Buckets:
| # After notes: teacher placement probe -> teacher cache -> E3 -> E2o -> E2s -> E4 latency. Logs in logs/. | |
| set -e | |
| cd /workspace/think-decide | |
| export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True | |
| for p in before after; do | |
| python src/teacher_cache.py --placement $p --limit 1500 --probe-only > logs/teacher_probe_$p.log 2>&1 | |
| done | |
| PLACE=$(python - <<'PY' | |
| import json, re | |
| def keep(p): | |
| return json.load(open("runs/e3/teacher_probe_%s.json" % p))["report"]["ALL"]["keep_rate"] | |
| b, a = keep("before"), keep("after") | |
| print("after" if a >= b - 0.02 else "before") | |
| PY | |
| ) | |
| echo "placement=$PLACE" | tee logs/teacher_placement.txt | |
| python src/teacher_cache.py --placement $PLACE > logs/teacher_cache.log 2>&1 | |
| while pgrep -f "src/eval.py" > /dev/null; do sleep 10; done # let E2/S1 evals finish before a 94%-VRAM run | |
| scripts/run_arm.sh e3 --n-think 32 --k-choices 1,2,4,8 --budget 42000 --distill runs/e3/teacher_cache.pt > logs/e3_run.out 2>&1 | |
| scripts/run_arm.sh e2o --n-think 32 --k-choices 1,2,4,8 --budget 42000 --gate-init 0.014 > logs/e2o_run.out 2>&1 | |
| scripts/run_arm.sh e2s --n-think 32 --k-choices 1,2,4,8 --budget 36000 --fb-mode soft --gate-init 1.0 --gate-fixed > logs/e2s_run.out 2>&1 | |
| python src/latency.py --model runs/e2/model.safetensors > logs/latency.log 2>&1 | |
| echo QUEUE_DONE | |
Xet Storage Details
- Size:
- 1.3 kB
- Xet hash:
- 25aca50bb6f5762d1895f4fa80dd1bdeb1a4b305192afa697aa2cd7834e70b06
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.