From ad4622e42abf2a55a37c5e522a6f8302e2e95cfd Mon Sep 17 00:00:00 2001 From: Oscar Wan Date: Tue, 21 Jul 2026 15:28:06 -0700 Subject: Add portable 135M BP sweep handoff Match the active EP configuration and support four-GPU BP candidates from collaborators' data locations. Co-authored-by: Cursor --- ep_run/runs/fw135m_bp_sweep.sh | 48 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 48 insertions(+) create mode 100644 ep_run/runs/fw135m_bp_sweep.sh (limited to 'ep_run/runs') diff --git a/ep_run/runs/fw135m_bp_sweep.sh b/ep_run/runs/fw135m_bp_sweep.sh new file mode 100644 index 0000000..e9715b4 --- /dev/null +++ b/ep_run/runs/fw135m_bp_sweep.sh @@ -0,0 +1,48 @@ +#!/usr/bin/env bash +# Run one EP-matched 135M BP LR-sweep candidate. +# +# Usage: +# ./runs/fw135m_bp_sweep.sh <7e-4|1e-3|1.4e-3> [seed] +# Set GPUS=4 to launch a four-A6000 NCCL data-parallel run with torchrun. + +set -euo pipefail + +LR="${1:?Usage: $0 <7e-4|1e-3|1.4e-3> [seed]}" +SEED="${2:-1}" +case "${LR}" in + 7e-4|1e-3|1.4e-3) ;; + *) echo "Unsupported LR: ${LR}" >&2; exit 2 ;; +esac + +RUN_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)" +PYTHON_BIN="${PYTHON_BIN:-python}" +WAND_PROJECT="${WAND_PROJECT:-ept-fineweb-135M}" +GPUS="${GPUS:-1}" +export EPT_DATA_ROOT="${EPT_DATA_ROOT:-${RUN_DIR}/data}" + +LR_TAG="${LR//-/m}" +LR_TAG="${LR_TAG//./p}" +TAG="fw135m_bp_lr${LR_TAG}_s${SEED}" + +cd "${RUN_DIR}" +if (( GPUS > 1 )); then + LAUNCH=(torchrun --standalone --nproc_per_node="${GPUS}") +else + LAUNCH=("${PYTHON_BIN}") +fi + +exec "${LAUNCH[@]}" casc_bp_train.py \ + --tag "${TAG}" \ + --L 12 --C 768 --H 12 --T 256 --B 24 \ + --steps 440000 \ + --lr "${LR}" \ + --warmup 1000 \ + --amp --olmo2 \ + --wd 0.1 \ + --opt muon --muon_lr 0.02 \ + --cosine --lr_min_ratio 0.1 \ + --data fineweb_edu \ + --seed "${SEED}" \ + --save_every 5000 --log 100 \ + --wandb "${WAND_PROJECT}" \ + --wandb_run "${TAG}" -- cgit v1.2.3