patches

Patches

Patches against the pinned upstream commits, one directory per engine:

llama/          llama.cpp, seven patches that rebuild the ported tree
shared-metal/   the Metal backend for whisper.cpp and stable-diffusion.cpp
whisper/        whisper.cpp, outside the Metal backend
image/          stable-diffusion.cpp, outside the Metal backend

scripts/build-engines.sh discovers them per engine and applies them in numeric order of the file name, recursively, so the number is the apply order: never reuse one, and keep it when moving a file.

llama/

0001 to 0007 are the port to upstream ca3d5a3e1, split by the files they touch. They are a baseline, not somewhere to add things: a new change gets its own numbered patch after them, one patch per change, so each one can be read, measured and reverted on its own.

	area
0001-metal-kernels	ggml/src/ggml-metal/kernels/
0002-metal-backend-host	the rest of ggml/src/ggml-metal/
0003-moe-cache	tosh-moe/, turbo quant
0004-ggml-core	ggml/ outside the Metal backend
0005-llama-core	src/
0006-common-and-spec	common/
0007-tools-and-build	tools/, tests/, top level CMake

Only touch one of the seven when the change belongs to the port itself. Together with everything numbered after them they reconstruct the built tree byte for byte, which is what the round-trip check below verifies.

The seven replaced the old per-area folders (llama/metal/, llama/core/, llama/model/, llama/server/) at the port; the running numbering they used did not go away.

shared-metal/

Whisper and stable-diffusion.cpp are pinned to commits from before upstream split ggml-metal.metal into kernels/, so they cannot take llama/0001 and llama/0002. The 0001-* series here is that same backend against the older layout, and both engines apply it unchanged.

The image engine's ggml is older still and exactly four hunks cannot land on it. They are applied from patches/image/0002 in their adapted form, and ggml-metal-impl.h needs -C2 because its decode block does not match at wider context. The build script asserts which hunks are expected to reject and stops instead of shipping a backend missing one. Do not remove that check.

Regenerating one

The vendor tree carries every patch at once, so git diff there is all of them. Because the llama/ sets are disjoint by file, one patch is its own files diffed out of the live tree:

f=patches/llama/0002-metal-backend-host.patch
files=(${(f)"$(grep '^diff --git' $f | sed 's|.* b/||')"})
git -C vendor/llama.cpp diff -U8 -- $files > $f

Three rules that have each cost real time:

• -U8 for llama/. Verified: at that width 0002, 0004, 0005 and 0006 regenerate byte for byte, and anything narrower rewrites every hunk header in the file. 0001, 0003 and 0007 also add files, which git diff will not carry on its own. Never the git diff default of three: with it a hunk lands in a different kernel and the build still exits 0.
• Check the tree is the one you think it is before regenerating. If the files modified in the vendor tree do not match the files the series touches, stop: an experiment was left applied, or a patch was left reverted, and regenerating will overwrite it.

  diff <(cat patches/llama/000*.patch | grep '^diff --git' | sed 's|.* b/||' | sort -u) \
       <(git -C vendor/llama.cpp status --short | awk '{print $2}' | sort -u)

• Verify the round trip. Add a worktree at the pinned commit, apply the whole series into it, and diff it against the live tree. No source file may differ.

scripts/build-engines.sh resets the vendor tree and re-applies the patch files, so working tree edits are discarded: regenerate before building, never after.