Update pinned llama.cpp

.github/workflows/pr_builds.yml

@@ -538,14 +538,14 @@ jobs:
             ${{ env.CACHE_NAMESPACE }}-${{ runner.os }}-skippy-ci-smoke-models-
 
       - name: SDK and API crate tests
-        if: ${{ matrix.group == 'sdk-api' && (needs.changes.outputs.all_rust == 'true' || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-client') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-api-client') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-api-server') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-config') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-commands') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-events') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-hardware-profile') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-runtime-install') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-sdk') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-cli') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-tui') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-embedded-runtime') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-console-server') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-ffi') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-nodejs')) }}
+        if: ${{ matrix.group == 'sdk-api' && (needs.changes.outputs.all_rust == 'true' || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-client') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-api-client') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-api-server') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-config') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-commands') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-events') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-hardware-profile') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-runtime-install') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-native-runtime') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-routing') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-types') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-sdk') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-cli') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-tui') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-embedded-runtime') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-console-server') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-ffi') || contains(fromJson(needs.changes.outputs.affected_crates || '[]'), 'mesh-llm-nodejs')) }}
         run: |
           should_test() {
             local crate="$1"
             [ "$ALL_RUST" = "true" ] || jq -e --arg crate "$crate" 'index($crate) != null' <<<"$AFFECTED" >/dev/null
           }
 
-          for c in mesh-llm-client mesh-llm-api-client mesh-llm-api-server mesh-llm-config mesh-llm-commands mesh-llm-events mesh-llm-hardware-profile mesh-llm-runtime-install mesh-llm-cli mesh-llm-tui mesh-llm-embedded-runtime mesh-llm-sdk mesh-llm-console-server mesh-llm-ffi mesh-llm-nodejs; do
+          for c in mesh-llm-client mesh-llm-api-client mesh-llm-api-server mesh-llm-config mesh-llm-commands mesh-llm-events mesh-llm-hardware-profile mesh-llm-runtime-install mesh-llm-native-runtime mesh-llm-routing mesh-llm-types mesh-llm-cli mesh-llm-tui mesh-llm-embedded-runtime mesh-llm-sdk mesh-llm-console-server mesh-llm-ffi mesh-llm-nodejs; do
             if should_test "$c"; then
               cargo test -p "$c"
             else

third_party/llama.cpp/patches/0001-Add-Skippy-ABI-and-package-writer-foundation.patch

@@ -1,4 +1,4 @@
-From 5c6e3bea6ac4126a8528a250bd270f72ee2cd61b Mon Sep 17 00:00:00 2001
+From 21cd005d8e0a4e5b1bb3bc0768969d4cb1bedafd Mon Sep 17 00:00:00 2001
 From: Mesh-LLM CI <ci@mesh-llm.local>
 Date: Fri, 26 Jun 2026 13:47:52 +1000
 Subject: [PATCH 01/10] Add Skippy ABI and package writer foundation
@@ -334,10 +334,10 @@ index 00000000..7ab93d35
 +
 +#endif // SKIPPY_ABI_H
 diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
-index d15ccfd9..2526f45c 100644
+index 320784c3..5298c337 100644
 --- a/src/CMakeLists.txt
 +++ b/src/CMakeLists.txt
-@@ -35,6 +35,7 @@ add_library(llama
+@@ -36,6 +36,7 @@ add_library(llama
              llama-model.cpp
              llama-quant.cpp
              llama-sampler.cpp
@@ -346,11 +346,11 @@ index d15ccfd9..2526f45c 100644
              unicode-data.cpp
              unicode.cpp
 diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
-index 68c9e606..c39487b8 100644
+index 4c86e43c..118987b3 100644
 --- a/src/llama-graph.cpp
 +++ b/src/llama-graph.cpp
-@@ -19,6 +19,20 @@
- #include <sstream>
+@@ -21,6 +21,20 @@
+ #include <string>
  #include <unordered_set>
 
 +static thread_local skippy_graph_filter g_skippy_graph_filter;
@@ -370,7 +370,7 @@ index 68c9e606..c39487b8 100644
  // dedup helpers
 
  static ggml_tensor * build_attn_inp_kq_mask(
-@@ -1855,28 +1869,32 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
+@@ -2164,28 +2178,32 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
      {
          auto & cur = inps[0];
 
@@ -421,10 +421,10 @@ index 68c9e606..c39487b8 100644
      }
 
 diff --git a/src/llama-graph.h b/src/llama-graph.h
-index a6e8c398..00170956 100644
+index 4b5b75c6..b943320e 100644
 --- a/src/llama-graph.h
 +++ b/src/llama-graph.h
-@@ -28,6 +28,18 @@ class llama_memory_recurrent_context;
+@@ -30,6 +30,18 @@ class llama_memory_recurrent_context;
  class llama_memory_hybrid_context;
  class llama_memory_hybrid_iswa_context;
 
@@ -444,7 +444,7 @@ index a6e8c398..00170956 100644
  enum llm_graph_type {
      LLM_GRAPH_TYPE_DEFAULT,
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index 474cabdf..86909230 100644
+index e07b0d23..9b0e00ae 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
 @@ -17,6 +17,16 @@ static const size_t kiB = 1024;
@@ -464,7 +464,7 @@ index 474cabdf..86909230 100644
  const char * llama_file_version_name(llama_fver version) {
      switch (version) {
          case GGUF_FILE_VERSION_V1: return "GGUF V1 (support until nov 2023)";
-@@ -1100,6 +1110,40 @@ struct ggml_tensor * llama_model_loader::create_tensor(
+@@ -1103,6 +1113,40 @@ struct ggml_tensor * llama_model_loader::create_tensor(
              throw std::runtime_error(format("missing tensor info mapping for %s", tn.str().c_str()));
          }
 

third_party/llama.cpp/patches/0002-Add-early-staged-model-family-and-chat-support.patch

@@ -1,4 +1,4 @@
-From 459e05b46cc375894b3386f6f732f6028a9a6699 Mon Sep 17 00:00:00 2001
+From 5625209c975d14399629caa50adf3cbb810dbcf5 Mon Sep 17 00:00:00 2001
 From: Mesh-LLM CI <ci@mesh-llm.local>
 Date: Fri, 26 Jun 2026 13:47:52 +1000
 Subject: [PATCH 02/10] Add early staged model-family and chat support
@@ -32,11 +32,11 @@ Squashes the first Qwen MoE staged-runtime support, native KV page basics, sessi
  create mode 100644 common/stage-chat.cpp
 
 diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
-index fc16b21c..4e5a28c0 100644
+index 4cf580a0..b06bfeda 100644
 --- a/common/CMakeLists.txt
 +++ b/common/CMakeLists.txt
-@@ -100,6 +100,7 @@ add_library(${TARGET}
-     regex-partial.h
+@@ -98,6 +98,7 @@ add_library(${TARGET}
+     reasoning-budget.h
      sampling.cpp
      sampling.h
 +    stage-chat.cpp
@@ -319,18 +319,18 @@ index 7ab93d35..23a46e0f 100644
          const char * path,
          struct skippy_model_info ** out_info,
 diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
-index c39487b8..126261f2 100644
+index 118987b3..dfde71c8 100644
 --- a/src/llama-graph.cpp
 +++ b/src/llama-graph.cpp
-@@ -20,6 +20,7 @@
+@@ -22,6 +22,7 @@
  #include <unordered_set>
 
  static thread_local skippy_graph_filter g_skippy_graph_filter;
 +static thread_local skippy_activation_tokens g_skippy_activation_tokens;
 
  void skippy_graph_set_filter(const skippy_graph_filter & filter) {
      g_skippy_graph_filter = filter;
-@@ -33,6 +34,18 @@ const skippy_graph_filter & skippy_graph_get_filter() {
+@@ -35,6 +36,18 @@ const skippy_graph_filter & skippy_graph_get_filter() {
      return g_skippy_graph_filter;
  }
 
@@ -349,7 +349,7 @@ index c39487b8..126261f2 100644
  // dedup helpers
 
  static ggml_tensor * build_attn_inp_kq_mask(
-@@ -153,6 +166,18 @@ bool llm_graph_input_embd_h::can_reuse(const llm_graph_params & params) {
+@@ -155,6 +168,18 @@ bool llm_graph_input_embd_h::can_reuse(const llm_graph_params & params) {
      return res;
  }
 
@@ -368,7 +368,7 @@ index c39487b8..126261f2 100644
  void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
      if (ubatch->pos && pos) {
          const int64_t n_tokens = ubatch->n_tokens;
-@@ -680,17 +705,22 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
+@@ -967,17 +992,22 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
  }
 
  void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) {
@@ -396,7 +396,7 @@ index c39487b8..126261f2 100644
      }
 
      const int64_t n_rs = mctx->get_recr()->get_n_rs();
-@@ -733,9 +763,14 @@ bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
+@@ -1020,9 +1050,14 @@ bool llm_graph_input_mem_hybrid::can_reuse(const llm_graph_params & params) {
  // Instead of creating a hybrid input, the graph can simply create 2 separate inputs.
  // Refactoring is required in the future.
  void llm_graph_input_mem_hybrid_k::set_input(const llama_ubatch * ubatch) {
@@ -414,10 +414,10 @@ index c39487b8..126261f2 100644
      const int64_t n_rs = mctx->get_recr()->get_n_rs();
 
 diff --git a/src/llama-graph.h b/src/llama-graph.h
-index 00170956..bc1e10e2 100644
+index b943320e..31083518 100644
 --- a/src/llama-graph.h
 +++ b/src/llama-graph.h
-@@ -36,9 +36,17 @@ struct skippy_graph_filter {
+@@ -38,9 +38,17 @@ struct skippy_graph_filter {
      bool include_output = false;
  };
 
@@ -435,7 +435,7 @@ index 00170956..bc1e10e2 100644
 
  // certain models (typically multi-modal) can produce different types of graphs
  enum llm_graph_type {
-@@ -152,6 +160,18 @@ public:
+@@ -154,6 +162,18 @@ public:
      const int64_t n_embd = 0;
  };
 
@@ -455,7 +455,7 @@ index 00170956..bc1e10e2 100644
  public:
      llm_graph_input_pos(uint32_t n_pos_per_embd) : n_pos_per_embd(n_pos_per_embd) {}
 diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
-index 2802103b..400a9d29 100644
+index 12bf5c37..906e9eac 100644
 --- a/src/llama-kv-cache.cpp
 +++ b/src/llama-kv-cache.cpp
 @@ -4,6 +4,7 @@
@@ -466,8 +466,8 @@ index 2802103b..400a9d29 100644
 
  #include <algorithm>
  #include <cassert>
-@@ -1220,6 +1221,352 @@ ggml_type llama_kv_cache::type_v() const {
-     return layers[0].v->type;
+@@ -1240,6 +1241,352 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
+     return layers[ikv].k;
  }
 
 +bool llama_kv_cache::stage_export_kv_page(
@@ -820,7 +820,7 @@ index 2802103b..400a9d29 100644
      uint32_t result = 0;
 
 diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
-index 3d68f98c..83cee5f9 100644
+index 531d99db..d9301c19 100644
 --- a/src/llama-kv-cache.h
 +++ b/src/llama-kv-cache.h
 @@ -6,12 +6,14 @@
@@ -838,9 +838,9 @@ index 3d68f98c..83cee5f9 100644
 
  //
  // llama_kv_cache
-@@ -161,6 +163,23 @@ public:
-     ggml_type type_k() const;
-     ggml_type type_v() const;
+@@ -164,6 +166,23 @@ public:
+     std::vector<uint32_t> get_layer_ids() const;
+     ggml_tensor * get_k_storage(int32_t il) const;
 
 +    bool stage_export_kv_page(
 +            int32_t layer_start,
@@ -863,10 +863,10 @@ index 3d68f98c..83cee5f9 100644
      // graph_build API
      //
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index 86909230..622b228e 100644
+index 9b0e00ae..561e9850 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
-@@ -1088,13 +1088,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(
+@@ -1091,13 +1091,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(
      };
 
      auto buft_for_tensor = [&](ggml_tensor * t_meta) -> ggml_backend_buffer_type_t {
@@ -880,7 +880,7 @@ index 86909230..622b228e 100644
          // some models use the token embedding tensor as the output, but since these are used in different layers and with different ops
          // the tensor is duplicated
          // to handle this, we check if the tensor is duplicated, and if so, we assume that it is being loaded as the output tensor
-@@ -1128,22 +1121,38 @@ struct ggml_tensor * llama_model_loader::create_tensor(
+@@ -1131,22 +1124,38 @@ struct ggml_tensor * llama_model_loader::create_tensor(
                      break;
              }
 
@@ -929,10 +929,10 @@ index 86909230..622b228e 100644
          if (info.op == GGML_OP_NONE || (flags & TENSOR_SKIP)) {
              const size_t nbytes = ggml_nbytes(t_meta);
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index d041a9ce..fc13d414 100644
+index d58ebac2..3196973f 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1360,13 +1360,13 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
+@@ -1367,13 +1367,13 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
              }
 
              // MoE expert weight scales (per-expert, shape {n_expert})

third_party/llama.cpp/patches/0003-Add-staged-sampling-checkpoints-and-part-loading.patch

@@ -1,4 +1,4 @@
-From 3bb375d5797de655cdee37f0a8eb272cfb609812 Mon Sep 17 00:00:00 2001
+From a4f23f17019419c4e1f7acd2bea9bd9ee18159f9 Mon Sep 17 00:00:00 2001
 From: Mesh-LLM CI <ci@mesh-llm.local>
 Date: Fri, 26 Jun 2026 13:47:52 +1000
 Subject: [PATCH 03/10] Add staged sampling, checkpoints, and part loading
@@ -245,18 +245,18 @@ index 23a46e0f..4dd985da 100644
          size_t output_text_capacity,
          size_t * out_text_bytes,
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index 622b228e..16cc0b06 100644
+index 561e9850..97372c4c 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
-@@ -525,6 +525,7 @@ llama_model_loader::llama_model_loader(
+@@ -528,6 +528,7 @@ llama_model_loader::llama_model_loader(
          void * set_tensor_data_ud,
          const std::string & fname,
          std::vector<std::string> & splits,
 +        bool ordered_parts,
          FILE * file,
          bool use_mmap,
          bool use_direct_io,
-@@ -593,11 +594,50 @@ llama_model_loader::llama_model_loader(
+@@ -596,11 +597,50 @@ llama_model_loader::llama_model_loader(
              n_bytes    += ggml_nbytes(cur);
              weights_map.emplace(tensor_name, llama_tensor_weight(files.back().get(), 0, metadata, cur));
          }

third_party/llama.cpp/patches/0004-Add-lanes-external-media-and-chat-grammar-support.patch

@@ -1,4 +1,4 @@
-From 0e34ce3acca749ba8e949d9b3b9e24dfc1537933 Mon Sep 17 00:00:00 2001
+From 29baeca7b5f406c2df3966e98c1e5e4b4ef10258 Mon Sep 17 00:00:00 2001
 From: Mesh-LLM CI <ci@mesh-llm.local>
 Date: Fri, 26 Jun 2026 13:47:52 +1000
 Subject: [PATCH 04/10] Add lanes, external media, and chat grammar support
@@ -8,11 +8,11 @@ Squashes Qwen3 MoE staged execution, KV cache type and generation signal ABI, ba
  common/stage-chat.cpp    |  230 ++++++++
  include/skippy-signals.h |   44 ++
  include/skippy.h         |  165 +++---
- src/llama-kv-cache.cpp   |  346 ------------
+ src/llama-kv-cache.cpp   |  361 +------------
  src/llama-kv-cache.h     |   19 -
  src/models/qwen3moe.cpp  |   18 +-
  src/skippy.cpp           | 1079 +++++++++++++++++++++++---------------
- 7 files changed, 1013 insertions(+), 888 deletions(-)
+ 7 files changed, 1019 insertions(+), 897 deletions(-)
  create mode 100644 include/skippy-signals.h
 
 diff --git a/common/stage-chat.cpp b/common/stage-chat.cpp
@@ -573,11 +573,40 @@ index 4dd985da..3ffdb879 100644
          const char * path,
          struct skippy_model_info ** out_info,
 diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
-index 400a9d29..024bac7a 100644
+index 906e9eac..4bdefc22 100644
 --- a/src/llama-kv-cache.cpp
 +++ b/src/llama-kv-cache.cpp
-@@ -1221,352 +1221,6 @@ ggml_type llama_kv_cache::type_v() const {
-     return layers[0].v->type;
+@@ -212,12 +212,10 @@ llama_kv_cache::llama_kv_cache(
+             n_embd_head_k_all = -1;
+         }
+
+-        if (!is_mla) {
+-            if (n_embd_head_v_all == 0) {
+-                n_embd_head_v_all = (int32_t) hparams.n_embd_head_v(il);
+-            } else if (n_embd_head_v_all > 0 && n_embd_head_v_all != (int32_t) hparams.n_embd_head_v(il)) {
+-                n_embd_head_v_all = -1;
+-            }
++        if (n_embd_head_v_all == 0) {
++            n_embd_head_v_all = (int32_t) hparams.n_embd_head_v(il);
++        } else if (n_embd_head_v_all > 0 && n_embd_head_v_all != (int32_t) hparams.n_embd_head_v(il)) {
++            n_embd_head_v_all = -1;
+         }
+
+         // [TAG_V_CACHE_VARIABLE]
+@@ -339,9 +337,8 @@ llama_kv_cache::llama_kv_cache(
+             ggml_is_quantized(type_k) &&
+             hparams.n_embd_head_k() % 64 == 0;
+
+-        // always create Hadamard rotation tensors for DeepSeek lightning indexers
+-        if ((model.arch == LLM_ARCH_DEEPSEEK32 || model.arch == LLM_ARCH_DEEPSEEK4) &&
+-                hparams.n_embd_head_k_full == hparams.indexer_head_size) {
++        // always create Hadamard rotation tensors for DeepSeek V3.2 DSA lightning indexer
++        if (model.arch == LLM_ARCH_DEEPSEEK32 && hparams.n_embd_head_k_full == hparams.indexer_head_size) {
+             attn_rot_k = true;
+         }
+
+@@ -1241,352 +1238,6 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
+     return layers[ikv].k;
  }
 
 -bool llama_kv_cache::stage_export_kv_page(
@@ -930,7 +959,7 @@ index 400a9d29..024bac7a 100644
      uint32_t result = 0;
 
 diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
-index 83cee5f9..a3a1ec96 100644
+index d9301c19..92394d4c 100644
 --- a/src/llama-kv-cache.h
 +++ b/src/llama-kv-cache.h
 @@ -13,8 +13,6 @@ struct llama_cparams;
@@ -942,9 +971,9 @@ index 83cee5f9..a3a1ec96 100644
  //
  // llama_kv_cache
  //
-@@ -163,23 +161,6 @@ public:
-     ggml_type type_k() const;
-     ggml_type type_v() const;
+@@ -166,23 +164,6 @@ public:
+     std::vector<uint32_t> get_layer_ids() const;
+     ggml_tensor * get_k_storage(int32_t il) const;
 
 -    bool stage_export_kv_page(
 -            int32_t layer_start,