diff --git a/configs/mcli/mitchish7.yaml b/configs/mcli/mitchish7.yaml
new file mode 100644
index 000000000..c099da378
--- /dev/null
+++ b/configs/mcli/mitchish7.yaml
@@ -0,0 +1,61 @@
+name: olmo-7b
+image: mosaicml/pytorch:2.1.2_cu121-python3.10-ubuntu20.04
+compute:
+  cluster: r15z1p1
+  gpus: 32
+  gpu_type: h100_80gb
+integrations:
+  - integration_type: git_repo
+    git_repo: allenai/OLMo
+    git_branch: train-olmo-large
+    pip_install: -e .[train]
+    ssh_clone: true
+env_variables:
+  PIP_DISABLE_PIP_VERSION_CHECK: "1"
+  OMP_NUM_THREADS: "8"
+  LOG_FILTER_TYPE: local_rank0_only
+command: |-
+  # Install AWS CLI (for download unsharded checkpoints).
+  #apt-get update
+  #apt-get install zip unzip
+  #curl "https://awscli.amazonaws.com/awscli-exe-linux-x86_64.zip" -o "awscliv2.zip"
+  #unzip awscliv2.zip
+  #sudo ./aws/install
+
+  # Make sure we have a recent flash-attn.
+  # NOTE: only pinning flash-attn here to future proof it.
+  pip install flash-attn==2.5.3 --no-build-isolation
+
+  # Show packages for debugging.
+  pip freeze
+
+  # Prepare environment.
+  mkdir -p /root/.cache/torch
+  # warm up huggingface cache
+  pushd /root/.cache
+  curl "https://storage.googleapis.com/dirkgr-public/huggingface_cache_v2.tar.gz" | tar -xzf -
+  popd
+  export HF_DATASETS_OFFLINE=1
+
+  cd OLMo
+
+  torchrun \
+  --master_addr "$MASTER_ADDR" \
+  --master_port "$MASTER_PORT" \
+  --nnodes "$NUM_NODES" \
+  --node_rank "$NODE_RANK" \
+  --nproc_per_node 8 \
+  scripts/train.py configs/mitchish7-s3.yaml \
+    --run_name=mitchish7 \
+    --wandb.group=mitchish7 \
+    --model.flash_attention=true \
+    --fsdp.wrapping_strategy=by_block_and_size \
+    --fsdp.sharding_strategy=SHARD_GRAD_OP \
+    --save_folder=runs/ \
+    --activation_checkpointing=fine_grained \
+    --fused_loss=true \
+    --device_train_microbatch_size=2 \
+    --global_train_batch_size=1024 \
+    '--load_path=${path.last_checkpoint:${remote_save_folder}}' \
+    --save_overwrite \
+    --load_path=s3://ai2-llm/checkpoints/OLMo-medium/mitchish7/step51450-unsharded
diff --git a/scripts/beaker/mitchish70-launch.sh b/scripts/beaker/mitchish70-launch.sh
new file mode 100755
index 000000000..0b2c88872
--- /dev/null
+++ b/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 ${NUM_NODES}"
diff --git a/scripts/beaker/mitchish70-loadtest-launch.sh b/scripts/beaker/mitchish70-loadtest-launch.sh
new file mode 100755
index 000000000..8718f9446
--- /dev/null
+++ b/scripts/beaker/mitchish70-loadtest-launch.sh
@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+
+set -ex
+
+NUM_NODES=4
+
+gantry run \
+  --workspace ai2/dirkg \
+  --task-name mitchish70-loadtest \
+  --description "OLMo large - 70B - loadtest" \
+  --priority high \
+  --stop-preemptible \
+  --beaker-image petew/olmo-torch2-gantry \
+  --cluster ai2/pluto-cirrascale \
+  --gpus 8 \
+  --replicas "${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-loadtest.sh \$BEAKER_LEADER_REPLICA_HOSTNAME ${NUM_NODES}"
diff --git a/scripts/beaker/mitchish70-loadtest.sh b/scripts/beaker/mitchish70-loadtest.sh
new file mode 100755
index 000000000..22468d343
--- /dev/null
+++ b/scripts/beaker/mitchish70-loadtest.sh
@@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+set -exuo pipefail
+IFS=$'\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=$BEAKER_LEADER_REPLICA_HOSTNAME:29400 \
+  scripts/train.py \
+  configs/mitchish70-s3.yaml \
+    --run_name=mitchish70-loadtest \
+    --wandb.name=mitchish70-loadtest \
+    --model.flash_attention=true \
+    --fsdp.wrapping_strategy=by_block_and_size \
+    --save_folder=runs/ \
+    --fused_loss=true \
+    --device_train_microbatch_size=2 \
+    --global_train_batch_size=512 \
+    --save_overwrite \
+    --remote_save_folder=null \
+    --load_path=s3://ai2-llm/checkpoints/OLMo-large/mitchish70-002/step32300-unsharded
\ No newline at end of file
diff --git a/scripts/beaker/mitchish70.sh b/scripts/beaker/mitchish70.sh
index a5b279ebf..06f837877 100755
--- a/scripts/beaker/mitchish70.sh
+++ b/scripts/beaker/mitchish70.sh
@@ -1,36 +1,31 @@
 #!/usr/bin/env bash
+set -exuo pipefail
+IFS=$'\n\t'
 
-set -ex
+BEAKER_LEADER_REPLICA_HOSTNAME=$1
+shift
 
-CONFIG_PATH=configs/mitchish70-s3.yaml
-NUM_NODES=4
-RUN_NAME="mitchish70-001"
-ARGS="--run_name=${RUN_NAME} --device_train_microbatch_size=4"
+NUM_NODES=$1
+shift
 
-gantry run \
-  --allow-dirty \
-  --workspace ai2/llm-testing \
-  --task-name mitchish70 \
-  --description "OLMo large - 70B" \
-  --priority high \
-  --stop-preemptible \
-  --beaker-image olmo-torch2-gantry \
-  --cluster ai2/general-cirrascale-a100-80g-ib \
-  --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 \
-  --shared-memory 10GiB \
-  --venv base \
-  --yes \
-  --timeout=-1 \
-  -- /bin/bash -c "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 ${CONFIG_PATH} ${ARGS}"
+# 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
\ No newline at end of file