diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-21 08:26:31 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-21 08:26:31 +0200 |
| commit | 445e0ff83bcf13f04d0a5b6f89f89fe6ada8bebf (patch) | |
| tree | 09c7e853df4dd6fece5e3758709aad7bd6ea9503 /devtools | |
| parent | a5f67f2f828a7ac5c541ec54abf758597efcfbb6 (diff) | |
| download | ffwd-445e0ff83bcf13f04d0a5b6f89f89fe6ada8bebf.tar.gz ffwd-445e0ff83bcf13f04d0a5b6f89f89fe6ada8bebf.zip | |
build: rehome check scripts and modernize all.bash
Diffstat (limited to 'devtools')
| -rwxr-xr-x | devtools/check_code_quality.sh | 181 | ||||
| -rwxr-xr-x | devtools/container-build-test.sh | 87 | ||||
| -rwxr-xr-x | devtools/verify_builds.sh | 330 |
3 files changed, 598 insertions, 0 deletions
diff --git a/devtools/check_code_quality.sh b/devtools/check_code_quality.sh new file mode 100755 index 0000000..305ba50 --- /dev/null +++ b/devtools/check_code_quality.sh @@ -0,0 +1,181 @@ +#!/usr/bin/env bash + +set -euo pipefail + +ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd -P)" +cd "$ROOT" + +# Lint and analyze every C source under libffwd/, tests/, and bench/. Each +# platform keeps its own GPU backend and skips the other's, which it does not +# build (macOS builds MLX, Linux builds CUDA). +SYS="$(uname -s)" +SRC_SOURCES=() +for f in libffwd/*.c; do + [[ "$SYS" == Darwin && "$f" == libffwd/backend_cuda.c ]] && continue + [[ "$SYS" != Darwin && ( "$f" == libffwd/mlx.c || "$f" == libffwd/backend_mlx.c ) ]] && continue + SRC_SOURCES+=("$f") +done + +TEST_C_SOURCES=(tests/*.c) +BENCH_C_SOURCES=(bench/*.c) + +# CUDA backend sources are C++17 and only built on Linux. Parse them with the +# clang CUDA front end in host-only mode so no GPU or device toolchain is +# required; macOS has no CUDA toolkit, so skip them there. +CU_SOURCES=() +if [[ "$SYS" != Darwin ]]; then + for f in libffwd/*.cu; do + [[ -e "$f" ]] && CU_SOURCES+=("$f") + done +fi + +ANALYZE_SOURCES=("${SRC_SOURCES[@]}") +TIDY_SOURCES=("${SRC_SOURCES[@]}" "${TEST_C_SOURCES[@]}" "${BENCH_C_SOURCES[@]}") + +ANALYZE_COMMON=( + clang --analyze -Xanalyzer -analyzer-output=text + -Wall -Wextra + -I. -Ilibffwd -Ideps/ae +) + +PLATFORM_CFLAGS=() +if [[ "$(uname -s)" == "Darwin" ]]; then + PLATFORM_CFLAGS=( + -DUSE_BLAS -DACCELERATE_NEW_LAPACK -DUSE_MLX + -I/opt/homebrew/opt/mlx-c/include + -I/opt/homebrew/include + -I/usr/local/include + ) +else + PLATFORM_CFLAGS=( + -DUSE_BLAS -DUSE_OPENBLAS + -I/usr/include/openblas + ) +fi + +# Flags for the CUDA backend front end. CUDA_HOME matches the Makefile default. +# --cuda-host-only parses host code only so device codegen and libdevice are not +# needed; the macros mirror the nvcc command in the Makefile. +CUDA_HOME="${CUDA_HOME:-/usr/local/cuda}" +CUDA_CFLAGS=( + -x cuda --cuda-host-only "--cuda-path=$CUDA_HOME" + -std=c++17 + -DUSE_BLAS -DUSE_OPENBLAS -DUSE_GPU + -I"$CUDA_HOME/include" + -I/usr/include/openblas +) + +print_cmd() { + printf '+' + printf ' %q' "$@" + printf '\n' +} + +run() { + print_cmd "$@" + "$@" +} + +run_analyze() { + local output + local status + + print_cmd "$@" + + set +e + output="$("$@" 2>&1)" + status=$? + set -e + + if [[ -n "$output" ]]; then + printf '%s\n' "$output" + fi + + if (( status != 0 )); then + return "$status" + fi + + if grep -q 'warning:' <<<"$output"; then + echo "clang analyzer warnings found" >&2 + return 1 + fi +} + +have() { + command -v "$1" >/dev/null 2>&1 +} + +find_clang_tidy() { + local brew_path="/opt/homebrew/opt/llvm@21/bin/clang-tidy" + + if command -v clang-tidy >/dev/null 2>&1; then + command -v clang-tidy + return 0 + fi + + if [[ -x "$brew_path" ]]; then + printf '%s\n' "$brew_path" + return 0 + fi + + return 1 +} + +main() { + local clang_tidy + + if ! have clang; then + echo "clang not found" >&2 + return 1 + fi + + run git --no-pager diff --check + run make cpu CC=clang + + run_analyze "${ANALYZE_COMMON[@]}" "${PLATFORM_CFLAGS[@]}" \ + "${ANALYZE_SOURCES[@]}" + run_analyze "${ANALYZE_COMMON[@]}" "${PLATFORM_CFLAGS[@]}" \ + -I. -Ilibffwd -Itests "${TEST_C_SOURCES[@]}" + run_analyze "${ANALYZE_COMMON[@]}" "${PLATFORM_CFLAGS[@]}" \ + -I. -Ilibffwd -Ibench "${BENCH_C_SOURCES[@]}" + + if (( ${#CU_SOURCES[@]} > 0 )); then + run_analyze "${ANALYZE_COMMON[@]}" "${CUDA_CFLAGS[@]}" \ + "${CU_SOURCES[@]}" + fi + + if clang_tidy="$(find_clang_tidy)"; then + run \ + "$clang_tidy" \ + --config-file=.clang-tidy \ + "${TIDY_SOURCES[@]}" \ + -- \ + -std=c17 \ + "${PLATFORM_CFLAGS[@]}" \ + -I. \ + -Ilibffwd \ + -Itests \ + -Ibench \ + -Ideps/ae + + if (( ${#CU_SOURCES[@]} > 0 )); then + run \ + "$clang_tidy" \ + --config-file=.clang-tidy \ + "${CU_SOURCES[@]}" \ + -- \ + "${CUDA_CFLAGS[@]}" \ + -I. \ + -Ilibffwd \ + -Ideps/ae + fi + else + echo "clang-tidy not found; skipping" + fi + + if [[ "$(uname -s)" == "Darwin" ]]; then + run make gpu CC=clang + fi +} + +main "$@" diff --git a/devtools/container-build-test.sh b/devtools/container-build-test.sh new file mode 100755 index 0000000..4dcc627 --- /dev/null +++ b/devtools/container-build-test.sh @@ -0,0 +1,87 @@ +#!/usr/bin/env bash +# Build, test and bench the OpenBLAS backend in a throwaway Ubuntu container to +# catch Linux portability breakage a macOS/Accelerate build hides. Needs Apple's +# `container` CLI running. Run from anywhere in the repo: it finds the repo root +# via git, bind-mounts it at /work, installs the Runpod toolchain, re-execs +# itself inside with --in-container, and runs make cpu/test/bench/bench-server- +# utils/debug, then `make clean` (skip with KEEP=1). Override the image with +# IMAGE=ubuntu:22.04. Exits non-zero at the first failing step. +set -euo pipefail + +IMAGE="${IMAGE:-ubuntu:24.04}" + +# --- inside the container: bootstrap toolchain, then build/test --------------- +run_in_container() { + export DEBIAN_FRONTEND=noninteractive + + echo "==================== apt bootstrap ====================" + apt-get update -qq + apt-get install -y --no-install-recommends \ + build-essential ca-certificates libcjson-dev libopenblas-dev \ + pkg-config git >/dev/null + + echo "==================== system / toolchain ====================" + uname -a + cc --version | sed -n '1p' + echo "libcjson $(pkg-config --modversion libcjson 2>/dev/null || echo MISSING) :: $(pkg-config --cflags --libs libcjson 2>/dev/null)" + echo "openblas $(pkg-config --modversion openblas 2>/dev/null || echo MISSING) :: $(pkg-config --cflags --libs openblas 2>/dev/null)" + + cd /work + # /work is a bind mount; git refuses to operate on it without this opt-in. + git config --global --add safe.directory /work 2>/dev/null || true + make clean >/dev/null 2>&1 || true + + echo "==================== make cpu ====================" + make cpu + ls -la ffwd-cli ffwd-server + ./ffwd-cli --version || true + + echo "==================== make test ====================" + make test + + echo "==================== make bench ====================" + make bench + + echo "==================== make bench-server-utils ====================" + make bench-server-utils + + echo "==================== make debug (sanitizers) ====================" + make debug + + if [ -n "${KEEP:-}" ]; then + echo "==================== KEEP set: leaving artifacts in tree ====================" + else + echo "==================== make clean ====================" + make clean + residual=$(ls -d build ffwd-cli ffwd-server 2>/dev/null | wc -l | tr -d ' ') + echo "residual build artifacts: $residual (want 0)" + fi + + echo "==================== ALL STEPS COMPLETED ====================" +} + +# --- host side: launch the container and re-exec this script inside it -------- +launch_on_host() { + command -v container >/dev/null 2>&1 || { + echo "error: Apple 'container' CLI not found on PATH" >&2; exit 1; } + + local repo_root script_rel + repo_root="$(git rev-parse --show-toplevel)" + # Path of this script relative to the repo root, so it resolves under /work. + script_rel="$(cd "$(dirname "$0")" && pwd)/$(basename "$0")" + script_rel="${script_rel#"$repo_root"/}" + + echo "repo: $repo_root" + echo "image: $IMAGE" + exec container run --rm \ + --env "KEEP=${KEEP:-}" \ + --mount "source=$repo_root,target=/work" \ + "$IMAGE" \ + bash "/work/$script_rel" --in-container +} + +if [ "${1:-}" = "--in-container" ]; then + run_in_container +else + launch_on_host +fi diff --git a/devtools/verify_builds.sh b/devtools/verify_builds.sh new file mode 100755 index 0000000..25347b5 --- /dev/null +++ b/devtools/verify_builds.sh @@ -0,0 +1,330 @@ +#!/usr/bin/env bash + +set -euo pipefail + +ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd -P)" + +DEFAULT_MODEL_GLOB='models--perplexity-ai--pplx-embed-v1-0.6b/snapshots/*' +DEFAULT_CONTEXT_MODEL_GLOB='models--perplexity-ai--pplx-embed-context-v1-0.6B/snapshots/*' + +model_dir="" +context_model_dir="" +binary="./ffwd-cli" +batch_size=4 +skip_build=0 +skip_mlx=0 +platform_skip_mlx=0 +skip_contextual=0 +context_runs=0 +late_model_dir="" +long=0 +skip_code_quality=0 +skip_hermetic=0 +skip_tokenizer=0 +skip_mlx_quantized=0 + +die() { + echo "error: $*" >&2 + exit 2 +} + +usage() { + cat <<EOF +Usage: $(basename "$0") [options] + +Options: + --model-dir DIR 0.6B model snapshot; defaults to the HF cache + --context-model-dir DIR context 0.6B snapshot; defaults to the HF cache + --binary PATH binary to test, default: ./ffwd-cli + --batch-size N batch size, default: 4 + --skip-build skip make cpu / make gpu + --skip-hermetic skip the model-free make test suite + --skip-mlx stop after the CPU build and CPU checks + --skip-contextual skip contextual parity checks + --skip-tokenizer skip Hugging Face tokenizer parity + --skip-mlx-quantized skip the MLX Q8 parity check + --context-runs N timed contextual batches per backend after parity + --late-model-dir DIR optional late-interaction 0.6B snapshot to verify + --long also run the long CPU attention parity check + --skip-code-quality skip the final compiler/analyzer/tidy CI gate + -h, --help show this help +EOF +} + +find_cached_snapshot() { + local glob="$1" + local label="$2" + local hub="$HOME/.cache/huggingface/hub" + local last="" + local p + + while IFS= read -r p; do + if [[ -d "$p" && -f "$p/config.json" ]]; then + last="$p" + fi + done < <({ compgen -G "$hub/$glob" || true; } | sort) + + if [[ -z "$last" ]]; then + echo "no cached $label snapshot found; pass the model dir" >&2 + exit 1 + fi + + printf '%s\n' "$last" +} + +find_default_model() { + find_cached_snapshot "$DEFAULT_MODEL_GLOB" "pplx-embed-v1-0.6b" +} + +find_default_context_model() { + find_cached_snapshot "$DEFAULT_CONTEXT_MODEL_GLOB" "pplx-embed-context-v1-0.6b" +} + +run() { + printf '+' + printf ' %q' "$@" + printf '\n' + (cd "$ROOT" && "$@") +} + +run_parity() { + local binary_arg="$1" + local model_dir_arg="$2" + local backend="$3" + local batch_size_arg="$4" + local use_long="$5" + + local cmd=( + "$ROOT/all.bash" check-batch-parity + --binary "$binary_arg" + --model-dir "$model_dir_arg" + --backend "$backend" + --batch-size "$batch_size_arg" + ) + + if (( use_long )); then + cmd+=(--long) + fi + + run "${cmd[@]}" +} + +run_contextual() { + local model_dir_arg="$1" + local backend="$2" + local runs="$3" + + run \ + "$ROOT/all.bash" check-context-batch-parity \ + --model-dir "$model_dir_arg" \ + --backend "$backend" \ + --runs "$runs" +} + +run_late() { + local model_dir_arg="$1" + local backend="$2" + + run \ + "$ROOT/all.bash" check-late-interaction \ + --model-dir "$model_dir_arg" \ + --backend "$backend" +} + +while (($#)); do + case "$1" in + --model-dir) + (($# >= 2)) || die "--model-dir requires a value" + model_dir="$2" + shift 2 + ;; + --model-dir=*) + model_dir="${1#*=}" + shift + ;; + + --context-model-dir) + (($# >= 2)) || die "--context-model-dir requires a value" + context_model_dir="$2" + shift 2 + ;; + --context-model-dir=*) + context_model_dir="${1#*=}" + shift + ;; + + --binary) + (($# >= 2)) || die "--binary requires a value" + binary="$2" + shift 2 + ;; + --binary=*) + binary="${1#*=}" + shift + ;; + + --batch-size) + (($# >= 2)) || die "--batch-size requires a value" + batch_size="$2" + shift 2 + ;; + --batch-size=*) + batch_size="${1#*=}" + shift + ;; + + --skip-build) + skip_build=1 + shift + ;; + --skip-hermetic) + skip_hermetic=1 + shift + ;; + --skip-mlx) + skip_mlx=1 + shift + ;; + --skip-contextual) + skip_contextual=1 + shift + ;; + --skip-tokenizer) + skip_tokenizer=1 + shift + ;; + --skip-mlx-quantized) + skip_mlx_quantized=1 + shift + ;; + + --context-runs) + (($# >= 2)) || die "--context-runs requires a value" + context_runs="$2" + shift 2 + ;; + --context-runs=*) + context_runs="${1#*=}" + shift + ;; + + --late-model-dir) + (($# >= 2)) || die "--late-model-dir requires a value" + late_model_dir="$2" + shift 2 + ;; + --late-model-dir=*) + late_model_dir="${1#*=}" + shift + ;; + + --long) + long=1 + shift + ;; + --skip-code-quality) + skip_code_quality=1 + shift + ;; + + -h|--help) + usage + exit 0 + ;; + + *) + die "unknown argument: $1" + ;; + esac +done + +if [[ "$(uname -s)" != "Darwin" && $skip_mlx -eq 0 ]]; then + echo "info: MLX checks require macOS; skipping on $(uname -s)" + skip_mlx=1 + platform_skip_mlx=1 +fi + +if [[ -z "$model_dir" ]]; then + model_dir="$(find_default_model)" +fi + +if (( skip_contextual )); then + context_model_dir="" +elif [[ -z "$context_model_dir" ]]; then + context_model_dir="$(find_default_context_model)" +fi + +if (( ! skip_hermetic )); then + run make test +fi + +if (( ! skip_build )); then + run make cpu +fi + +run "$ROOT/all.bash" check-kernel-golden --blas + +if (( ! skip_tokenizer )); then + run "$ROOT/all.bash" check-tokenizer-parity --model-dir "$model_dir" +fi + +run "$ROOT/all.bash" check-workspace-api --model-dir "$model_dir" + +run_parity "$binary" "$model_dir" cpu "$batch_size" 0 + +if (( long )); then + run_parity "$binary" "$model_dir" cpu "$batch_size" 1 +fi + +if [[ -n "$context_model_dir" ]]; then + run_contextual "$context_model_dir" cpu "$context_runs" +fi + +if [[ -n "$late_model_dir" ]]; then + run_late "$late_model_dir" cpu +fi + +if (( skip_mlx )); then + if (( platform_skip_mlx && ! skip_code_quality )); then + run "$ROOT/all.bash" check-code-quality + fi + if (( skip_build )); then + echo "ok: CPU checks passed; builds and MLX checks skipped" + else + echo "ok: CPU build verified; MLX build skipped" + fi + exit 0 +fi + +if (( ! skip_build )); then + run make gpu +fi + +run_parity "$binary" "$model_dir" mlx "$batch_size" 0 + +run \ + "$ROOT/all.bash" check-mlx-model-limits \ + --binary "$binary" \ + --model-dir "$model_dir" + +if (( ! skip_mlx_quantized )); then + run \ + "$ROOT/all.bash" check-mlx-quantized-parity \ + --binary "$binary" \ + --model-dir "$model_dir" \ + --batch-size "$batch_size" +fi + +if [[ -n "$context_model_dir" ]]; then + run_contextual "$context_model_dir" mlx "$context_runs" +fi + +if [[ -n "$late_model_dir" ]]; then + run_late "$late_model_dir" mlx +fi + +if (( ! skip_code_quality )); then + run "$ROOT/all.bash" check-code-quality +fi + +echo "verify ok" |
