whispercpp 1.3.3 → 1.3.5

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

@@ -1,6 +1,7 @@
 #pragma once
 
 #include "llama-arch.h"
+#include "llama-batch.h"
 #include "llama-hparams.h"
 #include "llama-adapter.h"
 
@@ -9,18 +10,18 @@
 #include <memory>
 #include <set>
 #include <functional>
+#include <map>
 
 struct ggml_cgraph;
 struct ggml_context;
 struct ggml_tensor;
 
-struct llama_ubatch;
 struct llama_cparams;
 
 struct llama_memory_context_i;
 
-class llama_kv_cache_unified_context;
-class llama_kv_cache_unified_iswa_context;
+class llama_kv_cache_context;
+class llama_kv_cache_iswa_context;
 class llama_memory_recurrent_context;
 class llama_memory_hybrid_context;
 
@@ -39,6 +40,7 @@ enum llm_ffn_op_type {
     LLM_FFN_SWIGLU,
     LLM_FFN_GEGLU,
     LLM_FFN_REGLU,
+    LLM_FFN_SWIGLU_OAI_MOE,
 };
 
 enum llm_ffn_gate_type {
@@ -69,20 +71,38 @@ struct llama_cross {
     std::vector<std::set<llama_seq_id>> seq_ids_enc;
 };
 
+struct llm_graph_params;
+
 //
 // llm_graph_input
 //
 
 class llm_graph_input_i {
 public:
+    llm_graph_input_i() {
+        const char * LLAMA_GRAPH_INPUT_DEBUG = getenv("LLAMA_GRAPH_INPUT_DEBUG");
+        debug = LLAMA_GRAPH_INPUT_DEBUG ? atoi(LLAMA_GRAPH_INPUT_DEBUG) : 0;
+    }
+
     virtual ~llm_graph_input_i() = default;
 
     virtual void set_input(const llama_ubatch * ubatch) = 0;
+
+    // return true if the resulting input tensors using the provided graph parameters would be
+    //   the same as the previous input tensors that we have currently stored in the object
+    virtual bool can_reuse(const llm_graph_params & params) {
+        // returning false here by default will prevent from reusing the graph if the check
+        //   for the input type has not been implemented yet
+        GGML_UNUSED(params);
+        return false;
+    }
+protected:
+    // env: LLAMA_GRAPH_INPUT_DEBUG
+    int debug = 0;
 };
 
 using llm_graph_input_ptr = std::unique_ptr<llm_graph_input_i>;
 
-
 class llm_graph_input_embd : public llm_graph_input_i {
 public:
     llm_graph_input_embd()          = default;
@@ -90,6 +110,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * tokens = nullptr; // I32 [n_batch]
     ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
 };
@@ -101,6 +123,8 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * pos = nullptr; // I32 [n_batch]
 
     const uint32_t n_pos_per_embd = 1;
@@ -109,8 +133,8 @@ public:
 // temperature tuning, used by llama4
 class llm_graph_input_attn_temp : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
-        : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
+    llm_graph_input_attn_temp(uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale, float f_attn_temp_offset)
+        : n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale), f_attn_temp_offset(f_attn_temp_offset) {}
     virtual ~llm_graph_input_attn_temp() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -119,6 +143,7 @@ public:
 
     const uint32_t n_attn_temp_floor_scale;
     const float    f_attn_temp_scale;
+    const float    f_attn_temp_offset;
 };
 
 class llm_graph_input_pos_bucket : public llm_graph_input_i {
@@ -130,23 +155,23 @@ public:
 
     ggml_tensor * pos_bucket = nullptr; // I32 [n_batch, n_batch]
 
-    const llama_hparams & hparams;
+    const llama_hparams hparams;
 };
 
 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_context * mctx) : hparams(hparams), mctx(mctx) {}
+            const llama_kv_cache_context * mctx) : hparams(hparams), mctx(mctx) {}
     virtual ~llm_graph_input_pos_bucket_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * pos_bucket = nullptr; // I32 [n_kv, n_batch]
 
-    const llama_hparams & hparams;
+    const llama_hparams hparams;
 
-    const llama_kv_cache_unified_context * mctx;
+    const llama_kv_cache_context * mctx;
 };
 
 class llm_graph_input_out_ids : public llm_graph_input_i {
@@ -154,17 +179,19 @@ public:
     llm_graph_input_out_ids(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            int32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}
+            uint32_t n_outputs) : hparams(hparams), cparams(cparams), n_outputs(n_outputs) {}
     virtual ~llm_graph_input_out_ids() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
     ggml_tensor * out_ids; // I32 [n_outputs]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
-    const int32_t n_outputs;
+    const uint32_t n_outputs;
 };
 
 class llm_graph_input_mean : public llm_graph_input_i {
@@ -176,19 +203,20 @@ public:
 
     ggml_tensor * mean; // F32 [n_batch, n_batch]
 
-    const llama_cparams & cparams;
+    const llama_cparams cparams;
 };
 
 class llm_graph_input_cls : public llm_graph_input_i {
 public:
-    llm_graph_input_cls(const llama_cparams & cparams) : cparams(cparams) {}
+    llm_graph_input_cls(const llama_cparams & cparams, const llm_arch arch) : cparams(cparams), arch(arch) {}
     virtual ~llm_graph_input_cls() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
     ggml_tensor * cls; // I32 [n_batch]
 
-    const llama_cparams & cparams;
+    const llama_cparams cparams;
+    const llm_arch arch;
 };
 
 class llm_graph_input_rs : public llm_graph_input_i {
@@ -198,9 +226,20 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * s_copy; // I32 [kv_size]
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * s_copy;  // I32 [n_rs]
+
+    // views of s_copy, computed once per graph
+    // and shared across layers which use build_rs
+    ggml_tensor * s_copy_main;   // I32 [n_seqs]
+    ggml_tensor * s_copy_extra;  // I32 [n_rs - n_seqs]
 
     const llama_memory_recurrent_context * mctx;
+
+    // used in view offsets, need to match for valid graph reuse
+    uint32_t head;
+    int32_t rs_z;
 };
 
 class llm_graph_input_cross_embd : public llm_graph_input_i {
@@ -226,66 +265,93 @@ public:
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * get_kq_mask() const { return kq_mask_cnv; }
+    ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
+    ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
-    ggml_tensor * kq_mask     = nullptr; // F32 [n_tokens, n_batch]
-    ggml_tensor * kq_mask_cnv = nullptr; //     [n_tokens, n_batch]
+    // n_tokens == n_batch
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_tokens, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_tokens, n_batch/n_stream, 1, n_stream]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 };
 
-class llm_graph_input_attn_kv_unified : public llm_graph_input_i {
+class llm_graph_input_attn_kv : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_kv_unified(
+    llm_graph_input_attn_kv(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_context * mctx) :
+            const llama_kv_cache_context * mctx) :
         hparams(hparams),
         cparams(cparams),
         mctx(mctx) {
     }
-    ~llm_graph_input_attn_kv_unified() = default;
+    ~llm_graph_input_attn_kv() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * get_k_idxs() const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs() const { return self_v_idxs; }
+
     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]
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+
+    // note: these have to be copies because in order to be able to reuse a graph, its inputs
+    //       need to carry these parameters with them. otherwise, they can point to freed
+    //       llm_graph_params from a previous batch, causing stack-use-after-return
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
-    const llama_kv_cache_unified_context * mctx;
+    const llama_kv_cache_context * mctx;
 };
 
-class llm_graph_input_attn_kv_unified_iswa : public llm_graph_input_i {
+class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
-    llm_graph_input_attn_kv_unified_iswa(
+    llm_graph_input_attn_kv_iswa(
             const llama_hparams & hparams,
             const llama_cparams & cparams,
-            const llama_kv_cache_unified_iswa_context * mctx) :
+            const llama_kv_cache_iswa_context * mctx) :
         hparams(hparams),
         cparams(cparams),
         mctx(mctx) {
     }
-    ~llm_graph_input_attn_kv_unified_iswa() = default;
+    ~llm_graph_input_attn_kv_iswa() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * get_k_idxs()     const { return self_k_idxs; }
+    ggml_tensor * get_v_idxs()     const { return self_v_idxs; }
+    ggml_tensor * get_k_idxs_swa() const { return self_k_idxs_swa; }
+    ggml_tensor * get_v_idxs_swa() const { return self_v_idxs_swa; }
+
     ggml_tensor * get_kq_mask()     const { return self_kq_mask_cnv; }
     ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
-    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch]
+    ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs     = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
+    ggml_tensor * self_k_idxs_swa = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+
+    const llama_hparams hparams;
+    const llama_cparams cparams;
 
-    const llama_kv_cache_unified_iswa_context * mctx;
+    const llama_kv_cache_iswa_context * mctx;
 };
 
 class llm_graph_input_attn_cross : public llm_graph_input_i {
@@ -297,8 +363,8 @@ public:
 
     ggml_tensor * get_kq_mask_cross() const { return cross_kq_mask_cnv; }
 
-    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch]
-    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch]
+    ggml_tensor * cross_kq_mask     = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
+    ggml_tensor * cross_kq_mask_cnv = nullptr; // F32 [n_outputs_enc, n_batch, 1, 1]
 
     const llama_cross * cross = nullptr;
 };
@@ -306,39 +372,41 @@ public:
 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),
+            std::unique_ptr<llm_graph_input_attn_kv> inp_attn,
+            std::unique_ptr<llm_graph_input_rs>      inp_rs,
+            const llama_memory_hybrid_context *      mctx) :
+        inp_attn(std::move(inp_attn)),
+        inp_rs(std::move(inp_rs)),
         cparams(cparams),
-        mctx(mctx) {
-    }
+        mctx(mctx) { }
     virtual ~llm_graph_input_mem_hybrid() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
 
-    ggml_tensor * s_copy; // I32 [kv_size]
+    bool can_reuse(const llm_graph_params & params) override;
 
-    ggml_tensor * get_kq_mask() const { return self_kq_mask_cnv; }
+    std::unique_ptr<llm_graph_input_attn_kv> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>      inp_rs;
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch]
+    llm_graph_input_attn_kv * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs      * get_recr() const { return inp_rs.get(); }
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    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 {
+class llm_graph_input_sampling : public llm_graph_input_i {
 public:
-    llm_graph_input_one() {}
-    virtual ~llm_graph_input_one() = default;
+    llm_graph_input_sampling(std::map<llama_seq_id, llama_sampler *> samplers) :
+        samplers(std::move(samplers)) { }
+    virtual ~llm_graph_input_sampling() = default;
 
-    void set_input(const llama_ubatch *) override;
+    void set_input(const llama_ubatch * ubatch) override;
+    bool can_reuse(const llm_graph_params & params) override;
 
-    ggml_tensor * one = nullptr; // F32
+    std::map<llama_seq_id, llama_sampler *> samplers;
 };
 
 //
@@ -351,40 +419,149 @@ public:
 // along with the input tensors, the object also provides commonly used outputs tensors, such as logits, embeddings, etc.
 //   these are used by the llama_context to extact the relevant data, based on the compute parameters
 
-class llm_graph_result_i {
-public:
-    virtual ~llm_graph_result_i() = default;
+// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
+using llm_graph_cb = std::function<void(const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il)>;
 
-    virtual ggml_tensor * get_tokens()      = 0;
-    virtual ggml_tensor * get_logits()      = 0;
-    virtual ggml_tensor * get_embd()        = 0;
-    virtual ggml_tensor * get_embd_pooled() = 0;
+class llm_graph_result;
 
-    virtual void set_inputs(const llama_ubatch * ubatch) = 0;
-};
+struct llm_graph_params {
+    llm_arch arch = LLM_ARCH_UNKNOWN;
 
-using llm_graph_result_ptr = std::unique_ptr<llm_graph_result_i>;
+    llama_hparams hparams;
+    llama_cparams cparams;
 
+    llama_ubatch ubatch; // note: intentionally make a copy
 
-class llm_graph_result : public llm_graph_result_i {
-public:
-    virtual ~llm_graph_result() = default;
+    llm_graph_type gtype;
+
+    ggml_backend_sched_t sched;
+    ggml_backend_t backend_cpu;
+
+    const llama_adapter_cvec     * cvec;
+    const llama_adapter_loras    * loras;
+    const llama_memory_context_i * mctx;
+    const llama_cross            * cross;
 
-    ggml_tensor * get_tokens()      override { return t_tokens; }
-    ggml_tensor * get_logits()      override { return t_logits; }
-    ggml_tensor * get_embd()        override { return t_embd; }
-    ggml_tensor * get_embd_pooled() override { return t_embd_pooled; }
+    std::map<llama_seq_id, llama_sampler *> samplers;
 
-    void set_inputs(const llama_ubatch * ubatch) override {
-        for (auto & input : inputs) {
-            input->set_input(ubatch);
+    static bool samplers_equal(
+          const std::map<llama_seq_id, llama_sampler *> & lhs,
+          const std::map<llama_seq_id, llama_sampler *> & rhs) {
+        if (lhs.size() != rhs.size()) {
+            return false;
+        }
+        for (const auto & [seq_id, sampler] : lhs) {
+            auto it = rhs.find(seq_id);
+            if (it == rhs.end() || it->second != sampler) {
+                return false;
+            }
         }
+        return true;
     }
 
-    llm_graph_input_i * add_input(llm_graph_input_ptr input) {
-        inputs.emplace_back(std::move(input));
-        return inputs.back().get();
+    uint32_t n_outputs;
+
+    llm_graph_cb cb;
+
+    llm_graph_result * res;
+
+    // return true if the "other" params would result in a graph with the same topology as with the current params
+    //   having the same topology allows us to reuse the graph in some cases
+    bool allow_reuse(const llm_graph_params & other) const {
+        // first check the ubatch
+        bool can_reuse_ubatch =
+            ubatch.equal_seqs() == other.ubatch.equal_seqs() &&
+            ubatch.n_tokens     == other.ubatch.n_tokens &&
+            ubatch.n_seq_tokens == other.ubatch.n_seq_tokens &&
+            ubatch.n_seqs       == other.ubatch.n_seqs &&
+            ubatch.n_seqs_unq   == other.ubatch.n_seqs_unq &&
+            (
+                (!ubatch.token && !other.ubatch.token) ||
+                (!ubatch.embd  && !other.ubatch.embd)
+            );
+
+        // when we split the batch using "equal_seqs" we have to verify that the participating sequences are the same
+        //   the reason is because the set of attention streams would be different for different sequences
+        if (can_reuse_ubatch && ubatch.equal_seqs()) {
+            if (!ubatch.data) {
+                // if the old ubatch does not own it's data, then we cannot guarantee that it is still alive, and
+                //   therefore we cannot perform the sequence id check. normally should never happen
+                can_reuse_ubatch = false;
+            } else {
+                for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
+                    can_reuse_ubatch &= ubatch.seq_id_unq[s] == other.ubatch.seq_id_unq[s];
+                }
+            }
+        }
+
+        if (!can_reuse_ubatch) {
+            return false;
+        }
+
+        if (n_outputs != other.n_outputs) {
+            return false;
+        }
+
+        if (!samplers_equal(samplers, other.samplers)) {
+            return false;
+        }
+
+        if (samplers.size() > 0) {
+            if (!ubatch.data || !other.ubatch.data) {
+                return false;
+            }
+
+            // check that the outputs are the same for all samplers
+            for (uint32_t i = 0; i < ubatch.n_tokens; ++i) {
+                if (ubatch.output[i]    != other.ubatch.output[i] ||
+                    ubatch.seq_id[i][0] != other.ubatch.seq_id[i][0]) {
+                    return false;
+                }
+            }
+        }
+
+        return
+            cparams.embeddings  == other.cparams.embeddings  &&
+            cparams.causal_attn == other.cparams.causal_attn &&
+            arch  == other.arch  &&
+            gtype == other.gtype &&
+            cvec  == other.cvec  &&
+            loras == other.loras &&
+            cross == other.cross;
     }
+};
+
+class llm_graph_result {
+public:
+    llm_graph_result(int64_t max_nodes);
+
+    virtual ~llm_graph_result() = default;
+
+    ggml_tensor * get_tokens()      const { return t_tokens; }
+    ggml_tensor * get_logits()      const { return t_logits; }
+    ggml_tensor * get_embd()        const { return t_embd; }
+    ggml_tensor * get_embd_pooled() const { return t_embd_pooled; }
+
+    ggml_cgraph  * get_gf()  const { return gf; }
+    ggml_context * get_ctx() const { return ctx_compute.get(); }
+
+    int64_t get_max_nodes() const;
+
+    void reset();
+
+    void set_inputs(const llama_ubatch * ubatch);
+    void set_outputs();
+
+    // try to update the existing graph result using the new graph parameters in order to reuse it
+    // this can only be done if we determine that the resulting graph using the new graph parameters
+    //   would be identical to the existing graph. in that case, we simply have to update the memory
+    //   contexts of the input tensors of the graph and we can reuse it for another computation
+    // return true if the graph was updated and can be reused
+    bool can_reuse(const llm_graph_params & params);
+
+    llm_graph_input_i * add_input(llm_graph_input_ptr input);
+
+    void set_params(const llm_graph_params & params);
 
     // important graph nodes
     ggml_tensor * t_tokens      = nullptr;
@@ -392,37 +569,40 @@ public:
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;
 
+    std::map<llama_seq_id, ggml_tensor*> t_sampled_logits;
+    std::map<llama_seq_id, ggml_tensor*> t_candidates;
+    std::map<llama_seq_id, ggml_tensor*> t_sampled;
+    std::map<llama_seq_id, ggml_tensor*> t_sampled_probs;
+
     std::vector<llm_graph_input_ptr> inputs;
-};
 
-//
-// llm_graph_context
-//
+    ggml_context_ptr ctx_compute;
 
-// callback that allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.)
-using llm_graph_cb = std::function<void(const llama_ubatch & ubatch, ggml_tensor * cur, const char * name, int il)>;
+    // memory buffers used to evaluate the model
+    std::vector<uint8_t> buf_compute_meta;
 
-struct llm_graph_params {
-    ggml_context * ctx;
+    ggml_cgraph * gf;
 
-    const llm_arch arch;
+    int64_t max_nodes;
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
-    const llama_ubatch  & ubatch;
+private:
+    // keep a copy of the previous graph parameters
+    // we will use this to determine whether the graph can be reused by comparing them with the new parameters
+    // note: these are updated after constructing the new graph
+    llm_graph_params params;
 
-    ggml_backend_sched_t sched;
-    ggml_backend_t backend_cpu;
+    // env: LLAMA_GRAPH_RESULT_DEBUG
+    int debug = 0;
+};
 
-    const llama_adapter_cvec     * cvec;
-    const llama_adapter_loras    * loras;
-    const llama_memory_context_i * mctx;
-    const llama_cross            * cross;
+using llm_graph_result_ptr = std::unique_ptr<llm_graph_result>;
 
-    uint32_t n_outputs;
+//
+// llm_graph_context
+//
 
-    const llm_graph_cb & cb;
-};
+// used in build_rs to properly order writes and avoid unnecessary copies
+using llm_graph_get_rows_fn = std::function<ggml_tensor * (ggml_context *, ggml_tensor * states, ggml_tensor * ids)>;
 
 struct llm_graph_context {
     const llm_arch arch;
@@ -460,8 +640,6 @@ struct llm_graph_context {
     const enum llama_pooling_type pooling_type;
     const enum llama_rope_type    rope_type;
 
-    ggml_context * ctx0 = nullptr;
-
     ggml_backend_sched_t sched;
 
     ggml_backend_t backend_cpu; // TODO: needed by build_attn_mha, figure out a way to remove?
@@ -471,9 +649,14 @@ struct llm_graph_context {
     const llama_memory_context_i * mctx;
     const llama_cross            * cross;
 
+    std::map<llama_seq_id, llama_sampler *> samplers;
+
     const llm_graph_cb & cb_func;
 
-    std::unique_ptr<llm_graph_result> res;
+    llm_graph_result * res;
+
+    ggml_context * ctx0 = nullptr;
+    ggml_cgraph  * gf   = nullptr;
 
     llm_graph_context(const llm_graph_params & params);
     virtual ~llm_graph_context() = default;
@@ -522,6 +705,7 @@ struct llm_graph_context {
        llm_ffn_gate_type   type_gate,
                      int   il) const;
 
+    // build MoE FFN without bias tensors
     ggml_tensor * build_moe_ffn(
              ggml_tensor * cur,
              ggml_tensor * gate_inp,
@@ -536,7 +720,29 @@ struct llm_graph_context {
                     bool   scale_w,
                    float   w_scale,
             llama_expert_gating_func_type gating_op,
-                     int   il) const;
+                     int   il,
+             ggml_tensor * probs_in = nullptr) const;
+
+    ggml_tensor * 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 = nullptr) const;
 
     //
     // inputs
@@ -554,64 +760,63 @@ struct llm_graph_context {
     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
     //
 
     ggml_tensor * build_attn_mha(
-             ggml_cgraph * gf,
-             ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
-             ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
-             ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
-             ggml_tensor * kq_b,
-             ggml_tensor * kq_mask,
-             ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
-                   float   kq_scale) const;
+            ggml_tensor * q,       // [n_embd_head_q, n_head_q, n_tokens]
+            ggml_tensor * k,       // [n_embd_head_k, n_head_k, n_tokens]
+            ggml_tensor * v,       // [n_embd_head_v, n_head_v, n_tokens] (v_trans == false)
+            ggml_tensor * kq_b,
+            ggml_tensor * kq_mask,
+            ggml_tensor * sinks,   // [n_head_q]
+            ggml_tensor * v_mla,   // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
+                  float   kq_scale,
+                    int   il) const;
 
     llm_graph_input_attn_no_cache * build_attn_inp_no_cache() const;
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_no_cache * 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 * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
-    llm_graph_input_attn_kv_unified * build_attn_inp_kv_unified() const;
+    llm_graph_input_attn_kv * build_attn_inp_kv() const;
 
     ggml_tensor * 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, // [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 * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
 
-    llm_graph_input_attn_kv_unified_iswa * build_attn_inp_kv_unified_iswa() const;
+    llm_graph_input_attn_kv_iswa * build_attn_inp_kv_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,
+            llm_graph_input_attn_kv_iswa * inp,
             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] optional
             ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] optional
             ggml_tensor * kq_b,
+            ggml_tensor * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   float   kq_scale,
                     int   il) const;
@@ -620,90 +825,85 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_cross * 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 * sinks, // [n_head_q]
             ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v]
                   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
     //
 
-    // 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
+    //       this is analogous to llama_kv_cache::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,
+            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 = false) const;
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) 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;
+            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) const;
 
     ggml_tensor * build_rwkv_token_shift_load(
         llm_graph_input_rs * inp,
-               ggml_cgraph * gf,
         const llama_ubatch & ubatch,
-                     int   il) const;
+                       int   il) const;
 
     ggml_tensor * build_rwkv_token_shift_store(
              ggml_tensor * token_shift,
       const llama_ubatch & ubatch,
                      int   il) const;
+    //
+    // hybrid
+    //
+
+    llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
 
     //
     // pooling
     //
 
     void build_pooling(
-            ggml_cgraph * gf,
             ggml_tensor * cls,
             ggml_tensor * cls_b,
             ggml_tensor * cls_out,
             ggml_tensor * cls_out_b) const;
+
+    //
+    // sampling (backend sampling)
+    //
+
+    void build_sampling() const;
+
+    //
+    // dense (out)
+    //
+
+    void build_dense_out(
+            ggml_tensor * dense_2,
+            ggml_tensor * dense_3) const;
 };
 
 // TODO: better name