llama.cpp: revert

To: pkgsrc-wip-changes%NetBSD.org@localhost
Subject: llama.cpp: revert
From: Thomas Klausner (via pkgsrc-wip) <commit-notify%pkgsrc.org@localhost>
Date: Tue, 11 Aug 2026 11:10:02 +0000
Module Name:	pkgsrc-wip
Committed By:	Thomas Klausner <wiz%NetBSD.org@localhost>
Pushed By:	wiz
Date:		Tue Aug 11 13:09:13 2026 +0200
Changeset:	0e2a18744a921e32003e69afd25b26c6e478516f

Modified Files:
	llama.cpp/Makefile
	llama.cpp/PLIST
Removed Files:
	llama.cpp/AGENTS.md

Log Message:
llama.cpp: revert

This was part of a commit for a different package, reverts
changes to the package from the last weeks, and added an
autogenerated AGENTS.md file.

To see a diff of this commit:
https://wip.pkgsrc.org/cgi-bin/gitweb.cgi?p=pkgsrc-wip.git;a=commitdiff;h=0e2a18744a921e32003e69afd25b26c6e478516f

Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.

diffstat:
 llama.cpp/AGENTS.md | 125 ----------------------------------------------------
 llama.cpp/Makefile  |  20 +--------
 llama.cpp/PLIST     |  32 --------------
 3 files changed, 2 insertions(+), 175 deletions(-)

diffs:
diff --git a/llama.cpp/AGENTS.md b/llama.cpp/AGENTS.md
deleted file mode 100644
index 011840ef20..0000000000
--- a/llama.cpp/AGENTS.md
+++ /dev/null
@@ -1,125 +0,0 @@
-# AGENTS.md
-
-This is the **NetBSD pkgsrc WIP packaging** for llama.cpp (`wip/llama.cpp`), not the upstream
-source tree. The upstream llama.cpp source lives inside `work/` and is untracked
-build state: **never edit it, and never commit or reference it**. Edits made to
-files in `work/` are wiped on the next `make clean` / extract. All actual
-package source lives in the pkgsrc metadata files at the top level:
-`Makefile`, `Makefile.common`, `distinfo`, `PLIST`, `DESCR`, `TODO`, `patches/`.
-
-Upstream is https://github.com/ggml-org/llama.cpp (the project moved from
-`ggerganov/` to `ggml-org/`; only `HOMEPAGE` in `Makefile` still cites the old
-org). See `work/llama.cpp-b10276/AGENTS.md` for upstream contribution rules,
-which apply only if you are contributing upstream, not for packaging work.
-
-## What this package is
-
-A pkgsrc WIP (work-in-progress) package. Key facts:
-
-- **Version scheme is unusual**: upstream publishes "builds" tagged `b10276`,
-  but the pkgsrc `PKGNAME` must be a dotted version, so `Makefile` does
-  `PKGNAME= ${DISTNAME:S/-b/-${LLAMA_CPP_VAR}./}` with
-  `LLAMA_CPP_VAR= 0.0.2` (the comment explains: "because 'b' is the second
-  letter of the alphabet"). Result: `llama.cpp-0.0.2.10276`.
-- `Makefile.common` is **shared**: it is included by `wip/llama.cpp/Makefile`
-  and `wip/py-llama-cpp-scripts/Makefile`. Do not move variables out of it into
-  `Makefile` without checking the other consumer.
-- `LLAMA_BUILD_NUMBER` and `LLAMA_BUILD_COMMIT` (in `Makefile.common`) must match
-  the exact upstream release being packaged; they are compiled into the binary
-  (`-DLLAMA_BUILD_NUMBER/COMMIT` cmake args). Bumping the release means bumping
-  these three things together: `GITHUB_TAG`/`LLAMA_BUILD_NUMBER`,
-  `LLAMA_BUILD_COMMIT`, and `distinfo` checksums.
-- **Two distfiles**: the source tarball plus a separate web UI tarball
-  `llama-${GITHUB_TAG}-ui.tar.gz` fetched from the GitHub release page
-  (`SITES.llama-...=` line in `Makefile`). Both must be added to `distinfo`
-  (BLAKE2s, SHA512, Size).
-- `post-extract` symlinks `${WRKDIR}/llama-${GITHUB_TAG}` into
-  `${WRKSRC}/tools/ui/dist`, so the web UI must be extracted in `WRKDIR` (it
-  can't be a plain DISTFILES target with its own subdir without adjusting this).
-- Build uses **cmake** (not make). **Bundled ggml is used** (`LLAMA_USE_SYSTEM_GGML`
-  is NOT set): ggml 0.18.1 is built from the tarball's `ggml/` subdir so llama
-  b10276 gets the exact API it expects. This means the package *ships its own*
-  `libggml*.so` (see PLIST) and depends on **OpenBLAS** (`math/blas`), not on
-  `math/ggml`. History: a `-DLLAMA_USE_SYSTEM_GGML=YES` era (against pkgsrc
-  `math/ggml-0.15.3`) was reverted because llama b10276's bundled ggml is 0.18.1
-  and the old system ggml lacks `GGML_TYPE_Q2_0`, `ggml_lightning_indexer`,
-  `ggml_dsv4_hc_*`, which llama calls unconditionally (DeepSeek-32/4, GLM-DSA,
-  Q1_0/Q2_0 quant types). Keep the two in sync: if you bump to a llama build
-  whose bundled ggml is >0.18.x, no separate ggml package bump is needed since
-  it's bundled.
-- Shared libraries are versioned with the build number:
-  `lib/libllama.so.0.0.${LLAMA_BUILD_NUMBER}` etc. `Makefile` uses
-  `PRINT_PLIST_AWK` + `PLIST_SUBST` to keep `PLIST` stable across bumps, so
-  the literal `PLIST` line still says `${LLAMA_BUILD_NUMBER}`; do not hand-edit
-  it to a real number. **Exception**: the bundled ggml libs keep their *own*
-  ggml version (e.g. `0.18.1`), hardcoded in PLIST, not the llama build number.
-- Installs **many** binaries (about 40 `bin/llama-*` tools, plus bundled ggml
-  headers/libs, see `PLIST`).
-
-## Commands
-
-Run from `/bd/pkgsrc/wip/llama.cpp` (the package dir). `work/` is gitignored
-(via `wip/.gitignore`), so changes inside it never show in `git status`.
-
-- `bmake` — build. The work directory already has completed `configure`,
-  `build`, `depends` phases (`.configure_done`, `.buildlink_*` markers), so
-  incremental rebuilds are fast. Must be run from a shell where `bmake` and the
-  NetBSD tools are available.
-- `bmake clean` / `bmake clean CONFIGURED=yes` — clean; also wipes `work/`
-  (the extracted upstream source).
-- `bmake package` / `bmake stage-install` — package staging.
-- `bmake install` / `bmake deinstall` — install/uninstall the binary package.
-- `bmake check` — run pkg_install checks (with tests disabled via
-  `LLAMA_BUILD_TESTS=NO` there are no unit tests).
-- `bmake print-PLIST` — regenerate PLIST (uses the `PRINT_PLIST_AWK`/`PLIST_SUBST`
-  machinery above; verify the `${LLAMA_BUILD_NUMBER}` substitution still applies).
-- `bmake url2pkg` — not useful here (package predates it).
-- `bmake resolve-depends` — check dependency availability.
-- After editing `Makefile` concerns: `bmake show-vars` to inspect what's set.
-- Committing is done from the `wip` tree root (`/bd/pkgsrc/wip` is a git repo;
-  the package dir itself is not). Commit messages follow pkgsrc wip style, e.g.
-  `llama.cpp: update to 0.0.2.10276.` — see `git log --oneline -- llama.cpp/`.
-
-## Updating the package (release bump workflow)
-
-1. Pick the new upstream tag `bN` and commit hash from
-   https://github.com/ggml-org/llama.cpp/releases.
-2. In `Makefile.common`: update `LLAMA_BUILD_NUMBER` and `LLAMA_BUILD_COMMIT`.
-   `GITHUB_TAG` derives from the number automatically.
-3. Fetch the new source tarball and UI tarball; run `bmake makesum` (or update
-   `distinfo` manually with BLAKE2s/SHA512/Size for both files).
-4. `bmake clean && bmake` — rebuild from scratch.
-5. Check `PLIST` for new/removed binaries with `bmake print-PLIST`; update
-   `PLIST` as needed (keeping `${LLAMA_BUILD_NUMBER}` unexpanded in llama lib
-   names, but hardcoding the bundled ggml version for `libggml*`, e.g.
-   `0.18.1`).
-6. If upstream moved/removed a `cmake` option or changed bundled ggml usage,
-   adjust `CMAKE_CONFIGURE_ARGS` (the bundled ggml builds from `ggml/`; no
-   `math/ggml` dependency is used).
-7. Commit from `/bd/pkgsrc/wip` with a message like
-   `llama.cpp: update to 0.0.2.10276.`
-
-## Gotchas / non-obvious things
-
-- **Do not touch `work/`.** It contains the full upstream source (with its own
-  `AGENTS.md`/`CLAUDE.md`) plus extracted `llama-*-ui` and the generated
-  `_gzip/` PWA app. Any patch you want to ship must go in `patches/` and be
-  registered in `distinfo` (SHA1) — see `patches/patch-vendor_sheredom_subprocess.h`,
-  which fixes a NetBSD `posix_spawn_file_actions_addchdir` build failure.
-- The NetBSD patch is the only reason the package builds on NetBSD: upstream
-  gates `posix_spawn_file_actions_addchdir` behind Apple-only `#if`, the patch
-  adds `defined(__NetBSD__)` (upstream PR sheredom/subprocess.h#102).
-- The `TODO` file lists 12 known CVEs (e.g. CVE-2026-17500) with no fixes yet;
-  they are the reason the package keeps getting updated. Reference them in
-  commit messages when bumping (repo history does this).
-- `Makefile` sets `CATEGORIES= devel` and `LICENSE= mit`; upstream's full
-  license check: llama.cpp is MIT, but `gguf-py` and the web UI ship
-  separately — do not add their licenses to `LICENSE` for the C++ package.
-- The `work/` UI tarball contains PWA assets (`apple-splash-*`), which have
-  their own licenses; they are installed as static assets under the server, not
-  listed in `PLIST`.
-- `git status`/`git log` only work from `/bd/pkgsrc/wip`, not from this dir.
-- `PKGNAME` in `Makefile` is deliberately squatting the `0.0.2` prefix so the
-  `libllama*.so` ABI version resolves correctly; don't "simplify" it.
-- `LLAMA_CPP_VAR` is defined in `Makefile` (not `Makefile.common`) — it's a hack
-  and must stay in sync manually if you ever move it.
diff --git a/llama.cpp/Makefile b/llama.cpp/Makefile
index 48a04c7b06..19494aecf3 100644
--- a/llama.cpp/Makefile
+++ b/llama.cpp/Makefile
@@ -18,36 +18,20 @@ USE_TOOLS+=		pkg-config
 USE_LANGUAGES=		c c++
 USE_CXX_FEATURES=	c++17
 
-BLAS_INDEX64=		yes
-BLAS_ACCEPTED=		openblas_pthread #openblas_openmp
-BLAS_C_INTERFACE=	yes
-
 PKGCONFIG_OVERRIDE+=	cmake/llama.pc.in
-PKGCONFIG_OVERRIDE+=	ggml/ggml.pc
 
 CMAKE_CONFIGURE_ARGS+=	-DLLAMA_BUILD_NUMBER=${LLAMA_BUILD_NUMBER}
 CMAKE_CONFIGURE_ARGS+=	-DLLAMA_BUILD_COMMIT=${LLAMA_BUILD_COMMIT}
-CMAKE_CONFIGURE_ARGS+=	-DGGML_BLAS=ON
-CMAKE_CONFIGURE_ARGS+=	-DGGML_BLAS_VENDOR=OpenBLAS
-CMAKE_CONFIGURE_ARGS+=	-DBLAS_LIBRARIES=${CBLAS_LIBS:Q}
 CMAKE_CONFIGURE_ARGS+=	-DLLAMA_BUILD_TESTS=NO
-CMAKE_CONFIGURE_ARGS+=	-DGGML_OPENCL_EMBED_KERNELS=OFF
-CMAKE_CONFIGURE_ARGS+=	-DGGML_OPENCL_PROFILING=OFF
-CMAKE_CONFIGURE_ARGS+=	-DGGML_OPENCL_USE_ADRENO_KERNELS=OFF
+CMAKE_CONFIGURE_ARGS+=	-DLLAMA_USE_SYSTEM_GGML=YES
 
 PRINT_PLIST_AWK+=	/lib\/.*\.so/ { sub(/${LLAMA_BUILD_NUMBER}/, "$${LLAMA_BUILD_NUMBER}") }
 PLIST_SUBST+=		LLAMA_BUILD_NUMBER=${LLAMA_BUILD_NUMBER}
 
-SUBST_CLASSES+=		findblas
-SUBST_STAGE.findblas=	pre-configure
-SUBST_MESSAGE.findblas=	Fixing libpci soname
-SUBST_FILES.findblas+=	ggml/src/ggml-blas/CMakeLists.txt
-SUBST_SED.findblas+=	-e 's,DepBLAS openblas64,DepBLAS ${BLAS_PC},'
-
 post-extract:
 	${LN} -s ${WRKDIR}/llama-${GITHUB_TAG} ${WRKSRC}/tools/ui/dist
 
 .include "../../www/curl/buildlink3.mk"
+.include "../../math/ggml/buildlink3.mk"
 .include "../../devel/cmake/build.mk"
-.include "../../mk/blas.buildlink3.mk"
 .include "../../mk/bsd.pkg.mk"
diff --git a/llama.cpp/PLIST b/llama.cpp/PLIST
index 6cbf8befd4..a8570b1856 100644
--- a/llama.cpp/PLIST
+++ b/llama.cpp/PLIST
@@ -66,35 +66,3 @@ lib/libmtmd.so
 lib/libmtmd.so.0
 lib/libmtmd.so.0.0.${LLAMA_BUILD_NUMBER}
 lib/pkgconfig/llama.pc
-include/ggml-alloc.h
-include/ggml-backend.h
-include/ggml-blas.h
-include/ggml-cann.h
-include/ggml-cpp.h
-include/ggml-cpu.h
-include/ggml-cuda.h
-include/ggml-metal.h
-include/ggml-openvino.h
-include/ggml-opt.h
-include/ggml-rpc.h
-include/ggml-sycl.h
-include/ggml-virtgpu.h
-include/ggml-vulkan.h
-include/ggml-webgpu.h
-include/ggml-zendnn.h
-include/ggml.h
-include/gguf.h
-lib/cmake/ggml/ggml-config.cmake
-lib/cmake/ggml/ggml-version.cmake
-lib/libggml-base.so
-lib/libggml-base.so.0
-lib/libggml-base.so.0.18.1
-lib/libggml-blas.so
-lib/libggml-blas.so.0
-lib/libggml-blas.so.0.18.1
-lib/libggml-cpu.so
-lib/libggml-cpu.so.0
-lib/libggml-cpu.so.0.18.1
-lib/libggml.so
-lib/libggml.so.0
-lib/libggml.so.0.18.1
Prev by Date: Update misc/CodeWhale to v.0.9.5
Next by Date: bmaptool: removed; imported into base
Previous by Thread: Update misc/CodeWhale to v.0.9.5
Next by Thread: bmaptool: removed; imported into base
Indexes:
Home | Main Index | Thread Index | Old Index