feat(campaign): protected-cast applier, recipe markers, KLD gate runner
apply-protected-cast.sh splices derive-protected-experts.py output into the recipe's PROTECTED-BEGIN/END section (refuses to re-derive, trashes the pre-edit recipe). kld-gate-swift15.sh boots the candidate on the serving pin in --kld-ref mode (pool fully resident, no disk tier) and enforces mean KL<=0.03 / top-1>=95% from the JSON report.
This commit is contained in:
Frank
Executable
+43
@@ -0,0 +1,43 @@
|
||||
#!/usr/bin/env bash
|
||||
# apply-protected-cast.sh — RUNS ON DANKSERVER. Rewrite the recipe's
|
||||
# PROTECTED-BEGIN/END section from Swift's OWN calibration grams: run
|
||||
# derive-protected-experts.py over the calib dir, splice its cast lines into
|
||||
# the marked section of the recipe in place, keep the pre-edit recipe in the
|
||||
# campaign trash. Refuses to run twice (marker text says derived).
|
||||
set -euo pipefail
|
||||
CALIB=${1:-/home/dank/corpora/calib}
|
||||
RECIPE=${2:-/home/dank/serverAppConfigs-swift15/scripts/recipes/qwen4exp-swift15-w4dn8-i8-hc8m.recipe}
|
||||
PY=/home/dank/venvs/radcorpus/bin/python
|
||||
DERIVE=/home/dank/serverAppConfigs-swift15/scripts/derive-protected-experts.py
|
||||
TS=$(date +%Y%m%d-%H%M%S)
|
||||
|
||||
grep -q "derive-protected-experts.py rewrote" "$RECIPE" && {
|
||||
echo "recipe already carries derived cast lines; refusing to re-derive" >&2; exit 1; }
|
||||
grep -q "PROTECTED-BEGIN" "$RECIPE" || { echo "no PROTECTED markers in $RECIPE" >&2; exit 1; }
|
||||
|
||||
"$PY" "$DERIVE" "$CALIB" > "/tmp/protected-cast-$TS.txt"
|
||||
N=$(grep -c 'cast dtype=bf16' "/tmp/protected-cast-$TS.txt")
|
||||
[ "$N" -ge 2 ] || { echo "derive produced $N cast lines; refusing" >&2; exit 1; }
|
||||
echo "derived $N protected experts -> /tmp/protected-cast-$TS.txt"
|
||||
|
||||
cp -f "$RECIPE" "/home/dank/.AGENTTRASH/$TS-qwen4exp-swift15-w4dn8-i8-hc8m.recipe"
|
||||
"$PY" - "$RECIPE" "/tmp/protected-cast-$TS.txt" <<'PYEOF'
|
||||
import sys
|
||||
recipe, castfile = sys.argv[1], sys.argv[2]
|
||||
lines = open(recipe).read().splitlines()
|
||||
cast = [l for l in open(castfile).read().splitlines() if 'cast dtype=bf16' in l]
|
||||
out, insec = [], False
|
||||
for l in lines:
|
||||
if 'PROTECTED-BEGIN' in l:
|
||||
insec = True
|
||||
out.append(l)
|
||||
out.extend(cast)
|
||||
elif 'PROTECTED-END' in l:
|
||||
insec = False
|
||||
out.append(l)
|
||||
out.append('# (derive-protected-experts.py rewrote this section from Swift\'s own grams)')
|
||||
elif not insec:
|
||||
out.append(l)
|
||||
open(recipe, 'w').write('\n'.join(out) + '\n')
|
||||
print(f'spliced {len(cast)} cast lines into {recipe}')
|
||||
PYEOF
|
||||
Executable
+86
@@ -0,0 +1,86 @@
|
||||
#!/usr/bin/env bash
|
||||
# kld-gate-swift15.sh — RUNS ON DANKSERVER. Boot the w4dn8 candidate on the
|
||||
# serving pin (1.0.13) in --kld-ref mode: score it against the bf16 reference
|
||||
# recorded by run-swift15-bf16.sh kld, on the held-out 1.8M-token corpus.
|
||||
# Gate (settled design): mean KL <= 0.03 AND top-1 agreement >= 95%.
|
||||
# On fail the escalation is gate_up w4nl -> w5nl (+3-4 h), not a re-tune.
|
||||
#
|
||||
# Shape: the candidate's experts are ~37.6 GiB a rank at w4dn8, under the
|
||||
# 40960 MiB pool, so the tier is fully resident — no disk tier here; the
|
||||
# record scored the bf16 model at its own (disk-tier) shape, which is a
|
||||
# placement choice, not a numerics choice (weights resident are identical).
|
||||
set -uo pipefail
|
||||
|
||||
PIN_1013=sha256:4c3f484564bb2872784e46b0a59cbd71becc9c1e746e301a3596ffae253c322e
|
||||
IMAGE="stilldeadcode/radiance@$PIN_1013"
|
||||
RAD=/home/dank/models/rad/qwen4exp-swift15-w4dn8-i8-hc8m.rad
|
||||
C=swift15gate
|
||||
LOG=/home/dank/logs/swift15_gate.log
|
||||
REPORT=/home/dank/corpora/kld-cand.json
|
||||
MAX_KL=${FRANK_KLD_GATE_MAX_KL:-0.03}
|
||||
MIN_TOP1=${FRANK_KLD_GATE_MIN_TOP1:-95}
|
||||
|
||||
restore() { sudo systemctl start frank-primary >/dev/null 2>&1 || true; }
|
||||
trap restore EXIT
|
||||
sudo systemctl stop frank-primary >/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 "$C" \
|
||||
--network host --init \
|
||||
--security-opt seccomp=unconfined \
|
||||
--ulimit memlock=-1:-1 \
|
||||
--device /dev/kfd --device /dev/dri \
|
||||
-v /home/dank/models:/models:ro \
|
||||
-v /home/dank/radiance-state:/data \
|
||||
-v /home/dank/corpora/kld-ref:/data/kld-ref:ro \
|
||||
-v /home/dank/corpora/out/kld.jsonl:/data/kld.jsonl:ro \
|
||||
-v /home/dank/corpora:/out \
|
||||
"$IMAGE" \
|
||||
--model /models/rad/$(basename "$RAD") \
|
||||
--tp 2 --placement expert_tiered --host-pool-mib 40960 \
|
||||
--deterministic --gpu-headroom-mib 768 \
|
||||
--num-speculative-tokens 0 \
|
||||
--kv-cache-dtype fp8 --max-model-len 16384 \
|
||||
--ngram-placement ram \
|
||||
--override-chat-template /models/tpl.jinja \
|
||||
--kld-ref /data/kld-ref --kld-corpus /data/kld.jsonl \
|
||||
--vram-kv-mib 3072 --max-num-seqs 4 --max-num-batched-tokens 32768 \
|
||||
--kld-out /out/kld-cand.json 2>&1 | tee "$LOG"
|
||||
rc=${PIPESTATUS[0]}
|
||||
[ "$rc" -eq 0 ] || { echo "kld run rc=$rc; log: $LOG" >&2; exit "$rc"; }
|
||||
|
||||
python3 - "$REPORT" "$MAX_KL" "$MIN_TOP1" <<'PYEOF'
|
||||
import json, sys, math
|
||||
report, max_kl, min_top1 = sys.argv[1], float(sys.argv[2]), float(sys.argv[3])
|
||||
d = json.load(open(report))
|
||||
print(json.dumps(d, indent=1)[:4000])
|
||||
flat = {}
|
||||
def walk(o, path=""):
|
||||
if isinstance(o, dict):
|
||||
for k, v in o.items(): walk(v, f"{path}.{k}" if path else k)
|
||||
elif isinstance(o, list):
|
||||
pass
|
||||
else:
|
||||
flat[path] = o
|
||||
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)
|
||||
ok = kl <= max_kl and t1 >= min_top1
|
||||
print("GATE_PASS" if ok else "GATE_FAIL")
|
||||
sys.exit(0 if ok else 1)
|
||||
PYEOF
|
||||
echo "report: $REPORT rows: $REPORT.rows"
|
||||
@@ -30,6 +30,7 @@
|
||||
# law -- an expert whose share of its layer's down-input energy is at least 3% and at least three
|
||||
# times its share of the routing, an even count on each rank's half of the layer (the four-bit GEMM
|
||||
# takes the rest as two equal tables by parity) by adding that half's next most energetic expert.
|
||||
# PROTECTED-BEGIN (rewritten from Swift's OWN grams by derive-protected-experts.py)
|
||||
blk.34.ffn_*_exps.407.weight cast dtype=bf16
|
||||
blk.34.ffn_*_exps.496.weight cast dtype=bf16
|
||||
blk.44.ffn_*_exps.292.weight cast dtype=bf16
|
||||
@@ -40,6 +41,7 @@ blk.47.ffn_*_exps.122.weight cast dtype=bf16
|
||||
blk.47.ffn_*_exps.143.weight cast dtype=bf16
|
||||
blk.47.ffn_*_exps.399.weight cast dtype=bf16
|
||||
blk.47.ffn_*_exps.445.weight cast dtype=bf16
|
||||
# PROTECTED-END
|
||||
|
||||
# Every other routed expert, trunk and MTP layer alike: the down projection as E4M3 codes with a
|
||||
# bf16 scale a 128x128 block -- one amax per block, scale = amax/448 rounded to bf16, the exact
|
||||
|
||||
Reference in New Issue
Block a user