CI(consolidated): replace prebuilt-zip llama.cpp smoke with install_llama_cpp build
The previous step downloaded ggml-org/llama.cpp's release asset matching `bin-ubuntu-x64.*\.zip$` and ran the bundled binary. ggml-org changed their asset naming (the regex stopped matching), so the step was silently exiting 0 with "no ubuntu-x64 prebuilt asset on the latest llama.cpp release; skipping smoke" -- a hidden no-op. Use the canonical `unsloth_zoo.llama_cpp.install_llama_cpp` flow instead. That function clones ggml-org/llama.cpp into ~/.unsloth/llama.cpp, builds the LLAMA_CPP_TARGETS list (llama-cli, llama-quantize, llama-mtmd-cli, llama-gguf-split, llama-server) via cmake, copies build/bin/llama-* to the install root, and returns (quantizer_path, converter_script_path). It is the same path users hit at runtime via `model.save_pretrained_gguf` and friends, so the smoke now exercises the production code path instead of an unrelated prebuilt-asset download. Pre-install build deps (build-essential, cmake, libssl-dev, libcurl4-openssl-dev, libgomp1, git, curl) up-front so install_llama_cpp's check_build_requirements step is a no-op. Then verify both `llama-cli --help` and `llama-quantize --help` produce recognizable help text. Wall-time: ~3-5 min cold, dominated by cmake of 5 targets on the runner's 4 cores; well within the 35-min job timeout.
This commit is contained in:
Daniel Han
parent
d08380d6a8
commit
d45a3b4de2
1 changed files with 82 additions and 43 deletions
125
.github/workflows/consolidated-tests-ci.yml
vendored
125
.github/workflows/consolidated-tests-ci.yml
vendored
|
|
@ -599,52 +599,91 @@ jobs:
|
|||
python -m pytest -q --tb=short tests/_tiled_mlp_check_shim.py -s
|
||||
rm -f tests/_tiled_mlp_check_shim.py
|
||||
|
||||
- name: llama.cpp install + `llama-cli --help` smoke
|
||||
# The user asked to confirm llama.cpp installs and the CLI runs.
|
||||
# Studio uses prebuilt llama.cpp binaries via studio/install_llama_prebuilt.py;
|
||||
# we mirror that flow here at a smaller scale: pull the upstream prebuilt
|
||||
# release tarball for ubuntu-latest x86_64 and run `llama-cli --help`
|
||||
# to confirm the binary is intact and runnable.
|
||||
- name: llama.cpp install via unsloth_zoo.llama_cpp + `llama-cli --help` smoke
|
||||
# Exercise the canonical `unsloth_zoo.llama_cpp.install_llama_cpp`
|
||||
# flow that GGUF export uses at runtime: clone ggml-org/llama.cpp
|
||||
# into ~/.unsloth/llama.cpp, build the LLAMA_CPP_TARGETS list
|
||||
# (llama-quantize, llama-cli, llama-mtmd-cli, llama-gguf-split,
|
||||
# llama-server) via cmake, then run `llama-cli --help`.
|
||||
#
|
||||
# This replaces the previous "download upstream prebuilt zip"
|
||||
# approach, which silently exited 0 with the message
|
||||
# "no ubuntu-x64 prebuilt asset" when ggml-org's release-asset
|
||||
# naming drifted (the regex `bin-ubuntu-x64.*\.zip$` no longer
|
||||
# matched their current asset names). The build path is the same
|
||||
# one Unsloth users hit in production via `model.save_pretrained_gguf`.
|
||||
#
|
||||
# Wall-time budget: ~3-5 min cold, dominated by cmake build of
|
||||
# 5 targets on the runner's 4 cores. Apt-package install is
|
||||
# handled by `install_llama_cpp` itself via its
|
||||
# `check_build_requirements` -> `install_package` chain.
|
||||
run: |
|
||||
set -euxo pipefail
|
||||
# Try the prebuilt download path Studio uses on Linux.
|
||||
# ggml-org/llama.cpp releases ship `llama-bXXXX-bin-ubuntu-x64.zip`.
|
||||
DEST="$RUNNER_TEMP/llama-cpp"
|
||||
mkdir -p "$DEST"
|
||||
# Pin to whatever Studio's installer would resolve; for the smoke
|
||||
# we just take the latest release.
|
||||
URL=$(curl -fsSL https://api.github.com/repos/ggml-org/llama.cpp/releases/latest \
|
||||
| python -c "import sys,json,re; d=json.load(sys.stdin); \
|
||||
urls=[a['browser_download_url'] for a in d['assets'] \
|
||||
if re.search(r'bin-ubuntu-x64.*\.zip$', a['name'])]; \
|
||||
print(urls[0] if urls else '', end='')")
|
||||
if [ -z "$URL" ]; then
|
||||
echo "::warning::no ubuntu-x64 prebuilt asset on the latest llama.cpp release; skipping smoke"
|
||||
exit 0
|
||||
fi
|
||||
echo "Downloading $URL"
|
||||
curl -fsSL -o "$DEST/llama.zip" "$URL"
|
||||
(cd "$DEST" && unzip -q llama.zip)
|
||||
# Find the llama-cli binary inside the extracted tree.
|
||||
CLI=$(find "$DEST" -type f -name 'llama-cli' -perm -u+x | head -1)
|
||||
if [ -z "$CLI" ]; then
|
||||
echo "::warning::llama-cli not found in extracted prebuilt; layout drifted"
|
||||
ls -R "$DEST" | head -50
|
||||
exit 0
|
||||
fi
|
||||
# Some prebuilds need libgomp1 / libcurl4 / libssl - install just in case.
|
||||
# libssl-dev / libcurl4-openssl-dev are needed by llama.cpp's
|
||||
# cmake build for HTTPS support; install up-front so the
|
||||
# `install_llama_cpp` requirement-check is a no-op.
|
||||
sudo apt-get update -qq
|
||||
sudo apt-get install -y -qq libgomp1 libcurl4 libssl3 || true
|
||||
chmod +x "$CLI"
|
||||
# `--help` exits non-zero on some llama-cli builds; tolerate it as
|
||||
# long as the help text contains the expected sentinel.
|
||||
if "$CLI" --help 2>&1 | grep -q -E 'usage|--help|-m,|--model'; then
|
||||
echo "OK: llama-cli --help responded"
|
||||
else
|
||||
echo "::error::llama-cli --help did not produce recognizable output"
|
||||
"$CLI" --help 2>&1 | head -20 || true
|
||||
exit 1
|
||||
fi
|
||||
sudo apt-get install -y -qq build-essential cmake git curl \
|
||||
libgomp1 libssl-dev libcurl4-openssl-dev
|
||||
python <<'PY'
|
||||
import os, shutil, subprocess, sys
|
||||
from unsloth_zoo.llama_cpp import (
|
||||
install_llama_cpp,
|
||||
LLAMA_CPP_DEFAULT_DIR,
|
||||
LLAMA_CPP_TARGETS,
|
||||
)
|
||||
print(f"Unsloth llama.cpp default dir: {LLAMA_CPP_DEFAULT_DIR}")
|
||||
print(f"Build targets: {LLAMA_CPP_TARGETS}")
|
||||
# install_llama_cpp returns (quantizer_path, converter_script_path).
|
||||
# The quantizer's directory is the `llama.cpp` install root, which
|
||||
# also holds llama-cli after build/bin/llama-* gets copied up
|
||||
# (llama_cpp.py:867-871).
|
||||
quantizer, converter = install_llama_cpp(print_output=True)
|
||||
assert quantizer and os.path.exists(quantizer), (
|
||||
f"install_llama_cpp returned quantizer={quantizer!r} but file missing"
|
||||
)
|
||||
assert converter and os.path.isfile(converter), (
|
||||
f"install_llama_cpp returned converter={converter!r} but missing"
|
||||
)
|
||||
install_root = os.path.dirname(quantizer)
|
||||
cli = os.path.join(install_root, "llama-cli")
|
||||
assert os.path.exists(cli), (
|
||||
f"llama-cli not found at {cli!r} after build. Build root contents: "
|
||||
f"{sorted(p for p in os.listdir(install_root) if p.startswith('llama-'))[:20]}"
|
||||
)
|
||||
assert os.access(cli, os.X_OK), f"{cli!r} not executable"
|
||||
# `llama-cli --help` exits non-zero on some builds; the contract
|
||||
# is that recognizable help text appears on stdout/stderr.
|
||||
proc = subprocess.run(
|
||||
[cli, "--help"], capture_output=True, text=True, timeout=30,
|
||||
)
|
||||
combined = (proc.stdout or "") + (proc.stderr or "")
|
||||
print("--- llama-cli --help (first 30 lines) ---")
|
||||
print("\n".join(combined.splitlines()[:30]))
|
||||
assert any(
|
||||
tok in combined.lower()
|
||||
for tok in ("usage", "--help", "--model", "-m,")
|
||||
), (
|
||||
f"llama-cli --help produced no recognizable help text. "
|
||||
f"exit={proc.returncode}\nstdout: {proc.stdout[:400]!r}\n"
|
||||
f"stderr: {proc.stderr[:400]!r}"
|
||||
)
|
||||
# Also exercise the quantizer the way GGUF export does: --help
|
||||
# round-trip on the binary that does the actual heavy lifting.
|
||||
q = subprocess.run(
|
||||
[quantizer, "--help"], capture_output=True, text=True, timeout=15,
|
||||
)
|
||||
q_combined = (q.stdout or "") + (q.stderr or "")
|
||||
assert "usage" in q_combined.lower() or "type" in q_combined.lower(), (
|
||||
f"llama-quantize --help produced no help text. "
|
||||
f"exit={q.returncode}\nstdout: {q.stdout[:400]!r}\n"
|
||||
f"stderr: {q.stderr[:400]!r}"
|
||||
)
|
||||
print(
|
||||
f"\nOK: install_llama_cpp produced a working llama-cli at {cli} "
|
||||
f"and llama-quantize at {quantizer}."
|
||||
)
|
||||
PY
|
||||
|
||||
- name: Summary
|
||||
if: always()
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue