mirror of
https://github.com/Druthulu/BFM-decomp
synced 2026-09-26 13:33:34 -04:00
Drew T
6678c6868a
- Fable5Max agent (Agent model=fable) matched a 248-ins reach-134 GIANT on the
§20/§10 store-vs-load wall (22 phases "CONFIRMED unsteerable") by reading the
gcc-2.7.2 source (tools/reference/gcc-papermario) + RTL -da dumps. Leaf
MATCH(248 ins) -> whole-binary banked:1 -> dedup_propagate ×134. Verified:
check-all 136/136 byte-identical, dedup-check 1780 validated/0 failed.
- 3 byte-proven idioms -> cookbook §30 (corrects §29's "not a bigger model"):
(1) store-vs-load is a deterministic MEM_IN_STRUCT_P /s aliasing flag, not a
scheduler tie-break; steer via ((struct{s32 f;}*)p)->f (anon struct keeps
/s AND propagates ×134) to grant, *p to deny
(2) def-side return-type wall has a MACRO escape: widen a discarding caller
macro's extern void->s32 (byte-neutral, check-all-verified) -- extends §29
(3) birthing-boost prologue-order lever: __asm__("":"=r"(x):"0"(x)) re-tie in a
later bb kills sched.c's REG_N_SETS==1 priority boost
- tools/glm_parallel.sh: K concurrent OpenRouter/GLM cloud drafters (parallel
api_draft), key read from .env at runtime
- §10/§20 "store-vs-load unsteerable" backlog now re-test candidates:
func_8014F2E0, func_80150528, func_8014EA4C
58 lines
2.5 KiB
Bash
58 lines
2.5 KiB
Bash
#!/usr/bin/env bash
|
||
# glm_parallel.sh — parallel cloud-drafting for the GLM (OpenRouter) matching tier.
|
||
#
|
||
# Splits a targets JSON into K contiguous slices and runs K concurrent api_draft.py
|
||
# workers against the same OpenRouter endpoint (GLM is latency-bound, so N concurrent
|
||
# HTTP calls ~= N× throughput; OpenRouter fans out over ~6 providers, no hard cap).
|
||
# All workers write <fn>.c + <fn>.reasoning.txt into ONE shared out dir (distinct fn
|
||
# names => no collision). The whole-binary byte-gate (gate_stage) is the sole arbiter
|
||
# afterward (G3/P9) — this script only DRAFTS.
|
||
#
|
||
# The OpenRouter key is read from .env (key `open_router_key=`) at runtime and passed
|
||
# via env to the workers only — never echoed, never written to a file.
|
||
#
|
||
# Usage: tools/glm_parallel.sh <targets.json> <out_dir> [K=6] [iters=2] [model=z-ai/glm-5.2] [maxtok=8000]
|
||
set -euo pipefail
|
||
cd "$(dirname "$0")/.." # repo root
|
||
|
||
T="${1:?usage: glm_parallel.sh <targets.json> <out_dir> [K] [iters] [model] [maxtok]}"
|
||
OUT="${2:?out_dir required}"
|
||
K="${3:-6}"
|
||
ITERS="${4:-2}"
|
||
MODEL_ID="${5:-z-ai/glm-5.2}"
|
||
MAXTOK_V="${6:-8000}"
|
||
|
||
KEY="$(sed -n 's/^open_router_key=//p' .env | tr -d '"'\''\r ')"
|
||
[ -n "$KEY" ] || { echo "glm_parallel: no open_router_key in .env" >&2; exit 1; }
|
||
|
||
mkdir -p "$OUT"
|
||
SLICEDIR="$OUT/.slices"; rm -rf "$SLICEDIR"; mkdir -p "$SLICEDIR"
|
||
|
||
NSLICES=$(python3 - "$T" "$K" "$SLICEDIR" <<'PY'
|
||
import json, sys, math
|
||
targets = json.load(open(sys.argv[1])); K = int(sys.argv[2]); sd = sys.argv[3]
|
||
n = len(targets); per = max(1, math.ceil(n / K)); w = 0
|
||
for i in range(0, n, per):
|
||
json.dump(targets[i:i+per], open('%s/slice_%d.json' % (sd, w), 'w')); w += 1
|
||
print(w)
|
||
PY
|
||
)
|
||
echo "glm_parallel: $T -> $NSLICES workers (iters=$ITERS, model=$MODEL_ID) -> $OUT"
|
||
|
||
pids=()
|
||
for f in "$SLICEDIR"/slice_*.json; do
|
||
API_BASE=https://openrouter.ai/api/v1 MODEL="$MODEL_ID" API_KEY="$KEY" \
|
||
LEAN=1 MAXTOK="$MAXTOK_V" REASON=1 TEMP=0.3 \
|
||
.venv/bin/python tools/api_draft.py --targets "$f" --out "$OUT" --iters "$ITERS" \
|
||
> "$OUT/$(basename "$f" .json).log" 2>&1 &
|
||
pids+=($!)
|
||
done
|
||
echo "glm_parallel: launched ${#pids[@]} workers (pids ${pids[*]})"
|
||
fail=0
|
||
for p in "${pids[@]}"; do wait "$p" || fail=$((fail+1)); done
|
||
echo "glm_parallel: all workers done ($fail failed)"
|
||
|
||
DRAFTS=$(ls "$OUT"/*.c 2>/dev/null | wc -l)
|
||
COST=$(grep -ohE 'cost \$[0-9.]+' "$OUT"/slice_*.log 2>/dev/null | grep -oE '[0-9.]+' | paste -sd+ | bc -l 2>/dev/null || echo 0)
|
||
echo "glm_parallel: $DRAFTS drafts written; summed reported cost \$${COST:-0}"
|