whispercpp 1.3.2 → 1.3.3

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

@@ -17,10 +17,12 @@ struct ggml_tensor;
 struct llama_ubatch;
 struct llama_cparams;
 
-class llama_memory_i;
-class llama_kv_cache_unified;
-class llama_kv_cache_unified_iswa;
-class llama_kv_cache_recurrent;
+struct llama_memory_context_i;
+
+class llama_kv_cache_unified_context;
+class llama_kv_cache_unified_iswa_context;
+class llama_memory_recurrent_context;
+class llama_memory_hybrid_context;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -35,6 +37,8 @@ enum llm_ffn_op_type {
     LLM_FFN_RELU,
     LLM_FFN_RELU_SQR,
     LLM_FFN_SWIGLU,
+    LLM_FFN_GEGLU,
+    LLM_FFN_REGLU,
 };
 
 enum llm_ffn_gate_type {
@@ -92,14 +96,14 @@ public:
 
 class llm_graph_input_pos : public llm_graph_input_i {
 public:
-    llm_graph_input_pos(int64_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
+    llm_graph_input_pos(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
     virtual ~llm_graph_input_pos() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * pos = nullptr; // I32 [n_batch]
 
-    const int64_t n_pos_per_embd = 1;
+    const uint32_t n_pos_per_embd = 1;
 };
 
 // temperature tuning, used by llama4
@@ -133,7 +137,7 @@ class llm_graph_input_pos_bucket_kv : public llm_graph_input_i {
 public:
     llm_graph_input_pos_bucket_kv(
             const llama_hparams & hparams,
-            const llama_kv_cache_unified * kv_self) : hparams(hparams), kv_self(kv_self) {}
+            const llama_kv_cache_unified_context * mctx) : hparams(hparams), mctx(mctx) {}
     virtual ~llm_graph_input_pos_bucket_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -141,7 +145,8 @@ public:
     ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
 
     const llama_hparams & hparams;
-    const llama_kv_cache_unified * kv_self;
+
+    const llama_kv_cache_unified_context * mctx;
 };
 
 class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -186,28 +191,16 @@ public:
     const llama_cparams & cparams;
 };
 
-class llm_graph_input_s_copy : public llm_graph_input_i {
+class llm_graph_input_rs : public llm_graph_input_i {
 public:
-    llm_graph_input_s_copy(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {}
-    virtual ~llm_graph_input_s_copy() = default;
+    llm_graph_input_rs(const llama_memory_recurrent_context * mctx) : mctx(mctx) {}
+    virtual ~llm_graph_input_rs() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * s_copy; // I32 [kv_size]
 
-    const llama_kv_cache_recurrent * kv_self;
-};
-
-class llm_graph_input_s_mask : public llm_graph_input_i {
-public:
-    llm_graph_input_s_mask(const llama_kv_cache_recurrent * kv_self) : kv_self(kv_self) {}
-    virtual ~llm_graph_input_s_mask() = default;
-
-    void set_input(const llama_ubatch * ubatch) override;
-
-    ggml_tensor * s_mask; // F32 [1, n_kv]
-
-    const llama_kv_cache_recurrent * kv_self;
+    const llama_memory_recurrent_context * mctx;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -247,10 +240,10 @@ public:
     llm_graph_input_attn_kv_unified(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified * kv_self) :
+            const llama_kv_cache_unified_context * mctx) :
         hparams(hparams),
         cparams(cparams),
-        kv_self(kv_self) {
+        mctx(mctx) {
     }
     ~llm_graph_input_attn_kv_unified() = default;
 
@@ -264,7 +257,7 @@ public:
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const llama_kv_cache_unified * kv_self;
+    const llama_kv_cache_unified_context * mctx;
 };
 
 class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
@@ -272,10 +265,10 @@ public:
     llm_graph_input_attn_kv_unified_iswa(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_iswa * kv_self) :
+            const llama_kv_cache_unified_iswa_context * mctx) :
         hparams(hparams),
         cparams(cparams),
-        kv_self(kv_self) {
+        mctx(mctx) {
     }
     ~llm_graph_input_attn_kv_unified_iswa() = default;
 
@@ -292,7 +285,7 @@ public:
     const llama_hparams & hparams;
     const llama_cparams & cparams;
 
-    const llama_kv_cache_unified_iswa * kv_self;
+    const llama_kv_cache_unified_iswa_context * mctx;
 };
 
 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -310,6 +303,44 @@ public:
     const llama_cross * cross = nullptr;
 };
 
+class llm_graph_input_mem_hybrid : public llm_graph_input_i {
+public:
+    llm_graph_input_mem_hybrid(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_memory_hybrid_context * mctx) :
+        hparams(hparams),
+        cparams(cparams),
+        mctx(mctx) {
+    }
+    virtual ~llm_graph_input_mem_hybrid() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    ggml_tensor * s_copy; // I32 [kv_size]
+
+    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
+
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
+
+    const llama_hparams & hparams;
+    const llama_cparams & cparams;
+
+    const llama_memory_hybrid_context * mctx;
+};
+
+// TODO: remove this when ggml_scale_add is implemented
+class llm_graph_input_one : public llm_graph_input_i {
+public:
+    llm_graph_input_one() {}
+    virtual ~llm_graph_input_one() = default;
+
+    void set_input(const llama_ubatch *) override;
+
+    ggml_tensor * one = nullptr; // F32
+};
+
 //
 // llm_graph_result
 //
@@ -383,12 +414,12 @@ struct llm_graph_params {
     ggml_backend_sched_t sched;
     ggml_backend_t backend_cpu;
 
-    const llama_adapter_cvec  * cvec;
-    const llama_adapter_loras * loras;
-    const llama_memory_i      * memory;
-    const llama_cross         * cross;
+    const llama_adapter_cvec     * cvec;
+    const llama_adapter_loras    * loras;
+    const llama_memory_context_i * mctx;
+    const llama_cross            * cross;
 
-    int32_t n_outputs;
+    uint32_t n_outputs;
 
     const llm_graph_cb & cb;
 };
@@ -422,8 +453,8 @@ struct llm_graph_context {
     const float norm_eps;
     const float norm_rms_eps;
 
-    const int32_t n_tokens;
-    const int32_t n_outputs;
+    const int64_t n_tokens;
+    const int64_t n_outputs;
     const int32_t n_ctx_orig; // yarn
 
     const enum llama_pooling_type pooling_type;
@@ -435,18 +466,17 @@ struct llm_graph_context {
 
     ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
 
-    const llama_adapter_cvec  * cvec;
-    const llama_adapter_loras * loras;
-    const llama_memory_i      * memory;
-    const llama_cross         * cross;
+    const llama_adapter_cvec     * cvec;
+    const llama_adapter_loras    * loras;
+    const llama_memory_context_i * mctx;
+    const llama_cross            * cross;
 
     const llm_graph_cb & cb_func;
 
     std::unique_ptr<llm_graph_result> res;
 
     llm_graph_context(const llm_graph_params & params);
-
-    int64_t n_pos_per_embd() const;
+    virtual ~llm_graph_context() = default;
 
     void cb(ggml_tensor * cur, const char * name, int il) const;
 
@@ -518,14 +548,14 @@ struct llm_graph_context {
     ggml_tensor * build_inp_out_ids() const;
     ggml_tensor * build_inp_mean() const;
     ggml_tensor * build_inp_cls() const;
-    ggml_tensor * build_inp_s_copy() const;
-    ggml_tensor * build_inp_s_mask() const;
 
     ggml_tensor * build_inp_cross_embd() const;
     ggml_tensor * build_inp_pos_bucket_enc() const;
     ggml_tensor * build_inp_pos_bucket_dec() const;
     ggml_tensor * build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const;
 
+    llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
+
     //
     // attention
     //
@@ -572,14 +602,15 @@ struct llm_graph_context {
 
     llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
 
+    // note: if k_cur or v_cur are not provided, they will not be stored in the memory
     ggml_tensor * build_attn(
             llm_graph_input_attn_kv_unified_iswa * inp,
             ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
-            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
-            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] optional
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
             ggml_tensor * kq_b,
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
@@ -600,23 +631,62 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
+    ggml_tensor * build_attn(
+            llm_graph_input_mem_hybrid * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens]
+            ggml_tensor * kq_b,
+            ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
     //
     // recurrent
     //
 
-    ggml_tensor * build_copy_mask_state(
-             ggml_cgraph * gf,
-             ggml_tensor * s,
-             ggml_tensor * state_copy,
-             ggml_tensor * state_mask,
-                 int32_t   n_state,
-                 int32_t   n_seqs) const;
+    // TODO: avoid notion of "kv"
+    // TODO: move this implementation to llama_memory_recurrent.
+    //       this is analogous to llama_kv_cache_unified::cpy_k / cpy_v
+    //       when moving, avoid passing `ggml_cgraph` - only pass `ggml_context`. would likely need to split the
+    //         implementation in 2 separate methods. the goal is to avoid calling `ggml_build_forward_expand` in
+    //         `llama_memory_recurrent`
+    ggml_tensor * build_rs(
+            ggml_cgraph * gf,
+            ggml_tensor * s,
+            ggml_tensor * state_copy,
+                int32_t   state_size,
+                int32_t   n_seqs,
+               uint32_t   n_kv,
+               uint32_t   kv_head,
+               uint32_t   kv_size,
+                int32_t   rs_zero,
+                   bool   avoid_copies = false) const;
+
+    llm_graph_input_rs * build_rs_inp() const;
+
+    ggml_tensor * build_rs(
+            llm_graph_input_rs * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * s,
+                int32_t   state_size,
+                int32_t   n_seqs,
+                   bool   avoid_copies = false) const;
+
+    ggml_tensor * build_rs(
+            llm_graph_input_mem_hybrid * inp,
+            ggml_cgraph * gf,
+            ggml_tensor * s,
+                int32_t   state_size,
+                int32_t   n_seqs,
+                   bool   avoid_copies = false) const;
 
     ggml_tensor * build_rwkv_token_shift_load(
-             ggml_cgraph * gf,
-             ggml_tensor * state_copy,
-             ggml_tensor * state_mask,
-      const llama_ubatch & ubatch,
+        llm_graph_input_rs * inp,
+               ggml_cgraph * gf,
+        const llama_ubatch & ubatch,
                      int   il) const;
 
     ggml_tensor * build_rwkv_token_shift_store(

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

@@ -65,7 +65,7 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     return n_embd_head_v * n_head_kv;
 }
 
-uint32_t llama_hparams::n_embd_k_s() const {
+uint32_t llama_hparams::n_embd_r() const {
     if (wkv_head_size != 0) {
         // for RWKV models
         return token_shift_count * n_embd;
@@ -76,7 +76,7 @@ uint32_t llama_hparams::n_embd_k_s() const {
     return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner;
 }
 
-uint32_t llama_hparams::n_embd_v_s() const {
+uint32_t llama_hparams::n_embd_s() const {
     if (wkv_head_size != 0) {
         // corresponds to RWKV's wkv_states size
         return n_embd * wkv_head_size;
@@ -86,6 +86,14 @@ uint32_t llama_hparams::n_embd_v_s() const {
     return ssm_d_state * ssm_d_inner;
 }
 
+bool llama_hparams::is_recurrent(uint32_t il) const {
+    return recurrent_layer_arr[il];
+}
+
+uint32_t llama_hparams::n_pos_per_embd() const {
+    return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1;
+}
+
 bool llama_hparams::is_swa(uint32_t il) const {
     if (il < n_layer) {
         return swa_layers[il];

data/ext/sources/examples/talk-llama/llama-hparams.h

@@ -115,6 +115,9 @@ struct llama_hparams {
     uint32_t ssm_d_state = 0;
     uint32_t ssm_dt_rank = 0;
 
+    // for hybrid state space models
+    std::array<bool, LLAMA_MAX_LAYERS> recurrent_layer_arr;
+
     bool ssm_dt_b_c_rms = false;
 
     float f_clamp_kqv      = 0.0f;
@@ -131,12 +134,21 @@ struct llama_hparams {
     bool attn_soft_cap = false;
     bool use_kq_norm   = true;
 
+    // for Classifiers
+    uint32_t n_cls_out = 1;
+
     // llama4
     uint32_t n_moe_layer_step        = 0;
     uint32_t n_no_rope_layer_step    = 4;
     uint32_t n_attn_temp_floor_scale = 8192;
     float    f_attn_temp_scale       = 0.1;
 
+    // gemma3n altup
+    uint32_t n_altup      = 4; // altup_num_inputs
+    uint32_t i_altup_act  = 0; // altup_active_idx
+    uint32_t laurel_rank  = 64;
+    uint32_t n_embd_altup = 256;
+
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
     // ref: https://github.com/ggerganov/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
@@ -178,10 +190,15 @@ struct llama_hparams {
 
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
-    uint32_t n_embd_k_s() const;
+    uint32_t n_embd_r() const;
 
     // dimension of the recurrent state embeddings
-    uint32_t n_embd_v_s() const;
+    uint32_t n_embd_s() const;
+
+    // whether or not the given layer is recurrent (for hybrid models)
+    bool is_recurrent(uint32_t il) const;
+
+    uint32_t n_pos_per_embd() const;
 
     bool is_swa(uint32_t il) const;
 };

data/ext/sources/examples/talk-llama/llama-kv-cache-unified-iswa.cpp

@@ -0,0 +1,279 @@
+#include "llama-kv-cache-unified-iswa.h"
+
+#include "llama-impl.h"
+#include "llama-batch.h"
+#include "llama-model.h"
+
+#include <algorithm>
+#include <cassert>
+
+//
+// llama_kv_cache_unified_iswa
+//
+
+llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
+        const llama_model & model,
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   offload,
+                     bool   swa_full,
+                 uint32_t   kv_size,
+                 uint32_t   n_seq_max,
+                 uint32_t   n_ubatch,
+                 uint32_t   n_pad) : hparams(model.hparams) {
+    llama_kv_cache_unified::layer_filter_cb filter_base = [&](int32_t il) { return !model.hparams.is_swa(il); };
+    llama_kv_cache_unified::layer_filter_cb filter_swa  = [&](int32_t il) { return  model.hparams.is_swa(il); };
+
+    const uint32_t size_base = kv_size;
+
+    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*n_seq_max + n_ubatch, n_pad));
+
+    // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size
+    if (swa_full) {
+        LLAMA_LOG_WARN("%s: using full-size SWA cache (ref: %s)\n",
+                __func__, "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
+
+        size_swa = size_base;
+    }
+
+    LLAMA_LOG_INFO("%s: creating non-SWA KV cache, size = %u cells\n", __func__, size_base);
+
+    kv_base = std::make_unique<llama_kv_cache_unified>(
+            model, std::move(filter_base), type_k, type_v,
+            v_trans, offload, size_base, n_seq_max, n_pad,
+            0, LLAMA_SWA_TYPE_NONE);
+
+    LLAMA_LOG_INFO("%s: creating     SWA KV cache, size = %u cells\n", __func__, size_swa);
+
+    kv_swa = std::make_unique<llama_kv_cache_unified>(
+            model, std::move(filter_swa), type_k, type_v,
+            v_trans, offload, size_swa, n_seq_max, n_pad,
+            hparams.n_swa, hparams.swa_type);
+}
+
+void llama_kv_cache_unified_iswa::clear(bool data) {
+    kv_base->clear(data);
+    kv_swa ->clear(data);
+}
+
+bool llama_kv_cache_unified_iswa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    bool res = true;
+
+    res = res & kv_base->seq_rm(seq_id, p0, p1);
+    res = res & kv_swa ->seq_rm(seq_id, p0, p1);
+
+    return res;
+}
+
+void llama_kv_cache_unified_iswa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    kv_base->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    kv_swa ->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_kv_cache_unified_iswa::seq_keep(llama_seq_id seq_id) {
+    kv_base->seq_keep(seq_id);
+    kv_swa ->seq_keep(seq_id);
+}
+
+void llama_kv_cache_unified_iswa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    kv_base->seq_add(seq_id, p0, p1, shift);
+    kv_swa ->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_kv_cache_unified_iswa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    kv_base->seq_div(seq_id, p0, p1, d);
+    kv_swa ->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_kv_cache_unified_iswa::seq_pos_min(llama_seq_id seq_id) const {
+    // the base cache is a superset of the SWA cache, so we can just check the SWA cache
+    return kv_swa->seq_pos_min(seq_id);
+}
+
+llama_pos llama_kv_cache_unified_iswa::seq_pos_max(llama_seq_id seq_id) const {
+    return kv_swa->seq_pos_max(seq_id);
+}
+
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_allocr & balloc, uint32_t n_ubatch, bool embd_all) {
+    GGML_UNUSED(embd_all);
+
+    // first try simple split
+    do {
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_simple(n_ubatch);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        auto heads_base = kv_base->prepare(ubatches);
+        if (heads_base.empty()) {
+            break;
+        }
+
+        auto heads_swa = kv_swa->prepare(ubatches);
+        if (heads_swa.empty()) {
+            break;
+        }
+
+        assert(heads_base.size() == heads_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+                this, std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    } while (false);
+
+    // if it fails, try equal split
+    do {
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = balloc.split_equal(n_ubatch);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch)); // NOLINT
+        }
+
+        auto heads_base = kv_base->prepare(ubatches);
+        if (heads_base.empty()) {
+            break;
+        }
+
+        auto heads_swa = kv_swa->prepare(ubatches);
+        if (heads_swa.empty()) {
+            break;
+        }
+
+        assert(heads_base.size() == heads_swa.size());
+
+        return std::make_unique<llama_kv_cache_unified_iswa_context>(
+                this, std::move(heads_base), std::move(heads_swa), std::move(ubatches));
+    } while (false);
+
+    // TODO: if we fail again, we should attempt different splitting strategies
+    //       but to do that properly, we first have to refactor the batches to be more flexible
+
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+}
+
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_full() {
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(this);
+}
+
+llama_memory_context_ptr llama_kv_cache_unified_iswa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_unified_iswa_context>(this, lctx, optimize);
+}
+
+bool llama_kv_cache_unified_iswa::get_can_shift() const {
+    return kv_base->get_size() == kv_swa->get_size();
+}
+
+void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+    kv_base->state_write(io, seq_id);
+    kv_swa ->state_write(io, seq_id);
+}
+
+void llama_kv_cache_unified_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+    kv_base->state_read(io, seq_id);
+    kv_swa ->state_read(io, seq_id);
+}
+
+llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_base() const {
+    return kv_base.get();
+}
+
+llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_swa() const {
+    return kv_swa.get();
+}
+
+//
+// llama_kv_cache_unified_iswa_context
+//
+
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(llama_memory_status status) : status(status) {}
+
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
+        llama_kv_cache_unified_iswa * kv) :
+    ctx_base(kv->get_base()->init_full()),
+    ctx_swa (kv->get_swa ()->init_full()),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
+        llama_kv_cache_unified_iswa * kv,
+        llama_context * lctx,
+        bool optimize) :
+    ctx_base(kv->get_base()->init_update(lctx, optimize)),
+    ctx_swa (kv->get_swa ()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_unified_iswa_context::llama_kv_cache_unified_iswa_context(
+        llama_kv_cache_unified_iswa * kv,
+        std::vector<uint32_t> heads_base,
+        std::vector<uint32_t> heads_swa,
+        std::vector<llama_ubatch> ubatches) :
+    ubatches(std::move(ubatches)),
+    // note: here we copy the ubatches. not sure if this is ideal
+    ctx_base(new llama_kv_cache_unified_context(kv->get_base(), std::move(heads_base), this->ubatches)),
+    ctx_swa (new llama_kv_cache_unified_context(kv->get_swa (), std::move(heads_swa),  this->ubatches)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_swa->get_status())) {
+}
+
+llama_kv_cache_unified_iswa_context:: ~llama_kv_cache_unified_iswa_context() = default;
+
+bool llama_kv_cache_unified_iswa_context::next() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    ctx_base->next();
+    ctx_swa ->next();
+
+    if (++i_next >= ubatches.size()) {
+        return false;
+    }
+
+    return true;
+}
+
+bool llama_kv_cache_unified_iswa_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
+
+    bool res = true;
+
+    res = res & ctx_base->apply();
+    res = res & ctx_swa ->apply();
+
+    return res;
+}
+
+llama_memory_status llama_kv_cache_unified_iswa_context::get_status() const {
+    return status;
+}
+
+const llama_ubatch & llama_kv_cache_unified_iswa_context::get_ubatch() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return ubatches[i_next];
+}
+
+const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_base() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_unified_context *>(ctx_base.get());
+}
+
+const llama_kv_cache_unified_context * llama_kv_cache_unified_iswa_context::get_swa()  const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_unified_context *>(ctx_swa.get());
+}