LLM Inference Benchmarking Workload

LLM Inference Benchmarking Workload#

This Helm chart submits a job to benchmark the performance of vLLM running a model in the same container.

Prerequisites#

  1. Helm: Install helm. Refer to the Helm documentation for instructions.

  2. MinIO Storage (optional): To use pre-downloaded model weights from MinIO storage, the following environment variables must be set, otherwise models will be downloaded from Hugging Face. MinIO storage is also used for saving benchmark results.

    • BUCKET_STORAGE_HOST

    • BUCKET_STORAGE_ACCESS_KEY

    • BUCKET_STORAGE_SECRET_KEY

    • BUCKET_MODEL_PATH

  3. HF Token (optional): If you need to download gated models from Hugging Face (e.g., Mistral and LLaMA 3.x) that are not available locally, ensure a secret named hf-token exists in the namespace.

Implementation#

Basic configurations are defined in the values.yaml file, with key settings:

  • env_vars.TESTOPT: Must be set to either “latency” or “throughput”

  • env_vars.USE_MAD: Controls whether to apply the MAD approach (see below)

Note: If the specified model cannot be found locally, the workload will attempt to download it from Hugging Face.

A. Scenario-specific approach#

In this approach (env_vars.USE_SCENARIO is not “false”), scenarios are defined in the mount/scenarios_{$TESTOPT}.csv file. Modify this file to specify models, parameters, and environment variables for benchmarking. Each column defines a parameter or variable, and each row represents a unique scenario to benchmark.

The default configuration benchmarks latency using benchmark_latency.py from vLLM. Setting env_vars.TESTOPT to “throughput” will use benchmark_throughput.py instead.

Example 1: Benchmark latency scenarios (default)

helm template . | kubectl apply -f -

Example 2: Benchmark throughput scenarios

helm template . --set env_vars.TESTOPT="throughput" | kubectl apply -f -

B. ROCm/MAD standalone approach#

When env_vars.USE_MAD is not “false”, the ROCm/MAD repository will be cloned. The specified model (env_vars.MAD_MODEL) will be benchmarked according to preset scripts.

Example 3: Benchmark using MAD standalone approach with override settings

helm template . -f overrides/methods/MAD-Qwen2.5_0.5B.yaml | kubectl apply -f -