model : support for DeepseekV32ForCausalLM with generic DeepSeek Sparse Attention (DSA) implementation (#23346)

* llama : support DeepSeek V3.2 model family (with DSA lightning indexer)

* convert : handle DeepseekV32ForCausalLM architecture

* ggml : support for f16 GGML_OP_FILL

* memory : separate hparams argument in llama_kv_cache constructor

* memory : add llama_kv_cache_dsa memory (KV cache + lightning indexer cache)

* llama : support for LLM_ARCH_DEEPSEEK32

* model : llama_model_deepseek32 implementation

* model : merge two scale operations into one in DSA lightning indexer implementation

* chore : remove unused code

* model : support NVFP4 in DeepSeek V3.2

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* memory : refactoring TODO

Co-authored-by: ggerganov <ggerganov@users.noreply.github.com>

---------

Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: ggerganov <ggerganov@users.noreply.github.com>

src/llama-graph.h

@@ -22,6 +22,7 @@ struct llama_layer;
 struct llama_memory_context_i;
 
 class llama_kv_cache_context;
+class llama_kv_cache_dsa_context;
 class llama_kv_cache_iswa_context;
 class llama_memory_recurrent_context;
 class llama_memory_hybrid_context;
@@ -373,6 +374,44 @@ public:
     const llama_kv_cache_context * mctx;
 };
 
+class llm_graph_input_attn_k_dsa : public llm_graph_input_i {
+public:
+    llm_graph_input_attn_k_dsa(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_dsa_context * mctx) :
+        hparams(hparams),
+        cparams(cparams),
+        mctx(mctx) {
+    }
+    ~llm_graph_input_attn_k_dsa() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * get_k_idxs_mla() const { return self_k_idxs_mla; }
+    ggml_tensor * get_k_idxs_lid() const { return self_k_idxs_lid; }
+
+    ggml_tensor * get_kq_mask_mla() const { return self_kq_mask_mla_cnv; }
+    ggml_tensor * get_kq_mask_lid() const { return self_kq_mask_lid; }
+
+    ggml_tensor * self_k_idxs_mla = nullptr; // I64 [n_batch]
+    ggml_tensor * self_k_idxs_lid = nullptr; // I64 [n_batch]
+
+    ggml_tensor * self_kq_mask_mla     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_mla_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_lid     = nullptr; // F32 [n_kv, n_batch/n_stream, 1, n_stream]
+    ggml_tensor * self_kq_mask_lid_cnv = nullptr; //     [n_kv, n_batch/n_stream, 1, n_stream]
+
+    ggml_tensor * self_k_rot_lid = nullptr;
+
+    const llama_hparams hparams;
+    const llama_cparams cparams;
+
+    const llama_kv_cache_dsa_context * mctx;
+};
+
 class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
     llm_graph_input_attn_kv_iswa(
@@ -973,6 +1012,23 @@ struct llm_graph_context {
                   float   kq_scale,
                     int   il) const;
 
+    llm_graph_input_attn_k_dsa * build_attn_inp_k_dsa() const;
+
+    ggml_tensor * build_attn(
+            llm_graph_input_attn_k_dsa * inp,
+            ggml_tensor * wo,
+            ggml_tensor * wo_b,
+            ggml_tensor * wo_s,
+            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]
+            ggml_tensor * top_k, // [n_indexer_top_k, n_tokens]
+                  float   kq_scale,
+                    int   il) 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