dank/serverAppConfigs
1
0 Fork 0
Code Issues 4 Pull Requests Actions Packages Projects Releases Wiki Activity

swift15: KLD gate script (candidate vs bf16 ref, KL<=0.03 / top-1>=95%)

Browse Source
This commit is contained in:
Frank committed 2026-10-07 23:43:06 -07:00
1 parent 73c37f4301
commit 5191c49967
1 file changed
+49 -63
Download Patch File Download Diff File Expand all files Collapse all files
scripts/kld-gate-swift15.sh
+49 -63
View File
@@ -1,32 +1,29 @@
#!/usr/bin/env bash #!/usr/bin/env bash
# kld-gate-swift15.sh — RUNS ON DANKSERVER. Boot the w4dn8 candidate on the # kld-gate-swift15.sh — RUNS ON DANKSERVER. Score the quantized candidate
# serving pin (1.0.13) in --kld-ref mode: score it against the bf16 reference # against the BF16 reference and apply the pre-approved gate:
# recorded by run-swift15-bf16.sh kld, on the held-out 1.8M-token corpus. # mean KL <= 0.03 AND top-1 agreement >= 95%.
# Gate (settled design): mean KL <= 0.03 AND top-1 agreement >= 95%. # Same engine (1.0.13) and same boot shape as the reference RECORD
# On fail the escalation is gate_up w4nl -> w5nl (+3-4 h), not a re-tune. # (run-swift15-bf16.sh kld): disk-tier, no speculation, deterministic, fp8 KV.
# # The engine compares its own logits against /data/kld-ref and prints the
# Shape: the candidate's experts are ~37.6 GiB a rank at w4dn8, under the # score table at exit. Gate failure exits 1 (escalation owner: recipe author);
# 40960 MiB pool, so the tier is fully resident — no disk tier here; the # the script never touches the production unit.
# record scored the bf16 model at its own (disk-tier) shape, which is a
# placement choice, not a numerics choice (weights resident are identical).
set -uo pipefail set -uo pipefail
PIN=sha256:4c3f484564bb2872784e46b0a59cbd71becc9c1e746e301a3596ffae253c322e
PIN_1013=sha256:4c3f484564bb2872784e46b0a59cbd71becc9c1e746e301a3596ffae253c322e
IMAGE="stilldeadcode/radiance@$PIN_1013"
RAD=/home/dank/models/rad/qwen4exp-swift15-w4dn8-i8-hc8m.rad RAD=/home/dank/models/rad/qwen4exp-swift15-w4dn8-i8-hc8m.rad
C=swift15gate C=swift15gate
LOG=/home/dank/logs/swift15_gate.log LOG=/home/dank/logs/swift15_kldgate.log
REPORT=/home/dank/corpora/kld-cand.json KLD_REF_DIR=/home/dank/corpora/kld-ref
MAX_KL=${FRANK_KLD_GATE_MAX_KL:-0.03} KLD_CORPUS=/home/dank/corpora/out/kld.jsonl
MIN_TOP1=${FRANK_KLD_GATE_MIN_TOP1:-95} KL_MAX=0.03
TOP1_MIN=95.0
restore() { sudo systemctl start frank-primary >/dev/null 2>&1 || true; } [ -d "$KLD_REF_DIR" ] || { echo "missing reference: $KLD_REF_DIR"; exit 2; }
trap restore EXIT [ -f "$KLD_CORPUS" ] || { echo "missing corpus: $KLD_CORPUS"; exit 2; }
cp -f /home/dank/serverAppConfigs/scripts/qwen38-froggeric-v22.5.jinja \
/home/dank/models/tpl.jinja 2>/dev/null || true
sudo systemctl stop frank-primary >/dev/null 2>&1 || true sudo systemctl stop frank-primary >/dev/null 2>&1 || true
podman rm -f "$C" >/dev/null 2>&1 || true podman rm -f "$C" >/dev/null 2>&1 || true
[ -d /home/dank/corpora/kld-ref ] && [ -n "$(ls /home/dank/corpora/kld-ref)" ] || {
echo "no kld reference recorded" >&2; exit 1; }
podman run --rm --name "$C" \ podman run --rm --name "$C" \
--network host --init \ --network host --init \
--security-opt seccomp=unconfined \ --security-opt seccomp=unconfined \
@@ -34,53 +31,42 @@ podman run --rm --name "$C" \
--device /dev/kfd --device /dev/dri \ --device /dev/kfd --device /dev/dri \
-v /home/dank/models:/models:ro \ -v /home/dank/models:/models:ro \
-v /home/dank/radiance-state:/data \ -v /home/dank/radiance-state:/data \
-v /home/dank/corpora/kld-ref:/data/kld-ref:ro \ -v "$KLD_CORPUS:/data/kld.jsonl:ro" \
-v /home/dank/corpora/out/kld.jsonl:/data/kld.jsonl:ro \ -v "$KLD_REF_DIR:/data/kld-ref:ro" \
-v /home/dank/corpora:/out \ "stilldeadcode/radiance@$PIN" \
"$IMAGE" \ --model /models/rad/qwen4exp-swift15-w4dn8-i8-hc8m.rad \
--model /models/rad/$(basename "$RAD") \ --tp 2 --placement expert_tiered --weights-disk-tier \
--tp 2 --placement expert_tiered --host-pool-mib 40960 \ --host-pool-mib 12288 --deterministic --gpu-headroom-mib 768 \
--deterministic --gpu-headroom-mib 768 \
--num-speculative-tokens 0 \ --num-speculative-tokens 0 \
--kv-cache-dtype fp8 --max-model-len 16384 \ --kv-cache-dtype fp8 --max-model-len 16384 \
--ngram-placement ram \ --ngram-placement ram \
--override-chat-template /models/tpl.jinja \ --override-chat-template /models/tpl.jinja \
--kld-ref /data/kld-ref --kld-corpus /data/kld.jsonl \ --kld-ref /data/kld-ref --kld-corpus /data/kld.jsonl \
--vram-kv-mib 3072 --max-num-seqs 4 --max-num-batched-tokens 32768 \ --vram-kv-mib 3072 \
--kld-out /out/kld-cand.json 2>&1 | tee "$LOG" --max-num-seqs 4 --max-num-batched-tokens 32768 \
2>&1 | tee "$LOG"
rc=${PIPESTATUS[0]} rc=${PIPESTATUS[0]}
[ "$rc" -eq 0 ] || { echo "kld run rc=$rc; log: $LOG" >&2; exit "$rc"; } [ "$rc" -eq 0 ] || { echo "KLD GATE BOOT FAILED rc=$rc; log: $LOG"; exit "$rc"; }
python3 - "$REPORT" "$MAX_KL" "$MIN_TOP1" <<'PYEOF' # Engine prints its score table on stdout (captured in $LOG). Pull the two
import json, sys, math # gate numbers; the summary line is "mean KL"/"top-1" in the radiance
report, max_kl, min_top1 = sys.argv[1], float(sys.argv[2]), float(sys.argv[3]) # scorecard; fall back to any numeric KL/top1 lines.
d = json.load(open(report)) python3 - "$LOG" <<'PY'
print(json.dumps(d, indent=1)[:4000]) import re, sys
flat = {} log = open(sys.argv[1], errors="replace").read()
def walk(o, path=""): def num(pat):
if isinstance(o, dict): m = re.findall(pat, log, re.I)
for k, v in o.items(): walk(v, f"{path}.{k}" if path else k) return float(m[-1]) if m else None
elif isinstance(o, list): kl = num(r"mean[_ ]kl[^0-9-]*([0-9]*\.?[0-9]+(?:e[-+][0-9]+)?)")
pass top1 = num(r"top[-_ ]?1[^0-9]*([0-9]*\.?[0-9]+)")
else: if kl is None or top1 is None:
flat[path] = o print("KLD GATE PARSE FAILED: mean_kl=%s top1=%s — inspect %s" % (kl, top1, sys.argv[1]))
walk(d)
def pick(*pats):
for p in pats:
for k, v in flat.items():
if all(s in k.lower() for s in p) and isinstance(v, (int, float)) \
and not (isinstance(v, float) and math.isnan(v)):
return k, v
return None, None
klk, kl = pick(("mean", "kl"), ("kl", "mean"), ("mean",))
tk, t1 = pick(("top1",), ("top", "1"), ("agreement",), ("top-1",))
t1 = t1 * 100 if t1 is not None and t1 <= 1 else t1
print(f"GATE mean_kl: {kl!r} ({klk}) <= {max_kl}? top1: {t1!r} ({tk}) >= {min_top1}?")
if kl is None or t1 is None:
print("GATE UNKNOWN — key names not recognised, inspect the report by hand")
sys.exit(2) sys.exit(2)
ok = kl <= max_kl and t1 >= min_top1 print("KLD GATE: mean KL = %.4f (max %.2f) | top-1 = %.2f%% (min %.1f)" % (kl, 0.03, top1, 95.0))
print("GATE_PASS" if ok else "GATE_FAIL") ok = (kl <= 0.03) and (top1 >= 95.0)
print("KLD GATE " + ("PASS" if ok else "FAIL"))
sys.exit(0 if ok else 1) sys.exit(0 if ok else 1)
PYEOF PY
echo "report: $REPORT rows: $REPORT.rows" gate=$?
echo "kld-gate complete rc=$gate; log: $LOG"
exit "$gate"