Refresh llama.cpp upstream patch queue

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

@@ -1,7 +1,7 @@
-From 34be3b6d89d94ac218678e7f6ada201a1bcba4c1 Mon Sep 17 00:00:00 2001
+From 7e292e39190998abf873f50a8445d50b701cdb69 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 01/45] Add Skippy ABI and package writer foundation
+Subject: [PATCH 01/47] Add Skippy ABI and package writer foundation
 
 Squashes the initial Skippy ABI headers, model info entrypoints, single-stage runtime baseline, runtime-slice tensor filtering, activation-frame execution, GGUF slice writing, and package part composition.
 ---
@@ -19,10 +19,10 @@ Squashes the initial Skippy ABI headers, model info entrypoints, single-stage ru
  create mode 100644 src/skippy.cpp
 
 diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 81f23d7e..50ad6e57 100644
+index 3df1d82d..ad7652a0 100644
 --- a/CMakeLists.txt
 +++ b/CMakeLists.txt
-@@ -245,6 +245,7 @@ set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location o
+@@ -254,6 +254,7 @@ set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location o
 
  set(LLAMA_PUBLIC_HEADERS
      ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h
@@ -346,7 +346,7 @@ index 320784c3..5298c337 100644
              unicode-data.cpp
              unicode.cpp
 diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
-index c8ecb0a2..660dee46 100644
+index 6d1c8f4e..137dc9da 100644
 --- a/src/llama-graph.cpp
 +++ b/src/llama-graph.cpp
 @@ -21,6 +21,20 @@
@@ -370,7 +370,7 @@ index c8ecb0a2..660dee46 100644
  // dedup helpers
 
  static ggml_tensor * build_attn_inp_kq_mask(
-@@ -2173,28 +2187,32 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
+@@ -2184,28 +2198,32 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
      {
          auto & cur = inps[0];
 
@@ -444,7 +444,7 @@ index 7ed490ce..a7986a87 100644
  enum llm_graph_type {
      LLM_GRAPH_TYPE_DEFAULT,
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index 43447f57..7f23dbbc 100644
+index 510586e9..6254771a 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
 @@ -18,6 +18,16 @@ static const size_t kiB = 1024;

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

@@ -1,7 +1,7 @@
-From 0f00b168cc09677e844bf8822040aa3d71f304b5 Mon Sep 17 00:00:00 2001
+From 4a3154e28533e87a29222f219f882ef0ec28dce7 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 02/45] Add early staged model-family and chat support
+Subject: [PATCH 02/47] Add early staged model-family and chat support
 
 Squashes the first Qwen MoE staged-runtime support, native KV page basics, session reset, token EOG exposure, batched token verification, chat-template ABI, and the first wider model-family support pass.
 ---
@@ -32,7 +32,7 @@ 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 99688f53..3928ba17 100644
+index 799d2275..a3cfaec1 100644
 --- a/common/CMakeLists.txt
 +++ b/common/CMakeLists.txt
 @@ -98,6 +98,7 @@ add_library(${TARGET}
@@ -42,7 +42,7 @@ index 99688f53..3928ba17 100644
 +    stage-chat.cpp
      speculative.cpp
      speculative.h
-     trie.cpp
+     subproc.cpp
 diff --git a/common/stage-chat.cpp b/common/stage-chat.cpp
 new file mode 100644
 index 00000000..39d6331a
@@ -319,7 +319,7 @@ 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 660dee46..cca99a8f 100644
+index 137dc9da..bed11342 100644
 --- a/src/llama-graph.cpp
 +++ b/src/llama-graph.cpp
 @@ -22,6 +22,7 @@
@@ -455,7 +455,7 @@ index a7986a87..33d57061 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 e25464c5..d7c7c220 100644
+index 44cb1668..638c9664 100644
 --- a/src/llama-kv-cache.cpp
 +++ b/src/llama-kv-cache.cpp
 @@ -4,6 +4,7 @@
@@ -466,7 +466,7 @@ index e25464c5..d7c7c220 100644
 
  #include <algorithm>
  #include <cassert>
-@@ -1224,6 +1225,352 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
+@@ -1337,6 +1338,352 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
      return layers[ikv].k;
  }
 
@@ -820,7 +820,7 @@ index e25464c5..d7c7c220 100644
      uint32_t result = 0;
 
 diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
-index 531d99db..d9301c19 100644
+index d5a92f44..e421bd1c 100644
 --- a/src/llama-kv-cache.h
 +++ b/src/llama-kv-cache.h
 @@ -6,12 +6,14 @@
@@ -863,7 +863,7 @@ index 531d99db..d9301c19 100644
      // graph_build API
      //
 diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
-index 7f23dbbc..c94aca78 100644
+index 6254771a..be9233c2 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
 @@ -1082,13 +1082,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(
@@ -929,10 +929,10 @@ index 7f23dbbc..c94aca78 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 b100f601..07d3f4f3 100644
+index be0a0df5..4152ee61 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1398,13 +1398,13 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
+@@ -1403,13 +1403,13 @@ 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

@@ -1,7 +1,7 @@
-From ab8b97e2a7169c1f6081a27c49f6bc64959f79a1 Mon Sep 17 00:00:00 2001
+From b2d816ab8938faef60abdd2d1f95fe3cc2a1e9d1 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 03/45] Add staged sampling, checkpoints, and part loading
+Subject: [PATCH 03/47] Add staged sampling, checkpoints, and part loading
 
 Squashes sampling configuration, batched verification frames, recurrent/native session checkpoints, logit bias, session trim, verification handoff fixes, split-metadata stripping, thinking toggle support, and ordered GGUF part loading.
 ---
@@ -49,10 +49,10 @@ index 39d6331a..60d53fc4 100644
 
      for (size_t i = 0; i < message_count; ++i) {
 diff --git a/include/llama.h b/include/llama.h
-index 9fab6931..e54b64dc 100644
+index 3c6d22be..99efc134 100644
 --- a/include/llama.h
 +++ b/include/llama.h
-@@ -514,6 +514,13 @@ extern "C" {
+@@ -515,6 +515,13 @@ extern "C" {
                                   size_t    n_paths,
                struct llama_model_params    params);
 
@@ -245,7 +245,7 @@ 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 c94aca78..ecafbdf2 100644
+index be9233c2..3070a8c6 100644
 --- a/src/llama-model-loader.cpp
 +++ b/src/llama-model-loader.cpp
 @@ -532,6 +532,7 @@ llama_model_loader::llama_model_loader(
@@ -321,10 +321,10 @@ index 92ad6430..f73fa907 100644
          llama_load_mode load_mode,
          bool check_tensors,
 diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
-index caf7733a..7a3d105a 100644
+index 92ebc11b..03a6c34f 100644
 --- a/src/llama-quant.cpp
 +++ b/src/llama-quant.cpp
-@@ -885,7 +885,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
+@@ -893,7 +893,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
      const llama_model_kv_override * kv_overrides = params->kv_overrides;
      std::vector<std::string> splits = {};
      llama_model_loader ml(/*metadata*/ nullptr, /*set_tensor_data*/ nullptr, /*set_tensor_data_ud*/ nullptr,
@@ -334,10 +334,10 @@ index caf7733a..7a3d105a 100644
 
      auto mparams = llama_model_default_params();
 diff --git a/src/llama.cpp b/src/llama.cpp
-index 11ac9656..48c578ee 100644
+index d22e4c81..37ddf52f 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -299,9 +299,9 @@ static bool llama_prepare_model_devices(const llama_model_params & params, llama
+@@ -302,9 +302,9 @@ static bool llama_prepare_model_devices(const llama_model_params & params, llama
 
  // Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
  static std::pair<int, llama_model *> llama_model_load(struct gguf_context * metadata, llama_model_set_tensor_data_t set_tensor_data, void * set_tensor_data_ud,
@@ -349,15 +349,15 @@ index 11ac9656..48c578ee 100644
              params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
 
          ml.print_info();
-@@ -366,6 +366,7 @@ static struct llama_model * llama_model_load_from_file_impl(
+@@ -369,6 +369,7 @@ static struct llama_model * llama_model_load_from_file_impl(
          void * set_tensor_data_ud,
          const std::string & path_model,
          std::vector<std::string> & splits,
 +        bool ordered_parts,
          FILE * file,
          struct llama_model_params params) {
      {
-@@ -408,7 +409,7 @@ static struct llama_model * llama_model_load_from_file_impl(
+@@ -411,7 +412,7 @@ static struct llama_model * llama_model_load_from_file_impl(
          };
      }
 
@@ -366,7 +366,7 @@ index 11ac9656..48c578ee 100644
      GGML_ASSERT(status <= 0);
      if (status < 0) {
          if (status == -1) {
-@@ -436,7 +437,7 @@ struct llama_model * llama_model_init_from_user(
+@@ -439,7 +440,7 @@ struct llama_model * llama_model_init_from_user(
      std::vector<std::string> splits = {};
      params.load_mode = LLAMA_LOAD_MODE_NONE;
      params.use_extra_bufts = false;
@@ -375,7 +375,7 @@ index 11ac9656..48c578ee 100644
  }
  // deprecated
  struct llama_model * llama_load_model_from_file(
-@@ -449,7 +450,7 @@ struct llama_model * llama_model_load_from_file(
+@@ -452,7 +453,7 @@ struct llama_model * llama_model_load_from_file(
          const char * path_model,
          struct llama_model_params params) {
      std::vector<std::string> splits = {};
@@ -384,7 +384,7 @@ index 11ac9656..48c578ee 100644
  }
 
  struct llama_model * llama_model_load_from_splits(
-@@ -465,7 +466,27 @@ struct llama_model * llama_model_load_from_splits(
+@@ -468,7 +469,27 @@ struct llama_model * llama_model_load_from_splits(
      for (size_t i = 0; i < n_paths; ++i) {
          splits.push_back(paths[i]);
      }
@@ -413,7 +413,7 @@ index 11ac9656..48c578ee 100644
  }
 
  struct llama_model * llama_model_load_from_file_ptr(FILE * file, struct llama_model_params params) {
-@@ -475,7 +496,7 @@ struct llama_model * llama_model_load_from_file_ptr(FILE * file, struct llama_mo
+@@ -478,7 +499,7 @@ struct llama_model * llama_model_load_from_file_ptr(FILE * file, struct llama_mo
      }
      std::string path_model;
      std::vector<std::string> splits = {};
@@ -422,7 +422,7 @@ index 11ac9656..48c578ee 100644
  }
 
  void llama_model_save_to_file(const struct llama_model * model, const char * path_model) {
-@@ -600,4 +621,3 @@ const char * llama_print_system_info(void) {
+@@ -603,4 +624,3 @@ const char * llama_print_system_info(void) {
 
      return s.c_str();
  }

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

@@ -1,7 +1,7 @@
-From 59cdfb7bf5cf5e112f229ebec9a4e0fa3d954547 Mon Sep 17 00:00:00 2001
+From aec3d7b74786f7434166c15be069d882cdb39ea8 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 04/45] Add lanes, external media, and chat grammar support
+Subject: [PATCH 04/47] Add lanes, external media, and chat grammar support
 
 Squashes Qwen3 MoE staged execution, KV cache type and generation signal ABI, backend device selection, external media prefill, shared execution lanes, batch/flash-attention config, tool-aware chat templates, and chat grammar sampling.
 ---
@@ -573,7 +573,7 @@ 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 d7c7c220..f74bb495 100644
+index 638c9664..41962096 100644
 --- a/src/llama-kv-cache.cpp
 +++ b/src/llama-kv-cache.cpp
 @@ -196,12 +196,10 @@ llama_kv_cache::llama_kv_cache(
@@ -593,7 +593,7 @@ index d7c7c220..f74bb495 100644
          }
 
          // [TAG_V_CACHE_VARIABLE]
-@@ -1225,352 +1223,6 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
+@@ -1338,352 +1336,6 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
      return layers[ikv].k;
  }
 
@@ -947,7 +947,7 @@ index d7c7c220..f74bb495 100644
      uint32_t result = 0;
 
 diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
-index d9301c19..92394d4c 100644
+index e421bd1c..6c92186d 100644
 --- a/src/llama-kv-cache.h
 +++ b/src/llama-kv-cache.h
 @@ -13,8 +13,6 @@ struct llama_cparams;

third_party/llama.cpp/patches/0005-Add-resident-prefix-cache-and-session-refinements.patch

@@ -1,12 +1,12 @@
-From 8d4cf568f56593d1fa8223ad12ca2965d680fee6 Mon Sep 17 00:00:00 2001
+From df9fea9e4a0d7bd91ce128ca3820ed9618aef466 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 05/45] Add resident prefix cache and session refinements
+Subject: [PATCH 05/47] Add resident prefix cache and session refinements
 
 Squashes resident prefix cache support, exact-prefix state restoration, borrowed prefix sessions, lazy generation signals, single-token allocation avoidance, server-like thread defaults, prefix cache borrowing, lane preservation, and recurrent-state remapping.
 ---
  include/skippy.h       | 119 +++++-
- src/llama-kv-cache.cpp | 369 +++++++++++++++++++
+ src/llama-kv-cache.cpp | 369 ++++++++++++++++++
  src/llama-kv-cache.h   |  21 ++
  src/skippy.cpp         | 821 ++++++++++++++++++++++++++++++++++++++++-
  4 files changed, 1320 insertions(+), 10 deletions(-)
@@ -196,10 +196,10 @@ index 3ffdb879..deb5151e 100644
          struct skippy_model * model,
          const char * text,
 diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp
-index f74bb495..41a4e0de 100644
+index 41962096..942a1a4a 100644
 --- a/src/llama-kv-cache.cpp
 +++ b/src/llama-kv-cache.cpp
-@@ -1223,6 +1223,375 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
+@@ -1336,6 +1336,375 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
      return layers[ikv].k;
  }
 
@@ -576,7 +576,7 @@ index f74bb495..41a4e0de 100644
      uint32_t result = 0;
 
 diff --git a/src/llama-kv-cache.h b/src/llama-kv-cache.h
-index 92394d4c..be51c3c7 100644
+index 6c92186d..beaf005e 100644
 --- a/src/llama-kv-cache.h
 +++ b/src/llama-kv-cache.h
 @@ -13,6 +13,8 @@ struct llama_cparams;

third_party/llama.cpp/patches/0006-Expand-staged-execution-across-dense-and-recurrent-f.patch

@@ -1,7 +1,7 @@
-From 7432fb6267d078b2e6c79df8cc301517913cf6c0 Mon Sep 17 00:00:00 2001
+From 95a491c6dd9fd92ba8b7b3406b632c8821cbc5a6 Mon Sep 17 00:00:00 2001
 From: Mesh-LLM CI <ci@mesh-llm.local>
 Date: Fri, 26 Jun 2026 13:47:53 +1000
-Subject: [PATCH 06/45] Expand staged execution across dense and recurrent
+Subject: [PATCH 06/47] Expand staged execution across dense and recurrent
  families
 
 Squashes the dense and decoder family support waves through LFM2, Mamba, Jamba, RWKV6, Qwen2MoE, and native sequence id exposure.
@@ -51,10 +51,10 @@ index deb5151e..84169398 100644
          const struct skippy_session * session);
 
 diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
-index cca99a8f..f2618710 100644
+index bed11342..ad8f2fc6 100644
 --- a/src/llama-graph.cpp
 +++ b/src/llama-graph.cpp
-@@ -2269,10 +2269,15 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
+@@ -2280,10 +2280,15 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
 
      res->t_inp_embd = cur;