Valentin Pollet
d1a2a4f7a8
The metadata kv dump in llama_model_loader was measured to be very expensive (~10 milliseconds on e.g. Mac M5) for certain models. This was not an issue prior to this bump in version of llama.cpp because we had a custom `llama_model_load_from_file_handle`. Upstream added support for it so we adopted it, but our custom was skipping it. The patch skips the dump using a pre-existing `trace` gate. Differential Revision: https://phabricator.services.mozilla.com/D308368
92 lines
2.6 KiB
YAML
92 lines
2.6 KiB
YAML
schema: 1
|
|
|
|
bugzilla:
|
|
product: Core
|
|
component: "xul"
|
|
|
|
origin:
|
|
name: llama.cpp
|
|
description: LLM inference in C/C++
|
|
|
|
url: https://github.com/ggml-org/llama.cpp
|
|
|
|
release: 74ade52741203e5c8f81eaf06a96cb1cfe15f2a3 (2026-06-16T18:24:28Z).
|
|
revision: 74ade52741203e5c8f81eaf06a96cb1cfe15f2a3
|
|
|
|
license: MIT
|
|
|
|
vendoring:
|
|
url: https://github.com/ggml-org/llama.cpp
|
|
source-hosting: github
|
|
tracking: commit
|
|
|
|
skip-vendoring-steps:
|
|
- update-moz-build
|
|
|
|
keep:
|
|
- generate_sources_mozbuild.py
|
|
- generate_sources_mozbuild.sh
|
|
- cmakeparser.py
|
|
- make_patch.py
|
|
- sources.mozbuild
|
|
- post_update.sh
|
|
- requirements.txt
|
|
- run_parser.py
|
|
- moz-overrides.h
|
|
|
|
exclude:
|
|
- "**"
|
|
- src/llama-quant.cpp
|
|
|
|
include:
|
|
- LICENSE
|
|
- src/*.cpp
|
|
- src/*.h
|
|
- src/models/*.cpp
|
|
- src/models/*.h
|
|
- src/CMakeLists.txt
|
|
- CMakeLists.txt
|
|
- include/llama.h
|
|
- ggml/src/*
|
|
- ggml/src/*.cpp
|
|
- ggml/src/*.h
|
|
- ggml/src/ggml-cpu/*.cpp
|
|
- ggml/src/ggml-cpu/*.c
|
|
- ggml/src/ggml-cpu/*.h
|
|
- ggml/src/ggml-cpu/amx/
|
|
- ggml/src/ggml-cpu/arch/*.cpp
|
|
- ggml/src/ggml-cpu/arch/*.h
|
|
- ggml/src/ggml-cpu/arch/arm/
|
|
- ggml/src/ggml-cpu/arch/x86/
|
|
- ggml/src/ggml-cpu/CMakeLists.txt
|
|
- ggml/src/cmake/*
|
|
- ggml/cmake/*
|
|
- ggml/CMakeLists.txt
|
|
- ggml/include/ggml.h
|
|
- ggml/include/ggml-alloc.h
|
|
- ggml/include/ggml-backend.h
|
|
- ggml/include/ggml-cpu.h
|
|
- ggml/include/ggml-cpp.h
|
|
- ggml/include/ggml-opt.h
|
|
- ggml/include/gguf.h
|
|
|
|
update-actions:
|
|
- action: run-script
|
|
script: '{yaml_dir}/generate_sources_mozbuild.sh'
|
|
cwd: '{yaml_dir}'
|
|
|
|
patches:
|
|
- exceptions.patch # disable exception usage
|
|
- fs.patch # disable std::fs usage we don't need and is too recent for linux
|
|
- no-future.patch # same for future, switch to sequential
|
|
- no-wstring.patch # same, charset conversion is too recent
|
|
- get-terminate.patch # remove forbidden function usage, that we don't need
|
|
- random_device.patch # use of std::random_device, also too recent on linux
|
|
- detect-chat-template.patch # don't crash on exception caught for control flow
|
|
- load-from-buffer-or-fd.patch # add an API to load a model from a buffer or a fd
|
|
- static-ctors.patch # remove some static ctors
|
|
- fix-RLIMIT_MEMLOCK-solaris.patch # fix the build on solaris
|
|
- profiler-hooks.patch # add thread callbacks for Firefox Profiler integration
|
|
- missing-includes.patch # missing include cerrno in gguf.cpp, include mutex in llama-model-loader.cpp
|
|
- skip-metadata-dump.patch # skip dumping metadata strings that proved costly in profiling
|