Merge commit '2fa51c19b0' into concedo_experimental

# Conflicts:
#	.github/actions/windows-setup-cuda/action.yml
#	.github/workflows/build-linux-cross.yml
#	.github/workflows/release.yml
#	README.md
#	docs/build-riscv64-spacemit.md
#	examples/model-conversion/logits.cpp
#	ggml/CMakeLists.txt
#	ggml/src/ggml-cpu/CMakeLists.txt
#	models/templates/Kimi-K2-Instruct.jinja
#	models/templates/Kimi-K2-Thinking.jinja
#	tests/test-chat.cpp
#	tools/server/README.md

src/llama-context.cpp

@@ -251,7 +251,10 @@ llama_context::llama_context(
 
         LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
 
-        const size_t max_nodes = this->graph_max_nodes();
+        const uint32_t n_seqs = cparams.n_seq_max;
+        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+
+        const size_t max_nodes = this->graph_max_nodes(n_tokens);
 
         LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);
 
@@ -309,9 +312,6 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
-        const uint32_t n_seqs = cparams.n_seq_max;
-        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
-
         // avoid reserving graphs with zero outputs - assume one output per sequence
         n_outputs = n_seqs;
 
@@ -1396,9 +1396,9 @@ void llama_context::output_reorder() {
 // graph
 //
 
-uint32_t llama_context::graph_max_nodes() const {
+uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
     if (model.arch == LLM_ARCH_QWEN3NEXT) {
-        return std::max<uint32_t>(8192u, 32u*model.n_tensors());
+        return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
     }
     return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }

src/llama-context.h

@@ -197,7 +197,7 @@ private:
     //
 
 public:
-    uint32_t graph_max_nodes() const;
+    uint32_t graph_max_nodes(uint32_t n_tokens) const;
 
     // can reuse the llm_graph_result instance of the context (for example to update a memory module)
     llm_graph_result * get_gf_res_reserve() const;

src/llama-model.cpp

@@ -1733,6 +1733,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
                 ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false);
 
+                // (optional) temperature tuning - used by mistral-large
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false);
+
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
                     case 60: type = LLM_TYPE_236B; break;

src/models/deepseek2.cpp

@@ -30,6 +30,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
     // {n_embd, n_tokens}
     inpL = build_inp_embd(model.tok_embd);
 
+    // (optional) temperature tuning - used by mistral-large
+    ggml_tensor * inp_attn_scale = nullptr;
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        inp_attn_scale = build_inp_attn_scale();
+    }
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -128,6 +134,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 ggml_tensor * Vcur = kv_cmpr;
                 cb(Vcur, "Vcur", il);
 
+                if (inp_attn_scale) {
+                    // apply llama 4 temperature scaling
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                    cb(Qcur, "Qcur_attn_temp_scaled", il);
+                }
+
                 // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
@@ -160,6 +172,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
                 cb(Kcur, "Kcur", il);
 
+                if (inp_attn_scale) {
+                    // apply llama 4 temperature scaling
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                    cb(Qcur, "Qcur_attn_temp_scaled", il);
+                }
+
                 // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
                 cur = build_attn(inp_attn,
                             model.layers[il].wo, NULL,