day35/scripts/run_gl.sh

#!/bin/bash
cd /workspace/RSI
export HF_HOME=/workspace/.hf_home
export LD_LIBRARY_PATH=/venv/main/lib/python3.10/site-packages/nvidia/cu13/lib:$LD_LIBRARY_PATH
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export VLLM_DETERMINISTIC=1
export VLLM_USE_FLASHINFER_SAMPLER=0
export PYTHONPATH=/workspace/RSI
export GPU_UTIL=0.45
export START_MODEL=${START_MODEL:-/workspace/RSI/expanded_models/gen2_B_grown}
export N_GENS=${N_GENS:-4}
export GROW_LAYERS=8
export FILL_STEPS=160
export FILL_RANK=32
export FILL_EPOCHS=3
export HOLD_SET=hard_holdout
export HOLD_N=60
export CEIL_K=8
export HARVEST_K=${HARVEST_K:-4}
export TRAIN_N=${TRAIN_N:-160}
export CORPUS_CAP=128
# speed knobs (day-35 diet: ~5h/gen -> ~1.5-2h/gen, no decision-quality loss)
export GL_PROBE_BASE=${GL_PROBE_BASE:-0}       # growth function-preservation verified 2x; skip per-gen base probe
export GL_ADAPTIVE_PROBE=${GL_ADAPTIVE_PROBE:-1}  # 2nd salted probe only on gate-boundary calls
export GL_PROBE_MARGIN=${GL_PROBE_MARGIN:-2.0}
# RATE knobs (the 1%/20min bar = >=+3%/h sustained -> ~1h fill-cycles)
export GL_FILLS_PER_GROW=${GL_FILLS_PER_GROW:-3}   # amortize the ~20min grow+merge over 3 fill-cycles
export GL_GPU_UTIL_SOLO=${GL_GPU_UTIL_SOLO:-0.80}  # harvest holds the GPU exclusively -> ~2x throughput
export GL_MAX_WEAK=${GL_MAX_WEAK:-2}               # weak cycles before regrow (twin saturation signal)
# calibration flags (written by gl_switchover.sh once 0.80-util probe EXACTLY reproduces a
# known result — only then does the sealed ruler run fast; proof before speed)
[ -f /workspace/RSI/outputs/gl_flags.env ] && source /workspace/RSI/outputs/gl_flags.env
export MAX_MODEL_LEN=4096
export MAXTOK=700
export VERIFY_TC=15
export VERIFY_TO=5
export GRADE_WORKERS=32
export WORKDIR=/workspace/RSI/expanded_models/gen_loop
export OUT=/workspace/RSI/outputs/gen_loop.jsonl
exec /venv/main/bin/python3 scripts/gen_loop.py