Skip to content

launch_vllm.py

Launches a vLLM server configured for hidden states extraction, used for online training or offline hidden states generation.

Basic Usage

python scripts/launch_vllm.py meta-llama/Llama-3.1-8B-Instruct

Arguments

Positional Arguments

  • model (str, required) Model name or path to extract hidden states from.

Speculators Arguments

  • --hidden-states-path (str, default: /tmp/hidden_states) The directory to initially cache hidden states to. Note: hidden states may then be moved or deleted by training/offline data generation.

  • --target-layer-ids (int list, default: auto-select) Space-separated list of integer layer IDs from which to capture hidden states. Note: if --include-last-layer is enabled (default), the model's last layer will be appended to this list. Default: [2, num_layers//2, num_layers-3]

Important: If set, you must also pass the same layer ids to the training script using --target-layer-ids, excluding the final layer — training takes the auxiliary layers only. For the full example below, that is --target-layer-ids 5 20 40.

  • --include-last-layer / --no-include-last-layer (flag, default: True) Append the last layer (num_hidden_layers) to target_layer_ids for verifier hidden states extraction.

  • --dry-run (flag) Print the command that would be executed without running it.

vLLM Arguments

All arguments after -- are passed directly to vLLM. Common vLLM arguments include:

  • --port: Server port (default: 8000)
  • --data-parallel-size: Number of data parallel instances
  • --tensor-parallel-size: Number of GPUs for tensor parallelism
  • --gpu-memory-utilization: GPU memory utilization (0.0 to 1.0)
  • --max-model-len: Maximum model context length
  • --trust-remote-code: Allow custom model code execution

See vLLM CLI documentation for full list of options.

Render throughput defaults

prepare_data.py --render-endpoint points at this server and drives its /v1/chat/completions/render endpoint. To keep that stage from serializing in vLLM's stock single-process, single-renderer front end, launch_vllm.py derives --api-server-count from the CPUs available to the process and defaults --renderer-num-workers to 2.

On the standard 384-CPU H100 node, this resolves to --api-server-count 18 --renderer-num-workers 2, paired with 72 preprocessing workers. The default intentionally leaves 25% of the CPU budget for native runtime threads and other application work; smaller hosts scale down automatically. Arguments passed after -- follow the defaults, so an explicit value still wins. No frontend defaults are added with --headless.

For non-headless launches, the script also defaults OMP_NUM_THREADS, OPENBLAS_NUM_THREADS, and MKL_NUM_THREADS to 1, and RAYON_NUM_THREADS to 2. These defaults prevent each API-server process from creating a host-sized native thread pool; values already present in the environment are preserved.

The training tutorial pins vLLM 0.27.1 for this path. These flags scale only the HTTP front end (chat-template application and tokenization); the engine and hidden-states connector are unaffected.

Full Example

python scripts/launch_vllm.py \
  meta-llama/Llama-3.1-70B-Instruct \
  --hidden-states-path /data/hidden_states \
  --target-layer-ids 5 20 40 80 \
  -- --data-parallel-size 2 --tensor-parallel-size 4 \
  --port 8000