exec-remote（远程执行）

Remote Execution Skill

This skill handles running code on remote GPU or TPU clusters via SkyPilot.

Defaults

The following defaults apply unless the user explicitly overrides them:

Use these values directly — do NOT ask the user to confirm or re-enter them unless they specify otherwise.

1. Determine Target Device

Identify the target device from the user's request:

If the user does not specify a device, ask them which one to use.

2. Prerequisites

• The cluster must already be provisioned. Check that the corresponding cluster name file (.cluster_name_gpu or .cluster_name_tpu) exists and is non-empty in the project root.
• If the file does not exist or is empty, provision the cluster using the appropriate method (see Section 3).

3. Cluster Provisioning

GPU (Standalone SkyPilot)

GPU clusters are provisioned using the standalone launch_gpu.sh script. Locate it in the scripts/ directory alongside this skill definition.

# Common accelerator types: H100:1, A100:1, L4:1
bash <absolute_path_to_launch_gpu.sh> <accelerator_type> <experiment_name>

The launch script automatically updates .cluster_name_gpu.

TPU

There are two provisioning paths for TPU:

Path A: GKE-based (via deploy-cluster skill) — Recommended

This path provisions TPU on GKE using the full pipeline: apply-resource -> deploy-cluster -> exec-remote.

Each TPU type gets its own SkyPilot cluster named <cluster>-<username>-<tpu_type>, allowing multiple topologies to run in parallel.

1. Use the deploy-cluster skill which will:

- Use default cluster/project/zone unless user overrides - Ensure the GKE cluster exists (via apply-resource) - Configure SkyPilot for GKE - Launch a per-TPU-type SkyPilot cluster - Save the cluster name to .cluster_name_tpu

/deploy-cluster

Supported TPU types: v6e-1, v6e-4, v6e-8, v6e-16, v6e-32, v6e-64, v6e-128, v6e-256

Path B: Standalone SkyPilot TPU VM

For quick, single-node TPU usage without GKE, use the standalone launch_tpu.sh script:

# Common accelerator types: tpu-v4-8, tpu-v4-16, tpu-v6e-1, tpu-v6e-4
bash <absolute_path_to_launch_tpu.sh> <accelerator_type> <experiment_name>

The launch script automatically updates .cluster_name_tpu.

Teardown

# GPU
sky down $(cat .cluster_name_gpu) -y

# TPU (tear down all per-TPU-type clusters)
sky down <CLUSTER_NAME>-<USERNAME>-v6e-1 -y
sky down <CLUSTER_NAME>-<USERNAME>-v6e-4 -y

For GKE-based TPU, also remove the GKE cluster via /apply-resource delete if no longer needed.

4. Execution Command

GPU

sky exec $(cat .cluster_name_gpu) --workdir . "export CUDA_VISIBLE_DEVICES=0; uv run --extra gpu python <PATH_TO_SCRIPT> [ARGS]"

• export CUDA_VISIBLE_DEVICES=0; ensures deterministic single-GPU execution. Adjust for multi-GPU jobs.
• --extra gpu activates GPU optional dependencies (e.g. jax[cuda]).

TPU

sky exec <CLUSTER_NAME>-<USERNAME>-<TPU_TYPE> --workdir . "uv run --extra tpu python <PATH_TO_SCRIPT> [ARGS]"

• --extra tpu activates TPU optional dependencies (e.g. jax[tpu]).
• Use the per-TPU-type cluster name (e.g. sglang-jax-agent-tests-hongmao-v6e-1).

Common flags

• --workdir. syncs the current local directory to the remote instance before running.
• For pytest, use python -m pytest <test_path> instead of calling pytest directly.

5. Usage Examples

Run a benchmark on GPU:

sky exec $(cat .cluster_name_gpu) --workdir . "export CUDA_VISIBLE_DEVICES=0; uv run --extra gpu python src/lynx/perf/benchmark_train.py"

Run tests on TPU (single type):

sky exec sglang-jax-agent-tests-hongmao-v6e-4 --workdir . "uv run --extra tpu python -m pytest src/lynx/test/"

Run CI tests on multiple TPU types in parallel:

# Deploy both types (sequential — config.yaml is global)
python <deploy-cluster>/scripts/deploy.py sglang-jax-agent-tests v6e-1 asia-northeast1-b
python <deploy-cluster>/scripts/deploy.py sglang-jax-agent-tests v6e-4 asia-northeast1-b

# Execute in parallel
sky exec sglang-jax-agent-tests-hongmao-v6e-1 --workdir . "python test/srt/run_suite.py --suite unit-test-tpu-v6e-1" &
sky exec sglang-jax-agent-tests-hongmao-v6e-4 --workdir . "python test/srt/run_suite.py --suite e2e-test-tpu-v6e-4" &
wait

6. Operational Notes

• Logs: SkyPilot streams stdout and stderr directly to the terminal.
• Interruption: Ctrl+C may not kill the remote process; check SkyPilot docs for cleanup if needed.

7. GKE TPU Full Pipeline Procedure (Path A)

When the user requests to run code on TPU and no .cluster_name_tpu exists (or the user explicitly wants a new cluster), follow this procedure to orchestrate the full pipeline: apply-resource -> deploy-cluster -> exec-remote.

All parameters use defaults unless the user explicitly overrides them — do NOT ask for confirmation.

7.1 Collect Parameters

Only ask the user for parameters they haven't specified. Use defaults for everything else:

7.2 Create GKE Cluster (apply-resource)

Check prerequisites, then create the GKE cluster:

which xpk && which gcloud && which kubectl

xpk cluster create-pathways \
  --cluster $CLUSTER_NAME \
  --num-slices=$NUM_SLICES \
  --tpu-type=$TPU_TYPE \
  --zone=$ZONE \
  --spot \
  --project=$PROJECT_ID

7.3 Wait for GKE Cluster Ready

Poll until the cluster status becomes RUNNING. Do NOT proceed to deploy SkyPilot while status is PROVISIONING or RECONCILING — it will fail with SSL errors.

gcloud container clusters list --project=$PROJECT_ID \
  --filter="name=$CLUSTER_NAME" --format="table(name,location,status)"

7.4 Deploy SkyPilot on GKE (deploy-cluster)

Run the deploy script for each required TPU type. Each call creates a separate SkyPilot cluster.

# Deploy each TPU type (must be sequential — config.yaml is global)
# Only tpu_type is required; cluster_name and zone use defaults
python <path-to-deploy-cluster>/scripts/deploy.py v6e-1
python <path-to-deploy-cluster>/scripts/deploy.py v6e-4

This creates:

• $CLUSTER_NAME-$USERNAME-v6e-1 — SkyPilot cluster for v6e-1 tests
• $CLUSTER_NAME-$USERNAME-v6e-4 — SkyPilot cluster for v6e-4 tests

After completion, verify:

sky status                  # Both clusters should show as UP

7.5 Execute User Code (exec-remote)

Determine num_nodes from the TPU type (v6e-N where total_chips = N, num_nodes = N / 4, minimum 1):

For single-node types (v6e-1, v6e-4), omit --num-nodes. For multi-node types, add --num-nodes <N>.

# Single-node (v6e-1, v6e-4) — use per-TPU-type cluster name
sky exec $CLUSTER_NAME-$USERNAME-v6e-1 --workdir . \
  "uv run --extra tpu python <PATH_TO_SCRIPT> [ARGS]"

# Multi-node (v6e-8+)
sky exec $CLUSTER_NAME-$USERNAME-v6e-8 --num-nodes 2 --workdir . \
  "uv run --extra tpu python <PATH_TO_SCRIPT> [ARGS]"

# Parallel execution across multiple TPU types
sky exec $CLUSTER_NAME-$USERNAME-v6e-1 --workdir . "..." &
sky exec $CLUSTER_NAME-$USERNAME-v6e-4 --workdir . "..." &
wait

7.6 Cleanup

When the user requests teardown, remove both layers:

# 1. Remove SkyPilot clusters (one per TPU type)
sky down $CLUSTER_NAME-$USERNAME-v6e-1 -y
sky down $CLUSTER_NAME-$USERNAME-v6e-4 -y

# 2. Remove GKE cluster (only for Path A / GKE-based)
xpk cluster delete \
  --cluster $CLUSTER_NAME \
  --zone=$ZONE \
  --project=$PROJECT_ID