@fugood/llama.node 1.0.3 → 1.0.4

package/src/llama.cpp/src/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"
 
@@ -14,7 +15,6 @@ struct ggml_cgraph;
 struct ggml_context;
 struct ggml_tensor;
 
-struct llama_ubatch;
 struct llama_cparams;
 
 struct llama_memory_context_i;
@@ -69,6 +69,8 @@ struct llama_cross {
     std::vector<std::set<llama_seq_id>> seq_ids_enc;
 };
 
+struct llm_graph_params;
+
 //
 // llm_graph_input
 //
@@ -78,11 +80,19 @@ public:
     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;
+    }
 };
 
 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 +100,8 @@ 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 +113,8 @@ 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;
@@ -154,17 +168,19 @@ 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 int32_t n_outputs;
+    const uint32_t n_outputs;
 };
 
 class llm_graph_input_mean : public llm_graph_input_i {
@@ -249,16 +265,18 @@ public:
 
     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_k_idxs = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
+    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
 
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+    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]
 
     const llama_hparams & hparams;
     const llama_cparams & cparams;
@@ -280,6 +298,8 @@ public:
 
     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; }
@@ -289,14 +309,14 @@ public:
     ggml_tensor * get_kq_mask_swa() const { return self_kq_mask_swa_cnv; }
 
     ggml_tensor * self_k_idxs     = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_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]
+    ggml_tensor * self_v_idxs_swa = nullptr; // I64 [n_batch] or [n_batch*n_embd_v_gqa]
 
-    ggml_tensor * self_kq_mask         = nullptr; // F32 [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_cnv     = nullptr; //     [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_swa     = nullptr; // F32 [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_swa_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+    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;
@@ -322,47 +342,25 @@ 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),
-        cparams(cparams),
-        mctx(mctx) {
-    }
+            std::unique_ptr<llm_graph_input_attn_kv_unified> 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)),
+        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_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_k_idxs = nullptr; // I64 [n_batch]
-    ggml_tensor * self_v_idxs = nullptr; // I64 [n_batch]
-
-    ggml_tensor * self_kq_mask     = nullptr; // F32 [n_kv, n_batch, 1, 1]
-    ggml_tensor * self_kq_mask_cnv = nullptr; //     [n_kv, n_batch, 1, 1]
+    std::unique_ptr<llm_graph_input_attn_kv_unified> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>              inp_rs;
 
-    const llama_hparams & hparams;
-    const llama_cparams & cparams;
+    llm_graph_input_attn_kv_unified * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs              * get_recr() const { return inp_rs.get(); }
 
     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 * ubatch) override;
-
-    ggml_tensor * one = nullptr; // F32
-};
-
 //
 // llm_graph_result
 //
@@ -373,40 +371,108 @@ 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_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; }
+    ggml_backend_sched_t sched;
+    ggml_backend_t backend_cpu;
 
-    void set_inputs(const llama_ubatch * ubatch) override {
-        for (auto & input : inputs) {
-            input->set_input(ubatch);
+    const llama_adapter_cvec     * cvec;
+    const llama_adapter_loras    * loras;
+    const llama_memory_context_i * mctx;
+    const llama_cross            * cross;
+
+    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)
+            );
+
+        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;
         }
-    }
 
-    llm_graph_input_i * add_input(llm_graph_input_ptr input) {
-        inputs.emplace_back(std::move(input));
-        return inputs.back().get();
+        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 &&
+            n_outputs == other.n_outputs;
     }
+};
+
+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);
+
+    // 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;
@@ -415,36 +481,31 @@ public:
     ggml_tensor * t_embd_pooled = nullptr;
 
     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;
-
-    const llama_adapter_cvec     * cvec;
-    const llama_adapter_loras    * loras;
-    const llama_memory_context_i * mctx;
-    const llama_cross            * cross;
+    // env: LLAMA_GRAPH_RESULT_DEBUG
+    int debug = 0;
+};
 
-    uint32_t n_outputs;
+using llm_graph_result_ptr = std::unique_ptr<llm_graph_result>;
 
-    const llm_graph_cb & cb;
-};
+//
+// llm_graph_context
+//
 
 // 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)>;
@@ -485,8 +546,6 @@ 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?
@@ -498,7 +557,10 @@ struct llm_graph_context {
 
     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;
@@ -579,14 +641,11 @@ 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)
@@ -599,7 +658,6 @@ struct llm_graph_context {
 
     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]
@@ -614,7 +672,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_attn(
             llm_graph_input_attn_kv_unified * inp,
-            ggml_cgraph * gf,
             ggml_tensor * wo,
             ggml_tensor * wo_b,
             ggml_tensor * q_cur, // [n_embd_head_q, n_head_q, n_tokens]
@@ -630,7 +687,6 @@ struct llm_graph_context {
     // 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]
@@ -645,7 +701,6 @@ 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]
@@ -656,18 +711,6 @@ 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
     //
@@ -679,7 +722,6 @@ struct llm_graph_context {
     //         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,
@@ -694,15 +736,6 @@ struct llm_graph_context {
 
     ggml_tensor * build_rs(
             llm_graph_input_rs * inp,
-            ggml_cgraph * gf,
-            ggml_tensor * s,
-                int32_t   state_size,
-                int32_t   n_seqs,
-            const llm_graph_get_rows_fn & get_state_rows = ggml_get_rows) 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,
@@ -710,21 +743,24 @@ struct llm_graph_context {
 
     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,

package/src/llama.cpp/src/llama-hparams.cpp

@@ -65,12 +65,57 @@ uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     return n_embd_head_v * n_head_kv;
 }
 
+bool llama_hparams::is_n_embd_k_gqa_variable() const {
+    const uint32_t val = n_embd_k_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (val != n_embd_k_gqa(il)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+bool llama_hparams::is_n_embd_v_gqa_variable() const {
+    const uint32_t val = n_embd_v_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        if (val != n_embd_v_gqa(il)) {
+            return true;
+        }
+    }
+
+    return false;
+}
+
+uint32_t llama_hparams::n_embd_k_gqa_max() const {
+    uint32_t val = n_embd_k_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        val = std::max(val, n_embd_k_gqa(il));
+    }
+
+    return val;
+}
+
+uint32_t llama_hparams::n_embd_v_gqa_max() const {
+    uint32_t val = n_embd_v_gqa();
+    for (uint32_t il = 0; il < n_layer; ++il) {
+        val = std::max(val, n_embd_v_gqa(il));
+    }
+
+    return val;
+}
+
 uint32_t llama_hparams::n_embd_r() const {
     if (wkv_head_size != 0) {
         // for RWKV models
         return token_shift_count * n_embd;
     }
 
+    if (n_shortconv_l_cache != 0) {
+        // for LFM2 models
+        return n_embd * (n_shortconv_l_cache - 1);
+    }
+
     // TODO: maybe support other convolution strides than 1
     // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed
     // Corresponds to Mamba's conv_states size

package/src/llama.cpp/src/llama-hparams.h

@@ -6,7 +6,7 @@
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
-#define LLAMA_MAX_EXPERTS 256  // DeepSeekV3
+#define LLAMA_MAX_EXPERTS 384  // Kimi-K2
 
 enum llama_expert_gating_func_type {
     LLAMA_EXPERT_GATING_FUNC_TYPE_NONE    = 0,
@@ -55,6 +55,8 @@ struct llama_hparams {
     struct llama_hparams_posnet   posnet;
     struct llama_hparams_convnext convnext;
 
+    uint32_t n_shortconv_l_cache  = 0;
+
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_head_kv_arr;
     std::array<uint32_t, LLAMA_MAX_LAYERS> n_ff_arr;
@@ -189,6 +191,14 @@ struct llama_hparams {
     // dimension of value embeddings across all k-v heads
     uint32_t n_embd_v_gqa(uint32_t il = 0) const;
 
+    // true if any layer has a different n_embd_k_gqa/n_embd_v_gqa
+    bool is_n_embd_k_gqa_variable() const;
+    bool is_n_embd_v_gqa_variable() const;
+
+    // return the maximum n_embd_k_gqa/n_embd_v_gqa across all layers
+    uint32_t n_embd_k_gqa_max() const;
+    uint32_t n_embd_v_gqa_max() const;
+
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
     uint32_t n_embd_r() const;

package/src/llama.cpp/src/llama-kv-cache-unified-iswa.cpp

@@ -18,16 +18,17 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
                      bool   v_trans,
                      bool   offload,
                      bool   swa_full,
+                     bool   unified,
                  uint32_t   kv_size,
                  uint32_t   n_seq_max,
                  uint32_t   n_ubatch,
-                 uint32_t   n_pad) : hparams(model.hparams) {
+                 uint32_t   n_pad) : hparams(model.hparams), unified(unified) {
     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));
+    uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + 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) {
@@ -41,14 +42,14 @@ llama_kv_cache_unified_iswa::llama_kv_cache_unified_iswa(
 
     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,
+            v_trans, offload, unified, 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,
+            v_trans, offload, unified, size_swa, n_seq_max, n_pad,
             hparams.n_swa, hparams.swa_type);
 }
 
@@ -100,6 +101,11 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
     // first try simple split
     do {
+        if (!unified) {
+            // requires equal splits, so we skip the simple split
+            break;
+        }
+
         balloc.split_reset();
 
         std::vector<llama_ubatch> ubatches;
@@ -140,7 +146,7 @@ llama_memory_context_ptr llama_kv_cache_unified_iswa::init_batch(llama_batch_all
 
         std::vector<llama_ubatch> ubatches;
         while (true) {
-            auto ubatch = balloc.split_equal(n_ubatch, false);
+            auto ubatch = balloc.split_equal(n_ubatch, !unified);
 
             if (ubatch.n_tokens == 0) {
                 break;

package/src/llama.cpp/src/llama-kv-cache-unified-iswa.h

@@ -20,6 +20,7 @@ public:
                          bool   v_trans,
                          bool   offload,
                          bool   swa_full,
+                         bool   unified,
                      uint32_t   kv_size,
                      uint32_t   n_seq_max,
                      uint32_t   n_ubatch,
@@ -68,6 +69,8 @@ public:
 private:
     const llama_hparams & hparams;
 
+    const bool unified;
+
     std::unique_ptr<llama_kv_cache_unified> kv_base;
     std::unique_ptr<llama_kv_cache_unified> kv_swa;
 };