whispercpp 1.3.3 → 1.3.4

data/ext/sources/examples/talk-llama/llama-graph.cpp

@@ -4,8 +4,8 @@
 #include "llama-batch.h"
 #include "llama-cparams.h"
 
-#include "llama-kv-cache-unified.h"
-#include "llama-kv-cache-unified-iswa.h"
+#include "llama-kv-cache.h"
+#include "llama-kv-cache-iswa.h"
 #include "llama-memory-hybrid.h"
 #include "llama-memory-recurrent.h"
 
@@ -28,6 +28,15 @@ void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+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;
@@ -50,6 +59,14 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+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;
@@ -71,7 +88,7 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {
         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
+        GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing
 
         int32_t * data = (int32_t *) pos_bucket->data;
 
@@ -118,6 +135,14 @@ void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+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;
@@ -163,38 +188,26 @@ void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
 
 void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
     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;
 
     if (cparams.embeddings && (
-            cparams.pooling_type == LLAMA_POOLING_TYPE_CLS ||
-            cparams.pooling_type == LLAMA_POOLING_TYPE_RANK
-        )) {
+        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));
 
-        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];
-
-                data[seq_idx] = i;
-            }
-        }
-    }
+        std::vector<int> target_pos(n_seqs_unq, -1);
+        std::vector<int> target_row(n_seqs_unq, -1);
 
-    if (cparams.embeddings && 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> last_pos(n_seqs_unq, -1);
-        std::vector<int> last_row(n_seqs_unq, -1);
+        const bool last = (
+             cparams.pooling_type == LLAMA_POOLING_TYPE_LAST ||
+            (cparams.pooling_type == LLAMA_POOLING_TYPE_RANK && arch == LLM_ARCH_QWEN3) // qwen3 reranking & embedding models use last token
+        );
 
         for (int i = 0; i < n_tokens; ++i) {
             const llama_pos pos = ubatch->pos[i];
@@ -203,16 +216,20 @@ void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) {
                 const llama_seq_id seq_id  = ubatch->seq_id[i][s];
                 const int32_t      seq_idx = ubatch->seq_idx[seq_id];
 
-                if (pos >= last_pos[seq_idx]) {
-                    last_pos[seq_idx] = pos;
-                    last_row[seq_idx] = i;
+                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 (last_row[s] >= 0) {
-                data[s] = last_row[s];
+            if (target_row[s] >= 0) {
+                data[s] = target_row[s];
             }
         }
     }
@@ -244,6 +261,36 @@ void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) {
     }
 }
 
+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;
@@ -253,6 +300,9 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
 
     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];
@@ -263,37 +313,90 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
                 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 (s0 == s1 && (!cparams.causal_attn || ubatch->pos[i0] <= ubatch->pos[i1])) {
-                        if (hparams.use_alibi) {
-                            f = -std::abs(ubatch->pos[i0] - ubatch->pos[i1]);
-                        } else {
-                            f = 0.0f;
-                        }
-                        break;
+                    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_unified::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        mctx->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
-    }
+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);
 }
 
-void llm_graph_input_attn_kv_unified_iswa::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        mctx->get_base()->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);
 
-    if (self_kq_mask_swa) {
-        mctx->get_swa()->set_input_kq_mask(self_kq_mask_swa, ubatch, cparams.causal_attn);
-    }
+    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) {
@@ -303,7 +406,7 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
     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
+    GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing
 
     float * data = (float *) cross_kq_mask->data;
 
@@ -333,27 +436,93 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
 }
 
 void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
-    if (self_kq_mask) {
-        mctx->get_attn()->set_input_kq_mask(self_kq_mask, ubatch, cparams.causal_attn);
+    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);
     }
+}
 
-    const int64_t n_rs = mctx->get_recr()->get_n_rs();
+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__);
+        }
 
-    if (s_copy) {
-        GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer));
-        int32_t * data = (int32_t *) s_copy->data;
+        return false;
+    }
 
-        // 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->get_recr()->s_copy(i);
+    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_input_one::set_input(const llama_ubatch *) {
-    GGML_ASSERT(one && ggml_nelements(one) == 1);
-    float f_one = 1.0f;
-    ggml_backend_tensor_set(one, &f_one, 0, sizeof(float));
+void llm_graph_result::set_params(const llm_graph_params & params) {
+    this->params = params;
 }
 
 //
@@ -390,7 +559,6 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     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),
@@ -398,7 +566,10 @@ llm_graph_context::llm_graph_context(const llm_graph_params & params) :
     mctx             (params.mctx),
     cross            (params.cross),
     cb_func          (params.cb),
-    res              (std::make_unique<llm_graph_result>()) {
+    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 {
@@ -613,6 +784,8 @@ ggml_tensor * llm_graph_context::build_ffn(
                 cur = ggml_reglu(ctx0, cur);
                 cb(cur, "ffn_reglu", il);
             } break;
+        default:
+            GGML_ABORT("fatal error");
     }
 
     if (gate && type_gate == LLM_FFN_PAR) {
@@ -622,8 +795,8 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     if (down) {
         cur = build_lora_mm(down, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
+        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);
         }
     }
@@ -658,13 +831,64 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 bool   scale_w,
                float   w_scale,
          llama_expert_gating_func_type gating_op,
-                 int   il) const {
+                 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 = build_lora_mm(gate_inp, cur); // [n_expert, n_tokens]
-    cb(logits, "ffn_moe_logits", il);
+    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) {
@@ -676,6 +900,10 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             {
                 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");
     }
@@ -695,15 +923,36 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         selection_probs = logits;
     }
 
+    if (arch == LLM_ARCH_GROVEMOE) {
+        selection_probs = ggml_sigmoid(ctx0, logits); // [n_expert, n_tokens]
+        cb(selection_probs, "ffn_moe_probs_biased", il);
+    }
+
     // 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]
+    if (arch == LLM_ARCH_GROVEMOE && n_expert != hparams.n_expert) {
+        // TODO: Use scalar div instead when/if implemented
+        ggml_tensor * f_sel = ggml_cast(ctx0, selected_experts, GGML_TYPE_F32);
+        selected_experts = ggml_cast(ctx0, ggml_scale(ctx0, f_sel, 1.0f / float(hparams.n_group_experts)), GGML_TYPE_I32);
+        probs = ggml_reshape_3d(ctx0, probs, 1, hparams.n_expert, n_tokens);
+    } else {
+        probs = ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens);
+    }
+
+    ggml_tensor * weights = ggml_get_rows(ctx0, probs, 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);
 
@@ -720,6 +969,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cb(weights, "ffn_moe_weights_scaled", il);
     }
 
+    //call early so that topk-moe can be used
+    ggml_build_forward_expand(gf, weights);
+
     cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens);
 
     if (weight_before_ffn) {
@@ -732,6 +984,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     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]
@@ -740,6 +997,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         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) {
@@ -757,6 +1019,22 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                 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");
     }
@@ -764,25 +1042,38 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
     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
-    ggml_tensor * moe_out = nullptr;
-    for (int i = 0; i < n_expert_used; ++i) {
-        ggml_tensor * cur_expert = ggml_view_2d(ctx0, experts, n_embd, n_tokens,
-                experts->nb[2], i*experts->nb[1]);
+    // 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];
 
-        if (i == 0) {
-            moe_out = cur_expert;
-        } else {
-            moe_out = ggml_add(ctx0, moe_out, cur_expert);
-        }
+    for (uint32_t i = 1; i < hparams.n_expert_used; ++i) {
+        moe_out = ggml_add(ctx0, moe_out, cur_experts[i]);
     }
 
-    if (n_expert_used == 1) {
+    if (hparams.n_expert_used == 1) {
         // avoid returning a non-contiguous tensor
         moe_out = ggml_cont(ctx0, moe_out);
     }
@@ -906,7 +1197,7 @@ ggml_tensor * llm_graph_context::build_inp_mean() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_cls() const {
-    auto inp = std::make_unique<llm_graph_input_cls>(cparams);
+    auto inp = std::make_unique<llm_graph_input_cls>(cparams, arch);
 
     auto & cur = inp->cls;
 
@@ -956,7 +1247,7 @@ ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const {
 }
 
 ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+    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);
 
@@ -987,51 +1278,28 @@ ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_t
     return pos_bias;
 }
 
-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 = std::make_unique<llm_graph_input_mem_hybrid>(hparams, cparams, mctx_cur);
-
-    {
-        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Hybrid recurrent is not supported with SWA attention layers");
-
-        const auto n_kv = inp->mctx->get_attn()->get_n_kv();
-
-        inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
-        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;
-    }
-
-    {
-        const auto n_rs = mctx_cur->get_recr()->get_n_rs();
-
-        inp->s_copy = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_rs);
-        ggml_set_input(inp->s_copy);
-    }
-
-    return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
-}
-
 ggml_tensor * llm_graph_context::build_attn_mha(
-         ggml_cgraph * gf,
          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) const {
+               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_view_4d(ctx0, q, q->ne[0], q->ne[1], q->ne[2]/n_stream, n_stream, q->nb[1], q->nb[2], q->nb[3]/n_stream, 0);
+
     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_tokens = q->ne[1];
-    const auto n_head   = q->ne[2];
-    const auto n_kv     = k->ne[1];
+    const auto n_kv = k->ne[1];
 
     ggml_tensor * cur;
 
@@ -1054,8 +1322,10 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         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_set_prec(cur, GGML_PREC_F32);
+        ggml_flash_attn_ext_add_sinks(cur, sinks);
+        ggml_flash_attn_ext_set_prec (cur, GGML_PREC_F32);
 
         if (v_mla) {
 #if 0
@@ -1068,14 +1338,16 @@ ggml_tensor * llm_graph_context::build_attn_mha(
             // 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]*n_head, n_tokens);
+        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
@@ -1083,42 +1355,54 @@ ggml_tensor * llm_graph_context::build_attn_mha(
 
         if (arch == LLM_ARCH_GROK) {
             // need to do the following:
-            // multiply by attn_output_multiplyer of 0.08838834764831845
+            // multiply by attn_output_multiplier
             // and then :
             // kq = 30 * tanh(kq / 30)
             // before the softmax below
 
-            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, 0.08838834764831845f/30.0f));
-            kq = ggml_scale(ctx0, kq, 30);
+            kq = ggml_tanh(ctx0, ggml_scale(ctx0, kq, hparams.f_attn_out_scale / hparams.f_attn_logit_softcapping));
+            cb(kq, "kq_tanh", il);
+            kq = ggml_scale(ctx0, kq, hparams.f_attn_logit_softcapping);
+            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);
 
-        cur = ggml_cont_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
+        // 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
@@ -1135,8 +1419,7 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
     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_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-    //cb(inp_kq_mask, "KQ_mask", -1);
+    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;
@@ -1146,13 +1429,13 @@ llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() con
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_no_cache * inp,
-        ggml_cgraph * gf,
         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 {
@@ -1166,11 +1449,16 @@ ggml_tensor * llm_graph_context::build_attn(
 
     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
+    //       but it might not be worth it: https://github.com/ggml-org/llama.cpp/pull/15636
+    //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(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    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) {
@@ -1188,35 +1476,51 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+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_unified>(hparams, cparams, 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_unified_iswa for SWA");
+        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;
 
-        const auto n_kv = mctx_cur->get_n_kv();
+        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_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        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 (llm_graph_input_attn_kv_unified *) res->add_input(std::move(inp));
+    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_unified * inp,
-        ggml_cgraph * gf,
+        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 {
@@ -1226,12 +1530,15 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_build_forward_expand(gf, k_cur);
     ggml_build_forward_expand(gf, v_cur);
 
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_context *>(mctx);
+    const auto * mctx_cur = inp->mctx;
 
     // store to KV cache
     {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il));
+        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();
@@ -1240,13 +1547,13 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    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) {
-            // GLM4 seems to have numerical issues with half-precision accumulators
+        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);
         }
     }
@@ -1259,14 +1566,14 @@ ggml_tensor * llm_graph_context::build_attn(
 }
 
 ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_attn_kv_unified_iswa * inp,
-        ggml_cgraph * gf,
+        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 {
@@ -1282,7 +1589,7 @@ ggml_tensor * llm_graph_context::build_attn(
         ggml_build_forward_expand(gf, v_cur);
     }
 
-    const auto * mctx_iswa = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
+    const auto * mctx_iswa = inp->mctx;
 
     const bool is_swa = hparams.is_swa(il);
 
@@ -1290,11 +1597,15 @@ ggml_tensor * llm_graph_context::build_attn(
 
     // optionally store to KV cache
     if (k_cur) {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
+        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) {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, il));
+        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();
@@ -1303,7 +1614,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = mctx_cur->get_k(ctx0, il);
     ggml_tensor * v = mctx_cur->get_v(ctx0, il);
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    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) {
@@ -1326,7 +1637,7 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
     const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train;
 
-    inp->cross_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
+    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;
@@ -1336,13 +1647,13 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
 
 ggml_tensor * llm_graph_context::build_attn(
         llm_graph_input_attn_cross * inp,
-        ggml_cgraph * gf,
         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 {
@@ -1358,7 +1669,7 @@ ggml_tensor * llm_graph_context::build_attn(
     ggml_tensor * k = k_cur;
     ggml_tensor * v = v_cur;
 
-    ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
+    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) {
@@ -1376,171 +1687,124 @@ ggml_tensor * llm_graph_context::build_attn(
     return cur;
 }
 
-ggml_tensor * llm_graph_context::build_attn(
-        llm_graph_input_mem_hybrid * inp,
-        ggml_cgraph * gf,
-        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 * 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 = static_cast<const llama_memory_hybrid_context *>(mctx)->get_attn();
-
-    // store to KV cache
-    {
-        ggml_build_forward_expand(gf, mctx_cur->cpy_k(ctx0, k_cur, il));
-        ggml_build_forward_expand(gf, mctx_cur->cpy_v(ctx0, v_cur, 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(gf, q, k, v, kq_b, kq_mask, v_mla, kq_scale);
-    cb(cur, "kqv_out", il);
-
-    if (wo) {
-        cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4) {
-            // GLM4 seems 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;
-}
+// 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);
 
-llm_graph_input_attn_kv_unified_iswa * llm_graph_context::build_attn_inp_kv_unified_iswa() const {
-    const auto * mctx_cur = static_cast<const llama_kv_cache_unified_iswa_context *>(mctx);
+    auto inp = std::make_unique<llm_graph_input_attn_kv_iswa>(hparams, cparams, mctx_cur);
 
-    auto inp = std::make_unique<llm_graph_input_attn_kv_unified_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_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask, "KQ_mask", -1);
+        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_unified for non-SWA");
+        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_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
-        //cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
+        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_unified_iswa *) res->add_input(std::move(inp));
+    return (llm_graph_input_attn_kv_iswa *) res->add_input(std::move(inp));
 }
 
 ggml_tensor * llm_graph_context::build_rs(
-        ggml_cgraph * gf,
         ggml_tensor * s,
-        ggml_tensor * state_copy,
+        ggml_tensor * state_copy_main,
+        ggml_tensor * state_copy_extra,
             int32_t   state_size,
             int32_t   n_seqs,
-           uint32_t   n_kv,
-           uint32_t   kv_head,
-           uint32_t   kv_size,
+           uint32_t   n_rs,
+           uint32_t   rs_head,
+           uint32_t   rs_size,
             int32_t   rs_zero,
-               bool   avoid_copies) const {
+        const llm_graph_get_rows_fn & get_state_rows) const {
 
-    ggml_tensor * states = ggml_reshape_2d(ctx0, s, state_size, kv_size);
+    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));
 
-    ggml_tensor * output_states;
-
-    if (!avoid_copies) {
-        // copy states
-        // NOTE: assuming the copy destinations are ALL contained between kv_head and kv_head + n_kv
-        // {state_size, kv_size} -> {state_size, n_seqs}
-        output_states = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_seqs, 0));
-        ggml_build_forward_expand(gf, output_states);
-    } else {
-        // FIXME: make the gathering operation happen before the copy below
-        //        (maybe with an optional lambda function passed as a parameter instead of `avoid_copies`?)
-        output_states = states;
-    }
+    // 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_kv)
-    ggml_tensor * states_extra = ggml_get_rows(ctx0, states, ggml_view_1d(ctx0, state_copy, n_kv - n_seqs, n_seqs*state_copy->nb[0]));
+    // 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_kv - n_seqs), (kv_head + n_seqs)*state_size*ggml_element_size(s))));
+            ggml_view_1d(ctx0, s, state_size*(n_rs - n_seqs), (rs_head + n_seqs)*state_size*ggml_element_size(s))));
 
     return output_states;
 }
 
-llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
-    const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
+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 auto n_rs = mctx_cur->get_n_rs();
+    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);
 
-    return (llm_graph_input_rs *) res->add_input(std::move(inp));
+    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;
 }
 
-ggml_tensor * llm_graph_context::build_rs(
-        llm_graph_input_rs * inp,
-        ggml_cgraph * gf,
-        ggml_tensor * s,
-            int32_t   state_size,
-            int32_t   n_seqs,
-               bool   avoid_copies) const {
+llm_graph_input_rs * llm_graph_context::build_rs_inp() const {
     const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, mctx_cur->get_n_rs(), mctx_cur->get_head(), mctx_cur->get_size(), mctx_cur->get_rs_z(), avoid_copies);
+    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_mem_hybrid * inp,
-        ggml_cgraph * gf,
+        llm_graph_input_rs * inp,
         ggml_tensor * s,
             int32_t   state_size,
             int32_t   n_seqs,
-               bool   avoid_copies) const {
-    const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx)->get_recr();
+        const llm_graph_get_rows_fn & get_state_rows) const {
+    const auto * kv_state = inp->mctx;
 
-    return build_rs(gf, s, inp->s_copy, state_size, n_seqs, mctx_cur->get_n_rs(), mctx_cur->get_head(), mctx_cur->get_size(), mctx_cur->get_rs_z(), avoid_copies);
+    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,
-           ggml_cgraph * gf,
     const llama_ubatch & ubatch,
-                 int   il) const {
+                   int   il) const {
     const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx);
 
     const auto token_shift_count = hparams.token_shift_count;
@@ -1550,7 +1814,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
     ggml_tensor * token_shift_all = mctx_cur->get_r_l(il);
 
     ggml_tensor * token_shift = build_rs(
-            inp, gf, token_shift_all,
+            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);
@@ -1578,8 +1842,18 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
     );
 }
 
+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_cgraph * gf,
         ggml_tensor * cls,
         ggml_tensor * cls_b,
         ggml_tensor * cls_out,
@@ -1623,34 +1897,32 @@ void llm_graph_context::build_pooling(
         case LLAMA_POOLING_TYPE_RANK:
             {
                 ggml_tensor * inp_cls = build_inp_cls();
-                inp = ggml_get_rows(ctx0, inp, inp_cls);
+                cur = ggml_get_rows(ctx0, inp, inp_cls);
 
+                // classification head
+                // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
                 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);
+                    cur = ggml_mul_mat(ctx0, cls, cur);
                     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);
+                // 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
+                // Single layer classification head (direct projection)
+                // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476
+                if (cls_out) {
+                    cur = ggml_mul_mat(ctx0, cls_out, cur);
                     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");
+                }
+
+                // softmax for qwen3 reranker
+                if (arch == LLM_ARCH_QWEN3) {
+                    cur = ggml_soft_max(ctx0, cur);
                 }
             } break;
         default: