cui-llama.rn 1.4.6 → 1.5.0

package/cpp/llama-graph.cpp

@@ -0,0 +1,1695 @@
+#include "llama-graph.h"
+
+#include "llama-impl.h"
+#include "llama-batch.h"
+#include "llama-cparams.h"
+#include "llama-kv-cache.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstring>
+
+static 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;
+}
+
+void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
+    if (ubatch->token) {
+        const int64_t n_tokens = ubatch->n_tokens;
+
+        lm_ggml_backend_tensor_set(tokens, ubatch->token, 0, n_tokens*lm_ggml_element_size(tokens));
+    }
+
+    if (ubatch->embd) {
+        const int64_t n_embd   = embd->ne[0];
+        const int64_t n_tokens = ubatch->n_tokens;
+
+        lm_ggml_backend_tensor_set(embd, ubatch->embd, 0, n_tokens*n_embd*lm_ggml_element_size(embd));
+    }
+}
+
+void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
+    if (ubatch->pos && pos) {
+        const int64_t n_tokens = ubatch->n_tokens;
+
+        lm_ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_token*lm_ggml_element_size(pos));
+    }
+}
+
+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;
+        }
+
+        lm_ggml_backend_tensor_set(attn_scale, attn_scale_data.data(), 0, n_tokens*n_pos_per_token*lm_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;
+
+        LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(pos_bucket->buffer));
+        LM_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) {
+        const int64_t n_tokens = ubatch->n_tokens;
+
+        LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(pos_bucket->buffer));
+        LM_GGML_ASSERT(!ubatch->equal_seqs); // TODO: use ubatch->n_seqs instead of failing
+
+        int32_t * data = (int32_t *) pos_bucket->data;
+
+        const int64_t n_kv = kv_self->n;
+
+        for (int h = 0; h < 1; ++h) {
+            for (int j = 0; j < n_tokens; ++j) {
+                for (int i = 0; i < n_kv; ++i) {
+                    data[h*(n_kv*n_tokens) + j*n_kv + i] = llama_relative_position_bucket(kv_self->cells[i].pos, ubatch->pos[j], hparams.n_rel_attn_bkts, false);
+                }
+            }
+        }
+    }
+}
+
+void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) {
+    if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        //LM_GGML_ASSERT(out_ids && "every model that can must skip unused outputs");
+
+        if (!out_ids) {
+            LLAMA_LOG_WARN("%s: 'out_ids' is not created\n", __func__);
+        } else {
+            const int64_t n_tokens = ubatch->n_tokens;
+
+            LM_GGML_ASSERT(lm_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;
+                }
+            } else if (ubatch->output) {
+                int32_t n_outputs = 0;
+                for (int i = 0; i < n_tokens; ++i) {
+                    if (ubatch->output[i]) {
+                        data[n_outputs++] = i;
+                    }
+                }
+                // the graph needs to have been passed the correct number of outputs
+                LM_GGML_ASSERT(n_outputs == n_outputs);
+            } else if (n_outputs == 1) {
+                // only keep last output
+                data[0] = n_tokens - 1;
+            } else {
+                LM_GGML_ASSERT(n_outputs == 0);
+            }
+        }
+    }
+}
+
+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       = ubatch->n_seqs;
+
+        LM_GGML_ASSERT(mean);
+        LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(mean->buffer));
+
+        float * data = (float *) mean->data;
+        memset(mean->data, 0, n_tokens * n_tokens * lm_ggml_element_size(mean));
+
+        std::vector<uint64_t> sum(n_tokens, 0);
+
+        for (int s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = ubatch->seq_id[s][0];
+
+            // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true
+            LM_GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
+
+            sum[seq_id] += ubatch->n_seq_tokens;
+        }
+
+        std::vector<float> div(n_tokens, 0.0f);
+        for (int i = 0; i < n_tokens; ++i) {
+            const uint64_t s = sum[i];
+            if (s > 0) {
+                div[i] = 1.0f/float(s);
+            }
+        }
+
+        for (int s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = ubatch->seq_id[s][0];
+
+            for (int i = 0; i < n_seq_tokens; ++i) {
+                data[seq_id*n_tokens + s*n_seq_tokens + i] = div[seq_id];
+            }
+        }
+    }
+}
+
+void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
+    if (cparams.embeddings && (
+                cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
+                cparams.pooling_type == LLAMA_POOLING_TYPE_RANK)) {
+        const int64_t n_tokens     = ubatch->n_tokens;
+        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+        const int64_t n_seqs       = ubatch->n_seqs;
+
+        LM_GGML_ASSERT(cls);
+        LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(cls->buffer));
+
+        uint32_t * data = (uint32_t *) cls->data;
+        memset(cls->data, 0, n_tokens * lm_ggml_element_size(cls));
+
+        for (int s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = ubatch->seq_id[s][0];
+
+            // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true
+            LM_GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS or RANK");
+
+            for (int i = 0; i < n_seq_tokens; ++i) {
+                const llama_pos pos = ubatch->pos[s*n_seq_tokens + i];
+
+                if (pos == 0) {
+                    data[seq_id] = s*n_seq_tokens + i;
+                }
+            }
+        }
+    }
+
+    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_LAST) {
+        const int64_t n_tokens     = ubatch->n_tokens;
+        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+        const int64_t n_seqs       = ubatch->n_seqs;
+
+        LM_GGML_ASSERT(cls);
+        LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(cls->buffer));
+
+        uint32_t * data = (uint32_t *) cls->data;
+        memset(cls->data, 0, n_tokens * lm_ggml_element_size(cls));
+
+        std::vector<int> last_pos(n_tokens, -1);
+        std::vector<int> last_row(n_tokens, -1);
+
+        for (int s = 0; s < n_seqs; ++s) {
+            const llama_seq_id seq_id = ubatch->seq_id[s][0];
+
+            // TODO: adapt limits to n_seqs when ubatch->equal_seqs is true
+            LM_GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == LAST");
+
+            for (int i = 0; i < n_seq_tokens; ++i) {
+                const llama_pos pos = ubatch->pos[s*n_seq_tokens + i];
+
+                if (pos >= last_pos[seq_id]) {
+                    last_pos[seq_id] = pos;
+                    last_row[seq_id] = s*n_seq_tokens + i;
+                }
+            }
+        }
+
+        for (int i = 0; i < n_tokens; ++i) {
+            if (last_row[i] >= 0) {
+                data[i] = last_row[i];
+            }
+        }
+    }
+}
+
+void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
+    LM_GGML_UNUSED(ubatch);
+
+    const int64_t n_kv = kv_self->n;
+
+    if (s_copy) {
+        LM_GGML_ASSERT(lm_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_kv; ++i) {
+            const uint32_t  cell_id = i + kv_self->head;
+
+            //////////////////////////////////////////////
+            // TODO: this should not mutate the KV cache !
+            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
+
+            // prevent out-of-bound sources
+            if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) {
+                kv_cell.src = cell_id;
+            }
+
+            data[i] = kv_cell.src;
+
+            // TODO: do not mutate the KV cache
+            // ensure copy only happens once
+            if (kv_cell.src != (int32_t) cell_id) {
+                kv_cell.src = cell_id;
+            }
+        }
+    }
+}
+
+void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
+    LM_GGML_UNUSED(ubatch);
+
+    const int64_t n_kv = kv_self->n;
+
+    if (s_mask) {
+        LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(s_mask->buffer));
+        float * data = (float *) s_mask->data;
+
+        // clear unused states
+        for (int i = 0; i < n_kv; ++i) {
+            const uint32_t  cell_id = i + kv_self->head;
+
+            //////////////////////////////////////////////
+            // TODO: this should not mutate the KV cache !
+            llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
+
+            data[i] = (float) (kv_cell.src >= 0);
+
+            // only clear once
+            if (kv_cell.src < 0) {
+                kv_cell.src = cell_id;
+            }
+        }
+    }
+}
+
+void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
+    LM_GGML_UNUSED(ubatch);
+
+    if (cross_embd && !cross->v_embd.empty()) {
+        assert(cross_embd->type == LM_GGML_TYPE_F32);
+
+        lm_ggml_backend_tensor_set(cross_embd, cross->v_embd.data(), 0, lm_ggml_nbytes(cross_embd));
+    }
+}
+
+void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
+    if (kq_mask) {
+        if (cparams.causal_attn) {
+            const int64_t n_kv         = ubatch->n_tokens;
+            const int64_t n_tokens     = ubatch->n_tokens;
+            const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+            const int64_t n_seqs       = ubatch->n_seqs;
+
+            LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(kq_mask->buffer));
+            float * data = (float *) kq_mask->data;
+
+            for (int h = 0; h < 1; ++h) {
+                for (int s1 = 0; s1 < n_seqs; ++s1) {
+                    const llama_seq_id seq_id = ubatch->seq_id[s1][0];
+
+                    for (int j = 0; j < n_seq_tokens; ++j) {
+                        const int32_t tj = s1*n_seq_tokens + j;
+
+                        for (int s0 = 0; s0 < n_seqs; ++s0) {
+                            for (int i = 0; i < n_seq_tokens; ++i) {
+                                const int32_t ti = s0*n_seq_tokens + i;
+                                float f = -INFINITY;
+
+                                for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
+                                    if (ubatch->seq_id[s0][s] == seq_id && ubatch->pos[ti] <= ubatch->pos[tj]) {
+                                        if (hparams.use_alibi) {
+                                            f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]);
+                                        } else {
+                                            f = 0.0f;
+                                        }
+                                        break;
+                                    }
+                                }
+
+                                data[h*(n_kv*n_tokens) + tj*n_kv + ti] = f;
+                            }
+                        }
+                    }
+                }
+            }
+        } else {
+            const int64_t n_tokens     = ubatch->n_tokens;
+            const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+            const int64_t n_seqs       = ubatch->n_seqs;
+            const int64_t n_stride     = ubatch->n_tokens;
+
+            LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(kq_mask->buffer));
+
+            float * data = (float *) kq_mask->data;
+
+            for (int h = 0; h < 1; ++h) {
+                for (int s1 = 0; s1 < n_seqs; ++s1) {
+                    const llama_seq_id seq_id = ubatch->seq_id[s1][0];
+
+                    for (int j = 0; j < n_seq_tokens; ++j) {
+                        const int32_t tj = s1*n_seq_tokens + j;
+
+                        for (int s0 = 0; s0 < n_seqs; ++s0) {
+                            for (int i = 0; i < n_seq_tokens; ++i) {
+                                const int32_t ti = s0*n_seq_tokens + i;
+                                float f = -INFINITY;
+
+                                for (int s = 0; s < ubatch->n_seq_id[s0]; ++s) {
+                                    if (ubatch->seq_id[s0][s] == seq_id) {
+                                        if (hparams.use_alibi) {
+                                            f = -std::abs(ubatch->pos[ti] - ubatch->pos[tj]);
+                                        } else {
+                                            f = 0.0f;
+                                        }
+                                        break;
+                                    }
+                                }
+
+                                data[h*(n_tokens*n_tokens) + tj*n_stride + ti] = f;
+                            }
+                        }
+
+                        for (int i = n_tokens; i < n_stride; ++i) {
+                            data[h*(n_tokens*n_tokens) + tj*n_stride + i] = -INFINITY;
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+void llm_graph_input_attn_kv_unified::set_input(const llama_ubatch * ubatch) {
+    if (self_kq_mask || self_kq_mask_swa) {
+        const int64_t n_kv         = kv_self->n;
+        const int64_t n_tokens     = ubatch->n_tokens;
+        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
+        const int64_t n_seqs       = ubatch->n_seqs;
+
+        float * data     = nullptr;
+        float * data_swa = nullptr;
+
+        if (self_kq_mask) {
+            LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(self_kq_mask->buffer));
+            data = (float *) self_kq_mask->data;
+        }
+
+        if (self_kq_mask_swa) {
+            LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(self_kq_mask_swa->buffer));
+            data_swa = (float *) self_kq_mask_swa->data;
+        }
+
+        // Use only the previous KV cells of the correct sequence for each token of the ubatch.
+        // It's assumed that if a token in the batch has multiple sequences, they are equivalent.
+        // Example with a cache of 10 tokens, 2 tokens populated in cache and 3 tokens in batch:
+        //   Causal mask:
+        //      xxx-------
+        //      xxxx------
+        //      xxxxx-----
+        //   Non-causal mask:
+        //      xxxxx-----
+        //      xxxxx-----
+        //      xxxxx-----
+        // To visualize the mask, see https://github.com/ggml-org/llama.cpp/pull/12615
+        for (int h = 0; h < 1; ++h) {
+            for (int s = 0; s < n_seqs; ++s) {
+                const llama_seq_id seq_id = ubatch->seq_id[s][0];
+
+                for (int j = 0; j < n_seq_tokens; ++j) {
+                    const llama_pos pos = ubatch->pos[s*n_seq_tokens + j];
+                    for (int i = 0; i < n_kv; ++i) {
+                        float f;
+                        // mask the token if:
+                        if (!kv_self->cells[i].has_seq_id(seq_id) // not the correct sequence
+                            || (cparams.causal_attn && kv_self->cells[i].pos > pos) // for causal, mask future tokens
+                        ) {
+                            f = -INFINITY;
+                        } else {
+                            if (hparams.use_alibi) {
+                                f = -std::abs(kv_self->cells[i].pos - pos);
+                            } else {
+                                f = 0.0f;
+                            }
+                        }
+
+                        if (data) {
+                            data[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
+                        }
+
+                        // may need to cut off old tokens for sliding window
+                        // TODO @ngxson : we are currently re-using the swa logic to store the chunked mask, we should rename SWA to something more generic like "aux mask"
+                        if (data_swa) {
+                            if (hparams.n_attn_chunk) {
+                                llama_pos pos_chunk_start = (pos / hparams.n_attn_chunk) * hparams.n_attn_chunk;
+                                if (kv_self->cells[i].pos < pos_chunk_start || pos < pos_chunk_start) {
+                                    f = -INFINITY;
+                                }
+                            } else {
+                                if (pos - kv_self->cells[i].pos >= (int32_t)hparams.n_swa) {
+                                    f = -INFINITY;
+                                }
+                            }
+                            data_swa[h*(n_kv*n_tokens) + s*(n_kv*n_seq_tokens) + j*n_kv + i] = f;
+                        }
+                    }
+                }
+            }
+
+            // mask padded tokens
+            if (data) {
+                for (int i = n_tokens; i < LM_GGML_PAD(n_tokens, LM_GGML_KQ_MASK_PAD); ++i) {
+                    for (int j = 0; j < n_kv; ++j) {
+                        data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
+                    }
+                }
+            }
+
+            // mask padded tokens
+            if (data_swa) {
+                for (int i = n_tokens; i < LM_GGML_PAD(n_tokens, LM_GGML_KQ_MASK_PAD); ++i) {
+                    for (int j = 0; j < n_kv; ++j) {
+                        data_swa[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
+                    }
+                }
+            }
+        }
+    }
+}
+
+void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
+    if (cross_kq_mask) {
+        const int64_t n_enc    = cross_kq_mask->ne[0];
+        const int64_t n_tokens = ubatch->n_tokens;
+
+        LM_GGML_ASSERT(lm_ggml_backend_buffer_is_host(cross_kq_mask->buffer));
+        LM_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 j = 0; j < n_tokens; ++j) {
+                for (int i = 0; i < n_enc; ++i) {
+                    float f = -INFINITY;
+                    for (int s = 0; s < ubatch->n_seq_id[j]; ++s) {
+                        const llama_seq_id seq_id = ubatch->seq_id[j][s];
+                        if (cross->seq_ids_enc[i].find(seq_id) != cross->seq_ids_enc[i].end()) {
+                            f = 0.0f;
+                        }
+                    }
+                    data[h*(n_enc*n_tokens) + j*n_enc + i] = f;
+                }
+            }
+
+            for (int i = n_tokens; i < LM_GGML_PAD(n_tokens, LM_GGML_KQ_MASK_PAD); ++i) {
+                for (int j = 0; j < n_enc; ++j) {
+                    data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY;
+                }
+            }
+        }
+    }
+}
+
+//
+// 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_ctx_per_seq    (cparams.n_ctx / cparams.n_seq_max),
+    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),
+    ctx0             (params.ctx),
+    sched            (params.sched),
+    backend_cpu      (params.backend_cpu),
+    cvec             (params.cvec),
+    loras            (params.loras),
+    memory           (params.memory),
+    cross            (params.cross),
+    cb_func          (params.cb),
+    res              (std::make_unique<llm_graph_result>()) {
+    }
+
+int64_t llm_graph_context::n_pos_per_token() const {
+    return arch == LLM_ARCH_QWEN2VL ? 4 : 1;
+}
+
+void llm_graph_context::cb(lm_ggml_tensor * cur, const char * name, int il) const {
+    if (cb_func) {
+        cb_func(ubatch, cur, name, il);
+    }
+}
+
+lm_ggml_tensor * llm_graph_context::build_cvec(
+         lm_ggml_tensor * cur,
+                 int   il) const {
+    return cvec->apply_to(ctx0, cur, il);
+}
+
+lm_ggml_tensor * llm_graph_context::build_lora_mm(
+          lm_ggml_tensor * w,
+          lm_ggml_tensor * cur) const {
+    lm_ggml_tensor * res = lm_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);
+
+        lm_ggml_tensor * ab_cur = lm_ggml_mul_mat(
+                ctx0, lw->b,
+                lm_ggml_mul_mat(ctx0, lw->a, cur)
+                );
+
+        ab_cur = lm_ggml_scale(ctx0, ab_cur, scale);
+        res = lm_ggml_add(ctx0, res, ab_cur);
+    }
+
+    return res;
+}
+
+lm_ggml_tensor * llm_graph_context::build_lora_mm_id(
+          lm_ggml_tensor * w,   // lm_ggml_tensor * as
+          lm_ggml_tensor * cur, // lm_ggml_tensor * b
+          lm_ggml_tensor * ids) const {
+    lm_ggml_tensor * res = lm_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;
+
+        lm_ggml_tensor * ab_cur = lm_ggml_mul_mat_id(
+                ctx0, lw->b,
+                lm_ggml_mul_mat_id(ctx0, lw->a, cur, ids),
+                ids
+                );
+
+        ab_cur = lm_ggml_scale(ctx0, ab_cur, scale);
+        res = lm_ggml_add(ctx0, res, ab_cur);
+    }
+
+    return res;
+}
+
+lm_ggml_tensor * llm_graph_context::build_norm(
+         lm_ggml_tensor * cur,
+         lm_ggml_tensor * mw,
+         lm_ggml_tensor * mb,
+       llm_norm_type   type,
+                 int   il) const {
+    switch (type) {
+        case LLM_NORM:       cur = lm_ggml_norm    (ctx0, cur, hparams.f_norm_eps);     break;
+        case LLM_NORM_RMS:   cur = lm_ggml_rms_norm(ctx0, cur, hparams.f_norm_rms_eps); break;
+        case LLM_NORM_GROUP:
+            {
+                cur = lm_ggml_reshape_3d(ctx0, cur, cur->ne[0], 1, cur->ne[1]);
+                cur = lm_ggml_group_norm(ctx0, cur, hparams.n_norm_groups, hparams.f_norm_group_eps);
+                cur = lm_ggml_reshape_2d(ctx0, cur, cur->ne[0],    cur->ne[2]);
+            } break;
+    }
+
+    if (mw || mb) {
+        cb(cur, "norm", il);
+    }
+
+    if (mw) {
+        cur = lm_ggml_mul(ctx0, cur, mw);
+        if (mb) {
+            cb(cur, "norm_w", il);
+        }
+    }
+
+    if (mb) {
+        cur = lm_ggml_add(ctx0, cur, mb);
+    }
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_ffn(
+         lm_ggml_tensor * cur,
+         lm_ggml_tensor * up,
+         lm_ggml_tensor * up_b,
+         lm_ggml_tensor * up_s,
+         lm_ggml_tensor * gate,
+         lm_ggml_tensor * gate_b,
+         lm_ggml_tensor * gate_s,
+         lm_ggml_tensor * down,
+         lm_ggml_tensor * down_b,
+         lm_ggml_tensor * down_s,
+         lm_ggml_tensor * act_scales,
+     llm_ffn_op_type   type_op,
+   llm_ffn_gate_type   type_gate,
+                 int   il) const {
+    lm_ggml_tensor * tmp = up ? build_lora_mm(up, cur) : cur;
+    cb(tmp, "ffn_up", il);
+
+    if (up_b) {
+        tmp = lm_ggml_add(ctx0, tmp, up_b);
+        cb(tmp, "ffn_up_b", il);
+    }
+
+    if (up_s) {
+        tmp = lm_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 = lm_ggml_add(ctx0, cur, gate_b);
+            cb(cur, "ffn_gate_b", il);
+        }
+
+        if (gate_s) {
+            cur = lm_ggml_mul(ctx0, cur, gate_s);
+            cb(cur, "ffn_gate_s", il);
+        }
+
+    } else {
+        cur = tmp;
+    }
+
+    switch (type_op) {
+        case LLM_FFN_SILU:
+            {
+                cur = lm_ggml_silu(ctx0, cur);
+                cb(cur, "ffn_silu", il);
+            } break;
+        case LLM_FFN_GELU:
+            {
+                cur = lm_ggml_gelu(ctx0, cur);
+                cb(cur, "ffn_gelu", il);
+                if (act_scales != NULL) {
+                    cur = lm_ggml_div(ctx0, cur, act_scales);
+                    cb(cur, "ffn_act", il);
+                }
+            } break;
+        case LLM_FFN_RELU:
+            {
+                cur = lm_ggml_relu(ctx0, cur);
+                cb(cur, "ffn_relu", il);
+            } break;
+        case LLM_FFN_RELU_SQR:
+            {
+                cur = lm_ggml_relu(ctx0, cur);
+                cb(cur, "ffn_relu", il);
+
+                cur = lm_ggml_sqr(ctx0, cur);
+                cb(cur, "ffn_sqr(relu)", il);
+            } break;
+        case LLM_FFN_SWIGLU:
+            {
+                // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
+                int64_t split_point = cur->ne[0] / 2;
+                lm_ggml_tensor * x0 = lm_ggml_cont(ctx0, lm_ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], 0));
+                lm_ggml_tensor * x1 = lm_ggml_cont(ctx0, lm_ggml_view_2d(ctx0, cur, split_point, cur->ne[1], cur->nb[1], split_point * lm_ggml_element_size(cur)));
+
+                x0 = lm_ggml_silu(ctx0, x0);
+                cb(cur, "ffn_silu", il);
+
+                cur = lm_ggml_mul(ctx0, x0, x1);
+                cb(cur, "ffn_mul", il);
+            } break;
+    }
+
+    if (type_gate == LLM_FFN_PAR) {
+        cur = lm_ggml_mul(ctx0, cur, tmp);
+        cb(cur, "ffn_gate_par", il);
+    }
+
+    if (down) {
+        cur = build_lora_mm(down, cur);
+    }
+
+    if (down_b) {
+        cb(cur, "ffn_down", il);
+    }
+
+    if (down_b) {
+        cur = lm_ggml_add(ctx0, cur, down_b);
+    }
+
+    if (down_s) {
+        cur = lm_ggml_mul(ctx0, cur, down_s);
+        cb(cur, "ffn_down_s", il);
+    }
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_moe_ffn(
+         lm_ggml_tensor * cur,
+         lm_ggml_tensor * gate_inp,
+         lm_ggml_tensor * up_exps,
+         lm_ggml_tensor * gate_exps,
+         lm_ggml_tensor * down_exps,
+         lm_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) 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
+
+    lm_ggml_tensor * logits = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens]
+    cb(logits, "ffn_moe_logits", il);
+
+    lm_ggml_tensor * probs = nullptr;
+    switch (gating_op) {
+        case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX:
+            {
+                probs = lm_ggml_soft_max(ctx0, logits); // [n_expert, n_tokens]
+            } break;
+        case LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID:
+            {
+                probs = lm_ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens]
+            } break;
+        default:
+            LM_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
+    lm_ggml_tensor * selection_probs = probs;
+    if (exp_probs_b != nullptr) {
+        selection_probs = lm_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
+    lm_ggml_tensor * selected_experts = lm_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);
+
+    lm_ggml_tensor * weights = lm_ggml_get_rows(ctx0,
+            lm_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 (norm_w) {
+        weights = lm_ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens);
+
+        lm_ggml_tensor * weights_sum = lm_ggml_sum_rows(ctx0, weights); // [1, n_tokens]
+        cb(weights_sum, "ffn_moe_weights_sum", il);
+
+        weights = lm_ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens]
+        cb(weights, "ffn_moe_weights_norm", il);
+
+        weights = lm_ggml_reshape_3d(ctx0, weights, 1, n_expert_used, n_tokens);
+    }
+    if (scale_w) {
+        weights = lm_ggml_scale(ctx0, weights, w_scale);
+        cb(weights, "ffn_moe_weights_scaled", il);
+    }
+
+    cur = lm_ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
+
+    if (weight_before_ffn) {
+        // TODO: this is a workaround as we don't yet have a repeat op that takes custom dim (lm_ggml_repeat_4d)
+        lm_ggml_tensor * repeated = lm_ggml_new_tensor_3d(ctx0, cur->type, n_embd, n_expert_used, n_tokens);
+        repeated = lm_ggml_repeat(ctx0, cur, repeated); // [n_embd, n_expert_used, n_tokens]
+        cur = lm_ggml_mul(ctx0, repeated, weights);
+        cb(cur, "ffn_moe_weighted", il);
+    }
+
+    lm_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);
+
+    lm_ggml_tensor * gate = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
+    cb(gate, "ffn_moe_gate", il);
+
+    switch (type_op) {
+        case LLM_FFN_SILU:
+            {
+                gate = lm_ggml_silu(ctx0, gate);
+                cb(gate, "ffn_moe_silu", il);
+            } break;
+        case LLM_FFN_GELU:
+            {
+                gate = lm_ggml_gelu(ctx0, gate);
+                cb(gate, "ffn_moe_gelu", il);
+            } break;
+        default:
+            LM_GGML_ABORT("fatal error");
+    }
+
+    lm_ggml_tensor * par = lm_ggml_mul(ctx0, up, gate); // [n_ff, n_expert_used, n_tokens]
+    cb(par, "ffn_moe_gate_par", il);
+
+    lm_ggml_tensor * experts = build_lora_mm_id(down_exps, par, selected_experts); // [n_embd, n_expert_used, n_tokens]
+    cb(experts, "ffn_moe_down", il);
+
+    if (!weight_before_ffn) {
+        experts = lm_ggml_mul(ctx0, experts, weights);
+        cb(cur, "ffn_moe_weighted", il);
+    }
+
+    // aggregate experts
+    lm_ggml_tensor * moe_out = nullptr;
+    for (int i = 0; i < n_expert_used; ++i) {
+        lm_ggml_tensor * cur_expert = lm_ggml_view_2d(ctx0, experts, n_embd, n_tokens,
+                experts->nb[2], i*experts->nb[1]);
+
+        if (i == 0) {
+            moe_out = cur_expert;
+        } else {
+            moe_out = lm_ggml_add(ctx0, moe_out, cur_expert);
+        }
+    }
+
+    if (n_expert_used == 1) {
+        // avoid returning a non-contiguous tensor
+        moe_out = lm_ggml_cont(ctx0, moe_out);
+    }
+
+    cb(moe_out, "ffn_moe_out", il);
+
+    return moe_out;
+}
+
+// input embeddings with optional lora
+lm_ggml_tensor * llm_graph_context::build_inp_embd(lm_ggml_tensor * tok_embd) const {
+    const int64_t n_embd = hparams.n_embd;
+
+    auto inp = std::make_unique<llm_graph_input_embd>();
+
+    lm_ggml_tensor * cur = nullptr;
+
+    if (ubatch.token) {
+        inp->tokens = lm_ggml_new_tensor_1d(ctx0, LM_GGML_TYPE_I32, ubatch.n_tokens);
+        //cb(inp->tokens, "inp_tokens", -1);
+        lm_ggml_set_input(inp->tokens);
+
+        cur = lm_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);
+
+            lm_ggml_tensor * inpL_delta = lm_ggml_scale(ctx0, lm_ggml_mul_mat(
+                        ctx0, lw->b, // non-transposed lora_b
+                        lm_ggml_get_rows(ctx0, lw->a, inp->tokens)
+                        ), scale);
+
+            cur = lm_ggml_add(ctx0, cur, inpL_delta);
+        }
+    } else {
+        inp->embd = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_F32, n_embd, ubatch.n_tokens);
+        lm_ggml_set_input(inp->embd);
+
+        cur = inp->embd;
+    }
+
+    // For Granite architecture
+    if (hparams.f_embedding_scale != 0.0f) {
+        cur = lm_ggml_scale(ctx0, cur, hparams.f_embedding_scale);
+    }
+
+    cb(cur, "inp_embd", -1);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_inp_pos() const {
+    auto inp = std::make_unique<llm_graph_input_pos>(n_pos_per_token());
+
+    auto & cur = inp->pos;
+
+    cur = lm_ggml_new_tensor_1d(ctx0, LM_GGML_TYPE_I32, n_tokens*n_pos_per_token());
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_inp_attn_scale() const {
+    auto inp = std::make_unique<llm_graph_input_attn_temp>(n_pos_per_token(), hparams.n_attn_temp_floor_scale, hparams.f_attn_temp_scale);
+
+    auto & cur = inp->attn_scale;
+
+    cur = lm_ggml_new_tensor_3d(ctx0, LM_GGML_TYPE_F32, 1, 1, n_tokens*n_pos_per_token());
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_inp_out_ids() const {
+    auto inp = std::make_unique<llm_graph_input_out_ids>(hparams, cparams, n_outputs);
+
+    auto & cur = inp->out_ids;
+
+    cur = lm_ggml_new_tensor_1d(ctx0, LM_GGML_TYPE_I32, n_outputs);
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_inp_mean() const {
+    auto inp = std::make_unique<llm_graph_input_mean>(cparams);
+
+    auto & cur = inp->mean;
+
+    cur = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_F32, n_tokens, n_tokens);
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_inp_cls() const {
+    auto inp = std::make_unique<llm_graph_input_cls>(cparams);
+
+    auto & cur = inp->cls;
+
+    cur = lm_ggml_new_tensor_1d(ctx0, LM_GGML_TYPE_I32, n_tokens);
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_inp_s_copy() const {
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+
+    auto inp = std::make_unique<llm_graph_input_s_copy>(kv_self);
+
+    const auto n_kv = kv_self->n;
+
+    auto & cur = inp->s_copy;
+
+    cur = lm_ggml_new_tensor_1d(ctx0, LM_GGML_TYPE_I32, n_kv);
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_inp_s_mask() const {
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+
+    auto inp = std::make_unique<llm_graph_input_s_mask>(kv_self);
+
+    const auto n_kv = kv_self->n;
+
+    auto & cur = inp->s_mask;
+
+    cur = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_F32, 1, n_kv);
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_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 = lm_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 = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_F32, n_embd, n_enc);
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_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 = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_I32, n_tokens, n_tokens);
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+
+    auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(hparams, kv_self);
+
+    const auto n_kv = kv_self->n;
+
+    auto & cur = inp->pos_bucket;
+
+    cur = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_I32, n_kv, n_tokens);
+    lm_ggml_set_input(cur);
+
+    res->add_input(std::move(inp));
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_pos_bias(lm_ggml_tensor * pos_bucket, lm_ggml_tensor * attn_rel_b) const {
+    lm_ggml_tensor * pos_bucket_1d = lm_ggml_reshape_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1]);
+    cb(pos_bucket_1d, "pos_bucket_1d", -1);
+
+    lm_ggml_tensor * pos_bias = lm_ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d);
+
+    pos_bias = lm_ggml_reshape_3d(ctx0, pos_bias, pos_bias->ne[0], pos_bucket->ne[0], pos_bucket->ne[1]);
+    pos_bias = lm_ggml_permute   (ctx0, pos_bias, 2, 0, 1, 3);
+    pos_bias = lm_ggml_cont      (ctx0, pos_bias);
+
+    cb(pos_bias, "pos_bias", -1);
+
+    return pos_bias;
+}
+
+lm_ggml_tensor * llm_graph_context::build_attn_mha(
+         lm_ggml_cgraph * gf,
+         lm_ggml_tensor * q,
+         lm_ggml_tensor * k,
+         lm_ggml_tensor * v,
+         lm_ggml_tensor * kq_b,
+         lm_ggml_tensor * kq_mask,
+             bool      v_trans,
+             float     kq_scale) const {
+  //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);
+
+  //const int64_t n_head    = hparams.n_head(il);
+  //const int64_t n_head_kv = hparams.n_head_kv(il);
+
+  //const auto & n_embd_head_k = hparams.n_embd_head_k;
+  //const auto & n_embd_head_v = hparams.n_embd_head_v;
+
+    const auto n_embd_head_v = v_trans ? v->ne[1] : v->ne[0];
+
+    const auto n_tokens = q->ne[1];
+    const auto n_head   = q->ne[2];
+    const auto n_kv     = k->ne[1];
+
+    lm_ggml_tensor * cur;
+
+    // TODO: replace hardcoded padding with ggml-provided padding
+    if (cparams.flash_attn && (n_kv % 256 == 0) && kq_b == nullptr) {
+        LM_GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");
+
+        if (v_trans) {
+            v = lm_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 == LM_GGML_TYPE_F32) {
+            k = lm_ggml_cast(ctx0, k, LM_GGML_TYPE_F16);
+        }
+
+        if (v->type == LM_GGML_TYPE_F32) {
+            v = lm_ggml_cast(ctx0, v, LM_GGML_TYPE_F16);
+        }
+
+        cur = lm_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);
+
+        lm_ggml_flash_attn_ext_set_prec(cur, LM_GGML_PREC_F32);
+
+        cur = lm_ggml_reshape_2d(ctx0, cur, n_embd_head_v*n_head, n_tokens);
+    } else {
+        lm_ggml_tensor * kq = lm_ggml_mul_mat(ctx0, k, q);
+
+        // 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
+        lm_ggml_mul_mat_set_prec(kq, LM_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 = lm_ggml_tanh(ctx0, lm_ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f));
+            kq = lm_ggml_scale(ctx0, kq, 30);
+        }
+
+        if (hparams.attn_soft_cap) {
+            kq = lm_ggml_scale(ctx0, kq, 1.0f / hparams.f_attn_logit_softcapping);
+            kq = lm_ggml_tanh (ctx0, kq);
+            kq = lm_ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
+        }
+
+        if (kq_b) {
+            kq = lm_ggml_add(ctx0, kq, kq_b);
+        }
+
+        kq = lm_ggml_soft_max_ext(ctx0, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
+
+        if (!v_trans) {
+            // note: avoid this branch
+            v = lm_ggml_cont(ctx0, lm_ggml_transpose(ctx0, v));
+        }
+
+        lm_ggml_tensor * kqv = lm_ggml_mul_mat(ctx0, v, kq);
+
+        lm_ggml_tensor * kqv_merged = lm_ggml_permute(ctx0, kqv, 0, 2, 1, 3);
+
+        cur = lm_ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens);
+
+        if (!cparams.offload_kqv) {
+            // all nodes between the KV store and the attention output are run on the CPU
+            lm_ggml_backend_sched_set_tensor_backend(sched, cur, backend_cpu);
+        }
+    }
+
+    lm_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 = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_F32, n_tokens, LM_GGML_PAD(n_tokens, LM_GGML_KQ_MASK_PAD));
+    //cb(inp_kq_mask, "KQ_mask", -1);
+    lm_ggml_set_input(inp->kq_mask);
+
+    inp->kq_mask_cnv = cparams.flash_attn ? lm_ggml_cast(ctx0, inp->kq_mask, LM_GGML_TYPE_F16) : inp->kq_mask;
+
+    return (llm_graph_input_attn_no_cache *) res->add_input(std::move(inp));
+}
+
+lm_ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_attn_no_cache * inp,
+        lm_ggml_cgraph * gf,
+        lm_ggml_tensor * wo,
+        lm_ggml_tensor * wo_b,
+        lm_ggml_tensor * q_cur,
+        lm_ggml_tensor * k_cur,
+        lm_ggml_tensor * v_cur,
+        lm_ggml_tensor * kq_b,
+            float     kq_scale,
+            int       il) const {
+    LM_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
+    lm_ggml_build_forward_expand(gf, q_cur);
+    lm_ggml_build_forward_expand(gf, k_cur);
+    lm_ggml_build_forward_expand(gf, v_cur);
+
+    const auto & kq_mask = inp->get_kq_mask();
+
+    lm_ggml_tensor * q = lm_ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
+    //cb(q, "q", il);
+
+    lm_ggml_tensor * k = lm_ggml_permute(ctx0, k_cur, 0, 2, 1, 3);
+    //cb(k, "k", il);
+
+    lm_ggml_tensor * v = lm_ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
+    //cb(k, "v", il);
+
+    lm_ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale);
+
+    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 = lm_ggml_add(ctx0, cur, wo_b);
+    }
+
+    return cur;
+}
+
+llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+
+    auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_self);
+
+    const auto n_kv = kv_self->n;
+
+    inp->self_kq_mask = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_F32, n_kv, LM_GGML_PAD(n_tokens, LM_GGML_KQ_MASK_PAD));
+    //cb(inp->self_kq_mask, "KQ_mask", -1);
+    lm_ggml_set_input(inp->self_kq_mask);
+
+    inp->self_kq_mask_cnv = cparams.flash_attn ? lm_ggml_cast(ctx0, inp->self_kq_mask, LM_GGML_TYPE_F16) : inp->self_kq_mask;
+
+    if (hparams.n_swa_pattern > 1) {
+        LM_GGML_ASSERT(hparams.n_swa > 0);
+
+        inp->self_kq_mask_swa = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_F32, n_kv, LM_GGML_PAD(n_tokens, LM_GGML_KQ_MASK_PAD));
+        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
+        lm_ggml_set_input(inp->self_kq_mask_swa);
+
+        inp->self_kq_mask_swa_cnv = cparams.flash_attn ? lm_ggml_cast(ctx0, inp->self_kq_mask_swa, LM_GGML_TYPE_F16) : inp->self_kq_mask_swa;
+    }
+
+    return (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
+}
+
+lm_ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_attn_kv_unified * inp,
+        lm_ggml_cgraph * gf,
+        lm_ggml_tensor * wo,
+        lm_ggml_tensor * wo_b,
+        lm_ggml_tensor * q_cur,
+        lm_ggml_tensor * k_cur,
+        lm_ggml_tensor * v_cur,
+        lm_ggml_tensor * kq_b,
+            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
+    lm_ggml_build_forward_expand(gf, q_cur);
+    lm_ggml_build_forward_expand(gf, k_cur);
+    lm_ggml_build_forward_expand(gf, v_cur);
+
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+    const auto & n_ctx = cparams.n_ctx;
+
+    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);
+
+    const auto n_tokens = q_cur->ne[2];
+
+    const bool v_trans = !cparams.flash_attn;
+
+    // store to KV cache
+    {
+        LM_GGML_ASSERT(!kv_self->recurrent);
+
+        const auto kv_head = kv_self->head;
+
+        LM_GGML_ASSERT(kv_self->size == n_ctx);
+
+        lm_ggml_tensor * k_cache_view = lm_ggml_view_1d(ctx0, kv_self->k_l[il], n_tokens*n_embd_k_gqa, lm_ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa)*kv_head);
+        //cb(k_cache_view, "k_cache_view", il);
+
+        // note: storing RoPE-ed version of K in the KV cache
+        lm_ggml_build_forward_expand(gf, lm_ggml_cpy(ctx0, k_cur, k_cache_view));
+
+        v_cur = lm_ggml_reshape_2d(ctx0, v_cur, n_embd_v_gqa, n_tokens);
+
+        lm_ggml_tensor * v_cache_view = nullptr;
+
+        if (!v_trans) {
+            v_cache_view = lm_ggml_view_1d(ctx0, kv_self->v_l[il], n_tokens*n_embd_v_gqa, lm_ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa)*kv_head);
+        } else {
+            // note: the V cache is transposed when not using flash attention
+            v_cache_view = lm_ggml_view_2d(ctx0, kv_self->v_l[il], n_tokens, n_embd_v_gqa,
+                    (  n_ctx)*lm_ggml_element_size(kv_self->v_l[il]),
+                    (kv_head)*lm_ggml_element_size(kv_self->v_l[il]));
+
+            v_cur = lm_ggml_transpose(ctx0, v_cur);
+        }
+        //cb(v_cache_view, "v_cache_view", il);
+
+        lm_ggml_build_forward_expand(gf, lm_ggml_cpy(ctx0, v_cur, v_cache_view));
+    }
+
+    const bool is_swa = hparams.is_swa(il);
+
+    const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask();
+
+    const auto n_kv = kv_self->n;
+
+    const int64_t n_head_kv = hparams.n_head_kv(il);
+
+    const auto & n_embd_head_k = hparams.n_embd_head_k;
+    const auto & n_embd_head_v = hparams.n_embd_head_v;
+
+    lm_ggml_tensor * q = lm_ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
+    //cb(q, "q", il);
+
+    lm_ggml_tensor * k =
+        lm_ggml_view_3d(ctx0, kv_self->k_l[il],
+                n_embd_head_k, n_kv, n_head_kv,
+                lm_ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
+                lm_ggml_row_size(kv_self->k_l[il]->type, n_embd_head_k),
+                0);
+    //cb(k, "k", il);
+
+    lm_ggml_tensor * v = !v_trans ?
+        lm_ggml_view_3d(ctx0, kv_self->v_l[il],
+                n_embd_head_v, n_kv, n_head_kv,
+                lm_ggml_row_size(kv_self->v_l[il]->type, n_embd_v_gqa),
+                lm_ggml_row_size(kv_self->v_l[il]->type, n_embd_head_v),
+                0) :
+        lm_ggml_view_3d(ctx0, kv_self->v_l[il],
+                n_kv, n_embd_head_v, n_head_kv,
+                lm_ggml_element_size(kv_self->v_l[il])*n_ctx,
+                lm_ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v,
+                0);
+
+    lm_ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_trans, kq_scale);
+    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 = lm_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 = lm_ggml_new_tensor_2d(ctx0, LM_GGML_TYPE_F32, n_enc, LM_GGML_PAD(n_tokens, LM_GGML_KQ_MASK_PAD));
+    lm_ggml_set_input(inp->cross_kq_mask);
+
+    inp->cross_kq_mask_cnv = cparams.flash_attn ? lm_ggml_cast(ctx0, inp->cross_kq_mask, LM_GGML_TYPE_F16) : inp->cross_kq_mask;
+
+    return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
+}
+
+lm_ggml_tensor * llm_graph_context::build_attn(
+        llm_graph_input_attn_cross * inp,
+        lm_ggml_cgraph * gf,
+        lm_ggml_tensor * wo,
+        lm_ggml_tensor * wo_b,
+        lm_ggml_tensor * q_cur,
+        lm_ggml_tensor * k_cur,
+        lm_ggml_tensor * v_cur,
+        lm_ggml_tensor * kq_b,
+            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
+    lm_ggml_build_forward_expand(gf, q_cur);
+    lm_ggml_build_forward_expand(gf, k_cur);
+    lm_ggml_build_forward_expand(gf, v_cur);
+
+    const auto & kq_mask = inp->get_kq_mask_cross();
+
+    lm_ggml_tensor * q = lm_ggml_permute(ctx0, q_cur, 0, 2, 1, 3);
+    //cb(q, "q", il);
+
+    lm_ggml_tensor * k = lm_ggml_permute(ctx0, k_cur, 0, 2, 1, 3);
+    //cb(k, "k", il);
+
+    lm_ggml_tensor * v = lm_ggml_permute(ctx0, v_cur, 0, 2, 1, 3);
+    //cb(k, "v", il);
+
+    lm_ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale);
+
+    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 = lm_ggml_add(ctx0, cur, wo_b);
+    }
+
+    return cur;
+}
+
+lm_ggml_tensor * llm_graph_context::build_copy_mask_state(
+         lm_ggml_cgraph * gf,
+         lm_ggml_tensor * s,
+         lm_ggml_tensor * state_copy,
+         lm_ggml_tensor * state_mask,
+             int32_t   n_state,
+             int32_t   n_seqs) const {
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+
+    const auto n_kv    = kv_self->n;
+    const auto kv_head = kv_self->head;
+
+    lm_ggml_tensor * states = lm_ggml_reshape_2d(ctx0, s, n_state, kv_self->size);
+
+    // copy states
+    // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
+    // this shrinks the tensors's ne[1] to n_kv
+    states = lm_ggml_get_rows(ctx0, states, state_copy);
+
+    // clear states of sequences which are starting at the beginning of this batch
+    // FIXME: zero-out NANs?
+    states = lm_ggml_mul(ctx0, states, state_mask);
+
+    // copy states which won't be changed further (between n_seqs and n_kv)
+    lm_ggml_build_forward_expand(gf,
+        lm_ggml_cpy(ctx0,
+            lm_ggml_view_1d(ctx0, states, n_state*(n_kv - n_seqs), (n_seqs          )*n_state*lm_ggml_element_size(states)),
+            lm_ggml_view_1d(ctx0, s,      n_state*(n_kv - n_seqs), (kv_head + n_seqs)*n_state*lm_ggml_element_size(s))));
+
+    // the part of the states that will be used and modified
+    return lm_ggml_view_2d(ctx0, states, n_state, n_seqs, states->nb[1], 0);
+}
+
+lm_ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
+         lm_ggml_cgraph * gf,
+         lm_ggml_tensor * state_copy,
+         lm_ggml_tensor * state_mask,
+  const llama_ubatch & ubatch,
+                 int   il) const {
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+
+    const auto token_shift_count = hparams.token_shift_count;
+
+    const int64_t n_seqs  = ubatch.n_seqs;
+
+    lm_ggml_tensor * token_shift_all = kv_self->k_l[il];
+
+    lm_ggml_tensor * token_shift = build_copy_mask_state(
+            gf, token_shift_all, state_copy, state_mask,
+            hparams.n_embd_k_s(), n_seqs);
+
+    token_shift = lm_ggml_reshape_3d(ctx0, token_shift, hparams.n_embd, token_shift_count, n_seqs);
+
+    return token_shift;
+}
+
+lm_ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
+         lm_ggml_tensor * token_shift,
+  const llama_ubatch & ubatch,
+                 int   il) const {
+    const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
+
+    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 = kv_self->head;
+
+    return lm_ggml_cpy(
+        ctx0,
+        lm_ggml_view_1d(ctx0, token_shift, n_embd * n_seqs * token_shift_count, 0),
+        lm_ggml_view_1d(ctx0, kv_self->k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * lm_ggml_element_size(kv_self->k_l[il]))
+    );
+}
+
+void llm_graph_context::build_pooling(
+        lm_ggml_cgraph * gf,
+        lm_ggml_tensor * cls,
+        lm_ggml_tensor * cls_b,
+        lm_ggml_tensor * cls_out,
+        lm_ggml_tensor * cls_out_b) const {
+    if (!cparams.embeddings) {
+        return;
+    }
+
+    lm_ggml_tensor * inp = res->t_embd;
+
+    //// find result_norm tensor for input
+    //for (int i = lm_ggml_graph_n_nodes(gf) - 1; i >= 0; --i) {
+    //    inp = lm_ggml_graph_node(gf, i);
+    //    if (strcmp(inp->name, "result_norm") == 0 || strcmp(inp->name, "result_embd") == 0) {
+    //        break;
+    //    }
+
+    //    inp = nullptr;
+    //}
+
+    LM_GGML_ASSERT(inp != nullptr && "missing result_norm/result_embd tensor");
+
+    lm_ggml_tensor * cur;
+
+    switch (pooling_type) {
+        case LLAMA_POOLING_TYPE_NONE:
+            {
+                cur = inp;
+            } break;
+        case LLAMA_POOLING_TYPE_MEAN:
+            {
+                lm_ggml_tensor * inp_mean = build_inp_mean();
+                cur = lm_ggml_mul_mat(ctx0, lm_ggml_cont(ctx0, lm_ggml_transpose(ctx0, inp)), inp_mean);
+            } break;
+        case LLAMA_POOLING_TYPE_CLS:
+        case LLAMA_POOLING_TYPE_LAST:
+            {
+                lm_ggml_tensor * inp_cls = build_inp_cls();
+                cur = lm_ggml_get_rows(ctx0, inp, inp_cls);
+            } break;
+        case LLAMA_POOLING_TYPE_RANK:
+            {
+                lm_ggml_tensor * inp_cls = build_inp_cls();
+                inp = lm_ggml_get_rows(ctx0, inp, inp_cls);
+
+                // classification head
+                // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
+                LM_GGML_ASSERT(cls   != nullptr);
+                LM_GGML_ASSERT(cls_b != nullptr);
+
+                cur = lm_ggml_add (ctx0, lm_ggml_mul_mat(ctx0, cls, inp), cls_b);
+                cur = lm_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) {
+                    LM_GGML_ASSERT(cls_out_b != nullptr);
+
+                    cur = lm_ggml_add (ctx0, lm_ggml_mul_mat(ctx0, cls_out, cur), cls_out_b);
+                }
+            } break;
+        default:
+            {
+                LM_GGML_ABORT("unknown pooling type");
+            }
+    }
+
+    cb(cur, "result_embd_pooled", -1);
+    res->t_embd_pooled = cur;
+
+    lm_ggml_build_forward_expand(gf, cur);
+}
+