Multiply replica_rank by num_devices (tp × pp) to prevent port collisions when scaling to 2+ replicas with TP≥2 or PP≥2.

Root Cause

PR #57771 fixed port collisions in python/ray/llm/_internal/serve/engines/vllm/kv_transfer/base.py for TP/PP by using Ray Serve's replica_rank for port offsets instead of defaulting to 0. However, the implementation doesn't account for port spacing needed when each replica spawns multiple workers - so it could still lead to overlap.

Main issue: Consecutive replicas get consecutive port offsets (0, 1, 2, ...), but each replica actually needs num_devices (tp × pp) consecutive ports for its workers. This causes port ranges to overlap between replicas.

Example: 2 replicas, TP=2

Current implementation in base.py:_compute_port_offset():

  return rc.rank  # Returns 0, 1, 2, ...

Port allocation:

  Replica 0 (rank=0): offset=0  → base port 50000 → workers use [50000, 50001]
  Replica 1 (rank=1): offset=1  → base port 50001 → workers use [50001, 50002]
                                                                   ^^^^^ Collision!

Replica 0 Worker 1 and Replica 1 Worker 0 both bind to port 50001.

Example: 2 replicas, TP=2, PP=2

Each replica spawns 4 workers (tp × pp = 2 × 2 = 4). Each worker needs a unique port.

Current implementation (incorrect):
  Replica 0 (rank=0): offset=0  → workers use [50000, 50001, 50002, 50003]
  Replica 1 (rank=1): offset=1  → workers use [50001, 50002, 50003, 50004]
                                              ^^^^^ Multiple collisions!

With fix (multiply by num_devices = 4):
  Replica 0 (rank=0): offset=0  → workers use [50000, 50001, 50002, 50003]
  Replica 1 (rank=1): offset=4  → workers use [50004, 50005, 50006, 50007]
                                              ✓ No collisions

Solution:

Space replicas by num_devices (tp × pp) ports to reserve room for all workers:

Replica 0 uses ports: [base, base+1, ..., base+(num_devices-1)]
Replica 1 uses ports: [base+num_devices, base+num_devices+1, ...]

The fix uses llm_config.get_engine_config().num_devices which correctly accounts for both TP and PP workers.

Impact:

• Fixes port collisions when autoscaling to 2+ replicas with TP≥2 or PP≥2
• Handles combined TP+PP scenarios correctly (e.g., TP=2, PP=2 requires 4 ports per replica)
• Backward compatible: TP=1, PP=1 multiplies by 1 (no-op)
• DP deployments unchanged: vLLM handles spacing internally
• Single replica deployments unchanged: no other replica to collide with

Note (about Data Parallel)

DP deployments don't need this fix because vLLM already multiplies data_parallel_rank by tp_size for the offset internally:

# vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py:641-642
side_channel_port = base + (data_parallel_rank × tensor_parallel_size)

So for DP, the spacing is automatic - but for replica_rank, we do the offset multiplication ourselves since vLLM doesn't know about Ray Serve's replica concept. The fix uses num_devices instead of just tp_size to ensure PP workers also get unique ports.

Related: PR #57771, #55775, #58072