Feature Request: P/D support

[lionelvillard]

The vllmd sidecar is already using vllm-sim for local e2e testing.

In order to test P/D the simulator must support

• max_completion_tokens (alias for max_tokens)
• P/D simulation. Details to follow as design is being worked out.

@nilig @mayabar

[lionelvillard]

For debugging purpose, P/D vLLM uses a shared filesystem to transfer KVs: https://github.com/vllm-project/vllm/blob/ad18a3b440a7403cfc7e465f0d4604b8111a0fc3/vllm/distributed/kv_transfer/kv_connector/v1/shared_storage_connector.py#L71

[mayabar]

@lionelvillard We added PD support to Inference-sim (nixl connector type only), can you please confirm that it contains all required features

[mayabar]

/close