llama: update llama.cpp to b9637 (#16609)

llama/compat/002-llama-cpp-ui-empty-assets.patch

@@ -0,0 +1,15 @@
+diff --git a/tools/ui/embed.cpp b/tools/ui/embed.cpp
+--- a/tools/ui/embed.cpp
++++ b/tools/ui/embed.cpp
+@@ -147,9 +147,9 @@ int main(int argc, char ** argv) {
+ 
+     const std::string out_cpp   = argv[1];
+     const std::string out_h     = argv[2];
+-    const std::string asset_dir = argv[3];
++    const std::string asset_dir = argc == 4 ? argv[3] : "";
+ 
+-    const bool        use_gzip = std::filesystem::exists(asset_dir + "/_gzip");
++    const bool        use_gzip = !asset_dir.empty() && std::filesystem::exists(asset_dir + "/_gzip");
+     const std::string in_dir   = use_gzip ? (asset_dir + "/_gzip") : asset_dir;
+ 
+     std::vector<asset_entry> assets;

llama/compat/README.md

@@ -23,11 +23,14 @@ intentionally skipped so a developer can iterate on a local llama.cpp tree.
 - `llama-ollama-compat-util.h`, `llama-ollama-compat-util.cpp` - helpers for
   KV edits, tensor renames, skip-prefix tracking, tensor load operations, and
   small tensor repacking primitives.
-- `llama-cpp-hooks.patch` - small additive call-site edits in llama.cpp files.
+- `001-llama-cpp-hooks.patch` - small additive call-site edits in llama.cpp files.
   It currently touches `src/llama-model-loader.cpp` and `tools/mtmd/clip.cpp`.
+- `002-llama-cpp-ui-empty-assets.patch` - lets the llama.cpp UI embed helper
+  generate an empty asset table when no UI assets are present.
 - `compat.cmake`, `apply-patch.cmake` - CMake glue and an idempotent applier
   (used by `llama/server/CMakeLists.txt`) that applies every `*.patch` under
-  this directory — the hooks patch plus each `models/` architecture patch.
+  this directory by numeric filename order — the hooks patch plus each
+  `models/` architecture patch.
 - `models/` - the sibling **new-architecture** layer: implementations of
   architectures llama.cpp doesn't support yet, each added via a small
   registration patch. (Those files *add* archs; the files above *translate*
@@ -116,7 +119,7 @@ cd /path/to/llama.cpp
 git diff -- \
     src/llama-model-loader.cpp \
     tools/mtmd/clip.cpp \
-    > /path/to/ollama/llama/compat/llama-cpp-hooks.patch
+    > /path/to/ollama/llama/compat/001-llama-cpp-hooks.patch
 ```
 
 ## Implementation Notes

llama/compat/apply-patch.cmake

@@ -3,10 +3,11 @@
 # Invocation (from a CMake PATCH_COMMAND):
 #   cmake -DPATCH_DIR=<dir of *.patch> -P apply-patch.cmake
 #
-# Every *.patch under PATCH_DIR is applied in the current working directory
-# (which ExternalProject / FetchContent sets to the fetched source's
-# SOURCE_DIR). A patch already applied — detected via `git apply --reverse
-# --check` — is skipped. This makes re-configuring and re-building safe.
+# Every *.patch under PATCH_DIR is applied in numeric filename order in the
+# current working directory (which ExternalProject / FetchContent sets to the
+# fetched source's SOURCE_DIR). A patch already applied — detected via
+# `git apply --reverse --check` — is skipped. This makes re-configuring and
+# re-building safe.
 
 if(NOT DEFINED PATCH_DIR)
     message(FATAL_ERROR "apply-patch.cmake: PATCH_DIR not set")
@@ -19,8 +20,17 @@ get_filename_component(_git_ceiling "${_patch_workdir}" DIRECTORY)
 set(_git_apply_env GIT_CEILING_DIRECTORIES=${_git_ceiling})
 
 file(GLOB_RECURSE _patches "${PATCH_DIR}/*.patch")
-list(SORT _patches)
+set(_patch_entries)
 foreach(PATCH_FILE IN LISTS _patches)
+    get_filename_component(_patch_name "${PATCH_FILE}" NAME)
+    list(APPEND _patch_entries "${_patch_name}|${PATCH_FILE}")
+endforeach()
+
+list(SORT _patch_entries)
+foreach(_patch_entry IN LISTS _patch_entries)
+    string(REGEX REPLACE "^[^|]*\\|" "" PATCH_FILE "${_patch_entry}")
+    file(RELATIVE_PATH _patch_rel "${PATCH_DIR}" "${PATCH_FILE}")
+
     # If the patch can be REVERSED cleanly, it's already applied. Skip.
     execute_process(
         COMMAND ${CMAKE_COMMAND} -E env ${_git_apply_env}
@@ -29,7 +39,7 @@ foreach(PATCH_FILE IN LISTS _patches)
         OUTPUT_QUIET ERROR_QUIET
     )
     if(_reverse_check EQUAL 0)
-        message(STATUS "llama/compat: patch already applied, skipping")
+        message(STATUS "llama/compat: ${_patch_rel} already applied, skipping")
         continue()
     endif()
 
@@ -41,10 +51,10 @@ foreach(PATCH_FILE IN LISTS _patches)
     )
     if(NOT _apply_result EQUAL 0)
         message(FATAL_ERROR
-            "llama/compat: failed to apply ${PATCH_FILE}\n"
+            "llama/compat: failed to apply ${_patch_rel}\n"
             "This usually means the pinned llama.cpp source has changed. "
             "Regenerate the patch against the pinned LLAMA_CPP_VERSION and retry.")
     endif()
 
-    message(STATUS "llama/compat: applied patch")
+    message(STATUS "llama/compat: applied ${_patch_rel}")
 endforeach()

llama/compat/compat.cmake

@@ -18,7 +18,7 @@
 # The compat layer consists of:
 #   1. Ollama-owned compat source files linked into the fetched llama.cpp
 #      targets from this directory.
-#   2. A small patch file that adds call-sites in llama.cpp loaders.
+#   2. A small ordered patch set that adds call-sites in llama.cpp loaders.
 
 set(_compat_dir ${CMAKE_CURRENT_LIST_DIR})
 
@@ -46,7 +46,7 @@ set(OLLAMA_LLAMA_CPP_COMPAT_DIR
 # Also export the individual paths in case callers want to do something
 # custom (e.g. emit a dependency on the patch so reconfigures re-apply).
 set(OLLAMA_LLAMA_CPP_COMPAT_PATCH_FILE
-    "${_compat_dir}/llama-cpp-hooks.patch"
+    "${_compat_dir}/001-llama-cpp-hooks.patch"
     CACHE INTERNAL "Path to the llama.cpp compat patch")
 
 set(OLLAMA_LLAMA_CPP_COMPAT_SOURCES

llama/compat/models/003-llama-cpp-laguna.patch

@@ -16,8 +16,8 @@ diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
 diff --git a/src/llama-arch.h b/src/llama-arch.h
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
-@@ -140,2 +140,3 @@ enum llm_arch {
-     LLM_ARCH_MELLUM,
+@@ -145,2 +145,3 @@ enum llm_arch {
+     LLM_ARCH_EAGLE3,
 +    LLM_ARCH_LAGUNA,
      LLM_ARCH_UNKNOWN,
 @@ -382,2 +383,3 @@ enum llm_tensor {

llama/compat/models/laguna.cpp

@@ -14,7 +14,7 @@ void llama_model_laguna::load_arch_hparams(llama_model_loader & ml) {
 
     ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
-    ml.get_key_or_arr("laguna.attention.layer_types", hparams.is_swa_impl, hparams.n_layer, false);
+    ml.get_key_or_arr("laguna.attention.layer_types", hparams.is_swa_impl, hparams.n_layer(), false);
 
     ml.get_key("laguna.rope.swa.dimension_count", hparams.n_rot_swa,                false);
     ml.get_key("laguna.rope.swa.freq_base",       hparams.rope_freq_base_train_swa, false);