Buckets:

Mercity/FluxDistill / scripts /run_scale256.sh
Pranav2748's picture
download
raw
1.04 kB
#!/bin/bash
# Scale the controlled ablation + deployable model to N=256, then recompute metrics.
# teacher/fq0/fq128 extend in place (skip existing); ours128_real regenerates (re-times).
set -e
export PYTHONPATH=.
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
N=256
echo "### teacher"
python3 -u scripts/32_gen_eval.py teacher outputs/eval/imgs/teacher 0 $N 512
echo "### plain NVFP4 r0"
python3 -u scripts/32_gen_eval.py fq:0 outputs/eval/imgs/pilot_fq0 0 $N 512
echo "### SVDQuant r128 (fake-q)"
python3 -u scripts/32_gen_eval.py fq:128 outputs/eval/imgs/pilot_fq128 0 $N 512
echo "### SVDQuant r128 (real kernel)"
python3 -u scripts/35_gen_real.py ours:128 outputs/eval/imgs/pilot_ours128_real ours128_256 0 $N 512
echo "### metrics @ N=256"
python3 -u scripts/34_metrics.py outputs/eval/imgs/teacher outputs/eval/imgs/mjhq_ref \
outputs/eval/imgs/pilot_ours128_real outputs/eval/imgs/pilot_fq128 outputs/eval/imgs/pilot_fq0
echo "### SCALE256 ALL DONE"

Xet Storage Details

Size:
1.04 kB
·
Xet hash:
9a6a2347c64aa336713bafbadccff5477aaa7c687b7d694e84ce540f6aa5af49

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.