Add arch-ab.sh: cross-rig KV-dtype A/B runner for #246 (lean tier)
One command per volunteer rig: runs the pilot variant once per KV-dtype arm (fresh symmetric boot each), verify-full + bench n=5 + NIAH ladder per arm -- soak and quality packs deliberately skipped (quality consolidates at promotion time per the #246 test plan; ~25-30 min/arm). Ends with a per-arm comparison table and ONE attachable tarball. Arms: e5m2 (control) / e4m3 (native FP8 KV, sm_89+) / nvfp4 (Blackwell opt-in, refused below sm 10.0) / fp8w (vllm/qwen-27b-dual-max STOCK -- FP8-weights checkpoints reject fp8 KV, so this arm measures the native-FP8-WEIGHTS lift instead; dual-rig only). Guard rails: variant auto-pick by GPU count (2+ -> vllm/dual, 1 -> vllm/minimal); KV arms are explicit KV_CACHE_DTYPE pins (unambiguous on any launcher version); the running container's --kv-cache-dtype is asserted live before any bench time is spent; --dry-run / --resume / --full escape hatch. test-arch-ab: hermetic plan/refusal contract via CLUB3090_FAKE_GPUS. Full scripts gate 65/65. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EfF565T9eSLaqGzidyJ1Pm
This commit is contained in:
noonghunna
195
scripts/arch-ab.sh
Executable file
195
scripts/arch-ab.sh
Executable file
@@ -0,0 +1,195 @@
|
||||
#!/usr/bin/env bash
|
||||
#
|
||||
# arch-ab.sh — cross-rig KV-dtype A/B runner for issue #246 (lean tier).
|
||||
#
|
||||
# Runs the SAME pilot variant once per KV-dtype arm, each arm on a fresh
|
||||
# symmetric boot, with the protocol bench + NIAH recall ladder — and bundles
|
||||
# everything into one tarball to attach to the issue. Deliberately NO quality
|
||||
# packs and NO soak: quality validation is consolidated at promotion time
|
||||
# (issue #246 test plan), so a full run is ~25-30 min/arm, not hours.
|
||||
#
|
||||
# Usage:
|
||||
# bash scripts/arch-ab.sh # arms e5m2,e4m3; variant by GPU count
|
||||
# bash scripts/arch-ab.sh --arms e5m2,e4m3,nvfp4 # Blackwell (sm>=10) opt-in third arm
|
||||
# bash scripts/arch-ab.sh --variant vllm/dual # force the variant
|
||||
# bash scripts/arch-ab.sh --dry-run # print the plan, run nothing
|
||||
# bash scripts/arch-ab.sh --resume # skip arms/steps with artifacts
|
||||
# bash scripts/arch-ab.sh --full # + soak + 8-pack (long; NOT the ask)
|
||||
#
|
||||
# Variant auto-pick: 2+ GPUs -> vllm/dual (TP=2), 1 GPU -> vllm/minimal.
|
||||
# Arms:
|
||||
# e5m2 fp8_e5m2 KV on the pilot variant (control — today's shipped default)
|
||||
# e4m3 fp8_e4m3 KV on the pilot variant (native FP8 KV compute on sm_89+)
|
||||
# nvfp4 nvfp4 KV on the pilot variant (Blackwell-only, UNVALIDATED — refuse
|
||||
# below sm 10.0)
|
||||
# fp8w vllm/qwen-27b-dual-max STOCK (FP8 weights + int8-PTH KV, no dtype
|
||||
# override; dual-rig only) — measures the native-FP8-WEIGHTS lift on
|
||||
# sm_89+ vs Ampere's Marlin-dequant path. NOTE: fp8 KV is rejected on
|
||||
# FP8-weights checkpoints, which is why the KV arms can't run there.
|
||||
# KV arms are EXPLICIT KV_CACHE_DTYPE pins, so what ran is never ambiguous
|
||||
# regardless of launcher-injection state.
|
||||
#
|
||||
# Outputs: results/rebench/246-ab-<arm>/ per arm + a single
|
||||
# results/rebench/246-ab-bundle-<host>-<date>.tgz to attach to issue #246.
|
||||
set -euo pipefail
|
||||
|
||||
ROOT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
|
||||
cd "$ROOT_DIR"
|
||||
# shellcheck source=lib/compose-meta.sh
|
||||
source "${ROOT_DIR}/scripts/lib/compose-meta.sh"
|
||||
|
||||
die() { echo "[arch-ab] ERROR: $*" >&2; exit 1; }
|
||||
|
||||
ARMS="e5m2,e4m3"
|
||||
VARIANT=""
|
||||
DRY_RUN=0
|
||||
RESUME=0
|
||||
FULL=0
|
||||
while [[ $# -gt 0 ]]; do
|
||||
case "$1" in
|
||||
--arms) ARMS="$2"; shift 2 ;;
|
||||
--arms=*) ARMS="${1#*=}"; shift ;;
|
||||
--variant) VARIANT="$2"; shift 2 ;;
|
||||
--variant=*) VARIANT="${1#*=}"; shift ;;
|
||||
--dry-run) DRY_RUN=1; shift ;;
|
||||
--resume) RESUME=1; shift ;;
|
||||
--full) FULL=1; shift ;;
|
||||
-h|--help) awk 'NR>1 && /^set -euo/{exit} NR>1{sub(/^# ?/,""); print}' "${BASH_SOURCE[0]}"; exit 0 ;;
|
||||
*) die "unknown argument: $1 (see --help)" ;;
|
||||
esac
|
||||
done
|
||||
|
||||
# KV arms pin a dtype on the pilot variant; the fp8w arm swaps the VARIANT
|
||||
# (stock config, no pin) and asserts its registry KV format is live instead.
|
||||
declare -A ARM_DTYPE=([e5m2]=fp8_e5m2 [e4m3]=fp8_e4m3 [nvfp4]=nvfp4 [fp8w]="")
|
||||
declare -A ARM_VARIANT_OVERRIDE=([fp8w]="vllm/qwen-27b-dual-max")
|
||||
declare -A ARM_ASSERT_DTYPE=([fp8w]="int8_per_token_head")
|
||||
|
||||
# --- detect GPUs -------------------------------------------------------------
|
||||
GPU_LINES="$(compose_hw_detect_gpus 2>/dev/null || true)"
|
||||
[[ -n "$GPU_LINES" ]] || die "no NVIDIA GPUs detected (nvidia-smi missing or empty)"
|
||||
GPU_COUNT=0
|
||||
MIN_SM=999
|
||||
while IFS=$'\t' read -r idx name mem sm; do
|
||||
[[ -z "$idx" ]] && continue
|
||||
GPU_COUNT=$((GPU_COUNT + 1))
|
||||
awk -v a="$sm" -v b="$MIN_SM" 'BEGIN{exit !(a<b)}' && MIN_SM="$sm"
|
||||
done <<< "$GPU_LINES"
|
||||
|
||||
# --- resolve variant + validate arms ----------------------------------------
|
||||
if [[ -z "$VARIANT" ]]; then
|
||||
if (( GPU_COUNT >= 2 )); then VARIANT="vllm/dual"; else VARIANT="vllm/minimal"; fi
|
||||
fi
|
||||
[[ "$VARIANT" == "vllm/dual" && "$GPU_COUNT" -lt 2 ]] \
|
||||
&& die "vllm/dual is TP=2 and needs 2 GPUs (detected ${GPU_COUNT}); use --variant vllm/minimal"
|
||||
|
||||
IFS=',' read -ra ARM_LIST <<< "$ARMS"
|
||||
[[ "${#ARM_LIST[@]}" -ge 1 ]] || die "no arms given"
|
||||
for arm in "${ARM_LIST[@]}"; do
|
||||
[[ -n "${ARM_DTYPE[$arm]+x}" ]] || die "unknown arm '${arm}' (valid: e5m2, e4m3, nvfp4, fp8w)"
|
||||
if [[ "$arm" == "nvfp4" ]]; then
|
||||
awk -v a="$MIN_SM" 'BEGIN{exit !(a>=10.0)}' \
|
||||
|| die "arm 'nvfp4' needs Blackwell (sm>=10.0); detected min sm_${MIN_SM}. Drop it: --arms e5m2,e4m3"
|
||||
fi
|
||||
if [[ "$arm" == "fp8w" && "$GPU_COUNT" -lt 2 ]]; then
|
||||
die "arm 'fp8w' runs vllm/qwen-27b-dual-max (TP=2) and needs 2 GPUs (detected ${GPU_COUNT})"
|
||||
fi
|
||||
done
|
||||
|
||||
# --- plan --------------------------------------------------------------------
|
||||
echo "[arch-ab] plan: variant=${VARIANT} arms=${ARMS} gpus=${GPU_COUNT} (min sm_${MIN_SM})"
|
||||
echo "[arch-ab] per arm: fresh boot -> verify-full + bench + verify-stress$( ((FULL)) && echo ' + soak + 8-pack both (--full)' || echo ' (soak/quality skipped — lean tier)')"
|
||||
for arm in "${ARM_LIST[@]}"; do
|
||||
arm_variant="${ARM_VARIANT_OVERRIDE[$arm]:-$VARIANT}"
|
||||
if [[ -n "${ARM_DTYPE[$arm]}" ]]; then
|
||||
echo "[arch-ab] arm ${arm}: ${arm_variant} + KV_CACHE_DTYPE=${ARM_DTYPE[$arm]} -> results/rebench/246-ab-${arm}/"
|
||||
else
|
||||
echo "[arch-ab] arm ${arm}: ${arm_variant} STOCK (no dtype pin) -> results/rebench/246-ab-${arm}/"
|
||||
fi
|
||||
done
|
||||
if (( DRY_RUN )); then
|
||||
echo "[arch-ab] dry run — nothing executed."
|
||||
exit 0
|
||||
fi
|
||||
|
||||
# --- container names for the live-dtype assertion (registry-derived) ---------
|
||||
declare -A VARIANT_DEFAULT_PORT=() VARIANTS=() VARIANT_STATUS=() VARIANT_STATUS_NOTE=() VARIANT_CONTAINER=()
|
||||
# shellcheck source=lib/registry-emit.sh
|
||||
source "${ROOT_DIR}/scripts/lib/registry-emit.sh"
|
||||
derive_switch_variant_tables "${ROOT_DIR}"
|
||||
|
||||
# --- run arms ----------------------------------------------------------------
|
||||
for arm in "${ARM_LIST[@]}"; do
|
||||
dtype="${ARM_DTYPE[$arm]}"
|
||||
arm_variant="${ARM_VARIANT_OVERRIDE[$arm]:-$VARIANT}"
|
||||
assert_dtype="${ARM_ASSERT_DTYPE[$arm]:-$dtype}"
|
||||
container="${VARIANT_CONTAINER[$arm_variant]:-}"
|
||||
[[ -n "$container" ]] || die "no container mapping for ${arm_variant} in the registry"
|
||||
tag="246-ab-${arm}"
|
||||
echo ""
|
||||
echo "[arch-ab] ===== arm ${arm}: ${arm_variant}$( [[ -n "$dtype" ]] && echo " KV_CACHE_DTYPE=${dtype}" || echo " (stock)") -> tag ${tag} ====="
|
||||
# Fresh symmetric boot per arm (same settle path for every arm). KV arms use
|
||||
# an explicit env pin so what ran is unambiguous on any launcher version;
|
||||
# the fp8w arm runs stock (its checkpoint REJECTS fp8 KV — see header).
|
||||
if [[ -n "$dtype" ]]; then
|
||||
KV_CACHE_DTYPE="$dtype" bash scripts/switch.sh "$arm_variant"
|
||||
else
|
||||
bash scripts/switch.sh "$arm_variant"
|
||||
fi
|
||||
# Assert the expected dtype is actually live before spending bench time on it.
|
||||
running_cmd="$(docker inspect "$container" --format '{{join .Config.Cmd " "}}' 2>/dev/null || true)"
|
||||
case "$running_cmd" in
|
||||
*"--kv-cache-dtype ${assert_dtype}"*) echo "[arch-ab] verified: container runs --kv-cache-dtype ${assert_dtype}" ;;
|
||||
*) die "arm ${arm}: container '${container}' is NOT running --kv-cache-dtype ${assert_dtype} — aborting before benching the wrong config" ;;
|
||||
esac
|
||||
rb_args=(--tag "$tag")
|
||||
if (( FULL )); then
|
||||
rb_args+=(--with-8pack-thinking=both)
|
||||
else
|
||||
rb_args+=(--skip soak)
|
||||
fi
|
||||
(( RESUME )) && rb_args+=(--resume)
|
||||
bash scripts/rebench-full.sh "${rb_args[@]}"
|
||||
done
|
||||
|
||||
# --- summary + bundle ---------------------------------------------------------
|
||||
echo ""
|
||||
echo "[arch-ab] ===== summary ====="
|
||||
ARMS="$ARMS" python3 - <<'PY'
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
from pathlib import Path
|
||||
|
||||
arms = [a for a in os.environ["ARMS"].split(",") if a]
|
||||
print(f"{'arm':<8} {'narr TPS':>9} {'code TPS':>9} {'TTFT ms':>8} NIAH ladder")
|
||||
for arm in arms:
|
||||
tag = Path(f"results/rebench/246-ab-{arm}")
|
||||
try:
|
||||
bench = json.loads((tag / "_internal.json").read_text()).get("bench") or {}
|
||||
narr = (bench.get("narrative") or {}).get("decode_tps_mean")
|
||||
code = (bench.get("code") or {}).get("decode_tps_mean")
|
||||
ttft = (bench.get("narrative") or {}).get("ttft_ms_mean")
|
||||
except OSError:
|
||||
narr = code = ttft = None
|
||||
ladder = "(no verify-stress artifact)"
|
||||
try:
|
||||
m = re.search(r"all (\d+) rungs passed — fillable to (\d+) tok",
|
||||
(tag / "verify-stress.log").read_text(errors="replace"))
|
||||
if m:
|
||||
ladder = f"{m.group(1)} rungs clean, fillable to {int(m.group(2)):,} tok"
|
||||
else:
|
||||
ladder = "LADDER DID NOT PASS — check verify-stress.log"
|
||||
except OSError:
|
||||
pass
|
||||
fmt = lambda v: f"{v:.2f}" if isinstance(v, (int, float)) else "—"
|
||||
print(f"{arm:<8} {fmt(narr):>9} {fmt(code):>9} {fmt(ttft):>8} {ladder}")
|
||||
PY
|
||||
|
||||
bundle="results/rebench/246-ab-bundle-$(hostname)-$(date +%Y%m%d).tgz"
|
||||
tar czf "$bundle" $(for arm in "${ARM_LIST[@]}"; do echo "results/rebench/246-ab-${arm}"; done)
|
||||
echo ""
|
||||
echo "[arch-ab] bundle written: ${bundle}"
|
||||
echo "[arch-ab] -> attach that ONE file to https://github.com/noonghunna/club-3090/issues/246"
|
||||
echo "[arch-ab] plus a sentence on anything that surprised or annoyed you — the friction"
|
||||
echo "[arch-ab] report is as valuable as the numbers."
|
||||
52
scripts/tests/test-arch-ab.sh
Normal file
52
scripts/tests/test-arch-ab.sh
Normal file
@@ -0,0 +1,52 @@
|
||||
#!/usr/bin/env bash
|
||||
# test-arch-ab — plan/refusal contract of scripts/arch-ab.sh (issue #246).
|
||||
# Hermetic: GPUs faked via CLUB3090_FAKE_GPUS; only --dry-run/--help paths run
|
||||
# (they exit before any docker/switch/rebench side effect).
|
||||
set -euo pipefail
|
||||
|
||||
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
|
||||
cd "$ROOT_DIR"
|
||||
|
||||
fail() { echo "FAIL: $1" >&2; exit 1; }
|
||||
assert_contains() { [[ "$1" == *"$2"* ]] || { echo "FAIL: missing '$2' in:" >&2; echo "$1" >&2; exit 1; }; }
|
||||
|
||||
GPU_4090="0:NVIDIA_GeForce_RTX_4090:24564:8.9"
|
||||
GPU_3090="0:NVIDIA_GeForce_RTX_3090:24576:8.6"
|
||||
GPU_5090x2="0:NVIDIA_GeForce_RTX_5090:32607:12.0,1:NVIDIA_GeForce_RTX_5090:32607:12.0"
|
||||
|
||||
# --- help exits 0 and names the arms -----------------------------------------
|
||||
out="$(bash scripts/arch-ab.sh --help)" || fail "--help nonzero"
|
||||
assert_contains "$out" "e5m2"
|
||||
assert_contains "$out" "fp8w"
|
||||
|
||||
# --- single-card auto-pick + default arms ------------------------------------
|
||||
out="$(CLUB3090_FAKE_GPUS="$GPU_4090" bash scripts/arch-ab.sh --dry-run)"
|
||||
assert_contains "$out" "variant=vllm/minimal"
|
||||
assert_contains "$out" "arm e5m2: vllm/minimal + KV_CACHE_DTYPE=fp8_e5m2"
|
||||
assert_contains "$out" "arm e4m3: vllm/minimal + KV_CACHE_DTYPE=fp8_e4m3"
|
||||
assert_contains "$out" "dry run — nothing executed"
|
||||
|
||||
# --- dual auto-pick + all four arms on 2x blackwell ---------------------------
|
||||
out="$(CLUB3090_FAKE_GPUS="$GPU_5090x2" bash scripts/arch-ab.sh --arms e5m2,e4m3,nvfp4,fp8w --dry-run)"
|
||||
assert_contains "$out" "variant=vllm/dual"
|
||||
assert_contains "$out" "arm nvfp4: vllm/dual + KV_CACHE_DTYPE=nvfp4"
|
||||
assert_contains "$out" "arm fp8w: vllm/qwen-27b-dual-max STOCK"
|
||||
|
||||
# --- refusals (fail-loud, each names the fix) ---------------------------------
|
||||
if out="$(CLUB3090_FAKE_GPUS="$GPU_3090" bash scripts/arch-ab.sh --arms e5m2,nvfp4 --dry-run 2>&1)"; then
|
||||
fail "nvfp4 on sm_8.6 must refuse"
|
||||
fi
|
||||
assert_contains "$out" "needs Blackwell"
|
||||
if out="$(CLUB3090_FAKE_GPUS="$GPU_4090" bash scripts/arch-ab.sh --arms fp8w --dry-run 2>&1)"; then
|
||||
fail "fp8w on 1 GPU must refuse"
|
||||
fi
|
||||
assert_contains "$out" "needs 2 GPUs"
|
||||
if out="$(CLUB3090_FAKE_GPUS="$GPU_4090" bash scripts/arch-ab.sh --variant vllm/dual --dry-run 2>&1)"; then
|
||||
fail "vllm/dual on 1 GPU must refuse"
|
||||
fi
|
||||
if out="$(CLUB3090_FAKE_GPUS="$GPU_4090" bash scripts/arch-ab.sh --arms bogus --dry-run 2>&1)"; then
|
||||
fail "unknown arm must refuse"
|
||||
fi
|
||||
assert_contains "$out" "unknown arm"
|
||||
|
||||
echo "test-arch-ab: ok"
|
||||
Reference in New Issue
Block a user