scripts/run_h2h.sh

#!/bin/bash
# Head-to-head N=512 generation, one model per process (no OOM), per-run logs.
# Models: A teacher | D plain-NVFP4 r0 | ours fake-q r128 | C ours real NVFP4 kernel | E BFL fp8.
set -e
export PYTHONPATH=. PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True HF_HUB_DISABLE_XET=1
N=512; H=outputs/eval/h2h; L=$H/logs
mkdir -p $L
echo "### A teacher";        python3 -u scripts/32_gen_eval.py teacher   $H/teacher 0 $N 512 > $L/teacher.log 2>&1
echo "### D plain-NVFP4 r0"; python3 -u scripts/32_gen_eval.py fq:0      $H/fq0     0 $N 512 > $L/fq0.log 2>&1
echo "### ours fake-q r128"; python3 -u scripts/32_gen_eval.py fq:128    $H/fq128   0 $N 512 > $L/fq128.log 2>&1
echo "### C ours REAL kernel"; python3 -u scripts/35_gen_real.py ours:128 $H/ours128 h2h_ours128 0 $N 512 > $L/ours128.log 2>&1
echo "### E BFL fp8";        python3 -u scripts/41_gen_bfl_fp8.py        $H/bfl_fp8 0 $N 512 > $L/bfl_fp8.log 2>&1
echo "### ALL H2H GEN DONE"