Your Name
0b5043a9f3
gates / consistency-and-conventions (push) Successful in 26s
User report after the llamacpp app install: 'installed llama.cpp unknown', valid flags rejected (randomly per run), 'Model not found' for the HF downloader's own layout, and a systemd user-bus failure over SSH. Detective (real b10822 binary, FACT) found four independent causes: - version: llama-server --version prints to STDERR; detect_llama_version's 2>/dev/null swallowed it -> always 'unknown'. Now captures 2>&1 + accepts semver/build tokens (incl. build 1.2.3 edge) - validation: printf|grep -q under pipefail -> SIGPIPE rc=141 race randomly rejected flags present in the 59 KB --help. Now pipe-less grep (no race); 20x determinism regression test - model resolution: resolve_model accepted files only, but the HF downloader creates <models>/<repo>/file.gguf dirs. Now expands a dir with exactly one *.gguf (never silently picks; multi-gguf lists + errs) - port: llama.cpp default 8080 vs tool/adapter 8088; validation reliability means --port is now always pinned in the unit - user bus: headless/SSH sessions lack XDG_RUNTIME_DIR -> ensure_user_bus in lib/common.sh pre-flights all three systemctl --user tools with remediation text; pos ai server --no-unit direct-run escape hatch (pidfile) for boxes with no bus - find_llamacpp narrowed to llama-server/llama-server-cuda (bare 'server' fallback hazard); installer post-install sanity (version+help execute, symlink targets resolve) Architect decisions DQ1-DQ6 recorded. Tester: 4 new regression files (version-from-stderr, 25x flag-validation determinism, model dir expansion, bus pre-flight + E2E) + 3 fixture updates; suite 16 files / 269 checks. Verified: make gen idempotent; make check OK; make lint 0 FAIL, 0 WARN; make test 269/269 (~49s); bash -n clean; git diff --check clean.
108 lines
5.5 KiB
Bash
108 lines
5.5 KiB
Bash
#!/usr/bin/env bash
|
|
set -euo pipefail
|
|
# t-ai-server-flags.sh — pos-ai-server argument generation: flag set from
|
|
# CLI-explicit vs config vs defaults; dedupe; every emitted flag comes from a
|
|
# validated source. Behavior asserted through the tool's own dry-run ExecStart
|
|
# output (no unit write, no systemd).
|
|
|
|
run_test() {
|
|
local sbin sandbox stubs cfg models
|
|
sandbox="$(mksandbox ai-server-flags)"
|
|
stubs="$sandbox/stubs"
|
|
cfg="$sandbox/cfg"
|
|
models="$sandbox/models"
|
|
mkdir -p "$stubs" "$cfg" "$models"
|
|
: > "$cfg/ai.env"
|
|
touch "$models/my-model.gguf"
|
|
|
|
# ── stub llama-server: version + full help; nvidia-smi forced to fail so
|
|
# GPU detection deterministically resolves to "cpu" → --n-gpu-layers 0.
|
|
cat > "$stubs/llama-server" <<'STUB'
|
|
#!/usr/bin/env bash
|
|
case "$1" in
|
|
--version) echo "llama.cpp 1.2.3" ;;
|
|
--help)
|
|
cat <<'HELP'
|
|
usage: llama-server [options]
|
|
options:
|
|
--host <addr> bind address
|
|
--port <port> server port
|
|
--n-gpu-layers <n> layers to offload
|
|
--ctx-size <n> context size
|
|
--threads <n> cpu threads
|
|
--mmap memory mapping
|
|
--tensor-split <n> tensor split
|
|
HELP
|
|
;;
|
|
esac
|
|
STUB
|
|
cat > "$stubs/nvidia-smi" <<'STUB'
|
|
#!/usr/bin/env bash
|
|
exit 1
|
|
STUB
|
|
# systemctl: succeed — F4's ensure_user_bus pre-flight must pass so the
|
|
# dry-run reaches the ExecStart output this test asserts on (bus-missing
|
|
# behavior is covered separately in t-ai-server-bus.sh).
|
|
printf '#!/usr/bin/env bash\nexit 0\n' > "$stubs/systemctl"
|
|
chmod +x "$stubs/llama-server" "$stubs/nvidia-smi" "$stubs/systemctl"
|
|
|
|
local server="$ROOT/bin/pos-ai-server"
|
|
local base_env=(PATH="$stubs:/usr/bin:/bin" DRY_RUN=1
|
|
CONFIG_FILE="$cfg/ai.env" USER_SYSTEMD_DIR="$sandbox/userunits")
|
|
|
|
# 1. defaults: no CLI flags, no config → the always-emitted default set.
|
|
test_run_env "${base_env[@]}" -- "$server" start "$models/my-model.gguf"
|
|
check_rc "start with no flags exits 0" 0 "$TR_RC"
|
|
local exec_line
|
|
exec_line="$(printf '%s\n' "$TR_OUT" | grep 'dry-run) ExecStart:' | sed 's/.*ExecStart: //')"
|
|
check_contains "dry-run ExecStart printed" "ExecStart:" "$TR_OUT"
|
|
check_contains "default --port 8088" "--port 8088" "$exec_line"
|
|
check_contains "default --host 127.0.0.1" "--host 127.0.0.1" "$exec_line"
|
|
check_contains "default --n-gpu-layers 0 (cpu)" "--n-gpu-layers 0" "$exec_line"
|
|
check_contains "default --ctx-size 4096" "--ctx-size 4096" "$exec_line"
|
|
check_contains "default --threads emitted" "--threads" "$exec_line"
|
|
check_contains "model path present via -m" "-m \"$models/my-model.gguf\"" "$exec_line"
|
|
|
|
# 2. CLI-explicit flags override defaults and are emitted.
|
|
test_run_env "${base_env[@]}" -- "$server" start "$models/my-model.gguf" --port 9090 --ctx-size 2048
|
|
check_rc "start with CLI flags exits 0" 0 "$TR_RC"
|
|
exec_line="$(printf '%s\n' "$TR_OUT" | grep 'dry-run) ExecStart:' | sed 's/.*ExecStart: //')"
|
|
check_contains "CLI --port 9090 emitted" "--port 9090" "$exec_line"
|
|
check_contains "CLI --ctx-size 2048 emitted" "--ctx-size 2048" "$exec_line"
|
|
check_not_contains "CLI --port wins over default 8088" "--port 8088" "$exec_line"
|
|
|
|
# 3. dedupe: --ctx and --ctx-size map to the same canonical flag → one token
|
|
# (validate_requested_flags dedupes alias-mapped flags).
|
|
test_run_env "${base_env[@]}" -- "$server" start "$models/my-model.gguf" --ctx 1024 --ctx-size 2048
|
|
check_rc "start with alias pair exits 0" 0 "$TR_RC"
|
|
exec_line="$(printf '%s\n' "$TR_OUT" | grep 'dry-run) ExecStart:' | sed 's/.*ExecStart: //')"
|
|
check_eq "--ctx-size emitted exactly once after dedupe" 1 "$(count_token --ctx-size "$exec_line")"
|
|
check_contains "dedupe keeps last value 2048" "--ctx-size 2048" "$exec_line"
|
|
|
|
# 4. config-sourced values are emitted (no CLI).
|
|
printf 'LLAMACPP_CTX_SIZE=512\nLLAMACPP_PORT=9999\n' > "$cfg/ai.env"
|
|
test_run_env "${base_env[@]}" -- "$server" start "$models/my-model.gguf"
|
|
check_rc "start with config exits 0" 0 "$TR_RC"
|
|
exec_line="$(printf '%s\n' "$TR_OUT" | grep 'dry-run) ExecStart:' | sed 's/.*ExecStart: //')"
|
|
check_contains "config LLAMACPP_CTX_SIZE=512 emitted" "--ctx-size 512" "$exec_line"
|
|
check_contains "config LLAMACPP_PORT=9999 emitted" "--port 9999" "$exec_line"
|
|
|
|
# 5. CLI beats config (precedence contract: CLI > env > file > defaults).
|
|
test_run_env "${base_env[@]}" -- "$server" start "$models/my-model.gguf" --port 1234
|
|
check_rc "start with CLI over config exits 0" 0 "$TR_RC"
|
|
exec_line="$(printf '%s\n' "$TR_OUT" | grep 'dry-run) ExecStart:' | sed 's/.*ExecStart: //')"
|
|
check_contains "CLI --port 1234 beats config 9999" "--port 1234" "$exec_line"
|
|
check_not_contains "config port 9999 omitted when CLI given" "--port 9999" "$exec_line"
|
|
|
|
# 6. only validated flags emitted: every default + requested flag appears
|
|
# exactly once in ExecStart.
|
|
test_run_env "${base_env[@]}" -- "$server" start "$models/my-model.gguf" --mmap
|
|
check_rc "start with --mmap exits 0" 0 "$TR_RC"
|
|
exec_line="$(printf '%s\n' "$TR_OUT" | grep 'dry-run) ExecStart:' | sed 's/.*ExecStart: //')"
|
|
check_eq "--port exactly once" 1 "$(count_token --port "$exec_line")"
|
|
check_eq "--host exactly once" 1 "$(count_token --host "$exec_line")"
|
|
check_eq "--n-gpu-layers exactly once" 1 "$(count_token --n-gpu-layers "$exec_line")"
|
|
check_eq "--ctx-size exactly once" 1 "$(count_token --ctx-size "$exec_line")"
|
|
check_eq "--threads exactly once" 1 "$(count_token --threads "$exec_line")"
|
|
check_eq "--mmap exactly once" 1 "$(count_token --mmap "$exec_line")"
|
|
} |