whispercpp 1.3.5 → 1.3.6

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

@@ -24,6 +24,7 @@ class llama_kv_cache_context;
 class llama_kv_cache_iswa_context;
 class llama_memory_recurrent_context;
 class llama_memory_hybrid_context;
+class llama_memory_hybrid_iswa_context;
 
 // certain models (typically multi-modal) can produce different types of graphs
 enum llm_graph_type {
@@ -105,7 +106,7 @@ 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;
+    llm_graph_input_embd(int64_t n_embd) : n_embd(n_embd) {}
     virtual ~llm_graph_input_embd() = default;
 
     void set_input(const llama_ubatch * ubatch) override;
@@ -114,6 +115,8 @@ public:
 
     ggml_tensor * tokens = nullptr; // I32 [n_batch]
     ggml_tensor * embd   = nullptr; // F32 [n_embd, n_batch]
+
+    const int64_t n_embd = 0;
 };
 
 class llm_graph_input_pos : public llm_graph_input_i {
@@ -314,6 +317,39 @@ public:
     const llama_kv_cache_context * mctx;
 };
 
+// V-less input for the KV cache
+// ref: https://github.com/ggml-org/llama.cpp/pull/19067
+class llm_graph_input_attn_k : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_k(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_context * mctx) :
+        hparams(hparams),
+        cparams(cparams),
+        mctx(mctx) {
+    }
+    ~llm_graph_input_attn_k() = 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_kq_mask() const { return self_kq_mask_cnv; }
+
+    ggml_tensor * self_k_idxs = nullptr; // I64 [n_batch]
+
+    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;
+
+    const llama_kv_cache_context * mctx;
+};
+
 class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
     llm_graph_input_attn_kv_iswa(
@@ -397,6 +433,62 @@ public:
     const llama_memory_hybrid_context * mctx;
 };
 
+class llm_graph_input_mem_hybrid_k : public llm_graph_input_i {
+public:
+    llm_graph_input_mem_hybrid_k(
+            const llama_cparams & cparams,
+            std::unique_ptr<llm_graph_input_attn_k> 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) { }
+    virtual ~llm_graph_input_mem_hybrid_k() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    std::unique_ptr<llm_graph_input_attn_k> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>      inp_rs;
+
+    llm_graph_input_attn_k * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs      * get_recr() const { return inp_rs.get(); }
+
+    const llama_cparams cparams;
+
+    const llama_memory_hybrid_context * mctx;
+};
+
+class llm_graph_input_mem_hybrid_iswa : public llm_graph_input_i {
+public:
+    llm_graph_input_mem_hybrid_iswa(
+            const llama_cparams & cparams,
+            std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn,
+            std::unique_ptr<llm_graph_input_rs>          inp_rs,
+            const llama_memory_hybrid_iswa_context *     mctx) :
+        inp_attn(std::move(inp_attn)),
+        inp_rs(std::move(inp_rs)),
+        cparams(cparams),
+        mctx(mctx) { }
+    virtual ~llm_graph_input_mem_hybrid_iswa() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    std::unique_ptr<llm_graph_input_attn_kv_iswa> inp_attn;
+    std::unique_ptr<llm_graph_input_rs>          inp_rs;
+
+    llm_graph_input_attn_kv_iswa * get_attn() const { return inp_attn.get(); }
+    llm_graph_input_rs           * get_recr() const { return inp_rs.get(); }
+
+    const llama_cparams cparams;
+
+    const llama_memory_hybrid_iswa_context * mctx;
+};
+
 class llm_graph_input_sampling : public llm_graph_input_i {
 public:
     llm_graph_input_sampling(std::map<llama_seq_id, llama_sampler *> samplers) :
@@ -537,7 +629,7 @@ public:
 
     virtual ~llm_graph_result() = default;
 
-    ggml_tensor * get_tokens()      const { return t_tokens; }
+    ggml_tensor * get_inp_tokens()  const { return t_inp_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; }
@@ -564,7 +656,8 @@ public:
     void set_params(const llm_graph_params & params);
 
     // important graph nodes
-    ggml_tensor * t_tokens      = nullptr;
+    ggml_tensor * t_inp_tokens  = nullptr;
+    ggml_tensor * t_inp_embd    = nullptr; // [n_embd_inp, n_tokens]
     ggml_tensor * t_logits      = nullptr;
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;
@@ -671,10 +764,11 @@ struct llm_graph_context {
              ggml_tensor * cur,
                      int   il) const;
 
-    // do mat_mul, while optionally apply lora
+    // do mat_mul, while optionally apply lora and per-tensor scale
     ggml_tensor * build_lora_mm(
               ggml_tensor * w,
-              ggml_tensor * cur) const;
+              ggml_tensor * cur,
+              ggml_tensor * w_s = nullptr) const;
 
     // do mat_mul_id, while optionally apply lora
     ggml_tensor * build_lora_mm_id(
@@ -717,11 +811,14 @@ struct llm_graph_context {
                  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;
+             ggml_tensor * probs_in = nullptr,
+             ggml_tensor * gate_up_exps = nullptr,
+             ggml_tensor * up_exps_s = nullptr,
+             ggml_tensor * gate_exps_s = nullptr,
+             ggml_tensor * down_exps_s = nullptr) const;
 
     ggml_tensor * build_moe_ffn(
              ggml_tensor * cur,
@@ -738,11 +835,15 @@ struct llm_graph_context {
                  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;
+             ggml_tensor * probs_in = nullptr,
+             ggml_tensor * gate_up_exps = nullptr,
+             ggml_tensor * gate_up_exps_b = nullptr,
+             ggml_tensor * up_exps_s = nullptr,
+             ggml_tensor * gate_exps_s = nullptr,
+             ggml_tensor * down_exps_s = nullptr) const;
 
     //
     // inputs
@@ -801,6 +902,21 @@ struct llm_graph_context {
             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] // TODO: remove
+                  float   kq_scale,
+                    int   il) const;
+
+    llm_graph_input_attn_k  * build_attn_inp_k() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_k * 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;
@@ -880,6 +996,9 @@ struct llm_graph_context {
     //
 
     llm_graph_input_mem_hybrid * build_inp_mem_hybrid() const;
+    llm_graph_input_mem_hybrid_k * build_inp_mem_hybrid_k() const;
+
+    llm_graph_input_mem_hybrid_iswa * build_inp_mem_hybrid_iswa() const;
 
     //
     // pooling
@@ -889,7 +1008,8 @@ struct llm_graph_context {
             ggml_tensor * cls,
             ggml_tensor * cls_b,
             ggml_tensor * cls_out,
-            ggml_tensor * cls_out_b) const;
+            ggml_tensor * cls_out_b,
+            ggml_tensor * cls_norm) const;
 
     //
     // sampling (backend sampling)
@@ -903,6 +1023,7 @@ struct llm_graph_context {
 
     void build_dense_out(
             ggml_tensor * dense_2,
+            ggml_tensor * dense_2_b,
             ggml_tensor * dense_3) const;
 };
 

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

@@ -62,6 +62,14 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const {
     return n_head/n_head_kv;
 }
 
+uint32_t llama_hparams::n_rot(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_rot_swa : n_rot_full;
+    }
+
+    GGML_ABORT("fatal error");
+}
+
 uint32_t llama_hparams::n_embd_inp() const {
     uint32_t n_embd_inp = n_embd;
 
@@ -72,20 +80,36 @@ uint32_t llama_hparams::n_embd_inp() const {
     return n_embd_inp;
 }
 
-uint32_t llama_hparams::get_n_embd_out() const {
-    return n_embd_out > 0 ? n_embd_out : n_embd;
+uint32_t llama_hparams::n_embd_out() const {
+    return n_embd_out_impl > 0 ? n_embd_out_impl : n_embd;
+}
+
+uint32_t llama_hparams::n_embd_head_k(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_embd_head_k_swa : n_embd_head_k_full;
+    }
+
+    GGML_ABORT("fatal error");
+}
+
+uint32_t llama_hparams::n_embd_head_v(uint32_t il) const {
+    if (il < n_layer) {
+        return is_swa(il) ? n_embd_head_v_swa : n_embd_head_v_full;
+    }
+
+    GGML_ABORT("fatal error");
 }
 
 uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
-    return n_embd_head_k * n_head_kv;
+    return n_embd_head_k(il) * n_head_kv;
 }
 
 uint32_t llama_hparams::n_embd_v_gqa(uint32_t il) const {
     const uint32_t n_head_kv = this->n_head_kv(il);
 
-    return n_embd_head_v * n_head_kv;
+    return n_embd_head_v(il) * n_head_kv;
 }
 
 bool llama_hparams::is_n_embd_k_gqa_variable() const {
@@ -139,6 +163,13 @@ uint32_t llama_hparams::n_embd_r() const {
         return n_embd * (n_shortconv_l_cache - 1);
     }
 
+    if (n_embd_head_kda != 0) {
+        // for Kimi KDA layers
+        // Conv state for Q, K, V: 3 * (d_conv - 1) * n_head * head_dim
+        const uint32_t d_inner = n_head() * n_embd_head_kda;  // 32 * 128 = 4096
+        return 3 * (ssm_d_conv > 0 ? ssm_d_conv - 1 : 3) * d_inner;
+    }
+
     // 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
@@ -151,6 +182,13 @@ uint32_t llama_hparams::n_embd_s() const {
         return n_embd * wkv_head_size;
     }
 
+    if (n_embd_head_kda != 0) {
+        // for Kimi KDA layers
+        // Full recurrent state: head_dim * head_dim * n_head
+        // h tensor shape for delta attention: [head_dim, head_dim, n_head]
+        return n_embd_head_kda * n_embd_head_kda * n_head();  // 128 * 128 * 32 = 524288
+    }
+
     // corresponds to Mamba's ssm_states size
     return ssm_d_state * ssm_d_inner;
 }
@@ -175,6 +213,21 @@ bool llama_hparams::is_swa(uint32_t il) const {
     GGML_ABORT("fatal error");
 }
 
+bool llama_hparams::is_mla() const {
+    assert((n_embd_head_k_mla_impl == 0 && n_embd_head_v_mla_impl == 0) ||
+           (n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0));
+
+    return n_embd_head_k_mla_impl != 0 && n_embd_head_v_mla_impl != 0;
+}
+
+uint32_t llama_hparams::n_embd_head_k_mla() const {
+    return is_mla() ? n_embd_head_k_mla_impl : n_embd_head_k();
+}
+
+uint32_t llama_hparams::n_embd_head_v_mla() const {
+    return is_mla() ? n_embd_head_v_mla_impl : n_embd_head_v();
+}
+
 bool llama_hparams::has_kv(uint32_t il) const {
     if (n_layer_kv_from_start >= 0) {
         if (il < (uint32_t) n_layer_kv_from_start) {
@@ -200,42 +253,6 @@ uint32_t llama_hparams::n_layer_kv() const {
     return res;
 }
 
-bool llama_hparams::is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
-    assert(p0 >= 0 && p1 >= 0);
-
-    switch (swa_type) {
-        case LLAMA_SWA_TYPE_NONE:
-            {
-            } break;
-        case LLAMA_SWA_TYPE_STANDARD:
-            {
-                if (p1 - p0 >= (int32_t) n_swa) {
-                    return true;
-                }
-            } break;
-        case LLAMA_SWA_TYPE_CHUNKED:
-            {
-                const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
-
-                if (p0 < pos_chunk_start) {
-                    return true;
-                }
-            } break;
-        case LLAMA_SWA_TYPE_SYMMETRIC:
-            {
-                const int32_t half_n_swa = (int32_t) n_swa / 2;
-                const int32_t pos_diff = p1 - p0;
-
-                // Mask if outside the symmetric window
-                if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
-                    return true;
-                }
-            } break;
-    }
-
-    return false;
-}
-
 bool llama_hparams::use_mrope() const {
     return rope_sections[0] > 0 && rope_sections[1] > 0;
 }

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

@@ -3,6 +3,7 @@
 #include "llama.h"
 
 #include <array>
+#include <cassert>
 
 // bump if necessary
 #define LLAMA_MAX_LAYERS  512
@@ -41,19 +42,25 @@ struct llama_hparams {
 
     uint32_t n_ctx_train; // context size the model was trained on
     uint32_t n_embd;
-    uint32_t n_embd_features = 0;
     uint32_t n_layer;
     int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache
-    uint32_t n_rot;
-    uint32_t n_embd_head_k; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
-    uint32_t n_embd_head_v; // dimension of values (d_v) aka n_embd_head
     uint32_t n_expert = 0;
     uint32_t n_expert_used = 0;
     uint32_t n_rel_attn_bkts = 0;
 
+    // different head size for full_attention and SWA layers
+    uint32_t n_embd_head_k_full; // dimension of keys (d_k). d_q is assumed to be the same, but there are n_head q heads, and only n_head_kv k-v heads
+    uint32_t n_embd_head_v_full; // dimension of values (d_v) aka n_embd_head
+    uint32_t n_embd_head_k_swa;
+    uint32_t n_embd_head_v_swa;
+
+    // different RoPE dimensions for full_attention and SWA layers
+    uint32_t n_rot_full;
+    uint32_t n_rot_swa;
+
     // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA
-    uint32_t n_embd_head_k_mla = 0;
-    uint32_t n_embd_head_v_mla = 0;
+    uint32_t n_embd_head_k_mla_impl = 0;
+    uint32_t n_embd_head_v_mla_impl = 0;
 
     // for WavTokenizer
     struct llama_hparams_posnet   posnet;
@@ -82,6 +89,7 @@ struct llama_hparams {
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
     uint32_t moe_every_n_layers   = 0;
+    uint32_t moe_latent_size      = 0;
     uint32_t nextn_predict_layers = 0;
 
     float f_norm_eps;
@@ -136,6 +144,9 @@ struct llama_hparams {
     uint32_t ssm_dt_rank = 0;
     uint32_t ssm_n_group = 0;
 
+    // for Kimi Linear KDA
+    uint32_t n_embd_head_kda = 0;
+
     // for hybrid state space models
     std::array<bool, LLAMA_MAX_LAYERS> recurrent_layer_arr;
 
@@ -163,7 +174,7 @@ struct llama_hparams {
     uint32_t n_cls_out = 1;
 
     // output embedding dimension (0 = use n_embd)
-    uint32_t n_embd_out = 0;
+    uint32_t n_embd_out_impl = 0;
 
     // llama4 smallthinker
     uint32_t n_moe_layer_step        = 0;
@@ -190,11 +201,16 @@ struct llama_hparams {
     std::array<float, LLAMA_MAX_LAYERS> xielu_beta;
     std::array<float, LLAMA_MAX_LAYERS> xielu_eps;
 
+    // DSA (deepseek sparse attention)
+    uint32_t indexer_n_head    = 0;
+    uint32_t indexer_head_size = 0;
+    uint32_t indexer_top_k     = 0;
+
     // qwen3vl deepstack
     uint32_t n_deepstack_layers = 0;
 
     // needed by encoder-decoder models (e.g. T5, FLAN-T5)
-    // ref: https://github.com/ggerganov/llama.cpp/pull/8141
+    // ref: https://github.com/ggml-org/llama.cpp/pull/8141
     llama_token dec_start_token_id = LLAMA_TOKEN_NULL;
     uint32_t    dec_n_layer        = 0;
 
@@ -202,6 +218,11 @@ struct llama_hparams {
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
     enum llama_rope_scaling_type rope_scaling_type_train = LLAMA_ROPE_SCALING_TYPE_NONE;
 
+
+    // Step35: optional per-layer clamps for (Swi)GLU
+    std::array<float, LLAMA_MAX_LAYERS> swiglu_clamp_exp; // clamping for expert FFN
+    std::array<float, LLAMA_MAX_LAYERS> swiglu_clamp_shexp; // shared expert
+
     // this value n_pattern means that every nth layer is dense (i.e. non-SWA)
     // dense_first means whether the pattern is start with a dense layer
     // note that if n_pattern == 0, all layers are SWA
@@ -234,11 +255,17 @@ struct llama_hparams {
 
     uint32_t n_gqa(uint32_t il = 0) const;
 
+    uint32_t n_rot(uint32_t il = 0) const;
+
     // dimension of main + auxiliary input embeddings
     uint32_t n_embd_inp() const;
 
     // dimension of output embeddings
-    uint32_t get_n_embd_out() const;
+    uint32_t n_embd_out() const;
+
+    // dimension of key/value embeddings for each head (per layer)
+    uint32_t n_embd_head_k(uint32_t il = 0) const;
+    uint32_t n_embd_head_v(uint32_t il = 0) const;
 
     // dimension of key embeddings across all k-v heads
     uint32_t n_embd_k_gqa(uint32_t il = 0) const;
@@ -268,15 +295,57 @@ struct llama_hparams {
 
     bool is_swa(uint32_t il) const;
 
+    // note: currently only support if either all or none of the layers are MLA
+    bool is_mla() const;
+
+    uint32_t n_embd_head_k_mla() const;
+    uint32_t n_embd_head_v_mla() const;
+
     bool has_kv(uint32_t il) const;
 
     // number of layers for which has_kv() returns true
     uint32_t n_layer_kv() const;
 
     // note that this function uses different SWA parameters from those in the hparams
+    // note: inlined on purpose for performance reasons
     // TODO: think of a better place for this function
     // TODO: pack the SWA params in a struct?
-    static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1);
+    static bool is_masked_swa(uint32_t n_swa, llama_swa_type swa_type, llama_pos p0, llama_pos p1) {
+        assert(p0 >= 0 && p1 >= 0);
+
+        switch (swa_type) {
+            case LLAMA_SWA_TYPE_NONE:
+                {
+                } break;
+            case LLAMA_SWA_TYPE_STANDARD:
+                {
+                    if (p1 - p0 >= (int32_t) n_swa) {
+                        return true;
+                    }
+                } break;
+            case LLAMA_SWA_TYPE_CHUNKED:
+                {
+                    const llama_pos pos_chunk_start = (p1 / n_swa) * n_swa;
+
+                    if (p0 < pos_chunk_start) {
+                        return true;
+                    }
+                } break;
+            case LLAMA_SWA_TYPE_SYMMETRIC:
+                {
+                    const int32_t half_n_swa = (int32_t) n_swa / 2;
+                    const int32_t pos_diff = p1 - p0;
+
+                    // Mask if outside the symmetric window
+                    if (pos_diff < -half_n_swa || pos_diff > half_n_swa) {
+                        return true;
+                    }
+                } break;
+        }
+
+        return false;
+    }
+
 
     bool use_mrope() const;
 };

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

@@ -100,18 +100,18 @@ std::string format(const char * fmt, ...) {
 
 std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {
     char buf[256];
-    snprintf(buf, sizeof(buf), "%5" PRId64, ne.at(0));
+    snprintf(buf, sizeof(buf), "%6" PRId64, ne.at(0));
     for (size_t i = 1; i < ne.size(); i++) {
-        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, ne.at(i));
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %6" PRId64, ne.at(i));
     }
     return buf;
 }
 
 std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
     char buf[256];
-    snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
+    snprintf(buf, sizeof(buf), "%6" PRId64, t->ne[0]);
     for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %6" PRId64, t->ne[i]);
     }
     return buf;
 }

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

@@ -49,6 +49,16 @@ struct time_meas {
     int64_t & t_acc;
 };
 
+template <typename T>
+struct buffer_view {
+    T * data;
+    size_t size = 0;
+
+    bool has_data() const {
+        return data && size > 0;
+    }
+};
+
 void replace_all(std::string & s, const std::string & search, const std::string & replace);
 
 // TODO: rename to llama_format ?
@@ -60,4 +70,6 @@ std::string llama_format_tensor_shape(const struct ggml_tensor * t);
 
 std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
 
-#define LLAMA_TENSOR_NAME_FATTN "__fattn__"
+#define LLAMA_TENSOR_NAME_FATTN   "__fattn__"
+#define LLAMA_TENSOR_NAME_FGDN_AR "__fgdn_ar__"
+#define LLAMA_TENSOR_NAME_FGDN_CH "__fgdn_ch__"

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

@@ -218,7 +218,9 @@ llama_memory_context_ptr llama_kv_cache_iswa::init_update(llama_context * lctx,
 }
 
 bool llama_kv_cache_iswa::get_can_shift() const {
-    return kv_base->get_size() == kv_swa->get_size();
+    return kv_base->get_can_shift() &&
+           kv_swa->get_can_shift() &&
+           kv_base->get_size() == kv_swa->get_size();
 }
 
 void llama_kv_cache_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {