Add P/D network dimension to Vidur config explorer

[Copilot]

Vidur’s config explorer could not sweep P/D deployment ratios and network parameters, preventing direct capacity and Pareto comparisons across P/D network configurations.

• Design-space generation

  • Add optional pd_networks dimension.
  • Map ratio, P/D bandwidth, RDMA bandwidth, NVLink bandwidth, and transfer dtype to existing ReplicaConfig arguments.

• Validation

  • Reject invalid ratios, non-positive bandwidths, unsupported dtypes, missing names, and duplicate points.
  • Skip TP/PP combinations that produce zero P or D replicas before simulation.

• Job identity

  • Include every explicit network parameter in deterministic job/cache keys.
  • Preserve the existing mixed-serving job and key when pd_networks is omitted.

• Configuration example

pd_networks:
  - name: pd-balanced-100g
    pd_node_ratio: 0.5
    pd_p2p_comm_bandwidth: 100
    rdma_bandwidth: 100
    nvlink_bandwidth: 900
    pd_p2p_comm_dtype: float16

Capacity search and Pareto analysis continue consuming generated runs without changes.

• Fixes [Feature] Sweep P/D deployment and network dimensions in Vidur config optimizer #296

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}