Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-19 09:15:04 +02:00
Code Issues Packages Projects Releases Wiki Activity

docs: add profiling progress notes (wip)

Browse Source
This commit is contained in:
Daniel Bevenius
2026-08-31 08:49:36 +02:00
parent d801fb4dc4
commit 7f34209f0d
8 changed files with 1000 additions and 247 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/build-profiling.md
+47 -4
View File
@@ -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.
```
docs/profiling-reports/baseline-report.txt
-239
View File
@@ -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.
docs/profiling-reports/profile-report-baseline-mac.txt
+239
View File
@@ -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.
docs/profiling-reports/profile-report-baseline-ubuntu.txt
+245
View File
@@ -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.
docs/profiling-reports/profile-report-full-baseline-ubuntu.txt
+238
View File
@@ -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.
docs/profiling-reports/profile-report-full-mac.txt
+228
View File
@@ -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.
scripts/build-profile-baseline.sh
+2 -3
View File
@@ -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"
scripts/build-profile-full.sh
+1 -1
View File
@@ -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