][s];
                const int32_t      seq_idx = ubatch->seq_idx[seq_id];

                if (
                    (target_pos[seq_idx] == -1) ||
                    ( last && pos >= target_pos[seq_idx]) ||
                    (!last && pos <  target_pos[seq_idx])
                ) {
                    target_pos[seq_idx] = pos;
                    target_row[seq_idx] = i;
                }
            }
        }

        for (int s = 0; s < n_seqs_unq; ++s) {
            if (target_row[s] >= 0) {
                data[s] = target_row[s];
            }
        }
    }
}

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

    const int64_t n_rs = mctx->get_n_rs();

    if (s_copy) {
        GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
        int32_t * data = (int32_t *) s_copy->data;

        // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
        for (uint32_t i = 0; i < n_rs; ++i) {
            data[i] = mctx->s_copy(i);
        }
    }
}

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

    if (cross_embd && !cross->v_embd.empty()) {
        assert(cross_embd->type == GGML_TYPE_F32);

        ggml_backend_tensor_set(cross_embd, cross->v_embd.data(), 0, ggml_nbytes(cross_embd));
    }
}

static void print_mask(float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) {
    LLAMA_LOG_DEBUG("%s: === Attention mask ===\n", __func__);
    const char * swa_type_str = (swa_type == LLAMA_SWA_TYPE_NONE) ? "LLAMA_SWA_TYPE_NONE" :
                          (swa_type == LLAMA_SWA_TYPE_STANDARD) ? "LLAMA_SWA_TYPE_STANDARD" :
                          (swa_type == LLAMA_SWA_TYPE_CHUNKED) ? "LLAMA_SWA_TYPE_CHUNKED" :
                          (swa_type == LLAMA_SWA_TYPE_SYMMETRIC) ? "LLAMA_SWA_TYPE_SYMMETRIC" : "unknown";
    LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n", __func__, (int)n_swa, (int)n_kv, swa_type_str);
    LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n", __func__);
    LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n", __func__);

    LLAMA_LOG_DEBUG("    ");
    for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
        LLAMA_LOG_DEBUG("%2d", j);
    }
    LLAMA_LOG_DEBUG("\n");

    for (int i = 0; i < std::min((int64_t)20, n_tokens); ++i) {
        LLAMA_LOG_DEBUG(" %2d ", i);
        for (int j = 0; j < std::min((int64_t)20, n_kv); ++j) {
            float val = data[i * n_kv + j];
            if (val == -INFINITY) {
                LLAMA_LOG_DEBUG(" ∞");
            } else {
                LLAMA_LOG_DEBUG(" 0");
            }
        }
        LLAMA_LOG_DEBUG("\n");
    }
}

void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
    const int64_t n_kv     = ubatch->n_tokens;
    const int64_t n_tokens = ubatch->n_tokens;

    GGML_ASSERT(kq_mask);
    GGML_ASSERT(ggml_backend_buffer_is_host(kq_mask->buffer));

    float * data = (float *) kq_mask->data;

    // [TAG_NO_CACHE_ISWA]
    GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "TODO: implement");

    for (int h = 0; h < 1; ++h) {
        for (int i1 = 0; i1 < n_tokens; ++i1) {
            const llama_seq_id s1 = ubatch->seq_id[i1][0];

            for (int i0 = 0; i0 < n_tokens; ++i0) {
                float f = -INFINITY;

                for (int s = 0; s < ubatch->n_seq_id[i0]; ++s) {
                    const llama_seq_id s0 = ubatch->seq_id[i0][0];

                    if (s0 != s1) {
                        continue; // skip different sequences
                    }

                    if (cparams.causal_attn && ubatch->pos[i0] > ubatch->pos[i1]) {
                        continue; // skip future tokens for causal attention
                    }

                    // TODO: this does not take into account that some layers are SWA and others are note (i.e. iSWA) [TAG_NO_CACHE_ISWA]
                    //if (hparams.is_masked_swa(ubatch->pos[i0], ubatch->pos[i1])) {
                    //    continue; // skip masked tokens for SWA
                    //}

                    // TODO: reimplement this like in llama_kv_cache_unified
                    if (hparams.use_alibi) {
                        f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
                    } else {
                        f = 0.0f;
                    }
                }
                data[h*(n_kv*n_tokens) + i1*n_kv + i0] = f;
            }
        }
    }
    if (debug) {
        print_mask(data, n_tokens, n_kv, hparams.n_swa, hparams.swa_type);
    }
}

void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) {
    mctx->set_input_k_idxs(self_k_idxs, ubatch);
    mctx->set_input_v_idxs(self_v_idxs, ubatch);

    mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
}

bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) {
    const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);

    this->mctx = mctx;

    bool res = true;

    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
  //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there

    res &= self_kq_mask->ne[0] == mctx->get_n_kv();
    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);

    return res;
}

void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) {
    mctx->get_base()->set_input_k_idxs(self_k_idxs, ubatch);
    mctx->get_base()->set_input_v_idxs(self_v_idxs, ubatch);

    mctx->get_base()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);

    mctx->get_swa()->set_input_k_idxs(self_k_idxs_swa, ubatch);
    mctx->get_swa()->set_input_v_idxs(self_v_idxs_swa, ubatch);

    mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
}

bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) {
    const auto * mctx = static_cast<const llama_kv_cache_iswa_context *>(params.mctx);

    this->mctx = mctx;

    bool res = true;

    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
  //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there

    res &= self_k_idxs_swa->ne[0] == params.ubatch.n_tokens;
  //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there

    res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv();
    res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);

    res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv();
    res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD);

    return res;
}

void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
    GGML_ASSERT(cross_kq_mask);

    const int64_t n_enc    = cross_kq_mask->ne[0];
    const int64_t n_tokens = ubatch->n_tokens;

    GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer));
    GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing

    float * data = (float *) cross_kq_mask->data;

    for (int h = 0; h < 1; ++h) {
        for (int i = 0; i < n_tokens; ++i) {
            for (int j = 0; j < n_enc; ++j) {
                float f = -INFINITY;

                for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
                    const llama_seq_id seq_id = ubatch->seq_id[i][s];

                    if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) {
                        f = 0.0f;
                    }
                }

                data[h*(n_enc*n_tokens) + i*n_enc + j] = f;
            }
        }

        for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
            for (int j = 0; j < n_enc; ++j) {
                data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY;
            }
        }
    }
}

void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
    inp_attn->set_input(ubatch);
    inp_rs->set_input(ubatch);
}

//
// llm_graph_result
//

llm_graph_result::llm_graph_result(int64_t max_nodes) : max_nodes(max_nodes) {
    reset();

    const char * LLAMA_GRAPH_RESULT_DEBUG = getenv("LLAMA_GRAPH_RESULT_DEBUG");
    debug = LLAMA_GRAPH_RESULT_DEBUG ? atoi(LLAMA_GRAPH_RESULT_DEBUG) : 0;
}

int64_t llm_graph_result::get_max_nodes() const {
    return max_nodes;
}

void llm_graph_result::reset() {
    t_tokens      = nullptr;
    t_logits      = nullptr;
    t_embd        = nullptr;
    t_embd_pooled = nullptr;

    params = {};

    inputs.clear();

    buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));

    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));

    gf = ggml_new_graph_custom(ctx_compute.get(), max_nodes, false);
}

void llm_graph_result::set_inputs(const llama_ubatch * ubatch) {
    for (auto & input : inputs) {
        input->set_input(ubatch);
    }
}

bool llm_graph_result::can_reuse(const llm_graph_params & params) {
    if (!this->params.allow_reuse(params)) {
        if (debug > 1) {
            LLAMA_LOG_DEBUG("%s: cannot reuse graph due to incompatible graph parameters\n", __func__);
        }

        return false;
    }

    if (debug > 1) {
        LLAMA_LOG_DEBUG("%s: checking compatibility of %d inputs:\n", __func__, (int) inputs.size());
    }

    bool res = true;

    for (auto & input : inputs) {
        const bool cur = input->can_reuse(params);

        if (debug > 1) {
            LLAMA_LOG_DEBUG("%s: can_reuse = %d\n", "placeholder", cur);
        }

        res = res && cur;
    }

    if (debug > 0) {
        LLAMA_LOG_DEBUG("%s: can reuse graph = %d\n", __func__, res);
    }

    return res;
}

llm_graph_input_i * llm_graph_result::add_input(llm_graph_input_ptr input) {
    inputs.emplace_back(std::move(input));
    return inputs.back().get();
}

void llm_graph_result::set_params(const llm_graph_params & params) {
    this->params = params;
}

//
// llm_graph_context
//

llm_graph_context::llm_graph_context(const llm_graph_params & params) :
    arch             (params.arch),
    hparams          (params.hparams),
    cparams          (params.cparams),
    ubatch           (params.ubatch),
    n_embd           (hparams.n_embd),
    n_layer          (hparams.n_layer),
    n_rot            (hparams.n_rot),
    n_ctx            (cparams.n_ctx),
    n_head           (hparams.n_head()),
    n_head_kv        (hparams.n_head_kv()),
    n_embd_head_k    (hparams.n_embd_head_k),
    n_embd_k_gqa     (hparams.n_embd_k_gqa()),
    n_embd_head_v    (hparams.n_embd_head_v),
    n_embd_v_gqa     (hparams.n_embd_v_gqa()),
    n_expert         (hparams.n_expert),
    n_expert_used    (cparams.warmup ? hparams.n_expert : hparams.n_expert_used),
    freq_base        (cparams.rope_freq_base),
    freq_scale       (cparams.rope_freq_scale),
    ext_factor       (cparams.yarn_ext_factor),
    attn_factor      (cparams.yarn_attn_factor),
    beta_fast        (cparams.yarn_beta_fast),
    beta_slow        (cparams.yarn_beta_slow),
    norm_eps         (hparams.f_norm_eps),
    norm_rms_eps     (hparams.f_norm_rms_eps),
    n_tokens         (ubatch.n_tokens),
    n_outputs        (params.n_outputs),
    n_ctx_orig       (cparams.n_ctx_orig_yarn),
    pooling_type     (cparams.pooling_type),
    rope_type        (hparams.rope_type),
    sched            (params.sched),
    backend_cpu      (params.backend_cpu),
    cvec             (params.cvec),
    loras            (params.loras),
    mctx             (params.mctx),
    cross            (params.cross),
    cb_func          (params.cb),
    res              (params.res),
    ctx0             (res->get_ctx()),
    gf               (res->get_gf()) {
        res->set_params(params);
    }

void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const {
    if (cb_func) {
        cb_func(ubatch, cur, name, il);
    }
}

ggml_tensor * llm_graph_context::build_cvec(
         ggml_tensor * cur,
                 int   il) const {
    return cvec->apply_to(ctx0, cur, il);
}

ggml_tensor * llm_graph_context::build_lora_mm(
          ggml_tensor * w,
          ggml_tensor * cur) const {
    ggml_tensor * res = ggml_mul_mat(ctx0, w, cur);

    for (const auto & lora : *loras) {
        llama_adapter_lora_weight * lw = lora.first->get_weight(w);
        if (lw == nullptr) {
            continue;
        }

        const float adapter_scale = lora.second;
        const float scale = lw->get_scale(lora.first->alpha, adapter_scale);

        ggml_tensor * ab_cur = ggml_mul_mat(
                ctx0, lw->b,
                ggml_mul_mat(ctx0, lw->a, cur)
                );

        ab_cur = ggml_scale(ctx0, ab_cur, scale);
        res = ggml_add(ctx0, res, ab_cur);
    }

    return res;
}

ggml_tensor * llm_graph_context::build_lora_mm_id(
          ggml_tensor * w,   // ggml_tensor * as
          ggml_tensor * cur, // ggml_tensor * b
          ggml_tensor * ids) const {
    ggml_tensor * res = ggml_mul_mat_id(ctx0, w, cur, ids);
    for (const auto & lora : *loras) {
        llama_adapter_lora_weight * lw = lora.first->get_weight(w);
        if (lw == nullptr) {
            continue;
        }

        const float alpha = lora.first->alpha;
        const float rank  = (float) lw->b->ne[0];
        const float scale = alpha ? lora.second * alpha / rank : lora.second;

        ggml_tensor * ab_cur = ggml_mul_mat_id(
                ctx0, lw->b,
                ggml_mul_mat_id(ctx0, lw->a, cur, ids),
                ids
                );

        ab_cur = ggml_scale(ctx0, ab_cur, scale);
        res = ggml_add(ctx0, res, ab_cur);
    }

    return res;
}

ggml_tensor * llm_graph_context::build_norm(
         ggml_tensor * cur,
         ggml_tensor * mw,
         ggml_tensor * mb,
       llm_norm_type   type,
                 int   il) const {
    switch (type) {
        case LLM_NORM:       cur = ggml_norm    (ctx0, cur, hparams.f_norm_eps);     break;
        case LLM_NORM_RMS:   cur = ggml_rms_norm(ctx0, cur, hparams.f_norm_rms_eps); break;
        case LLM_NORM_GROUP:
            {
                cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], 1, cur->ne[1]);
                cur = ggml_group_norm(ctx0, cur, hparams.n_norm_groups, hparams.f_norm_group_eps);
                cur = ggml_reshape_2d(ctx0, cur, cur->ne[0],    cur->ne[2]);
            } break;
    }

    if (mw || mb) {
        cb(cur, "norm", il);
    }

    if (mw) {
        cur = ggml_mul(ctx0, cur, mw);
        if (mb) {
            cb(cur, "norm_w", il);
        }
    }

    if (mb) {
        cur = ggml_add(ctx0, cur, mb);
    }

    return cur;
}

ggml_tensor * llm_graph_context::build_ffn(
         ggml_tensor * cur,
         ggml_tensor * up,
         ggml_tensor * up_b,
         ggml_tensor * up_s,
         ggml_tensor * gate,
         ggml_tensor * gate_b,
         ggml_tensor * gate_s,
         ggml_tensor * down,
         ggml_tensor * down_b,
         ggml_tensor * down_s,
         ggml_tensor * act_scales,
     llm_ffn_op_type   type_op,
   llm_ffn_gate_type   type_gate,
                 int   il) const {
    ggml_tensor * tmp = up ? build_lora_mm(up, cur) : cur;
    cb(tmp, "ffn_up", il);

    if (up_b) {
        tmp = ggml_add(ctx0, tmp, up_b);
        cb(tmp, "ffn_up_b", il);
    }

    if (up_s) {
        tmp = ggml_mul(ctx0, tmp, up_s);
        cb(tmp, "ffn_up_s", il);
    }

    if (gate) {
        switch (type_gate) {
            case LLM_FFN_SEQ:
                {
                    cur = build_lora_mm(gate, tmp);
                    cb(cur, "ffn_gate", il);
                } break;
            case LLM_FFN_PAR:
                {
                    cur = build_lora_mm(gate, cur);
                    cb(cur, "ffn_gate", il);
                } break;
        }

        if (gate_b) {
            cur = ggml_add(ctx0, cur, gate_b);
            cb(cur, "ffn_gate_b", il);
        }

        if (gate_s) {
            cur = ggml_mul(ctx0, cur, gate_s);
            cb(cur, "ffn_gate_s", il);
        }

    } else {
        cur = tmp;
    }

    switch (type_op) {
        case LLM_FFN_SILU:
            if (gate && type_gate == LLM_FFN_PAR) {
                cur = ggml_swiglu_split(ctx0, cur, tmp);
                cb(cur, "ffn_swiglu", il);
                type_gate = LLM_FFN_SEQ;
            } else {
                cur = ggml_silu(ctx0, cur);
                cb(cur, "ffn_silu", il);
            } break;
        case LLM_FFN_GELU:
            if (gate && type_gate == LLM_FFN_PAR) {
                cur = ggml_geglu_split(ctx0, cur, tmp);
                cb(cur, "ffn_geglu", il);
                type_gate = LLM_FFN_SEQ;
            } else {
                cur = ggml_gelu(ctx0, cur);
                cb(cur, "ffn_gelu", il);
                if (act_scales != NULL) {
                    cur = ggml_div(ctx0, cur, act_scales);
                    cb(cur, "ffn_act", il);
                }
            } break;
        case LLM_FFN_RELU:
            if (gate && type_gate == LLM_FFN_PAR) {
                cur = ggml_reglu_split(ctx0, cur, tmp);
                cb(cur, "ffn_reglu", il);
                type_gate = LLM_FFN_SEQ;
            } else {
                cur = ggml_relu(ctx0, cur);
                cb(cur, "ffn_relu", il);
            } break;
        case LLM_FFN_RELU_SQR:
            {
                cur = ggml_relu(ctx0, cur);
                cb(cur, "ffn_relu", il);

                cur = ggml_sqr(ctx0, cur);
                cb(cur, "ffn_sqr(relu)", il);
            } break;
        case LLM_FFN_SWIGLU:
            {
                cur = ggml_swiglu(ctx0, cur);
                cb(cur, "ffn_swiglu", il);
            } break;
        case LLM_FFN_GEGLU:
            {
                cur = ggml_geglu(ctx0, cur);
                cb(cur, "ffn_geglu", il);
            } break;
        case LLM_FFN_REGLU:
            {
                cur = ggml_reglu(ctx0, cur);
                cb(cur, "ffn_reglu", il);
            } break;
        default:
            GGML_ABORT("fatal error");
    }

    if (gate && type_gate == LLM_FFN_PAR) {
        cur = ggml_mul(ctx0, cur, tmp);
        cb(cur, "ffn_gate_par", il);
    }

    if (down) {
        cur = build_lora_mm(down, cur);
        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
        }
    }

    if (down_b) {
        cb(cur, "ffn_down", il);
    }

    if (down_b) {
        cur = ggml_add(ctx0, cur, down_b);
    }

    if (down_s) {
        cur = ggml_mul(ctx0, cur, down_s);
        cb(cur, "ffn_down_s", il);
    }

    return cur;
}

ggml_tensor * llm_graph_context::build_moe_ffn(
         ggml_tensor * cur,
         ggml_tensor * gate_inp,
         ggml_tensor * up_exps,
         ggml_tensor * gate_exps,
         ggml_tensor * down_exps,
         ggml_tensor * exp_probs_b,
             int64_t   n_expert,
             int64_t   n_expert_used,
     llm_ffn_op_type   type_op,
                bool   norm_w,
                bool   scale_w,
               float   w_scale,
         llama_expert_gating_func_type gating_op,
                 int   il,
         ggml_tensor * probs_in) const {
    return build_moe_ffn(
        cur,
        gate_inp,  /* gate_inp_b  */ nullptr,
        up_exps,   /* up_exps_b   */ nullptr,
        gate_exps, /* gate_exps_b */ nullptr,
        down_exps, /* down_exps_b */ nullptr,
        exp_probs_b,
        n_expert,
        n_expert_used,
        type_op,
        norm_w,
        scale_w,
        w_scale,
        gating_op,
        il,
        probs_in
    );
}

ggml_tensor * llm_graph_context::build_moe_ffn(
         ggml_tensor * cur,
         ggml_tensor * gate_inp,
         ggml_tensor * gate_inp_b,
         ggml_tensor * up_exps,
         ggml_tensor * up_exps_b,
         ggml_tensor * gate_exps,
         ggml_tensor * gate_exps_b,
         ggml_tensor * down_exps,
         ggml_tensor * down_exps_b,
         ggml_tensor * exp_probs_b,
             int64_t   n_expert,
             int64_t   n_expert_used,
     llm_ffn_op_type   type_op,
                bool   norm_w,
                bool   scale_w,
               float   w_scale,
        llama_expert_gating_func_type gating_op,
                 int   il,
         ggml_tensor * probs_in) const {
    const int64_t n_embd   = cur->ne[0];
    const int64_t n_tokens = cur->ne[1];
    const bool weight_before_ffn = arch == LLM_ARCH_LLAMA4; // for llama4, we apply the sigmoid-ed weights before the FFN

    ggml_tensor * logits = nullptr;

    if (probs_in == nullptr) {
        logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens]
        cb(logits, "ffn_moe_logits", il);
    } else {
        logits = probs_in;
    }

    if (gate_inp_b) {
        logits = ggml_add(ctx0, logits, gate_inp_b);
        cb(logits, "ffn_moe_logits_biased", il);
    }

    ggml_tensor * probs = nullptr;
    switch (gating_op) {
        case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX:
            {
                probs = ggml_soft_max(ctx0, logits); // [n_expert, n_tokens]
            } break;
        case LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID:
            {
                probs = ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens]
            } break;
        case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT:
            {
                probs = logits; // [n_expert, n_tokens]
            } break;
        default:
            GGML_ABORT("fatal error");
    }
    cb(probs, "ffn_moe_probs", il);

    // add experts selection bias - introduced in DeepSeek V3
    // leave probs unbiased as it's later used to get expert weights
    ggml_tensor * selection_probs = probs;
    if (exp_probs_b != nullptr) {
        selection_probs = ggml_add(ctx0, probs, exp_probs_b);
        cb(selection_probs, "ffn_moe_probs_biased", il);
    }

    // llama4 doesn't have exp_probs_b, and sigmoid is only used after top_k
    // see: https://github.com/meta-llama/llama-models/blob/699a02993512fb36936b1b0741e13c06790bcf98/models/llama4/moe.py#L183-L198
    if (arch == LLM_ARCH_LLAMA4) {
        selection_probs = logits;
    }

    // select experts
    ggml_tensor * selected_experts = ggml_top_k(ctx0, selection_probs, n_expert_used); // [n_expert_used, n_tokens]
    cb(selected_experts->src[0], "ffn_moe_argsort", il);
    cb(selected_experts, "ffn_moe_topk", il);

    ggml_tensor * weights = ggml_get_rows(ctx0,
            ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts); // [1, n_expert_used, n_tokens]
    cb(weights, "ffn_moe_weights", il);

    if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT) {
        weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
        weights = ggml_soft_max(ctx0, weights); // [n_expert_used, n_tokens]
        weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens);
        cb(weights, "ffn_moe_weights_softmax", il);
    }

    if (norm_w) {
        weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);

        ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens]
        cb(weights_sum, "ffn_moe_weights_sum", il);

        weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
        cb(weights, "ffn_moe_weights_norm", il);

        weights = ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens);
    }
    if (scale_w) {
        weights = ggml_scale(ctx0, weights, w_scale);
        cb(weights, "ffn_moe_weights_scaled", il);
    }

    cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);

    if (weight_before_ffn) {
        // repeat cur to [n_embd, n_expert_used, n_tokens]
        ggml_tensor * repeated = ggml_repeat_4d(ctx0, cur, n_embd, n_expert_used, n_tokens, 1);
        cur = ggml_mul(ctx0, repeated, weights);
        cb(cur, "ffn_moe_weighted", il);
    }

    ggml_tensor * up = build_lora_mm_id(up_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
    cb(up, "ffn_moe_up", il);

    if (up_exps_b) {
        up = ggml_add_id(ctx0, up, up_exps_b, selected_experts);
        cb(up, "ffn_moe_up_biased", il);
    }

    ggml_tensor * experts = nullptr;
    if (gate_exps) {
        cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
        cb(cur, "ffn_moe_gate", il);
    } else {
        cur = up;
    }

    if (gate_exps_b) {
        cur = ggml_add_id(ctx0, cur, gate_exps_b, selected_experts);
        cb(cur, "ffn_moe_gate_biased", il);
    }

    switch (type_op) {
        case LLM_FFN_SILU:
            if (gate_exps) {
                cur = ggml_swiglu_split(ctx0, cur, up);
                cb(cur, "ffn_moe_swiglu", il);
            } else {
                cur = ggml_silu(ctx0, cur);
                cb(cur, "ffn_moe_silu", il);
            } break;
        case LLM_FFN_GELU:
            if (gate_exps) {
                cur = ggml_geglu_split(ctx0, cur, up);
                cb(cur, "ffn_moe_geglu", il);
            } else {
                cur = ggml_gelu(ctx0, cur);
                cb(cur, "ffn_moe_gelu", il);
            } break;
        case LLM_FFN_SWIGLU_OAI_MOE:
            {
                // TODO: move to hparams?
                constexpr float alpha = 1.702f;
                constexpr float limit = 7.0f;
                cur = ggml_swiglu_oai(ctx0, cur, up, alpha, limit);
                cb(cur, "ffn_moe_swiglu_oai", il);
            } break;
        case LLM_FFN_RELU:
            if (gate_exps) {
                cur = ggml_reglu_split(ctx0, cur, up);
                cb(cur, "ffn_moe_reglu", il);
            } else {
                cur = ggml_relu(ctx0, cur);
                cb(cur, "ffn_moe_relu", il);
            } break;
        default:
            GGML_ABORT("fatal error");
    }

    experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
    cb(experts, "ffn_moe_down", il);

    if (down_exps_b) {
        experts = ggml_add_id(ctx0, experts, down_exps_b, selected_experts);
        cb(experts, "ffn_moe_down_biased", il);
    }

    if (!weight_before_ffn) {
        experts = ggml_mul(ctx0, experts, weights);
        cb(cur, "ffn_moe_weighted", il);
    }

    ggml_tensor * cur_experts[LLAMA_MAX_EXPERTS] = { nullptr };

    assert(n_expert_used > 0);

    // order the views before the adds
    for (uint32_t i = 0; i < hparams.n_expert_used; ++i) {
        cur_experts[i] = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]);

        ggml_build_forward_expand(gf, cur_experts[i]);
    }

    // aggregate experts
    // note: here we explicitly use hparams.n_expert_used instead of n_expert_used
    //       to avoid potentially a large number of add nodes during warmup
    //       ref: https://github.com/ggml-org/llama.cpp/pull/14753
    ggml_tensor * moe_out = cur_experts[0];

    for (uint32_t i = 1; i < hparams.n_expert_used; ++i) {
        moe_out = ggml_add(ctx0, moe_out, cur_experts[i]);
    }

    if (hparams.n_expert_used == 1) {
        // avoid returning a non-contiguous tensor
        moe_out = ggml_cont(ctx0, moe_out);
    }

    cb(moe_out, "ffn_moe_out", il);

    return moe_out;
}

// input embeddings with optional lora
ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
    const int64_t n_embd = hparams.n_embd;

    auto inp = std::make_unique<llm_graph_input_embd>();

    ggml_tensor * cur = nullptr;

    if (ubatch.token) {
        inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
        //cb(inp->tokens, "inp_tokens", -1);
        ggml_set_input(inp->tokens);
        res->t_tokens = inp->tokens;

        cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);

        // apply lora for embedding tokens if needed
        for (const auto & lora : *loras) {
            llama_adapter_lora_weight * lw = lora.first->get_weight(tok_embd);
            if (lw == nullptr) {
                continue;
            }

            const float adapter_scale = lora.second;
            const float scale = lw->get_scale(lora.first->alpha, adapter_scale);

            ggml_tensor * inpL_delta = ggml_scale(ctx0, ggml_mul_mat(
                        ctx0, lw->b, // non-transposed lora_b
                        ggml_get_rows(ctx0, lw->a, inp->tokens)
                        ), scale);

            cur = ggml_add(ctx0, cur, inpL_delta);
        }
    } else {
        inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ubatch.n_tokens);
        ggml_set_input(inp->embd);

        cur = inp->embd;
    }

    // For Granite architecture
    if (hparams.f_embedding_scale != 0.0f) {
        cur = ggml_scale(ctx0, cur, hparams.f_embedding_scale);
    }

    cb(cur, "inp_embd", -1);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_pos() const {
    auto inp = std::make_unique<llm_graph_input_pos>(hparams.n_pos_per_embd());

    auto & cur = inp->pos;

    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, (int64_t)n_tokens*hparams.n_pos_per_embd());
    ggml_set_input(cur);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
    auto inp = std::make_unique<llm_graph_input_attn_temp>(hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);

    auto & cur = inp->attn_scale;

    // this need to be 1x1xN for broadcasting
    cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, 1, n_tokens);
    ggml_set_input(cur);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_out_ids() const {
    // note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls,
    //       but this would make the graph topology depend on the number of output tokens, which can interere with
    //       features that require constant topology such as pipline parallelism
    //       ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471
    //if (n_outputs < n_tokens) {
    //    return nullptr;
    //}

    auto inp = std::make_unique<llm_graph_input_out_ids>(hparams, cparams, n_outputs);

    auto & cur = inp->out_ids;

    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_outputs);
    ggml_set_input(cur);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_mean() const {
    auto inp = std::make_unique<llm_graph_input_mean>(cparams);

    auto & cur = inp->mean;

    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, ubatch.n_seqs_unq);
    ggml_set_input(cur);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_cls() const {
    auto inp = std::make_unique<llm_graph_input_cls>(cparams);

    auto & cur = inp->cls;

    cur = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_seqs_unq);
    ggml_set_input(cur);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_cross_embd() const {
    auto inp = std::make_unique<llm_graph_input_cross_embd>(cross);

    auto & cur = inp->cross_embd;

    // if we have the output embeddings from the encoder, use them directly
    // TODO: needs more work to be correct, for now just use the tensor shape
    //if (cross->t_embd) {
    //    cur = ggml_view_tensor(ctx0, cross->t_embd);

    //    return cur;
    //}

    const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd;
    const auto n_enc  = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;

    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc);
    ggml_set_input(cur);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
    auto inp = std::make_unique<llm_graph_input_pos_bucket>(hparams);

    auto & cur = inp->pos_bucket;

    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_tokens, n_tokens);
    ggml_set_input(cur);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);

    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, mctx_cur);

    const auto n_kv = mctx_cur->get_n_kv();

    auto & cur = inp->pos_bucket;

    cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, n_tokens);
    ggml_set_input(cur);

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

    return cur;
}

ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const {
    ggml_tensor * pos_bucket_1d = ggml_reshape_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1]);
    cb(pos_bucket_1d, "pos_bucket_1d", -1);

    ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d);

    pos_bias = ggml_reshape_3d(ctx0, pos_bias, pos_bias->ne[0], pos_bucket->ne[0], pos_bucket->ne[1]);
    pos_bias = ggml_permute   (ctx0, pos_bias, 2, 0, 1, 3);
    pos_bias = ggml_cont      (ctx0, pos_bias);

    cb(pos_bias, "pos_bias", -1);

    return pos_bias;
}

ggml_tensor * llm_graph_context::build_attn_mha(
         ggml_tensor * q,
         ggml_tensor * k,
         ggml_tensor * v,
         ggml_tensor * kq_b,
         ggml_tensor * kq_mask,
         ggml_tensor * sinks,
         ggml_tensor * v_mla,
               float   kq_scale,
                 int   il) const {
    const bool v_trans = v->nb[1] > v->nb[2];

    // split the batch into streams if needed
    const auto n_stream = k->ne[3];

    q = ggml_reshape_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream);

    q = ggml_permute(ctx0, q, 0, 2, 1, 3);
    k = ggml_permute(ctx0, k, 0, 2, 1, 3);
    v = ggml_permute(ctx0, v, 0, 2, 1, 3);

    const auto n_kv = k->ne[1];

    ggml_tensor * cur;

    // TODO: replace hardcoded padding with ggml-provided padding
    if (cparams.flash_attn && (n_kv % 256 == 0) && kq_b == nullptr) {
        GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");

        if (v_trans) {
            v = ggml_transpose(ctx0, v);
        }

        // this can happen when KV cache is not used (e.g. an embedding model with non-causal attn)
        if (k->type == GGML_TYPE_F32) {
            k = ggml_cast(ctx0, k, GGML_TYPE_F16);
        }

        if (v->type == GGML_TYPE_F32) {
            v = ggml_cast(ctx0, v, GGML_TYPE_F16);
        }

        cur = ggml_flash_attn_ext(ctx0, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias,
                                  hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f);
        cb(cur, LLAMA_TENSOR_NAME_FATTN, il);

        ggml_flash_attn_ext_add_sinks(cur, sinks);
        ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);

        if (v_mla) {
#if 0
            // v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens.
            // However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient.
            cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
            cur = ggml_mul_mat(ctx0, v_mla, cur);
#else
            // It's preferable to do the calculation as a matrix-matrix multiplication with n_tokens in dimension 1.
            // The permutations are noops and only change how the tensor data is interpreted.
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            cur = ggml_mul_mat(ctx0, v_mla, cur);
            cb(cur, "fattn_mla", il);
            cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
            cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs.
#endif
        }

        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);
    } else {
        ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
        cb(kq, "kq", il);

        // note: this op tends to require high floating point range
        //       while for some models F16 is enough, for others it is not, so we default to F32 here
        ggml_mul_mat_set_prec(kq, GGML_PREC_F32);

        if (arch == LLM_ARCH_GROK) {
            // need to do the following:
            // multiply by attn_output_multiplyer of 0.08838834764831845
            // and then :
            // kq = 30 * tanh(kq / 30)
            // before the softmax below

            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f));
            cb(kq, "kq_tanh", il);
            kq = ggml_scale(ctx0, kq, 30);
            cb(kq, "kq_scaled", il);
        }

        if (hparams.attn_soft_cap) {
            kq = ggml_scale(ctx0, kq, 1.0f / hparams.f_attn_logit_softcapping);
            cb(kq, "kq_scaled_1", il);
            kq = ggml_tanh (ctx0, kq);
            cb(kq, "kq_tanh", il);
            kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
            cb(kq, "kq_scaled_2", il);
        }

        if (kq_b) {
            kq = ggml_add(ctx0, kq, kq_b);
            cb(kq, "kq_plus_kq_b", il);
        }

        kq = ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
        ggml_soft_max_add_sinks(kq, sinks);
        cb(kq, "kq_soft_max", il);

        if (!v_trans) {
            // note: avoid this branch
            v = ggml_cont(ctx0, ggml_transpose(ctx0, v));
            cb(v, "v_cont", il);
        }

        ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq);
        cb(kqv, "kqv", il);

        // for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA
        if (v_mla) {
            kqv = ggml_mul_mat(ctx0, v_mla, kqv);
            cb(kqv, "kqv_mla", il);
        }

        cur = ggml_permute(ctx0, kqv, 0, 2, 1, 3);

        // recombine streams
        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*cur->ne[1], cur->ne[2]*cur->ne[3]);

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

    ggml_build_forward_expand(gf, cur);

    return cur;
}

llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() const {
    auto inp = std::make_unique<llm_graph_input_attn_no_cache>(hparams, cparams);

    // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch
    inp->kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
    ggml_set_input(inp->kq_mask);

    inp->kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->kq_mask, GGML_TYPE_F16) : inp->kq_mask;

    return (llm_graph_input_attn_no_cache *) res->add_input(std::move(inp));
}

ggml_tensor * llm_graph_context::build_attn(
        llm_graph_input_attn_no_cache * inp,
        ggml_tensor * wo,
        ggml_tensor * wo_b,
        ggml_tensor * q_cur,
        ggml_tensor * k_cur,
        ggml_tensor * v_cur,
        ggml_tensor * kq_b,
        ggml_tensor * sinks,
        ggml_tensor * v_mla,
            float     kq_scale,
            int       il) const {
    GGML_UNUSED(n_tokens);

    // these nodes are added to the graph together so that they are not reordered
    // by doing so, the number of splits in the graph is reduced
    ggml_build_forward_expand(gf, q_cur);
    ggml_build_forward_expand(gf, k_cur);
    ggml_build_forward_expand(gf, v_cur);

    const auto & kq_mask = inp->get_kq_mask();

    // [TAG_NO_CACHE_PAD]
    // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams
    assert(!ubatch.equal_seqs() || (k_cur->ne[3] == 1 && k_cur->ne[3] == ubatch.n_seqs_unq));

    ggml_tensor * q = q_cur;
    ggml_tensor * k = k_cur;
    ggml_tensor * v = v_cur;

    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);

    if (wo) {
        cur = build_lora_mm(wo, cur);
    }

    if (wo_b) {
        //cb(cur, "kqv_wo", il);
    }

    if (wo_b) {
        cur = ggml_add(ctx0, cur, wo_b);
    }

    return cur;
}

static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl(
           ggml_context * ctx0,
     const llama_ubatch & ubatch,
    const llama_hparams & hparams,
    const llama_cparams & cparams,
    const llama_kv_cache_context * mctx_cur) {

    auto inp = std::make_unique<llm_graph_input_attn_kv>(hparams, cparams, mctx_cur);

    {
        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");

        const auto n_kv     = mctx_cur->get_n_kv();
        const auto n_tokens = ubatch.n_tokens;
        const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;

        inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx0, ubatch);
        inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx0, ubatch);

        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
        ggml_set_input(inp->self_kq_mask);

        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
    }

    return inp;
}

llm_graph_input_attn_kv * llm_graph_context::build_attn_inp_kv() const {
    const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx);

    auto inp = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur);

    return (llm_graph_input_attn_kv *) res->add_input(std::move(inp));
}

ggml_tensor * llm_graph_context::build_attn(
        llm_graph_input_attn_kv * inp,
        ggml_tensor * wo,
        ggml_tensor * wo_b,
        ggml_tensor * q_cur,
        ggml_tensor * k_cur,
        ggml_tensor * v_cur,
        ggml_tensor * kq_b,
        ggml_tensor * sinks,
        ggml_tensor * v_mla,
            float     kq_scale,
            int       il) const {
    // these nodes are added to the graph together so that they are not reordered
    // by doing so, the number of splits in the graph is reduced
    ggml_build_forward_expand(gf, q_cur);
    ggml_build_forward_expand(gf, k_cur);
    ggml_build_forward_expand(gf, v_cur);

    const auto * mctx_cur = inp->mctx;

    // store to KV cache
    {
        const auto & k_idxs = inp->get_k_idxs();
        const auto & v_idxs = inp->get_v_idxs();

        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
    }

    const auto & kq_mask = inp->get_kq_mask();

    ggml_tensor * q = q_cur;
    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
    ggml_tensor * v = mctx_cur->get_v(ctx0, il);

    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);

    if (wo) {
        cur = build_lora_mm(wo, cur);
        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
        }
    }

    if (wo_b) {
        cur = ggml_add(ctx0, cur, wo_b);
    }

    return cur;
}

ggml_tensor * llm_graph_context::build_attn(
        llm_graph_input_attn_kv_iswa * inp,
        ggml_tensor * wo,
        ggml_tensor * wo_b,
        ggml_tensor * q_cur,
        ggml_tensor * k_cur,
        ggml_tensor * v_cur,
        ggml_tensor * kq_b,
        ggml_tensor * sinks,
        ggml_tensor * v_mla,
            float     kq_scale,
            int       il) const {
    // these nodes are added to the graph together so that they are not reordered
    // by doing so, the number of splits in the graph is reduced
    ggml_build_forward_expand(gf, q_cur);

    if (k_cur) {
        ggml_build_forward_expand(gf, k_cur);
    }

    if (v_cur) {
        ggml_build_forward_expand(gf, v_cur);
    }

    const auto * mctx_iswa = inp->mctx;

    const bool is_swa = hparams.is_swa(il);

    const auto * mctx_cur = is_swa ? mctx_iswa->get_swa() : mctx_iswa->get_base();

    // optionally store to KV cache
    if (k_cur) {
        const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs();

        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, k_idxs, il));
    }

    if (v_cur) {
        const auto & v_idxs = is_swa ? inp->get_v_idxs_swa() : inp->get_v_idxs();

        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, v_idxs, il));
    }

    const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();

    ggml_tensor * q = q_cur;
    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
    ggml_tensor * v = mctx_cur->get_v(ctx0, il);

    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);

    if (wo) {
        cur = build_lora_mm(wo, cur);
    }

    if (wo_b) {
        //cb(cur, "kqv_wo", il);
    }

    if (wo_b) {
        cur = ggml_add(ctx0, cur, wo_b);
    }

    return cur;
}

llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
    auto inp = std::make_unique<llm_graph_input_attn_cross>(cross);

    const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;

    inp->cross_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1, 1);
    ggml_set_input(inp->cross_kq_mask);

    inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->cross_kq_mask, GGML_TYPE_F16) : inp->cross_kq_mask;

    return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
}

ggml_tensor * llm_graph_context::build_attn(
        llm_graph_input_attn_cross * inp,
        ggml_tensor * wo,
        ggml_tensor * wo_b,
        ggml_tensor * q_cur,
        ggml_tensor * k_cur,
        ggml_tensor * v_cur,
        ggml_tensor * kq_b,
        ggml_tensor * sinks,
        ggml_tensor * v_mla,
            float     kq_scale,
            int       il) const {
    // these nodes are added to the graph together so that they are not reordered
    // by doing so, the number of splits in the graph is reduced
    ggml_build_forward_expand(gf, q_cur);
    ggml_build_forward_expand(gf, k_cur);
    ggml_build_forward_expand(gf, v_cur);

    const auto & kq_mask = inp->get_kq_mask_cross();

    ggml_tensor * q = q_cur;
    ggml_tensor * k = k_cur;
    ggml_tensor * v = v_cur;

    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);

    if (wo) {
        cur = build_lora_mm(wo, cur);
    }

    if (wo_b) {
        //cb(cur, "kqv_wo", il);
    }

    if (wo_b) {
        cur = ggml_add(ctx0, cur, wo_b);
    }

    return cur;
}

// TODO: maybe separate the inner implementation into a separate function
//       like with the non-sliding window equivalent
//       once sliding-window hybrid caches are a thing.
llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const {
    const auto * mctx_cur = static_cast<const llama_kv_cache_iswa_context *>(mctx);

    auto inp = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, mctx_cur);

    const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq;

    {
        const auto n_kv = mctx_cur->get_base()->get_n_kv();

        inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx0, ubatch);
        inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx0, ubatch);

        inp->self_kq_mask = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
        ggml_set_input(inp->self_kq_mask);

        inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
    }

    {
        GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache for non-SWA");

        const auto n_kv = mctx_cur->get_swa()->get_n_kv();

        inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx0, ubatch);
        inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx0, ubatch);

        inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), 1, n_stream);
        ggml_set_input(inp->self_kq_mask_swa);

        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask_swa, GGML_TYPE_F16) : inp->self_kq_mask_swa;
    }

    return (llm_graph_input_attn_kv_iswa *) res->add_input(std::move(inp));
}

ggml_tensor * llm_graph_context::build_rs(
        ggml_tensor * s,
        ggml_tensor * state_copy_main,
        ggml_tensor * state_copy_extra,
            int32_t   state_size,
            int32_t   n_seqs,
           uint32_t   n_rs,
           uint32_t   rs_head,
           uint32_t   rs_size,
            int32_t   rs_zero,
        const llm_graph_get_rows_fn & get_state_rows) const {

    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, rs_size);

    // Clear a single state which will then be copied to the other cleared states.
    // Note that this is a no-op when the view is zero-sized.
    ggml_tensor * state_zero = ggml_view_1d(ctx0, states, state_size*(rs_zero >= 0), rs_zero*states->nb[1]*(rs_zero >= 0));
    ggml_build_forward_expand(gf, ggml_scale_inplace(ctx0, state_zero, 0));

    // copy states
    // NOTE: assuming the copy destinations are ALL contained between rs_head and rs_head + n_rs
    // {state_size, rs_size} -> {state_size, n_seqs}
    ggml_tensor * output_states = get_state_rows(ctx0, states, state_copy_main);
    ggml_build_forward_expand(gf, output_states);

    // copy extra states which won't be changed further (between n_seqs and n_rs)
    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, state_copy_extra);
    ggml_build_forward_expand(gf,
        ggml_cpy(ctx0,
            states_extra,
            ggml_view_1d(ctx0, s, state_size*(n_rs - n_seqs), (rs_head + n_seqs)*state_size*ggml_element_size(s))));

    return output_states;
}

static std::unique_ptr<llm_graph_input_rs> build_rs_inp_impl(
           ggml_context * ctx0,
     const llama_ubatch & ubatch,
    const llama_memory_recurrent_context * mctx_cur) {

    auto inp = std::make_unique<llm_graph_input_rs>(mctx_cur);

    const int64_t n_rs   = mctx_cur->get_n_rs();
    const int64_t n_seqs = ubatch.n_seqs;

    inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
    ggml_set_input(inp->s_copy);

    inp->s_copy_main  = ggml_view_1d(ctx0, inp->s_copy, n_seqs, 0);
    inp->s_copy_extra = ggml_view_1d(ctx0, inp->s_copy, n_rs - n_seqs, n_seqs * inp->s_copy->nb[0]);

    return inp;
}

llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);

    auto inp = build_rs_inp_impl(ctx0, ubatch, mctx_cur);

    return (llm_graph_input_rs *) res->add_input(std::move(inp));
}

ggml_tensor * llm_graph_context::build_rs(
        llm_graph_input_rs * inp,
        ggml_tensor * s,
            int32_t   state_size,
            int32_t   n_seqs,
        const llm_graph_get_rows_fn & get_state_rows) const {
    const auto * kv_state = inp->mctx;

    return build_rs(s, inp->s_copy_main, inp->s_copy_extra, state_size, n_seqs,
                    kv_state->get_n_rs(), kv_state->get_head(), kv_state->get_size(), kv_state->get_rs_z(),
                    get_state_rows);
}

ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
    llm_graph_input_rs * inp,
    const llama_ubatch & ubatch,
                   int   il) const {
    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);

    const auto token_shift_count = hparams.token_shift_count;

    const int64_t n_seqs  = ubatch.n_seqs;

    ggml_tensor * token_shift_all = mctx_cur->get_r_l(il);

    ggml_tensor * token_shift = build_rs(
            inp, token_shift_all,
            hparams.n_embd_r(), n_seqs);

    token_shift = ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);

    return token_shift;
}

ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
         ggml_tensor * token_shift,
  const llama_ubatch & ubatch,
                 int   il) const {
    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);

    const auto token_shift_count = hparams.token_shift_count;
    const auto n_embd = hparams.n_embd;

    const int64_t n_seqs = ubatch.n_seqs;

    const auto kv_head = mctx_cur->get_head();

    return ggml_cpy(
        ctx0,
        ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0),
        ggml_view_1d(ctx0, mctx_cur->get_r_l(il), hparams.n_embd_r()*n_seqs, hparams.n_embd_r()*kv_head*ggml_element_size(mctx_cur->get_r_l(il)))
    );
}

llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
    const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx);

    auto inp_rs   = build_rs_inp_impl(ctx0, ubatch, mctx_cur->get_recr());
    auto inp_attn = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur->get_attn());

    auto inp = std::make_unique<llm_graph_input_mem_hybrid>(std::move(inp_attn), std::move(inp_rs), mctx_cur);

    return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
}

void llm_graph_context::build_pooling(
        ggml_tensor * cls,
        ggml_tensor * cls_b,
        ggml_tensor * cls_out,
        ggml_tensor * cls_out_b) const {
    if (!cparams.embeddings) {
        return;
    }

    ggml_tensor * inp = res->t_embd;

    //// find result_norm tensor for input
    //for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) {
    //    inp = ggml_graph_node(gf, i);
    //    if (strcmp(inp->name, "result_norm") == 0 || strcmp(inp->name, "result_embd") == 0) {
    //        break;
    //    }

    //    inp = nullptr;
    //}

    GGML_ASSERT(inp != nullptr && "missing result_norm/result_embd tensor");

    ggml_tensor * cur;

    switch (pooling_type) {
        case LLAMA_POOLING_TYPE_NONE:
            {
                cur = inp;
            } break;
        case LLAMA_POOLING_TYPE_MEAN:
            {
                ggml_tensor * inp_mean = build_inp_mean();
                cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean);
            } break;
        case LLAMA_POOLING_TYPE_CLS:
        case LLAMA_POOLING_TYPE_LAST:
            {
                ggml_tensor * inp_cls = build_inp_cls();
                cur = ggml_get_rows(ctx0, inp, inp_cls);
            } break;
        case LLAMA_POOLING_TYPE_RANK:
            {
                ggml_tensor * inp_cls = build_inp_cls();
                inp = ggml_get_rows(ctx0, inp, inp_cls);

                if (cls) {
                    // classification head
                    // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
                    cur = ggml_mul_mat(ctx0, cls, inp);
                    if (cls_b) {
                        cur = ggml_add(ctx0, cur, cls_b);
                    }
                    cur = ggml_tanh(ctx0, cur);

                    // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
                    // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896
                    if (cls_out) {
                        cur = ggml_mul_mat(ctx0, cls_out, cur);
                        if (cls_out_b) {
                            cur = ggml_add(ctx0, cur, cls_out_b);
                        }
                    }
                } else if (cls_out) {
                    // Single layer classification head (direct projection)
                    // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
                    cur = ggml_mul_mat(ctx0, cls_out, inp);
                    if (cls_out_b) {
                        cur = ggml_add(ctx0, cur, cls_out_b);
                    }
                } else {
                    GGML_ABORT("RANK pooling requires either cls+cls_b or cls_out+cls_out_b");
                }
            } break;
        default:
            {
                GGML_ABORT("unknown pooling type");
            }
    }

    cb(cur, "result_embd_pooled", -1);
    res->t_embd_pooled = cur;

    ggml_build_forward_expand(gf, cur);
}

int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) {
    // TODO move to hparams if a T5 variant appears that uses a different value
    const int64_t max_distance = 128;

    if (bidirectional) {
        n_buckets >>= 1;
    }

    const int64_t max_exact = n_buckets >> 1;

    int32_t relative_position = x - y;
    int32_t relative_bucket = 0;

    if (bidirectional) {
        relative_bucket += (relative_position > 0) * n_buckets;
        relative_position = abs(relative_position);
    } else {
        relative_position = -std::min<int32_t>(relative_position, 0);
    }

    int32_t relative_position_if_large = floorf(max_exact + logf(1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(1.0 * max_distance / max_exact));
    relative_position_if_large = std::min<int32_t>(relative_position_if_large, n_buckets - 1);
    relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large);

    return relative_bucket;
}