diff --git a/docs/build-profiling.md b/docs/build-profiling.md
index 3ec8c7eee3..33251f21b6 100644
--- a/docs/build-profiling.md
+++ b/docs/build-profiling.md
@@ -13,7 +13,11 @@ brew install clang-build-analyzer
 
 Linux:
 ```console
-TODO:
+git clone https://github.com/aras-p/ClangBuildAnalyzer.git
+cd ClangBuildAnalyzer
+cmake -B build -DCMAKE_BUILD_TYPE=Release
+cmake --build build -j$(nproc)
+sudo cp build/ClangBuildAnalyzer /usr/local/bin/
 ```
 
 ### Usage
@@ -21,7 +25,46 @@ TODO:
 $ ./scripts/build-profile-baseline.sh
 ```
 
+### Linux (Ubuntu 24.04)
 
-### Baseline
-The initial run of this produces the following report:
-* [baseline-report.txt](./profiling-reports/baseline-report.txt)
+Environment:
+- Clang:     18.1.3 (Ubuntu clang version 18.1.3 (1ubuntu1))
+- libstdc++: GCC 13.3.0 (Ubuntu 13.3.0-6ubuntu2~24.04.1)
+- Target:    x86_64-pc-linux-gnu
+
+```console
++--------------------+-----+------------+------------+------------+
+| Build              | TUs | Frontend   | Backend    | Total      |
++--------------------+-----+------------+------------+------------+
+| Minimal, master    | 249 |   468.2 s  |   270.3 s  |   738.5 s  |
+| Minimal, with PCH  | 253 |   177.1 s  |   265.8 s  |   442.9 s  |
+| Full,    master    | 396 |   811.0 s  |   692.2 s  | 1,503.2 s  |
+| Full,    with PCH  | 404 |   418.3 s  |   663.9 s  | 1,082.2 s  |
++--------------------+-----+------------+------------+------------+
+
+PCH = precompiled header.
+Full = includes building Server, Tools, and Tests.
+```
+Note that the number of translation units (TUs) increases when using precompiled
+headers — each PCH target adds one extra TU for the precompilation step itself.
+
+### Mac (Apple M3)
+
+Environment:
+- Clang:  Apple clang version 17.0.0 (clang-1700.3.19.1)
+- libc++: ships with Apple clang 17.0.0 (Xcode toolchain)
+- Target: arm64-apple-macosx15.6
+
+```console
++--------------------+-----+------------+------------+------------+
+| Build              | TUs | Frontend   | Backend    | Total      |
++--------------------+-----+------------+------------+------------+
+| Minimal, master    | 256 |   154.5 s  |    94.8 s  |   249.3 s  |
+| Minimal, with PCH  | 261 |    65.9 s  |    90.0 s  |   155.9 s  |
+| Full,    master    | 407 |   265.7 s  |   209.7 s  |   475.4 s  |
+| Full,    with PCH  | 414 |   154.6 s  |   197.5 s  |   352.1 s  |
++--------------------+-----+------------+------------+------------+
+
+PCH = precompiled header.
+Full = includes building Server, Tools, and Tests.
+```
diff --git a/docs/profiling-reports/baseline-report.txt b/docs/profiling-reports/baseline-report.txt
deleted file mode 100644
index 7025545da9..0000000000
--- a/docs/profiling-reports/baseline-report.txt
+++ /dev/null
@@ -1,239 +0,0 @@
-Analyzing build trace from '/Users/danbev/work/ai/llama.cpp/build-profile-baseline/clang_analysis.bin'...
-**** Time summary:
-Compilation (257 times):
-  Parsing (frontend):          152.3 s
-  Codegen & opts (backend):     94.8 s
-
-**** Files that took longest to parse (compiler frontend):
-  2089 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat.cpp.o
-  2054 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/vendor/cpp-httplib/CMakeFiles/cpp-httplib.dir/httplib.cpp.o
-  1901 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/peg-parser.cpp.o
-  1857 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/json.cpp.o
-  1813 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/arg.cpp.o
-  1662 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat-diff-analyzer.cpp.o
-  1498 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/download.cpp.o
-  1477 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/jinja/value.cpp.o
-  1426 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-model.cpp.o
-  1356 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/json-schema-to-grammar.cpp.o
-
-**** Files that took longest to codegen (compiler backend):
-  6436 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat.cpp.o
-  6163 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/vendor/cpp-httplib/CMakeFiles/cpp-httplib.dir/httplib.cpp.o
-  5876 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/arg.cpp.o
-  3023 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat-diff-analyzer.cpp.o
-  2807 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/peg-parser.cpp.o
-  2798 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/json-schema-to-grammar.cpp.o
-  2784 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o
-  2511 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-model.cpp.o
-  2475 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/unicode.cpp.o
-  2439 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-kv-cache.cpp.o
-
-**** Templates that took longest to instantiate:
-  1085 ms: std::unordered_map<std::bitset<256>, std::vector<int>> (170 times, avg 6 ms)
-  1084 ms: std::vector<bool>::vector (173 times, avg 6 ms)
-   987 ms: std::unordered_map<int, int> (165 times, avg 5 ms)
-   930 ms: std::unordered_map<std::string, std::string> (174 times, avg 5 ms)
-   889 ms: std::__hash_table<std::__hash_value_type<std::bitset<256>, std::vect... (170 times, avg 5 ms)
-   830 ms: std::unordered_map<std::string, llama_adapter_lora_weight> (170 times, avg 4 ms)
-   806 ms: std::vector<bool>::__construct_at_end<std::__bit_iterator<std::vecto... (170 times, avg 4 ms)
-   793 ms: std::unordered_map<std::string, llama_model_kv_override> (156 times, avg 5 ms)
-   771 ms: std::basic_regex<char>::basic_regex (11 times, avg 70 ms)
-   769 ms: std::basic_regex<char>::__init<const char *> (11 times, avg 69 ms)
-   766 ms: std::basic_regex<char>::__parse<const char *> (11 times, avg 69 ms)
-   762 ms: std::__hash_table<std::__hash_value_type<int, int>, std::__unordered... (165 times, avg 4 ms)
-   743 ms: std::__hash_table<std::__hash_value_type<std::string, std::string>, ... (174 times, avg 4 ms)
-   741 ms: std::unordered_set<llama_adapter_lora *> (168 times, avg 4 ms)
-   682 ms: std::basic_regex<char>::basic_regex<std::char_traits<char>, std::all... (6 times, avg 113 ms)
-   681 ms: std::basic_regex<char>::__init<std::__wrap_iter<const char *>> (6 times, avg 113 ms)
-   677 ms: std::basic_regex<char>::__parse<std::__wrap_iter<const char *>> (6 times, avg 112 ms)
-   667 ms: std::basic_regex<char>::__parse_ecma_exp<const char *> (11 times, avg 60 ms)
-   663 ms: std::map<int, llama_sampler *> (170 times, avg 3 ms)
-   657 ms: std::basic_regex<char>::__parse_alternative<const char *> (11 times, avg 59 ms)
-   657 ms: std::basic_regex<char>::__parse_term<const char *> (11 times, avg 59 ms)
-   646 ms: std::basic_string<wchar_t> (245 times, avg 2 ms)
-   638 ms: std::__copy<std::__bit_iterator<std::vector<bool>, true>, std::__bit... (170 times, avg 3 ms)
-   637 ms: std::__hash_table<std::__hash_value_type<std::string, llama_adapter_... (170 times, avg 3 ms)
-   629 ms: std::__hash_table<std::__hash_value_type<std::string, llama_model_kv... (156 times, avg 4 ms)
-   592 ms: std::__hash_table<llama_adapter_lora *, std::hash<llama_adapter_lora... (168 times, avg 3 ms)
-   568 ms: std::__variant_detail::__copy_constructor<std::__variant_detail::__t... (8 times, avg 71 ms)
-   567 ms: std::__variant_detail::__ctor<std::__variant_detail::__traits<common... (8 times, avg 70 ms)
-   564 ms: std::map<unsigned int, llama_model_qwen4exp::llm_graph_input_qsa *> (152 times, avg 3 ms)
-   552 ms: std::basic_regex<char>::__parse_ecma_exp<std::__wrap_iter<const char... (6 times, avg 92 ms)
-
-**** Template sets that took longest to instantiate:
-  6644 ms: std::vector<$> (4706 times, avg 1 ms)
-  5031 ms: std::unordered_map<$> (905 times, avg 5 ms)
-  4783 ms: std::allocator_traits<$> (5799 times, avg 0 ms)
-  4625 ms: std::__hash_table<$> (1084 times, avg 4 ms)
-  3736 ms: std::map<$> (1105 times, avg 3 ms)
-  2878 ms: std::__tree<$> (1369 times, avg 2 ms)
-  2698 ms: std::unique_ptr<$> (2818 times, avg 0 ms)
-  2259 ms: std::function<$>::function<$> (466 times, avg 4 ms)
-  2234 ms: std::basic_string<$> (1010 times, avg 2 ms)
-  2213 ms: std::__function::__value_func<$>::__value_func<$> (466 times, avg 4 ms)
-  2159 ms: std::vector<$>::vector (487 times, avg 4 ms)
-  1775 ms: std::basic_string<$>::basic_string (997 times, avg 1 ms)
-  1724 ms: std::vector<$>::__construct_at_end<$> (492 times, avg 3 ms)
-  1716 ms: std::vector<$>::emplace_back<$> (468 times, avg 3 ms)
-  1613 ms: std::basic_string<$>::__init (979 times, avg 1 ms)
-  1565 ms: std::basic_regex<$>::__init<$> (18 times, avg 86 ms)
-  1558 ms: std::basic_regex<$>::__parse<$> (18 times, avg 86 ms)
-  1531 ms: std::unordered_map<$>::unordered_map (554 times, avg 2 ms)
-  1404 ms: std::vector<$>::~vector (670 times, avg 2 ms)
-  1393 ms: std::vector<$>::__swap_out_circular_buffer (627 times, avg 2 ms)
-  1315 ms: std::basic_regex<$>::__parse_ecma_exp<$> (18 times, avg 73 ms)
-  1294 ms: std::basic_regex<$>::__parse_alternative<$> (18 times, avg 71 ms)
-  1291 ms: std::basic_regex<$>::__parse_term<$> (18 times, avg 71 ms)
-  1215 ms: std::vector<$>::__emplace_back_slow_path<$> (365 times, avg 3 ms)
-  1199 ms: std::vector<$>::push_back (344 times, avg 3 ms)
-  1159 ms: std::pair<$> (1184 times, avg 0 ms)
-  1133 ms: std::__hash_table<$>::~__hash_table (546 times, avg 2 ms)
-  1106 ms: std::__hash_table<$>::__deallocate_node (543 times, avg 2 ms)
-  1090 ms: std::__copy<$> (377 times, avg 2 ms)
-   962 ms: std::vector<$>::__init_with_size<$> (264 times, avg 3 ms)
-
-**** Functions that took longest to compile:
-  1032 ms: common_params_parser_init(common_params&, llama_example, void (*)(in... (/Users/danbev/work/ai/llama.cpp/common/arg.cpp)
-   156 ms: common_chat_try_specialized_template(common_chat_template const&, st... (/Users/danbev/work/ai/llama.cpp/common/chat.cpp)
-   151 ms: llama_vocab::impl::load(llama_model_loader&, LLM_KV const&) (/Users/danbev/work/ai/llama.cpp/src/llama-vocab.cpp)
-   132 ms: llm_chat_apply_template(llm_chat_template, std::__1::vector<llama_ch... (/Users/danbev/work/ai/llama.cpp/src/llama-chat.cpp)
-   113 ms: llama_model_quantize_impl(std::__1::basic_string<char, std::__1::cha... (/Users/danbev/work/ai/llama.cpp/src/llama-quant.cpp)
-   107 ms: ggml_compute_forward_add_non_quantized (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
-   100 ms: ggml_compute_forward_div (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
-    99 ms: ggml_compute_forward_sub (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
-    94 ms: void set_input_kq_mask_impl<float>(args_set_input_kq_mask const&, fl... (/Users/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
-    92 ms: llama_model_base::load_tensors(llama_model_loader&) (/Users/danbev/work/ai/llama.cpp/src/llama-model.cpp)
-    92 ms: unicode_regex_split_custom(std::__1::basic_string<char, std::__1::ch... (/Users/danbev/work/ai/llama.cpp/src/unicode.cpp)
-    89 ms: void set_input_kq_mask_impl<unsigned short>(args_set_input_kq_mask c... (/Users/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
-    88 ms: ggml_compute_forward_mul (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
-    87 ms: ggml_compute_forward_dup (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/ops.cpp)
-    80 ms: ggml_compute_forward_out_prod (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/ops.cpp)
-    80 ms: common_models_handler_apply(common_models_handler&, common_params&, ... (/Users/danbev/work/ai/llama.cpp/common/arg.cpp)
-    79 ms: unicode_regex_split(std::__1::basic_string<char, std::__1::char_trai... (/Users/danbev/work/ai/llama.cpp/src/unicode.cpp)
-    75 ms: dsv4_build_comp_plans(std::__1::vector<llama_ubatch, std::__1::alloc... (/Users/danbev/work/ai/llama.cpp/src/llama-kv-cache-dsv4.cpp)
-    73 ms: common_schema_converter::visit(common_json const&, std::__1::basic_s... (/Users/danbev/work/ai/llama.cpp/common/json-schema-to-grammar.cpp)
-    71 ms: common_params_fit_impl(char const*, llama_model_params*, llama_conte... (/Users/danbev/work/ai/llama.cpp/common/fit.cpp)
-    70 ms: hf_cache::get_repo_files(std::__1::basic_string<char, std::__1::char... (/Users/danbev/work/ai/llama.cpp/common/hf-cache.cpp)
-    68 ms: llama_model_saver::add_kv_from_model() (/Users/danbev/work/ai/llama.cpp/src/llama-model-saver.cpp)
-    62 ms: ggml_compute_forward_flash_attn_ext (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/ops.cpp)
-    59 ms: common_sampler_init(llama_model const*, common_params_sampling&) (/Users/danbev/work/ai/llama.cpp/common/sampling.cpp)
-    57 ms: bool std::__1::basic_regex<char, std::__1::regex_traits<char> >::__m... (/Users/danbev/work/ai/llama.cpp/common/download.cpp)
-    56 ms: common_chat_peg_builder::python_style_tool_calls(common_json const&,... (/Users/danbev/work/ai/llama.cpp/common/chat-peg-parser.cpp)
-    52 ms: llama_meta_device_get_split_state(ggml_tensor const*, void*) (/Users/danbev/work/ai/llama.cpp/src/llama-model.cpp)
-    51 ms: common_schema_converter::_pattern_to_rule(std::__1::basic_string<cha... (/Users/danbev/work/ai/llama.cpp/common/json-schema-to-grammar.cpp)
-    50 ms: llama_model_loader::llama_model_loader(gguf_context*, void (*)(ggml_... (/Users/danbev/work/ai/llama.cpp/src/llama-model-loader.cpp)
-    50 ms: quantize_q6_K (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-quants.c)
-
-**** Function sets that took longest to compile / optimize:
-   286 ms: void (anonymous namespace)::tinyBLAS<$>::gemm<$>(long long, long lon... (36 times, avg 7 ms)
-   156 ms: common_chat_try_specialized_template(common_chat_template const&, st... (1 times, avg 156 ms)
-   153 ms: bool std::__1::basic_regex<$>::__match_at_start_posix_subs<$>(char c... (14 times, avg 10 ms)
-   150 ms: std::__1::__bracket_expression<$>::__exec(std::__1::__state<$>&) const (15 times, avg 10 ms)
-   146 ms: ggml::cpu::repack::tensor_traits<$>::forward_mul_mat(ggml_compute_pa... (16 times, avg 9 ms)
-   142 ms: bool std::__1::basic_regex<$>::__match_at_start_posix_nosubs<$>(char... (14 times, avg 10 ms)
-   132 ms: llm_chat_apply_template(llm_chat_template, std::__1::vector<$> const... (1 times, avg 132 ms)
-   116 ms: std::__1::__wrap_iter<$> std::__1::vector<$>::__insert_with_size[abi... (16 times, avg 7 ms)
-   113 ms: llama_model_quantize_impl(std::__1::basic_string<$> const&, std::__1... (1 times, avg 113 ms)
-   104 ms: ggml::cpu::repack::tensor_traits<$>::forward_mul_mat_id(ggml_compute... (16 times, avg 6 ms)
-    94 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, float*... (1 times, avg 94 ms)
-    92 ms: unicode_regex_split_custom(std::__1::basic_string<$> const&, std::__... (1 times, avg 92 ms)
-    89 ms: std::__1::vector<$>::__append(unsigned long) (69 times, avg 1 ms)
-    89 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, unsign... (1 times, avg 89 ms)
-    83 ms: void std::__1::__hash_table<$>::__do_rehash<$>(unsigned long) (56 times, avg 1 ms)
-    79 ms: unicode_regex_split(std::__1::basic_string<$> const&, std::__1::vect... (1 times, avg 79 ms)
-    75 ms: void std::__1::vector<$>::__assign_with_size[abi:ne200100]<$>(std::_... (18 times, avg 4 ms)
-    75 ms: dsv4_build_comp_plans(std::__1::vector<$> const&, unsigned int, bool... (1 times, avg 75 ms)
-    73 ms: common_schema_converter::visit(common_json const&, std::__1::basic_s... (1 times, avg 73 ms)
-    70 ms: hf_cache::get_repo_files(std::__1::basic_string<$> const&, std::__1:... (1 times, avg 70 ms)
-    70 ms: std::__1::__bracket_expression<$>::__add_range[abi:ne200100](std::__... (15 times, avg 4 ms)
-    67 ms: bool std::__1::basic_regex<$>::__match_at_start_ecma<$>(char const*,... (14 times, avg 4 ms)
-    64 ms: std::__1::deque<$>::__add_back_capacity() (16 times, avg 4 ms)
-    60 ms: std::__1::__loop<$>::__exec(std::__1::__state<$>&) const (15 times, avg 4 ms)
-    58 ms: void std::__1::__tree_balance_after_insert[abi:ne200100]<$>(std::__1... (44 times, avg 1 ms)
-    57 ms: std::__1::deque<$>::__add_front_capacity() (15 times, avg 3 ms)
-    51 ms: char const* std::__1::basic_regex<$>::__parse_expression_term<$>(cha... (11 times, avg 4 ms)
-    51 ms: common_schema_converter::_pattern_to_rule(std::__1::basic_string<cha... (1 times, avg 51 ms)
-    50 ms: llama_model_loader::llama_model_loader(gguf_context*, void (*)(ggml_... (1 times, avg 50 ms)
-    49 ms: common_chat_peg_builder::build_json_tools_nested_keys(common_json co... (1 times, avg 49 ms)
-
-**** Expensive headers:
-70838 ms: /Users/danbev/work/ai/llama.cpp/src/models/models.h (included 152 times, avg 466 ms), included via:
-  152x: <direct include>
-
-63362 ms: /Users/danbev/work/ai/llama.cpp/src/llama-model.h (included 168 times, avg 377 ms), included via:
-  151x: models.h 
-  17x: <direct include>
-
-47408 ms: /Users/danbev/work/ai/llama.cpp/src/llama-arch.h (included 173 times, avg 274 ms), included via:
-  151x: models.h llama-model.h 
-  3x: llama-model.h 
-  2x: llama-memory.h llama-graph.h 
-  2x: llama-model-loader.h 
-  2x: llama-context.h llama-graph.h 
-  1x: llama-graph.h 
-  ...
-
-36275 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/string (included 245 times, avg 148 ms), included via:
-  150x: models.h llama-model.h llama-arch.h 
-  13x: ggml-impl.h vector locale __locale 
-  3x: lexer.h utils.h 
-  3x: llama-impl.h 
-  3x: common.h list functional vector locale __locale 
-  3x: arg.h common.h list functional vector locale __locale 
-  ...
-
-18679 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/vector (included 241 times, avg 77 ms), included via:
-  151x: models.h llama-model.h llama-arch.h set functional 
-  13x: ggml-impl.h 
-  3x: common.h list functional 
-  3x: arg.h common.h list functional 
-  3x: llama-impl.h 
-  3x: lexer.h utils.h sstream istream ostream format queue deque functional 
-  ...
-
-17411 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/set (included 207 times, avg 84 ms), included via:
-  151x: models.h llama-model.h llama-arch.h 
-  7x: common.h 
-  4x: runtime.h value.h 
-  3x: llama-model.h llama-arch.h 
-  3x: arg.h common.h 
-  3x: <direct include>
-  ...
-
-13431 ms: /Users/danbev/work/ai/llama.cpp/src/llama-graph.h (included 170 times, avg 79 ms), included via:
-  151x: models.h llama-model.h 
-  4x: llama-model.h 
-  2x: llama-memory.h 
-  2x: llama-context.h 
-  1x: <direct include>
-  1x: llama-kv-cache-dsv4.h llama-kv-cache.h 
-  ...
-
-9203 ms: /Users/danbev/work/ai/llama.cpp/common/common.h (included 22 times, avg 418 ms), included via:
-  9x: <direct include>
-  3x: arg.h 
-  2x: sampling.h 
-  2x: chat-auto-parser-helpers.h chat-auto-parser.h chat.h 
-  1x: reasoning-budget.h 
-  1x: console.h 
-  ...
-
-7501 ms: /Users/danbev/work/ai/llama.cpp/src/llama-batch.h (included 170 times, avg 44 ms), included via:
-  151x: models.h llama-model.h llama-graph.h 
-  4x: llama-model.h llama-graph.h 
-  2x: llama-context.h llama-graph.h 
-  1x: llama-kv-cache-dsv4.h llama-kv-cache.h 
-  1x: llama-memory-hybrid.h 
-  1x: llama-kv-cache-iswa.h llama-kv-cache.h 
-  ...
-
-7320 ms: /Users/danbev/work/ai/llama.cpp/ggml/src/ggml-impl.h (included 30 times, avg 244 ms), included via:
-  21x: <direct include>
-  2x: amx.h ggml-cpu-impl.h 
-  2x: traits.h ggml-cpu-impl.h 
-  1x: vec.h 
-  1x: binary-ops.h common.h traits.h ggml-cpu-impl.h 
-  1x: unary-ops.h common.h traits.h ggml-cpu-impl.h 
-  ...
-
-  done in 0.0s.
diff --git a/docs/profiling-reports/profile-report-baseline-mac.txt b/docs/profiling-reports/profile-report-baseline-mac.txt
new file mode 100644
index 0000000000..313ce6a4ab
--- /dev/null
+++ b/docs/profiling-reports/profile-report-baseline-mac.txt
@@ -0,0 +1,239 @@
+Analyzing build trace from '/Users/danbev/work/ai/llama.cpp/build-profile-baseline/clang_analysis.bin'...
+**** Time summary:
+Compilation (256 times):
+  Parsing (frontend):          154.5 s
+  Codegen & opts (backend):     94.8 s
+
+**** Files that took longest to parse (compiler frontend):
+  2125 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/vendor/cpp-httplib/CMakeFiles/cpp-httplib.dir/httplib.cpp.o
+  2032 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat.cpp.o
+  1879 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/arg.cpp.o
+  1579 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-model.cpp.o
+  1568 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat-diff-analyzer.cpp.o
+  1523 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/json.cpp.o
+  1493 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/peg-parser.cpp.o
+  1474 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/download.cpp.o
+  1404 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/jinja/value.cpp.o
+  1214 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/preset.cpp.o
+
+**** Files that took longest to codegen (compiler backend):
+  6256 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/vendor/cpp-httplib/CMakeFiles/cpp-httplib.dir/httplib.cpp.o
+  5932 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat.cpp.o
+  5454 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/arg.cpp.o
+  3039 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/peg-parser.cpp.o
+  2881 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat-diff-analyzer.cpp.o
+  2860 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/ops.cpp.o
+  2720 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-model.cpp.o
+  2575 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-kv-cache.cpp.o
+  2530 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/unicode.cpp.o
+  2208 ms: /Users/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/json-schema-to-grammar.cpp.o
+
+**** Templates that took longest to instantiate:
+  1192 ms: std::unordered_map<std::bitset<256>, std::vector<int>> (170 times, avg 7 ms)
+  1064 ms: std::vector<bool>::vector (173 times, avg 6 ms)
+  1032 ms: std::unordered_map<std::string, std::string> (174 times, avg 5 ms)
+  1001 ms: std::unordered_map<int, int> (165 times, avg 6 ms)
+   958 ms: std::__hash_table<std::__hash_value_type<std::bitset<256>, std::vect... (170 times, avg 5 ms)
+   881 ms: std::unordered_map<std::string, llama_adapter_lora_weight> (170 times, avg 5 ms)
+   823 ms: std::unordered_map<std::string, llama_model_kv_override> (156 times, avg 5 ms)
+   815 ms: std::__hash_table<std::__hash_value_type<std::string, std::string>, ... (174 times, avg 4 ms)
+   789 ms: std::vector<bool>::__construct_at_end<std::__bit_iterator<std::vecto... (170 times, avg 4 ms)
+   786 ms: std::unordered_set<llama_adapter_lora *> (168 times, avg 4 ms)
+   784 ms: std::__hash_table<std::__hash_value_type<int, int>, std::__unordered... (165 times, avg 4 ms)
+   772 ms: std::basic_regex<char>::basic_regex (11 times, avg 70 ms)
+   771 ms: std::basic_regex<char>::__init<const char *> (11 times, avg 70 ms)
+   769 ms: std::basic_regex<char>::__parse<const char *> (11 times, avg 69 ms)
+   682 ms: std::basic_regex<char>::__parse_ecma_exp<const char *> (11 times, avg 62 ms)
+   675 ms: std::basic_regex<char>::__parse_alternative<const char *> (11 times, avg 61 ms)
+   675 ms: std::basic_regex<char>::__parse_term<const char *> (11 times, avg 61 ms)
+   670 ms: std::map<int, llama_sampler *> (170 times, avg 3 ms)
+   664 ms: std::__hash_table<std::__hash_value_type<std::string, llama_adapter_... (170 times, avg 3 ms)
+   656 ms: std::map<unsigned int, llama_model_qwen4exp::llm_graph_input_qsa *> (152 times, avg 4 ms)
+   643 ms: std::__hash_table<std::__hash_value_type<std::string, llama_model_kv... (156 times, avg 4 ms)
+   642 ms: std::__hash_table<llama_adapter_lora *, std::hash<llama_adapter_lora... (168 times, avg 3 ms)
+   640 ms: std::basic_string<wchar_t> (245 times, avg 2 ms)
+   622 ms: std::__copy<std::__bit_iterator<std::vector<bool>, true>, std::__bit... (170 times, avg 3 ms)
+   598 ms: std::basic_regex<char>::basic_regex<std::char_traits<char>, std::all... (6 times, avg 99 ms)
+   597 ms: std::basic_regex<char>::__init<std::__wrap_iter<const char *>> (6 times, avg 99 ms)
+   593 ms: std::basic_regex<char>::__parse<std::__wrap_iter<const char *>> (6 times, avg 98 ms)
+   541 ms: std::basic_string<char> (245 times, avg 2 ms)
+   532 ms: std::basic_string<char16_t> (245 times, avg 2 ms)
+   529 ms: std::unordered_map<std::string, std::string>::unordered_map (174 times, avg 3 ms)
+
+**** Template sets that took longest to instantiate:
+  6796 ms: std::vector<$> (4713 times, avg 1 ms)
+  5339 ms: std::unordered_map<$> (905 times, avg 5 ms)
+  5133 ms: std::allocator_traits<$> (6086 times, avg 0 ms)
+  4875 ms: std::__hash_table<$> (1084 times, avg 4 ms)
+  3940 ms: std::map<$> (1105 times, avg 3 ms)
+  3052 ms: std::__tree<$> (1369 times, avg 2 ms)
+  2865 ms: std::unique_ptr<$> (3024 times, avg 0 ms)
+  2326 ms: std::function<$>::function<$> (466 times, avg 4 ms)
+  2279 ms: std::__function::__value_func<$>::__value_func<$> (466 times, avg 4 ms)
+  2267 ms: std::basic_string<$> (1010 times, avg 2 ms)
+  2126 ms: std::vector<$>::vector (496 times, avg 4 ms)
+  1809 ms: std::basic_string<$>::basic_string (999 times, avg 1 ms)
+  1757 ms: std::vector<$>::emplace_back<$> (462 times, avg 3 ms)
+  1711 ms: std::vector<$>::__construct_at_end<$> (487 times, avg 3 ms)
+  1638 ms: std::basic_string<$>::__init (979 times, avg 1 ms)
+  1584 ms: std::unordered_map<$>::unordered_map (555 times, avg 2 ms)
+  1503 ms: std::basic_regex<$>::__init<$> (18 times, avg 83 ms)
+  1497 ms: std::basic_regex<$>::__parse<$> (18 times, avg 83 ms)
+  1395 ms: std::vector<$>::~vector (671 times, avg 2 ms)
+  1382 ms: std::vector<$>::__swap_out_circular_buffer (626 times, avg 2 ms)
+  1255 ms: std::basic_regex<$>::__parse_ecma_exp<$> (18 times, avg 69 ms)
+  1240 ms: std::basic_regex<$>::__parse_alternative<$> (18 times, avg 68 ms)
+  1238 ms: std::basic_regex<$>::__parse_term<$> (18 times, avg 68 ms)
+  1225 ms: std::vector<$>::__emplace_back_slow_path<$> (363 times, avg 3 ms)
+  1189 ms: std::vector<$>::push_back (338 times, avg 3 ms)
+  1152 ms: std::__hash_table<$>::~__hash_table (547 times, avg 2 ms)
+  1141 ms: std::pair<$> (1191 times, avg 0 ms)
+  1125 ms: std::__hash_table<$>::__deallocate_node (543 times, avg 2 ms)
+  1061 ms: std::__copy<$> (381 times, avg 2 ms)
+   970 ms: std::vector<$>::__init_with_size<$> (263 times, avg 3 ms)
+
+**** Functions that took longest to compile:
+   937 ms: common_params_parser_init(common_params&, llama_example, void (*)(in... (/Users/danbev/work/ai/llama.cpp/common/arg.cpp)
+   157 ms: common_chat_try_specialized_template(common_chat_template const&, st... (/Users/danbev/work/ai/llama.cpp/common/chat.cpp)
+   138 ms: llm_chat_apply_template(llm_chat_template, std::__1::vector<llama_ch... (/Users/danbev/work/ai/llama.cpp/src/llama-chat.cpp)
+   127 ms: ggml_compute_forward_sub (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+   123 ms: llama_vocab::impl::load(llama_model_loader&, LLM_KV const&) (/Users/danbev/work/ai/llama.cpp/src/llama-vocab.cpp)
+   109 ms: ggml_compute_forward_mul (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+   104 ms: void set_input_kq_mask_impl<float>(args_set_input_kq_mask const&, fl... (/Users/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
+    99 ms: llama_model_base::load_tensors(llama_model_loader&) (/Users/danbev/work/ai/llama.cpp/src/llama-model.cpp)
+    99 ms: ggml_compute_forward_div (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+    94 ms: ggml_compute_forward_add_non_quantized (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+    90 ms: void set_input_kq_mask_impl<unsigned short>(args_set_input_kq_mask c... (/Users/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
+    88 ms: llama_model_saver::add_kv_from_model() (/Users/danbev/work/ai/llama.cpp/src/llama-model-saver.cpp)
+    84 ms: unicode_regex_split(std::__1::basic_string<char, std::__1::char_trai... (/Users/danbev/work/ai/llama.cpp/src/unicode.cpp)
+    83 ms: common_models_handler_apply(common_models_handler&, common_params&, ... (/Users/danbev/work/ai/llama.cpp/common/arg.cpp)
+    83 ms: ggml_compute_forward_dup (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/ops.cpp)
+    80 ms: unicode_regex_split_custom(std::__1::basic_string<char, std::__1::ch... (/Users/danbev/work/ai/llama.cpp/src/unicode.cpp)
+    78 ms: ggml_compute_forward_out_prod (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/ops.cpp)
+    77 ms: common_params_fit_impl(char const*, llama_model_params*, llama_conte... (/Users/danbev/work/ai/llama.cpp/common/fit.cpp)
+    76 ms: llama_model_quantize_impl(std::__1::basic_string<char, std::__1::cha... (/Users/danbev/work/ai/llama.cpp/src/llama-quant.cpp)
+    72 ms: llama_model_loader::llama_model_loader(gguf_context*, void (*)(ggml_... (/Users/danbev/work/ai/llama.cpp/src/llama-model-loader.cpp)
+    71 ms: dsv4_build_comp_plans(std::__1::vector<llama_ubatch, std::__1::alloc... (/Users/danbev/work/ai/llama.cpp/src/llama-kv-cache-dsv4.cpp)
+    70 ms: common_schema_converter::visit(common_json const&, std::__1::basic_s... (/Users/danbev/work/ai/llama.cpp/common/json-schema-to-grammar.cpp)
+    61 ms: void std::__1::__introsort<std::__1::_ClassicAlgPolicy, common_chat_... (/Users/danbev/work/ai/llama.cpp/common/chat-peg-parser.cpp)
+    59 ms: ggml_compute_forward_flash_attn_ext (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/ops.cpp)
+    58 ms: common_memory_breakdown_print(llama_context const*) (/Users/danbev/work/ai/llama.cpp/common/fit.cpp)
+    55 ms: common_sampler_init(llama_model const*, common_params_sampling&) (/Users/danbev/work/ai/llama.cpp/common/sampling.cpp)
+    54 ms: hf_cache::get_repo_files(std::__1::basic_string<char, std::__1::char... (/Users/danbev/work/ai/llama.cpp/common/hf-cache.cpp)
+    53 ms: llama_meta_device_get_split_state(ggml_tensor const*, void*) (/Users/danbev/work/ai/llama.cpp/src/llama-model.cpp)
+    50 ms: llama_model_minimax_m3::graph::graph(llama_model const&, llm_graph_p... (/Users/danbev/work/ai/llama.cpp/src/models/minimax-m3.cpp)
+    47 ms: quantize_q6_K (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-quants.c)
+
+**** Function sets that took longest to compile / optimize:
+   250 ms: void (anonymous namespace)::tinyBLAS<$>::gemm<$>(long long, long lon... (36 times, avg 6 ms)
+   168 ms: bool std::__1::basic_regex<$>::__match_at_start_posix_nosubs<$>(char... (14 times, avg 12 ms)
+   164 ms: ggml::cpu::repack::tensor_traits<$>::forward_mul_mat(ggml_compute_pa... (16 times, avg 10 ms)
+   157 ms: common_chat_try_specialized_template(common_chat_template const&, st... (1 times, avg 157 ms)
+   157 ms: std::__1::__bracket_expression<$>::__exec(std::__1::__state<$>&) const (15 times, avg 10 ms)
+   138 ms: llm_chat_apply_template(llm_chat_template, std::__1::vector<$> const... (1 times, avg 138 ms)
+   123 ms: std::__1::__wrap_iter<$> std::__1::vector<$>::__insert_with_size[abi... (16 times, avg 7 ms)
+   111 ms: ggml::cpu::repack::tensor_traits<$>::forward_mul_mat_id(ggml_compute... (16 times, avg 6 ms)
+   104 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, float*... (1 times, avg 104 ms)
+    99 ms: bool std::__1::basic_regex<$>::__match_at_start_posix_subs<$>(char c... (14 times, avg 7 ms)
+    98 ms: bool std::__1::basic_regex<$>::__match_at_start_ecma<$>(char const*,... (14 times, avg 7 ms)
+    90 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, unsign... (1 times, avg 90 ms)
+    87 ms: std::__1::vector<$>::__append(unsigned long) (77 times, avg 1 ms)
+    84 ms: unicode_regex_split(std::__1::basic_string<$> const&, std::__1::vect... (1 times, avg 84 ms)
+    80 ms: unicode_regex_split_custom(std::__1::basic_string<$> const&, std::__... (1 times, avg 80 ms)
+    80 ms: void std::__1::__hash_table<$>::__do_rehash<$>(unsigned long) (56 times, avg 1 ms)
+    76 ms: llama_model_quantize_impl(std::__1::basic_string<$> const&, std::__1... (1 times, avg 76 ms)
+    72 ms: llama_model_loader::llama_model_loader(gguf_context*, void (*)(ggml_... (1 times, avg 72 ms)
+    71 ms: dsv4_build_comp_plans(std::__1::vector<$> const&, unsigned int, bool... (1 times, avg 71 ms)
+    70 ms: common_schema_converter::visit(common_json const&, std::__1::basic_s... (1 times, avg 70 ms)
+    67 ms: void std::__1::vector<$>::__assign_with_size[abi:ne200100]<$>(std::_... (18 times, avg 3 ms)
+    64 ms: void std::__1::__tree_balance_after_insert[abi:ne200100]<$>(std::__1... (44 times, avg 1 ms)
+    63 ms: std::__1::pair<$> std::__1::__hash_table<$>::__emplace_unique_key_ar... (15 times, avg 4 ms)
+    61 ms: void std::__1::__introsort<$>(std::__1::pair<$>*, std::__1::pair<$>*... (1 times, avg 61 ms)
+    60 ms: std::__1::deque<$>::__add_back_capacity() (16 times, avg 3 ms)
+    60 ms: std::__1::__bracket_expression<$>::__add_range[abi:ne200100](std::__... (15 times, avg 4 ms)
+    59 ms: char const* std::__1::basic_regex<$>::__parse_expression_term<$>(cha... (11 times, avg 5 ms)
+    57 ms: void (anonymous namespace)::tinyBLAS_Q0_ARM<$>::gemm<$>(long long, l... (18 times, avg 3 ms)
+    54 ms: hf_cache::get_repo_files(std::__1::basic_string<$> const&, std::__1:... (1 times, avg 54 ms)
+    50 ms: std::__1::deque<$>::__add_front_capacity() (15 times, avg 3 ms)
+
+**** Expensive headers:
+73736 ms: /Users/danbev/work/ai/llama.cpp/src/models/models.h (included 152 times, avg 485 ms), included via:
+  152x: <direct include>
+
+65981 ms: /Users/danbev/work/ai/llama.cpp/src/llama-model.h (included 168 times, avg 392 ms), included via:
+  151x: models.h 
+  17x: <direct include>
+
+49234 ms: /Users/danbev/work/ai/llama.cpp/src/llama-arch.h (included 173 times, avg 284 ms), included via:
+  151x: models.h llama-model.h 
+  3x: llama-model.h 
+  2x: llama-memory.h llama-graph.h 
+  2x: llama-model-loader.h 
+  2x: llama-context.h llama-graph.h 
+  1x: llama-model-saver.h 
+  ...
+
+37732 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/string (included 245 times, avg 154 ms), included via:
+  150x: models.h llama-model.h llama-arch.h 
+  13x: ggml-impl.h vector locale __locale 
+  3x: lexer.h utils.h 
+  3x: llama-impl.h 
+  3x: common.h list functional vector locale __locale 
+  3x: arg.h common.h list functional vector locale __locale 
+  ...
+
+19209 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/vector (included 241 times, avg 79 ms), included via:
+  151x: models.h llama-model.h llama-arch.h set functional 
+  13x: ggml-impl.h 
+  3x: common.h list functional 
+  3x: arg.h common.h list functional 
+  3x: llama-impl.h 
+  3x: lexer.h utils.h sstream istream ostream format queue deque functional 
+  ...
+
+17946 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/set (included 207 times, avg 86 ms), included via:
+  151x: models.h llama-model.h llama-arch.h 
+  7x: common.h 
+  4x: runtime.h value.h 
+  3x: llama-model.h llama-arch.h 
+  3x: arg.h common.h 
+  3x: <direct include>
+  ...
+
+14187 ms: /Users/danbev/work/ai/llama.cpp/src/llama-graph.h (included 170 times, avg 83 ms), included via:
+  151x: models.h llama-model.h 
+  4x: llama-model.h 
+  2x: llama-memory.h 
+  2x: llama-context.h 
+  1x: <direct include>
+  1x: llama-kv-cache-dsv4.h llama-kv-cache.h 
+  ...
+
+8229 ms: /Users/danbev/work/ai/llama.cpp/common/common.h (included 22 times, avg 374 ms), included via:
+  9x: <direct include>
+  3x: arg.h 
+  2x: sampling.h 
+  2x: chat-auto-parser-helpers.h chat-auto-parser.h chat.h 
+  1x: speculative.h 
+  1x: chat.h 
+  ...
+
+7954 ms: /Users/danbev/work/ai/llama.cpp/src/llama-batch.h (included 170 times, avg 46 ms), included via:
+  151x: models.h llama-model.h llama-graph.h 
+  4x: llama-model.h llama-graph.h 
+  2x: llama-context.h llama-graph.h 
+  1x: llama-kv-cache-dsv4.h llama-kv-cache.h 
+  1x: llama-kv-cache-dsa.h llama-kv-cache.h 
+  1x: llama-memory-hybrid-iswa.h 
+  ...
+
+7582 ms: /Users/danbev/work/ai/llama.cpp/ggml/src/ggml-impl.h (included 30 times, avg 252 ms), included via:
+  21x: <direct include>
+  2x: amx.h ggml-cpu-impl.h 
+  2x: traits.h ggml-cpu-impl.h 
+  1x: vec.h 
+  1x: unary-ops.h common.h traits.h ggml-cpu-impl.h 
+  1x: binary-ops.h common.h traits.h ggml-cpu-impl.h 
+  ...
+
+  done in 0.0s.
diff --git a/docs/profiling-reports/profile-report-baseline-ubuntu.txt b/docs/profiling-reports/profile-report-baseline-ubuntu.txt
new file mode 100644
index 0000000000..c807d2aab1
--- /dev/null
+++ b/docs/profiling-reports/profile-report-baseline-ubuntu.txt
@@ -0,0 +1,245 @@
+Analyzing build trace from '/home/danbev/work/ai/llama.cpp/build-profile-baseline/clang_analysis.bin'...
+**** Time summary:
+Compilation (249 times):
+  Parsing (frontend):          468.2 s
+  Codegen & opts (backend):    270.3 s
+
+**** Files that took longest to parse (compiler frontend):
+  6555 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat.cpp.o
+  6298 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/json.cpp.o
+  5897 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/arg.cpp.o
+  5076 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat-diff-analyzer.cpp.o
+  5037 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/preset.cpp.o
+  4819 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-model.cpp.o
+  4723 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/peg-parser.cpp.o
+  4223 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/download.cpp.o
+  4204 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/unicode-data.cpp.o
+  4104 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/json-schema-to-grammar.cpp.o
+
+**** Files that took longest to codegen (compiler backend):
+ 15027 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/arg.cpp.o
+ 14526 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/vendor/cpp-httplib/CMakeFiles/cpp-httplib.dir/httplib.cpp.o
+ 11187 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat.cpp.o
+  9741 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-kv-cache.cpp.o
+  9428 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/unicode.cpp.o
+  9045 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-model.cpp.o
+  8939 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/json-schema-to-grammar.cpp.o
+  8031 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/src/CMakeFiles/llama.dir/llama-sampler.cpp.o
+  7951 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/peg-parser.cpp.o
+  7631 ms: /home/danbev/work/ai/llama.cpp/build-profile-baseline/common/CMakeFiles/llama-common.dir/chat-diff-analyzer.cpp.o
+
+**** Templates that took longest to instantiate:
+  7849 ms: std::basic_regex<char>::_M_compile (15 times, avg 523 ms)
+  7549 ms: std::__detail::_Compiler<std::regex_traits<char>>::_Compiler (15 times, avg 503 ms)
+  6484 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_disjunction (15 times, avg 432 ms)
+  6438 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_alternative (15 times, avg 429 ms)
+  6327 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_term (15 times, avg 421 ms)
+  5344 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_atom (15 times, avg 356 ms)
+  5173 ms: std::basic_regex<char>::basic_regex (9 times, avg 574 ms)
+  3240 ms: std::unordered_map<int, int> (199 times, avg 16 ms)
+  3214 ms: std::unique_ptr<llm_graph_input_attn_kv> (170 times, avg 18 ms)
+  2940 ms: std::unique_ptr<llm_graph_input_attn_k_dsa> (170 times, avg 17 ms)
+  2830 ms: std::unique_ptr<llama_vocab::impl> (172 times, avg 16 ms)
+  2816 ms: std::unique_ptr<ggml_backend_buffer, ggml_backend_buffer_deleter> (171 times, avg 16 ms)
+  2795 ms: std::unique_ptr<llama_model::impl> (168 times, avg 16 ms)
+  2766 ms: std::__uniq_ptr_data<llm_graph_input_attn_kv, std::default_delete<ll... (170 times, avg 16 ms)
+  2738 ms: std::__uniq_ptr_impl<llm_graph_input_attn_kv, std::default_delete<ll... (170 times, avg 16 ms)
+  2737 ms: std::_Vector_base<std::unique_ptr<ggml_backend_buffer, ggml_backend_... (161 times, avg 17 ms)
+  2727 ms: std::unique_ptr<llm_graph_input_dsv4_raw> (170 times, avg 16 ms)
+  2725 ms: std::unique_ptr<llm_graph_input_attn_kv_iswa> (170 times, avg 16 ms)
+  2715 ms: std::unique_ptr<llm_graph_input_attn_k> (170 times, avg 15 ms)
+  2696 ms: std::unique_ptr<llm_graph_input_rs> (170 times, avg 15 ms)
+  2687 ms: std::unique_ptr<llama_file::impl> (159 times, avg 16 ms)
+  2681 ms: std::basic_regex<char>::basic_regex<std::char_traits<char>, std::all... (6 times, avg 446 ms)
+  2681 ms: std::unique_ptr<ggml_context, ggml_context_deleter> (173 times, avg 15 ms)
+  2675 ms: std::unique_ptr<llama_mlock::impl> (159 times, avg 16 ms)
+  2667 ms: std::unique_ptr<llama_mmap::impl> (159 times, avg 16 ms)
+  2550 ms: std::unique_ptr<llm_graph_context> (147 times, avg 17 ms)
+  2512 ms: std::unique_ptr<gguf_context, gguf_context_deleter> (158 times, avg 15 ms)
+  2436 ms: std::__uniq_ptr_data<ggml_backend_buffer, ggml_backend_buffer_deleter> (171 times, avg 14 ms)
+  2410 ms: std::__uniq_ptr_impl<ggml_backend_buffer, ggml_backend_buffer_deleter> (171 times, avg 14 ms)
+  2396 ms: std::__uniq_ptr_data<llama_vocab::impl, std::default_delete<llama_vo... (172 times, avg 13 ms)
+
+**** Template sets that took longest to instantiate:
+ 50618 ms: std::unique_ptr<$> (3044 times, avg 16 ms)
+ 42906 ms: std::__uniq_ptr_data<$> (3044 times, avg 14 ms)
+ 42431 ms: std::__uniq_ptr_impl<$> (3044 times, avg 13 ms)
+ 29501 ms: std::__and_<$> (24651 times, avg 1 ms)
+ 20580 ms: std::tuple<$> (3167 times, avg 6 ms)
+ 14925 ms: std::_Hashtable<$> (1666 times, avg 8 ms)
+ 13782 ms: std::vector<$> (4748 times, avg 2 ms)
+ 12460 ms: std::unordered_map<$> (938 times, avg 13 ms)
+ 10417 ms: std::_TupleConstraints<$>::__is_implicitly_constructible<$> (6459 times, avg 1 ms)
+  8269 ms: std::basic_regex<$>::_M_compile (16 times, avg 516 ms)
+  8014 ms: std::__or_<$> (10045 times, avg 0 ms)
+  7945 ms: std::__detail::_Compiler<$>::_Compiler (16 times, avg 496 ms)
+  7285 ms: std::_Vector_base<$> (4503 times, avg 1 ms)
+  6821 ms: std::__detail::_Compiler<$>::_M_disjunction (16 times, avg 426 ms)
+  6773 ms: std::__detail::_Compiler<$>::_M_alternative (16 times, avg 423 ms)
+  6656 ms: std::__detail::_Compiler<$>::_M_term (16 times, avg 416 ms)
+  6226 ms: std::_Tuple_impl<$> (3159 times, avg 1 ms)
+  6042 ms: std::map<$> (1120 times, avg 5 ms)
+  5669 ms: std::__detail::_Compiler<$>::_M_atom (16 times, avg 354 ms)
+  5659 ms: std::_Destroy<$> (1574 times, avg 3 ms)
+  5652 ms: std::vector<$>::~vector (1318 times, avg 4 ms)
+  5516 ms: std::tuple<$>::__assignable<$> (5521 times, avg 0 ms)
+  5470 ms: std::pair<$> (2005 times, avg 2 ms)
+  5203 ms: std::_Hashtable<$>::~_Hashtable (574 times, avg 9 ms)
+  5192 ms: std::unique_ptr<$>::~unique_ptr (1824 times, avg 2 ms)
+  5173 ms: std::basic_regex<$>::basic_regex (9 times, avg 574 ms)
+  4718 ms: std::allocator<$> (6850 times, avg 0 ms)
+  4297 ms: std::function<$>::function<$> (729 times, avg 5 ms)
+  4217 ms: std::vector<$>::emplace_back<$> (491 times, avg 8 ms)
+  4172 ms: std::is_move_constructible<$> (3399 times, avg 1 ms)
+
+**** Functions that took longest to compile:
+  2694 ms: common_params_parser_init(common_params&, llama_example, void (*)(in... (/home/danbev/work/ai/llama.cpp/common/arg.cpp)
+   458 ms: llama_model_quantize_impl(std::__cxx11::basic_string<char, std::char... (/home/danbev/work/ai/llama.cpp/src/llama-quant.cpp)
+   449 ms: llama_vocab::impl::load(llama_model_loader&, LLM_KV const&) (/home/danbev/work/ai/llama.cpp/src/llama-vocab.cpp)
+   414 ms: llama_model_base::load_tensors(llama_model_loader&) (/home/danbev/work/ai/llama.cpp/src/llama-model.cpp)
+   402 ms: void set_input_kq_mask_impl<unsigned short, true>(args_set_input_kq_... (/home/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
+   365 ms: void set_input_kq_mask_impl<float, false>(args_set_input_kq_mask con... (/home/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
+   351 ms: void set_input_kq_mask_impl<float, true>(args_set_input_kq_mask cons... (/home/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
+   350 ms: unicode_regex_split_custom(std::__cxx11::basic_string<char, std::cha... (/home/danbev/work/ai/llama.cpp/src/unicode.cpp)
+   350 ms: llm_chat_apply_template(llm_chat_template, std::vector<llama_chat_me... (/home/danbev/work/ai/llama.cpp/src/llama-chat.cpp)
+   349 ms: void set_input_kq_mask_impl<unsigned short, false>(args_set_input_kq... (/home/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
+   281 ms: common_params_fit_impl(char const*, llama_model_params*, llama_conte... (/home/danbev/work/ai/llama.cpp/common/fit.cpp)
+   271 ms: unicode_regex_split(std::__cxx11::basic_string<char, std::char_trait... (/home/danbev/work/ai/llama.cpp/src/unicode.cpp)
+   260 ms: ggml_compute_forward_mul (/home/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+   256 ms: llama_model_saver::add_kv_from_model() (/home/danbev/work/ai/llama.cpp/src/llama-model-saver.cpp)
+   256 ms: common_memory_breakdown_print(llama_context const*) (/home/danbev/work/ai/llama.cpp/common/fit.cpp)
+   251 ms: hf_cache::get_repo_files(std::__cxx11::basic_string<char, std::char_... (/home/danbev/work/ai/llama.cpp/common/hf-cache.cpp)
+   244 ms: llama_model_loader::llama_model_loader(gguf_context*, void (*)(ggml_... (/home/danbev/work/ai/llama.cpp/src/llama-model-loader.cpp)
+   213 ms: common_chat_try_specialized_template(common_chat_template const&, st... (/home/danbev/work/ai/llama.cpp/common/chat.cpp)
+   209 ms: llama_kv_cache::llama_kv_cache(llama_model const&, llama_hparams con... (/home/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
+   209 ms: llama_meta_device_get_split_state(ggml_tensor const*, void*) (/home/danbev/work/ai/llama.cpp/src/llama-model.cpp)
+   201 ms: ggml_compute_forward_div (/home/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+   185 ms: dsv4_build_comp_plans(std::vector<llama_ubatch, std::allocator<llama... (/home/danbev/work/ai/llama.cpp/src/llama-kv-cache-dsv4.cpp)
+   181 ms: common_download_file_single_online(std::__cxx11::basic_string<char, ... (/home/danbev/work/ai/llama.cpp/common/download.cpp)
+   169 ms: common_sampler_init(llama_model const*, common_params_sampling&) (/home/danbev/work/ai/llama.cpp/common/sampling.cpp)
+   165 ms: llm_tokenizer_plamo2::build(llama_vocab const&) (/home/danbev/work/ai/llama.cpp/src/llama-vocab.cpp)
+   164 ms: ggml_compute_forward_sub (/home/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+   162 ms: ggml_compute_forward_dup (/home/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/ops.cpp)
+   158 ms: common_docker_resolve_model(std::__cxx11::basic_string<char, std::ch... (/home/danbev/work/ai/llama.cpp/common/download.cpp)
+   158 ms: ggml_compute_forward_add_non_quantized (/home/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+   152 ms: llama_grammar_parser::parse_sequence(char const*, std::__cxx11::basi... (/home/danbev/work/ai/llama.cpp/src/llama-grammar.cpp)
+
+**** Function sets that took longest to compile / optimize:
+  1205 ms: bool std::__detail::_Compiler<$>::_M_expression_term<$>(std::__detai... (60 times, avg 20 ms)
+   933 ms: std::_Function_handler<$>::_M_manager(std::_Any_data&, std::_Any_dat... (645 times, avg 1 ms)
+   751 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, unsign... (2 times, avg 375 ms)
+   716 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, float*) (2 times, avg 358 ms)
+   715 ms: std::__detail::_BracketMatcher<$>::_BracketMatcher(std::__detail::_B... (60 times, avg 11 ms)
+   632 ms: void std::__detail::_Compiler<$>::_M_insert_bracket_matcher<$>(bool) (60 times, avg 10 ms)
+   559 ms: std::__detail::_Compiler<$>::_M_quantifier() (15 times, avg 37 ms)
+   487 ms: std::_Rb_tree<$>::_M_erase(std::_Rb_tree_node<$>*) (282 times, avg 1 ms)
+   464 ms: void std::__detail::_Compiler<$>::_M_insert_character_class_matcher<... (60 times, avg 7 ms)
+   458 ms: llama_model_quantize_impl(std::__cxx11::basic_string<$> const&, std:... (1 times, avg 458 ms)
+   456 ms: std::vector<$>::~vector() (193 times, avg 2 ms)
+   433 ms: std::__detail::_BracketMatcher<$>::_M_add_character_class(std::__cxx... (60 times, avg 7 ms)
+   430 ms: std::__detail::_BracketMatcher<$>::_M_make_range(char, char) (56 times, avg 7 ms)
+   428 ms: std::_Function_handler<$>::_M_invoke(std::_Any_data const&, jinja::f... (72 times, avg 5 ms)
+   424 ms: std::__detail::_BracketMatcher<$>::_M_add_equivalence_class(std::__c... (60 times, avg 7 ms)
+   423 ms: std::_Function_handler<$>::_M_invoke(std::_Any_data const&) (72 times, avg 5 ms)
+   360 ms: std::__detail::_Executor<$>::_M_dfs(std::__detail::_Executor<$>::_Ma... (32 times, avg 11 ms)
+   358 ms: void (anonymous namespace)::tinyBLAS<$>::gemm<$>(long, long, long) (27 times, avg 13 ms)
+   350 ms: unicode_regex_split_custom(std::__cxx11::basic_string<$> const&, std... (1 times, avg 350 ms)
+   350 ms: llm_chat_apply_template(llm_chat_template, std::vector<$> const&, st... (1 times, avg 350 ms)
+   335 ms: std::__detail::_StateSeq<$>::_M_clone() (15 times, avg 22 ms)
+   329 ms: std::__detail::_BracketMatcher<$>::_M_ready() (56 times, avg 5 ms)
+   320 ms: std::vector<$>::_M_default_append(unsigned long) (58 times, avg 5 ms)
+   313 ms: std::__detail::_BracketMatcher<$>::_M_add_collate_element(std::__cxx... (60 times, avg 5 ms)
+   308 ms: std::__detail::_Executor<$>::_M_lookahead(long) (32 times, avg 9 ms)
+   292 ms: std::__detail::_Compiler<$>::_M_atom() (15 times, avg 19 ms)
+   290 ms: std::_Function_handler<$>::_M_invoke(std::_Any_data const&, common_c... (17 times, avg 17 ms)
+   271 ms: unicode_regex_split(std::__cxx11::basic_string<$> const&, std::vecto... (1 times, avg 271 ms)
+   251 ms: hf_cache::get_repo_files(std::__cxx11::basic_string<$> const&, std::... (1 times, avg 251 ms)
+   245 ms: std::_Function_handler<$>::_M_invoke(std::_Any_data const&, char&&) (200 times, avg 1 ms)
+
+**** Expensive headers:
+225979 ms: /home/danbev/work/ai/llama.cpp/src/models/models.h (included 152 times, avg 1486 ms), included via:
+  151x: <direct include>
+  1x: llama-model.h 
+
+213012 ms: /home/danbev/work/ai/llama.cpp/src/llama-model.h (included 168 times, avg 1267 ms), included via:
+  151x: models.h 
+  2x: <direct include>
+  2x: llama-impl.h 
+  1x: llama-kv-cache-msa.h 
+  1x: llama-memory-recurrent.h 
+  1x: llama-context.h 
+  ...
+
+185681 ms: /home/danbev/work/ai/llama.cpp/include/llama.h (included 201 times, avg 923 ms), included via:
+  150x: models.h llama-model.h 
+  3x: arg.h common.h llama-cpp.h 
+  2x: sampling.h 
+  2x: llama-hparams.h 
+  2x: chat-auto-parser-helpers.h chat-auto-parser.h chat.h common.h llama-cpp.h 
+  2x: common.h llama-cpp.h 
+  ...
+
+68432 ms: /home/danbev/work/ai/llama.cpp/src/llama-graph.h (included 170 times, avg 402 ms), included via:
+  150x: models.h llama-model.h llama.h 
+  2x: llama-impl.h llama-model.h 
+  1x: <direct include>
+  1x: llama-memory.h llama.h 
+  1x: llama-model.h 
+  1x: llama-context.h llama.h 
+  ...
+
+44735 ms: /home/danbev/work/ai/llama.cpp/src/llama-arch.h (included 162 times, avg 276 ms), included via:
+  150x: models.h llama-model.h llama.h 
+  2x: llama-impl.h llama-model.h 
+  1x: models.h llama-model.h 
+  1x: llama-graph.h 
+  1x: <direct include>
+  1x: llama-model-saver.h gguf.h 
+  ...
+
+25323 ms: /home/danbev/work/ai/llama.cpp/src/llama-hparams.h (included 173 times, avg 146 ms), included via:
+  150x: models.h llama-model.h llama.h llama-graph.h 
+  2x: <direct include>
+  2x: llama-impl.h llama-model.h llama-graph.h 
+  1x: llama-memory-hybrid.h llama-batch.h llama-graph.h 
+  1x: llama-memory.h llama.h llama-graph.h 
+  1x: llama-memory-recurrent.h llama-batch.h llama-graph.h 
+  ...
+
+18443 ms: /usr/include/c++/13/unordered_map (included 199 times, avg 92 ms), included via:
+  150x: models.h llama-model.h llama.h llama-graph.h 
+  4x: runtime.h value.h cmath 
+  3x: arg.h common.h functional 
+  2x: chat-auto-parser-helpers.h chat-auto-parser.h chat.h common.h peg-parser.h 
+  2x: value.h cmath 
+  2x: <direct include>
+  ...
+
+18100 ms: /home/danbev/work/ai/llama.cpp/common/common.h (included 22 times, avg 822 ms), included via:
+  3x: arg.h 
+  2x: chat-auto-parser-helpers.h chat-auto-parser.h chat.h 
+  2x: <direct include>
+  2x: sampling.h llama.h 
+  1x: chat.h 
+  1x: chat-auto-parser.h chat.h 
+  ...
+
+14291 ms: /usr/include/c++/13/cmath (included 352 times, avg 40 ms), included via:
+  150x: models.h llama-model.h llama.h llama-graph.h llama-hparams.h 
+  133x: models.h llama-model.h 
+  6x: <direct include>
+  4x: runtime.h value.h 
+  2x: value.h 
+  2x: llama-hparams.h 
+  ...
+
+13663 ms: /home/danbev/work/ai/llama.cpp/ggml/include/ggml.h (included 222 times, avg 61 ms), included via:
+  150x: models.h llama-model.h llama.h 
+  4x: <direct include>
+  3x: arg.h common.h llama-cpp.h llama.h 
+  3x: ggml-backend.h 
+  2x: llama-impl.h 
+  2x: llama-hparams.h llama.h 
+  ...
+
+  done in 0.1s.
diff --git a/docs/profiling-reports/profile-report-full-baseline-ubuntu.txt b/docs/profiling-reports/profile-report-full-baseline-ubuntu.txt
new file mode 100644
index 0000000000..15ab2f544a
--- /dev/null
+++ b/docs/profiling-reports/profile-report-full-baseline-ubuntu.txt
@@ -0,0 +1,238 @@
+Analyzing build trace from '/home/danbev/work/ai/llama.cpp/build-profile-full/clang_analysis.bin'...
+**** Time summary:
+Compilation (396 times):
+  Parsing (frontend):          811.0 s
+  Codegen & opts (backend):    692.2 s
+
+**** Files that took longest to parse (compiler frontend):
+  9266 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/server-context.dir/server-context.cpp.o
+  6956 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/arg.cpp.o
+  6881 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/chat.cpp.o
+  6738 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/llama-server-impl.dir/server-models.cpp.o
+  6541 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/json.cpp.o
+  6446 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/server-context.dir/server-tools.cpp.o
+  6289 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/gguf-model-data.dir/gguf-model-data.cpp.o
+  6081 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-chat-peg-parser.dir/test-chat-peg-parser.cpp.o
+  5900 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/download.cpp.o
+  5874 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/src/CMakeFiles/llama.dir/unicode-data.cpp.o
+
+**** Files that took longest to codegen (compiler backend):
+ 42063 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-chat.dir/test-chat.cpp.o
+ 23793 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-jinja.dir/test-jinja.cpp.o
+ 23430 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-backend-ops.dir/test-backend-ops.cpp.o
+ 20616 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/arg.cpp.o
+ 17714 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-chat-auto-parser.dir/test-chat-auto-parser.cpp.o
+ 17636 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/chat.cpp.o
+ 16657 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/vendor/cpp-httplib/CMakeFiles/cpp-httplib.dir/httplib.cpp.o
+ 16022 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tools/mtmd/CMakeFiles/mtmd.dir/mtmd-helper.cpp.o
+ 15606 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-chat-peg-parser.dir/test-chat-peg-parser.cpp.o
+ 14043 ms: /home/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/server-context.dir/server-context.cpp.o
+
+**** Templates that took longest to instantiate:
+ 21982 ms: std::basic_regex<char>::_M_compile (42 times, avg 523 ms)
+ 21210 ms: std::__detail::_Compiler<std::regex_traits<char>>::_Compiler (42 times, avg 505 ms)
+ 18203 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_disjunction (42 times, avg 433 ms)
+ 18073 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_alternative (42 times, avg 430 ms)
+ 17744 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_term (42 times, avg 422 ms)
+ 15008 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_atom (42 times, avg 357 ms)
+ 14516 ms: std::basic_regex<char>::basic_regex<std::char_traits<char>, std::all... (29 times, avg 500 ms)
+  7480 ms: std::basic_regex<char>::basic_regex (14 times, avg 534 ms)
+  6331 ms: std::__detail::__variant::_Copy_ctor_base<false, common_peg_epsilon_... (34 times, avg 186 ms)
+  5343 ms: std::__detail::_Compiler<std::regex_traits<char>>::_M_insert_charact... (42 times, avg 127 ms)
+  4564 ms: std::unordered_map<int, int> (306 times, avg 14 ms)
+  3997 ms: std::__detail::__regex_algo_impl<__gnu_cxx::__normal_iterator<const ... (31 times, avg 128 ms)
+  3793 ms: std::__detail::__variant::__raw_idx_visit<(lambda at /usr/bin/../lib... (34 times, avg 111 ms)
+  3781 ms: std::__do_visit<std::__detail::__variant::__variant_idx_cookie, (lam... (34 times, avg 111 ms)
+  3659 ms: std::__detail::__variant::_Variant_storage<false, common_peg_epsilon... (34 times, avg 107 ms)
+  3619 ms: std::__detail::__variant::_Variant_storage<false, common_peg_epsilon... (34 times, avg 106 ms)
+  3591 ms: std::__do_visit<void, (lambda at /usr/bin/../lib/gcc/x86_64-linux-gn... (34 times, avg 105 ms)
+  3520 ms: std::__detail::__variant::__gen_vtable_impl<std::__detail::__variant... (34 times, avg 103 ms)
+  3498 ms: std::__detail::__variant::__gen_vtable_impl<std::__detail::__variant... (34 times, avg 102 ms)
+  3488 ms: std::_Hashtable<int, std::pair<const int, int>, std::allocator<std::... (306 times, avg 11 ms)
+  3425 ms: std::unique_ptr<ggml_context, ggml_context_deleter> (224 times, avg 15 ms)
+  3414 ms: std::__detail::__variant::__gen_vtable_impl<std::__detail::__variant... (34 times, avg 100 ms)
+  3390 ms: std::__detail::__variant::__gen_vtable_impl<std::__detail::__variant... (34 times, avg 99 ms)
+  3252 ms: std::unique_ptr<llm_graph_input_attn_kv> (172 times, avg 18 ms)
+  3062 ms: std::unique_ptr<llm_graph_input_attn_k_dsa> (172 times, avg 17 ms)
+  2979 ms: std::__uniq_ptr_data<ggml_context, ggml_context_deleter> (224 times, avg 13 ms)
+  2946 ms: std::__uniq_ptr_impl<ggml_context, ggml_context_deleter> (224 times, avg 13 ms)
+  2895 ms: std::unique_ptr<ggml_backend_buffer, ggml_backend_buffer_deleter> (176 times, avg 16 ms)
+  2894 ms: std::variant<common_peg_epsilon_parser, common_peg_start_parser, com... (34 times, avg 85 ms)
+  2862 ms: std::unique_ptr<llama_vocab::impl> (174 times, avg 16 ms)
+
+**** Template sets that took longest to instantiate:
+ 61404 ms: std::unique_ptr<$> (3737 times, avg 16 ms)
+ 52144 ms: std::__uniq_ptr_data<$> (3737 times, avg 13 ms)
+ 51565 ms: std::__uniq_ptr_impl<$> (3737 times, avg 13 ms)
+ 41853 ms: std::__and_<$> (36597 times, avg 1 ms)
+ 25749 ms: std::tuple<$> (3998 times, avg 6 ms)
+ 23336 ms: std::vector<$> (8051 times, avg 2 ms)
+ 22439 ms: std::basic_regex<$>::_M_compile (43 times, avg 521 ms)
+ 21640 ms: std::__detail::_Compiler<$>::_Compiler (43 times, avg 503 ms)
+ 18807 ms: std::_Hashtable<$> (2122 times, avg 8 ms)
+ 18567 ms: std::__detail::_Compiler<$>::_M_disjunction (43 times, avg 431 ms)
+ 18434 ms: std::__detail::_Compiler<$>::_M_alternative (43 times, avg 428 ms)
+ 18098 ms: std::__detail::_Compiler<$>::_M_term (43 times, avg 420 ms)
+ 15357 ms: std::__detail::_Compiler<$>::_M_atom (43 times, avg 357 ms)
+ 15271 ms: std::unordered_map<$> (1141 times, avg 13 ms)
+ 14973 ms: std::basic_regex<$>::basic_regex<$> (30 times, avg 499 ms)
+ 14313 ms: std::__or_<$> (17720 times, avg 0 ms)
+ 12757 ms: std::_TupleConstraints<$>::__is_implicitly_constructible<$> (7998 times, avg 1 ms)
+ 12237 ms: std::_Vector_base<$> (7717 times, avg 1 ms)
+ 10406 ms: std::vector<$>::emplace_back<$> (1332 times, avg 7 ms)
+ 10329 ms: std::_Destroy<$> (3890 times, avg 2 ms)
+  9801 ms: std::vector<$>::vector (1804 times, avg 5 ms)
+  9786 ms: std::vector<$>::~vector (2906 times, avg 3 ms)
+  9617 ms: std::function<$>::function<$> (1586 times, avg 6 ms)
+  9275 ms: std::__do_visit<$> (91 times, avg 101 ms)
+  9097 ms: std::pair<$> (3054 times, avg 2 ms)
+  8804 ms: std::vector<$>::push_back (1151 times, avg 7 ms)
+  8789 ms: std::__detail::__variant::__gen_vtable_impl<$>::_S_apply (91 times, avg 96 ms)
+  8729 ms: std::__detail::__variant::__gen_vtable_impl<$>::_S_apply_all_alts<$> (91 times, avg 95 ms)
+  8488 ms: std::__detail::__variant::__gen_vtable_impl<$>::_S_apply_single_alt<$> (1911 times, avg 4 ms)
+  8206 ms: std::map<$> (1526 times, avg 5 ms)
+
+**** Functions that took longest to compile:
+  4623 ms: common_params_parser_init(common_params&, llama_example, void (*)(in... (/home/danbev/work/ai/llama.cpp/common/arg.cpp)
+  4332 ms: test_template_output_peg_parsers(bool) (/home/danbev/work/ai/llama.cpp/tests/test-chat.cpp)
+  2044 ms: clip_model_loader::load_tensors(clip_ctx&) (/home/danbev/work/ai/llama.cpp/tools/mtmd/clip.cpp)
+  1533 ms: make_test_cases_eval() (/home/danbev/work/ai/llama.cpp/tests/test-backend-ops.cpp)
+  1502 ms: test_fuzzing(testing&) (/home/danbev/work/ai/llama.cpp/tests/test-jinja.cpp)
+  1111 ms: clip_encode(clip_ctx*, clip_encode_params*) (/home/danbev/work/ai/llama.cpp/tools/mtmd/clip.cpp)
+  1099 ms: test_array_methods(testing&) (/home/danbev/work/ai/llama.cpp/tests/test-jinja.cpp)
+  1023 ms: test_gbnf_generation(testing&) (/home/danbev/work/ai/llama.cpp/tests/peg-parser/test-gbnf-generation.cpp)
+   937 ms: make_test_cases_perf() (/home/danbev/work/ai/llama.cpp/tests/test-backend-ops.cpp)
+   915 ms: test_basic(testing&)::$_2::operator()(testing&) const (/home/danbev/work/ai/llama.cpp/tests/peg-parser/test-basic.cpp)
+   859 ms: test_filters(testing&) (/home/danbev/work/ai/llama.cpp/tests/test-jinja.cpp)
+   687 ms: server_schema::make_llama_cmpl_schema(common_params const&, task_par... (/home/danbev/work/ai/llama.cpp/tools/server/server-schema.cpp)
+   649 ms: server_task_result_cmpl_partial::to_json_oaicompat_resp() (/home/danbev/work/ai/llama.cpp/tools/server/server-task.cpp)
+   635 ms: server_chat_convert_anthropic_to_oai(common_json const&) (/home/danbev/work/ai/llama.cpp/tools/server/server-chat.cpp)
+   610 ms: test_python_dict_parser(testing&) (/home/danbev/work/ai/llama.cpp/tests/peg-parser/test-python-dict-parser.cpp)
+   580 ms: common_chat_try_specialized_template(common_chat_template const&, st... (/home/danbev/work/ai/llama.cpp/common/chat.cpp)
+   550 ms: server_task_result_cmpl_final::to_json_anthropic_stream() (/home/danbev/work/ai/llama.cpp/tools/server/server-task.cpp)
+   541 ms: server_chat_convert_responses_to_chatcmpl(common_json const&) (/home/danbev/work/ai/llama.cpp/tools/server/server-chat.cpp)
+   536 ms: analyze_template(std::__cxx11::basic_string<char, std::char_traits<c... (/home/danbev/work/ai/llama.cpp/tests/test-chat-analysis.cpp)
+   533 ms: test_complex_grammar() (/home/danbev/work/ai/llama.cpp/tests/test-grammar-integration.cpp)
+   498 ms: ma_flac_ds_read(void*, void*, unsigned long long, unsigned long long*) (/home/danbev/work/ai/llama.cpp/tools/mtmd/mtmd-helper.cpp)
+   471 ms: test_init(testing&) (/home/danbev/work/ai/llama.cpp/tests/test-batch-alloc.cpp)
+   470 ms: test_all(std::__cxx11::basic_string<char, std::char_traits<char>, st... (/home/danbev/work/ai/llama.cpp/tests/test-json-schema-to-grammar.cpp)
+   467 ms: ma_dr_flac__decode_samples_with_residual__rice__scalar(ma_dr_flac_bs... (/home/danbev/work/ai/llama.cpp/tools/mtmd/mtmd-helper.cpp)
+   464 ms: create_tools() (/home/danbev/work/ai/llama.cpp/tests/test-chat-peg-parser.cpp)
+   451 ms: llama_model_quantize_impl(std::__cxx11::basic_string<char, std::char... (/home/danbev/work/ai/llama.cpp/src/llama-quant.cpp)
+   449 ms: test_json_parser(testing&) (/home/danbev/work/ai/llama.cpp/tests/peg-parser/test-json-parser.cpp)
+   444 ms: test_unicode(testing&)::$_3::operator()(testing&) const (/home/danbev/work/ai/llama.cpp/tests/peg-parser/test-unicode.cpp)
+   441 ms: llama_model_base::load_tensors(llama_model_loader&) (/home/danbev/work/ai/llama.cpp/src/llama-model.cpp)
+   434 ms: test_json_schema() (/home/danbev/work/ai/llama.cpp/tests/test-grammar-integration.cpp)
+
+**** Function sets that took longest to compile / optimize:
+  2413 ms: bool std::__detail::_Compiler<$>::_M_expression_term<$>(std::__detai... (116 times, avg 20 ms)
+  1778 ms: std::_Function_handler<$>::_M_manager(std::_Any_data&, std::_Any_dat... (1039 times, avg 1 ms)
+  1371 ms: std::__detail::_BracketMatcher<$>::_BracketMatcher(std::__detail::_B... (116 times, avg 11 ms)
+  1274 ms: std::vector<$>::~vector() (478 times, avg 2 ms)
+  1242 ms: void std::__detail::_Compiler<$>::_M_insert_bracket_matcher<$>(bool) (116 times, avg 10 ms)
+  1135 ms: std::__detail::_Compiler<$>::_M_quantifier() (29 times, avg 39 ms)
+   920 ms: std::_Rb_tree<$>::_M_erase(std::_Rb_tree_node<$>*) (491 times, avg 1 ms)
+   912 ms: void std::__detail::_Compiler<$>::_M_insert_character_class_matcher<... (116 times, avg 7 ms)
+   878 ms: std::__detail::_Executor<$>::_M_dfs(std::__detail::_Executor<$>::_Ma... (70 times, avg 12 ms)
+   874 ms: std::__detail::_BracketMatcher<$>::_M_make_range(char, char) (112 times, avg 7 ms)
+   867 ms: std::_Function_handler<$>::_M_invoke(std::_Any_data const&, server_h... (48 times, avg 18 ms)
+   854 ms: std::__detail::_BracketMatcher<$>::_M_add_character_class(std::__cxx... (116 times, avg 7 ms)
+   827 ms: std::__detail::_BracketMatcher<$>::_M_add_equivalence_class(std::__c... (116 times, avg 7 ms)
+   794 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, unsign... (2 times, avg 397 ms)
+   776 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, float*) (2 times, avg 388 ms)
+   755 ms: std::_Function_handler<$>::_M_invoke(std::_Any_data const&, common_c... (25 times, avg 30 ms)
+   739 ms: std::__detail::_Executor<$>::_M_lookahead(long) (70 times, avg 10 ms)
+   680 ms: std::_Function_handler<$>::_M_invoke(std::_Any_data const&) (110 times, avg 6 ms)
+   675 ms: std::__detail::_BracketMatcher<$>::_M_ready() (112 times, avg 6 ms)
+   645 ms: std::__detail::_StateSeq<$>::_M_clone() (29 times, avg 22 ms)
+   637 ms: testing::print_result(std::__cxx11::basic_string<$> const&, int, int... (12 times, avg 53 ms)
+   622 ms: std::__detail::_BracketMatcher<$>::_M_add_collate_element(std::__cxx... (116 times, avg 5 ms)
+   591 ms: std::__detail::_Compiler<$>::_M_atom() (29 times, avg 20 ms)
+   580 ms: common_chat_try_specialized_template(common_chat_template const&, st... (1 times, avg 580 ms)
+   560 ms: std::vector<$>::_M_default_append(unsigned long) (93 times, avg 6 ms)
+   546 ms: void (anonymous namespace)::tinyBLAS<$>::gemm<$>(long, long, long) (27 times, avg 20 ms)
+   540 ms: void std::vector<$>::_M_range_insert<$>(__gnu_cxx::__normal_iterator... (48 times, avg 11 ms)
+   536 ms: analyze_template(std::__cxx11::basic_string<$> const&) (1 times, avg 536 ms)
+   533 ms: void std::vector<$>::_M_realloc_insert<$>(__gnu_cxx::__normal_iterat... (53 times, avg 10 ms)
+   531 ms: bool std::__detail::__regex_algo_impl<$>(__gnu_cxx::__normal_iterato... (32 times, avg 16 ms)
+
+**** Expensive headers:
+219629 ms: /home/danbev/work/ai/llama.cpp/src/models/models.h (included 152 times, avg 1444 ms), included via:
+  151x: <direct include>
+  1x: llama-model.h 
+
+208004 ms: /home/danbev/work/ai/llama.cpp/src/llama-model.h (included 169 times, avg 1230 ms), included via:
+  151x: models.h 
+  2x: <direct include>
+  2x: llama-impl.h 
+  2x: llama.h 
+  1x: llama-memory-hybrid-idx.h 
+  1x: llama-kv-cache-dsa-iswa.h 
+  ...
+
+205251 ms: /home/danbev/work/ai/llama.cpp/include/llama.h (included 320 times, avg 641 ms), included via:
+  150x: models.h llama-model.h 
+  43x: models.h clip-graph.h clip.h mtmd.h 
+  18x: arg.h common.h llama-cpp.h 
+  15x: <direct include>
+  6x: tests.h testing.h common.h llama-cpp.h 
+  5x: common.h llama-cpp.h 
+  ...
+
+68866 ms: /home/danbev/work/ai/llama.cpp/src/llama-graph.h (included 172 times, avg 400 ms), included via:
+  150x: models.h llama-model.h llama.h 
+  2x: llama-impl.h llama-model.h 
+  1x: <direct include>
+  1x: llama-context.h llama.h 
+  1x: llama.h llama-context.h 
+  1x: llama-memory.h llama.h 
+  ...
+
+66713 ms: /home/danbev/work/ai/llama.cpp/common/common.h (included 83 times, avg 803 ms), included via:
+  18x: arg.h 
+  9x: llama.h 
+  6x: <direct include>
+  6x: tests.h testing.h 
+  3x: chat-auto-parser-helpers.h chat-auto-parser.h chat.h 
+  3x: testing.h 
+  ...
+
+58734 ms: /home/danbev/work/ai/llama.cpp/tools/mtmd/models/models.h (included 44 times, avg 1334 ms), included via:
+  43x: <direct include>
+  1x: clip.h 
+
+48501 ms: /home/danbev/work/ai/llama.cpp/tools/mtmd/clip.h (included 48 times, avg 1010 ms), included via:
+  43x: models.h clip-graph.h 
+  2x: <direct include>
+  1x: mtmd-image.h ggml.h clip-model.h 
+  1x: mtmd-audio.h ggml.h clip-model.h 
+  1x: testing.h mtmd-image.h clip-model.h 
+
+43089 ms: /home/danbev/work/ai/llama.cpp/src/llama-arch.h (included 163 times, avg 264 ms), included via:
+  150x: models.h llama-model.h llama.h 
+  2x: <direct include>
+  2x: llama-impl.h llama-model.h 
+  1x: models.h llama-model.h 
+  1x: llama-graph.h 
+  1x: llama-model.h llama.h 
+  ...
+
+40560 ms: /home/danbev/work/ai/llama.cpp/common/chat.h (included 32 times, avg 1267 ms), included via:
+  6x: tests.h testing.h chat-peg-parser.h 
+  4x: <direct include>
+  3x: chat-auto-parser-helpers.h chat-auto-parser.h 
+  2x: chat-peg-parser.h 
+  2x: chat-auto-parser.h 
+  2x: server-task.h server-common.h log.h 
+  ...
+
+38337 ms: /home/danbev/work/ai/llama.cpp/ggml/include/ggml.h (included 344 times, avg 111 ms), included via:
+  150x: models.h llama-model.h llama.h 
+  43x: models.h clip-graph.h 
+  18x: arg.h common.h llama-cpp.h llama.h 
+  16x: <direct include>
+  14x: llama.h 
+  6x: tests.h testing.h common.h llama-cpp.h llama.h 
+  ...
+
+  done in 0.2s.
diff --git a/docs/profiling-reports/profile-report-full-mac.txt b/docs/profiling-reports/profile-report-full-mac.txt
new file mode 100644
index 0000000000..267de50b39
--- /dev/null
+++ b/docs/profiling-reports/profile-report-full-mac.txt
@@ -0,0 +1,228 @@
+Analyzing build trace from '/Users/danbev/work/ai/llama.cpp/build-profile-full/clang_analysis.bin'...
+**** Time summary:
+Compilation (407 times):
+  Parsing (frontend):          265.7 s
+  Codegen & opts (backend):    209.7 s
+
+**** Files that took longest to parse (compiler frontend):
+  2896 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tools/ui/CMakeFiles/llama-ui.dir/ui.cpp.o
+  2826 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/llama-server-impl.dir/server-models.cpp.o
+  2360 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/server-context.dir/server-context.cpp.o
+  2210 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/vendor/cpp-httplib/CMakeFiles/cpp-httplib.dir/httplib.cpp.o
+  2094 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/llama-server-impl.dir/server.cpp.o
+  2049 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/chat.cpp.o
+  1950 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/llama-server-impl.dir/server-http.cpp.o
+  1930 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/server-context.dir/server-tools.cpp.o
+  1925 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/arg.cpp.o
+  1879 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-peg-parser.dir/peg-parser/test-gbnf-generation.cpp.o
+
+**** Files that took longest to codegen (compiler backend):
+  8830 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-backend-ops.dir/test-backend-ops.cpp.o
+  7873 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-chat.dir/test-chat.cpp.o
+  6536 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/vendor/cpp-httplib/CMakeFiles/cpp-httplib.dir/httplib.cpp.o
+  5706 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/chat.cpp.o
+  5336 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/common/CMakeFiles/llama-common.dir/arg.cpp.o
+  4720 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-jinja.dir/test-jinja.cpp.o
+  4636 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tools/mtmd/CMakeFiles/mtmd.dir/mtmd-helper.cpp.o
+  4145 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tools/server/CMakeFiles/server-context.dir/server-context.cpp.o
+  4103 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-chat-auto-parser.dir/test-chat-auto-parser.cpp.o
+  4012 ms: /Users/danbev/work/ai/llama.cpp/build-profile-full/tests/CMakeFiles/test-chat-peg-parser.dir/test-chat-peg-parser.cpp.o
+
+**** Templates that took longest to instantiate:
+  3123 ms: std::basic_regex<char>::basic_regex<std::char_traits<char>, std::all... (29 times, avg 107 ms)
+  3117 ms: std::basic_regex<char>::__init<std::__wrap_iter<const char *>> (29 times, avg 107 ms)
+  3100 ms: std::basic_regex<char>::__parse<std::__wrap_iter<const char *>> (29 times, avg 106 ms)
+  2589 ms: std::__variant_detail::__copy_constructor<std::__variant_detail::__t... (34 times, avg 76 ms)
+  2582 ms: std::__variant_detail::__ctor<std::__variant_detail::__traits<common... (34 times, avg 75 ms)
+  2469 ms: std::basic_regex<char>::__parse_ecma_exp<std::__wrap_iter<const char... (29 times, avg 85 ms)
+  2416 ms: std::basic_regex<char>::__parse_alternative<std::__wrap_iter<const c... (29 times, avg 83 ms)
+  2405 ms: std::basic_regex<char>::__parse_term<std::__wrap_iter<const char *>> (29 times, avg 82 ms)
+  1476 ms: std::basic_regex<char>::__parse_atom<std::__wrap_iter<const char *>> (29 times, avg 50 ms)
+  1171 ms: std::unordered_map<std::bitset<256>, std::vector<int>> (172 times, avg 6 ms)
+  1096 ms: std::basic_regex<char>::basic_regex (17 times, avg 64 ms)
+  1092 ms: std::basic_regex<char>::__init<const char *> (17 times, avg 64 ms)
+  1089 ms: std::basic_regex<char>::__parse<const char *> (17 times, avg 64 ms)
+  1042 ms: std::basic_regex<char>::__push_lookahead (41 times, avg 25 ms)
+  1039 ms: std::__lookahead<char, std::regex_traits<char>>::__lookahead (41 times, avg 25 ms)
+  1035 ms: std::__lookahead<char, std::regex_traits<char>>::__exec (41 times, avg 25 ms)
+  1012 ms: std::vector<bool>::vector (176 times, avg 5 ms)
+   996 ms: std::unordered_map<std::string, std::string> (185 times, avg 5 ms)
+   971 ms: std::__hash_table<std::__hash_value_type<std::bitset<256>, std::vect... (172 times, avg 5 ms)
+   969 ms: std::unordered_map<int, int> (166 times, avg 5 ms)
+   956 ms: std::basic_regex<char>::__parse_ecma_exp<const char *> (17 times, avg 56 ms)
+   950 ms: std::__variant_detail::__visitation::__base::__visit_alt_at<(lambda ... (34 times, avg 27 ms)
+   947 ms: std::basic_regex<char>::__parse_alternative<const char *> (17 times, avg 55 ms)
+   947 ms: std::basic_regex<char>::__parse_term<const char *> (17 times, avg 55 ms)
+   940 ms: std::__variant_detail::__visitation::__base::__make_fdiagonal<(lambd... (34 times, avg 27 ms)
+   933 ms: std::__variant_detail::__visitation::__base::__make_fdiagonal_impl<(... (34 times, avg 27 ms)
+   926 ms: std::basic_regex<char>::__match_at_start_ecma<std::allocator<std::cs... (42 times, avg 22 ms)
+   892 ms: std::basic_string<wchar_t> (386 times, avg 2 ms)
+   855 ms: std::basic_string<char16_t> (386 times, avg 2 ms)
+   845 ms: std::unordered_map<std::string, llama_adapter_lora_weight> (172 times, avg 4 ms)
+
+**** Template sets that took longest to instantiate:
+ 11404 ms: std::vector<$> (7951 times, avg 1 ms)
+  7273 ms: std::allocator_traits<$> (8688 times, avg 0 ms)
+  5682 ms: std::unordered_map<$> (1002 times, avg 5 ms)
+  5545 ms: std::vector<$>::vector (1462 times, avg 3 ms)
+  5461 ms: std::__hash_table<$> (1252 times, avg 4 ms)
+  5198 ms: std::map<$> (1480 times, avg 3 ms)
+  4950 ms: std::function<$>::function<$> (889 times, avg 5 ms)
+  4846 ms: std::__function::__value_func<$>::__value_func<$> (889 times, avg 5 ms)
+  4555 ms: std::vector<$>::__construct_at_end<$> (1413 times, avg 3 ms)
+  4529 ms: std::vector<$>::emplace_back<$> (1243 times, avg 3 ms)
+  4335 ms: std::basic_regex<$>::__init<$> (47 times, avg 92 ms)
+  4314 ms: std::basic_regex<$>::__parse<$> (47 times, avg 91 ms)
+  4132 ms: std::__tree<$> (1895 times, avg 2 ms)
+  4028 ms: std::vector<$>::__init_with_size<$> (995 times, avg 4 ms)
+  3934 ms: std::unique_ptr<$> (4148 times, avg 0 ms)
+  3634 ms: std::__uninitialized_allocator_copy<$> (1226 times, avg 2 ms)
+  3597 ms: std::basic_string<$> (1638 times, avg 2 ms)
+  3537 ms: std::basic_regex<$>::__parse_ecma_exp<$> (47 times, avg 75 ms)
+  3483 ms: std::vector<$>::__swap_out_circular_buffer (1565 times, avg 2 ms)
+  3474 ms: std::basic_regex<$>::__parse_alternative<$> (47 times, avg 73 ms)
+  3463 ms: std::basic_regex<$>::__parse_term<$> (47 times, avg 73 ms)
+  3454 ms: std::vector<$>::__emplace_back_slow_path<$> (1009 times, avg 3 ms)
+  3447 ms: std::vector<$>::push_back (903 times, avg 3 ms)
+  3249 ms: std::basic_regex<$>::basic_regex<$> (30 times, avg 108 ms)
+  2989 ms: std::basic_string<$>::basic_string (1615 times, avg 1 ms)
+  2681 ms: std::basic_string<$>::__init (1545 times, avg 1 ms)
+  2600 ms: std::__variant_detail::__ctor<$>::__generic_construct<$> (35 times, avg 74 ms)
+  2589 ms: std::__variant_detail::__copy_constructor<$>::__copy_constructor (34 times, avg 76 ms)
+  2291 ms: std::__exception_guard_exceptions<$>::~__exception_guard_exceptions (2134 times, avg 1 ms)
+  2275 ms: std::__copy<$> (952 times, avg 2 ms)
+
+**** Functions that took longest to compile:
+  2744 ms: test_template_output_peg_parsers(bool) (/Users/danbev/work/ai/llama.cpp/tests/test-chat.cpp)
+  1008 ms: make_test_cases_eval() (/Users/danbev/work/ai/llama.cpp/tests/test-backend-ops.cpp)
+   957 ms: common_params_parser_init(common_params&, llama_example, void (*)(in... (/Users/danbev/work/ai/llama.cpp/common/arg.cpp)
+   567 ms: clip_model_loader::load_tensors(clip_ctx&) (/Users/danbev/work/ai/llama.cpp/tools/mtmd/clip.cpp)
+   404 ms: test_gbnf_generation(testing&) (/Users/danbev/work/ai/llama.cpp/tests/peg-parser/test-gbnf-generation.cpp)
+   354 ms: clip_encode(clip_ctx*, clip_encode_params*) (/Users/danbev/work/ai/llama.cpp/tools/mtmd/clip.cpp)
+   336 ms: test_basic(testing&)::$_2::operator()(testing&) const (/Users/danbev/work/ai/llama.cpp/tests/peg-parser/test-basic.cpp)
+   305 ms: test_fuzzing(testing&) (/Users/danbev/work/ai/llama.cpp/tests/test-jinja.cpp)
+   289 ms: make_test_cases_perf() (/Users/danbev/work/ai/llama.cpp/tests/test-backend-ops.cpp)
+   171 ms: test_python_dict_parser(testing&) (/Users/danbev/work/ai/llama.cpp/tests/peg-parser/test-python-dict-parser.cpp)
+   169 ms: test_array_methods(testing&) (/Users/danbev/work/ai/llama.cpp/tests/test-jinja.cpp)
+   157 ms: test_json_parser(testing&) (/Users/danbev/work/ai/llama.cpp/tests/peg-parser/test-json-parser.cpp)
+   146 ms: parse_cmd_params(int, char**) (/Users/danbev/work/ai/llama.cpp/tools/llama-bench/llama-bench.cpp)
+   144 ms: llama_server(common_params&, int, char**) (/Users/danbev/work/ai/llama.cpp/tools/server/server.cpp)
+   141 ms: ma_flac_ds_read(void*, void*, unsigned long long, unsigned long long*) (/Users/danbev/work/ai/llama.cpp/tools/mtmd/mtmd-helper.cpp)
+   139 ms: server_chat_convert_responses_to_chatcmpl(common_json const&) (/Users/danbev/work/ai/llama.cpp/tools/server/server-chat.cpp)
+   137 ms: common_chat_try_specialized_template(common_chat_template const&, st... (/Users/danbev/work/ai/llama.cpp/common/chat.cpp)
+   133 ms: test_handcrafted_file(unsigned int) (/Users/danbev/work/ai/llama.cpp/tests/test-gguf.cpp)
+   132 ms: server_chat_convert_anthropic_to_oai(common_json const&) (/Users/danbev/work/ai/llama.cpp/tools/server/server-chat.cpp)
+   130 ms: test_init(testing&) (/Users/danbev/work/ai/llama.cpp/tests/test-batch-alloc.cpp)
+   130 ms: llama_vocab::impl::load(llama_model_loader&, LLM_KV const&) (/Users/danbev/work/ai/llama.cpp/src/llama-vocab.cpp)
+   125 ms: void set_input_kq_mask_impl<float>(args_set_input_kq_mask const&, fl... (/Users/danbev/work/ai/llama.cpp/src/llama-kv-cache.cpp)
+   122 ms: llama_model_base::load_tensors(llama_model_loader&) (/Users/danbev/work/ai/llama.cpp/src/llama-model.cpp)
+   121 ms: llm_chat_apply_template(llm_chat_template, std::__1::vector<llama_ch... (/Users/danbev/work/ai/llama.cpp/src/llama-chat.cpp)
+   120 ms: llama_perplexity(int, char**) (/Users/danbev/work/ai/llama.cpp/tools/perplexity/perplexity.cpp)
+   117 ms: server_task_result_cmpl_partial::to_json_oaicompat_resp() (/Users/danbev/work/ai/llama.cpp/tools/server/server-task.cpp)
+   115 ms: llama_completion(int, char**) (/Users/danbev/work/ai/llama.cpp/tools/completion/completion.cpp)
+   115 ms: test_filters(testing&) (/Users/danbev/work/ai/llama.cpp/tests/test-jinja.cpp)
+   113 ms: ma_dr_mp3dec_decode_frame(ma_dr_mp3dec*, unsigned char const*, int, ... (/Users/danbev/work/ai/llama.cpp/tools/mtmd/mtmd-helper.cpp)
+   107 ms: ggml_compute_forward_add_non_quantized (/Users/danbev/work/ai/llama.cpp/ggml/src/ggml-cpu/binary-ops.cpp)
+
+**** Function sets that took longest to compile / optimize:
+   509 ms: bool std::__1::basic_regex<$>::__match_at_start_posix_nosubs<$>(char... (33 times, avg 15 ms)
+   343 ms: std::__1::__wrap_iter<$> std::__1::vector<$>::__insert_with_size[abi... (45 times, avg 7 ms)
+   285 ms: bool std::__1::basic_regex<$>::__match_at_start_posix_subs<$>(char c... (33 times, avg 8 ms)
+   278 ms: std::__1::__bracket_expression<$>::__exec(std::__1::__state<$>&) const (29 times, avg 9 ms)
+   246 ms: void (anonymous namespace)::tinyBLAS<$>::gemm<$>(long long, long lon... (36 times, avg 6 ms)
+   189 ms: bool std::__1::basic_regex<$>::__match_at_start_ecma<$>(char const*,... (33 times, avg 5 ms)
+   163 ms: std::__1::vector<$>::__append(unsigned long) (139 times, avg 1 ms)
+   161 ms: ggml::cpu::repack::tensor_traits<$>::forward_mul_mat(ggml_compute_pa... (16 times, avg 10 ms)
+   159 ms: void std::__1::__tree_balance_after_insert[abi:ne200100]<$>(std::__1... (119 times, avg 1 ms)
+   144 ms: std::__1::deque<$>::__add_back_capacity() (39 times, avg 3 ms)
+   138 ms: void std::__1::vector<$>::__assign_with_size[abi:ne200100]<$>(std::_... (38 times, avg 3 ms)
+   137 ms: common_chat_try_specialized_template(common_chat_template const&, st... (1 times, avg 137 ms)
+   129 ms: std::__1::vector<$>::~vector[abi:ne200100]() (145 times, avg 0 ms)
+   128 ms: std::__1::deque<$>::__add_front_capacity() (35 times, avg 3 ms)
+   127 ms: std::__1::__bracket_expression<$>::__add_range[abi:ne200100](std::__... (29 times, avg 4 ms)
+   125 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, float*... (1 times, avg 125 ms)
+   121 ms: llm_chat_apply_template(llm_chat_template, std::__1::vector<$> const... (1 times, avg 121 ms)
+   111 ms: void std::__1::__hash_table<$>::__do_rehash<$>(unsigned long) (85 times, avg 1 ms)
+   110 ms: ggml::cpu::repack::tensor_traits<$>::forward_mul_mat_id(ggml_compute... (16 times, avg 6 ms)
+   106 ms: std::__1::__loop<$>::__exec(std::__1::__state<$>&) const (29 times, avg 3 ms)
+   104 ms: test_all(std::__1::basic_string<$> const&, std::__1::function<$>) (1 times, avg 104 ms)
+   103 ms: void set_input_kq_mask_impl<$>(args_set_input_kq_mask const&, unsign... (1 times, avg 103 ms)
+    96 ms: std::__1::__exception_guard_exceptions<$>::~__exception_guard_except... (111 times, avg 0 ms)
+    95 ms: unicode_regex_split(std::__1::basic_string<$> const&, std::__1::vect... (1 times, avg 95 ms)
+    94 ms: char const* std::__1::basic_regex<$>::__parse_expression_term<$>(cha... (17 times, avg 5 ms)
+    93 ms: std::__1::__function::__func<$>::target(std::type_info const&) const (125 times, avg 0 ms)
+    92 ms: unicode_regex_split_custom(std::__1::basic_string<$> const&, std::__... (1 times, avg 92 ms)
+    90 ms: dsv4_build_comp_plans(std::__1::vector<$> const&, unsigned int, bool... (1 times, avg 90 ms)
+    88 ms: std::__1::vector<$>::assign(unsigned long, std::__1::sub_match<$> co... (34 times, avg 2 ms)
+    85 ms: std::__1::__murmur2_or_cityhash<unsigned long, 64ul>::operator()[abi... (29 times, avg 2 ms)
+
+**** Expensive headers:
+71656 ms: /Users/danbev/work/ai/llama.cpp/src/models/models.h (included 152 times, avg 471 ms), included via:
+  152x: <direct include>
+
+64040 ms: /Users/danbev/work/ai/llama.cpp/src/llama-model.h (included 169 times, avg 378 ms), included via:
+  151x: models.h 
+  18x: <direct include>
+
+47893 ms: /Users/danbev/work/ai/llama.cpp/src/llama-arch.h (included 176 times, avg 272 ms), included via:
+  151x: models.h llama-model.h 
+  4x: llama-model.h 
+  3x: llama-memory.h llama-graph.h 
+  2x: <direct include>
+  2x: llama-model-loader.h 
+  2x: llama-context.h llama-graph.h 
+  ...
+
+42695 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/string (included 386 times, avg 110 ms), included via:
+  150x: models.h llama-model.h llama-arch.h 
+  43x: models.h clip-graph.h clip.h mtmd.h map functional vector locale __locale 
+  19x: arg.h common.h list functional vector locale __locale 
+  16x: common.h list functional vector locale __locale 
+  15x: ggml-impl.h vector locale __locale 
+  10x: <direct include>
+  ...
+
+36571 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/functional (included 349 times, avg 104 ms), included via:
+  151x: models.h llama-model.h llama-arch.h set 
+  43x: models.h clip-graph.h clip.h mtmd.h map 
+  20x: common.h list 
+  19x: arg.h common.h list 
+  6x: llama-grammar.h map 
+  5x: tests.h json.h map 
+  ...
+
+30365 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/vector (included 377 times, avg 80 ms), included via:
+  151x: models.h llama-model.h llama-arch.h set functional 
+  43x: models.h clip-graph.h clip.h mtmd.h map functional 
+  19x: arg.h common.h list functional 
+  16x: common.h list functional 
+  15x: ggml-impl.h 
+  6x: <direct include>
+  ...
+
+29586 ms: /Users/danbev/work/ai/llama.cpp/common/common.h (included 84 times, avg 352 ms), included via:
+  25x: <direct include>
+  23x: arg.h 
+  7x: tests.h testing.h 
+  3x: chat-auto-parser-helpers.h chat-auto-parser.h chat.h 
+  3x: testing.h 
+  2x: server-task.h 
+  ...
+
+25740 ms: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk/usr/include/c++/v1/algorithm (included 386 times, avg 66 ms), included via:
+  150x: models.h llama-model.h llama-arch.h string string_view 
+  43x: models.h clip-graph.h clip.h mtmd.h map functional boyer_moore_searcher.h array 
+  19x: arg.h common.h list 
+  16x: common.h list 
+  15x: ggml-impl.h array 
+  9x: string string_view 
+  ...
+
+20790 ms: /Users/danbev/work/ai/llama.cpp/tools/mtmd/models/models.h (included 44 times, avg 472 ms), included via:
+  44x: <direct include>
+
+20453 ms: /Users/danbev/work/ai/llama.cpp/tools/mtmd/clip-graph.h (included 44 times, avg 464 ms), included via:
+  43x: models.h 
+  1x: <direct include>
+
+  done in 0.1s.
diff --git a/scripts/build-profile-baseline.sh b/scripts/build-profile-baseline.sh
index edf488d25c..4cbc992e23 100755
--- a/scripts/build-profile-baseline.sh
+++ b/scripts/build-profile-baseline.sh
@@ -3,7 +3,7 @@
 #
 # Requires ClangBuildAnalyzer:
 # Install mac: brew install clang-build-analyzer
-# Install linux: TODO
+# Install linux: https://github.com/aras-p/ClangBuildAnalyzer.git
 
 set -euo pipefail
 
@@ -89,8 +89,7 @@ find "${BUILD_DIR}" -name "*.json" ! -name "compile_commands.json" \
     | xargs ls -l 2>/dev/null \
     | awk 'NF>5 {print $5, $NF}' \
     | sort -rn \
-    | head -20 \
-    | awk '{printf "%8.1f KB  %s\n", $1/1024, $2}'
+    | awk 'NR<=20 {printf "%8.1f KB  %s\n", $1/1024, $2}'
 
 echo
 echo "ClangBuildAnalyzer report was generated: ${BUILD_DIR}/profile-report.txt"
diff --git a/scripts/build-profile-full.sh b/scripts/build-profile-full.sh
index 62c99adb11..bed258cfcb 100755
--- a/scripts/build-profile-full.sh
+++ b/scripts/build-profile-full.sh
@@ -3,7 +3,7 @@
 #
 # Requires ClangBuildAnalyzer:
 # Install mac: brew install clang-build-analyzer
-# Install linux: TODO
+# Install linux: https://github.com/aras-p/ClangBuildAnalyzer.git
 
 set -euo pipefail