Fix the mitchich70 launch scripts for Beaker

scripts/beaker/mitchish70-launch.sh

@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+
+set -ex
+
+NUM_NODES=4
+
+gantry run \
+  --workspace ai2/dirkg \
+  --task-name mitchish70 \
+  --description "OLMo large - 70B" \
+  --priority high \
+  --stop-preemptible \
+  --beaker-image petew/olmo-torch2-gantry \
+  --cluster ai2/pluto-cirrascale \
+  --gpus 8 \
+  --replicas "${NUM_NODES}" \
+  --leader-selection \
+  --host-networking \
+  --budget ai2/oe-training \
+  --nfs \
+  --mount /net/nfs.cirrascale/allennlp/petew/cache:/root/.cache \
+  --env LOG_FILTER_TYPE=local_rank0_only \
+  --env OMP_NUM_THREADS=8 \
+  --env OLMO_TASK=model \
+  --env-secret WANDB_API_KEY=WANDB_API_KEY \
+  --env-secret AWS_ACCESS_KEY_ID=AWS_ACCESS_KEY_ID \
+  --env-secret AWS_SECRET_ACCESS_KEY=AWS_SECRET_ACCESS_KEY \
+  --env-secret R2_ACCESS_KEY_ID=R2_ACCESS_KEY_ID \
+  --env-secret R2_SECRET_ACCESS_KEY=R2_SECRET_ACCESS_KEY \
+  --env-secret R2_ENDPOINT_URL=R2_ENDPOINT_URL \
+  --shared-memory 10GiB \
+  --venv base \
+  --yes \
+  --timeout=-1 \
+  -- /bin/bash -c "scripts/beaker/mitchish70.sh \$BEAKER_LEADER_REPLICA_HOSTNAME ${NUM_NODES}"

scripts/beaker/mitchish70.sh

@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+set -exuo pipefail
+IFS=$'\n\t'
+
+BEAKER_LEADER_REPLICA_HOSTNAME=$1
+shift
+
+NUM_NODES=$1
+shift
+
+# Warm HF cache
+mkdir -p /root/.cache
+pushd /root/.cache
+curl "https://storage.googleapis.com/dirkgr-public/huggingface_cache_v2.tar.gz" | tar --keep-newer-files -xzf -
+popd
+export HF_DATASETS_OFFLINE=1
+
+torchrun \
+  --nnodes ${NUM_NODES}:${NUM_NODES} \
+  --nproc-per-node 8 \
+  --rdzv_id=101 \
+  --rdzv_backend=c10d \
+  --rdzv_endpoint=$BEAKER_LEADER_REPLICA_HOSTNAME:29400 \
+  scripts/train.py \
+    configs/mitchish70-s3.yaml \
+      --run_name=mitchish70-002 \
+      --wandb.name=mitchish70-official \
+      --device_train_microbatch_size=3 \
+      --global_train_batch_size=1536 \
+      '--load_path=${path.last_checkpoint:${remote_save_folder}}' \
+      --save_overwrite