abidlabs/backprop-alternatives-repro / code /run_memory_matched.sh
abidlabs's picture
download
raw
610 Bytes
#!/usr/bin/env bash
set -euo pipefail
REPRO_SEED="${REPRO_SEED:-0}"
pip install --no-cache-dir datasets==4.8.4
python /work/code/repro_experiment.py \
--device cuda \
--train-size 8192 \
--eval-size 1024 \
--vocab-size 12000 \
--seq-len 64 \
--width 192 \
--depth 8 \
--rank 1 \
--batch-size 128 \
--batch-size-bp 48 \
--batch-size-fmad 16 \
--batch-size-zo 128 \
--steps 4000 \
--eval-every 200 \
--threshold 0.45 \
--perturbations 1 \
--lr-bp 0.003 \
--lr-fmad 0.001 \
--lr-zo 0.001 \
--seed "${REPRO_SEED}" \
--output-dir "/work/memory_matched_seed${REPRO_SEED}"

Xet Storage Details

Size:
610 Bytes
·
Xet hash:
a7fcc7aa4c9176a2fab7b0cbf5675a68ad18637b8675cad8c2115b223ee68df3

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