feat(pd): improve lifecycle supervision, health checks, and multimodal support

[hiworldwzj]

Summary

This PR improves Prefill/Decode (PD) disaggregated deployments, with a focus on container lifecycle management, topology-aware health reporting, connection liveness, KV transfer configuration, and multimodal correctness.

Process lifecycle and cleanup

• supervise the HTTP server and critical LightLLM submodule processes
• exit the launcher with a non-zero status when a managed process dies unexpectedly
• centralize signal handling and recursive process cleanup in SubmoduleManager
• add parent-process checks to PD KV manager, KV transfer, and status-reporting processes

PD connectivity and configuration

• add --pd_trans_mode with nccl and nixl options; default to nccl
• add a 15-second P/D heartbeat and remove nodes after a 30-second receive timeout
• raise the default LIGHTLLM_WEBSOCKET_MAX_SIZE to 128 MiB
• document proxy/NO_PROXY requirements for same-host or internal PD WebSocket traffic

Topology-aware health checks

• add --pd_master_mode
  • elastic: ready when at least one Prefill and one Decode node are registered
  • fixed topology such as 2p1d: ready only when the exact node counts are registered
• make PD Master /readiness reflect node readiness
• make fixed-topology PD Master /health combine inference progress, topology readiness, and P/D node health checks

Multimodal correctness

• initialize the tokenizer and generation configuration on PD Master
• preload multimodal inputs at PD Master and reuse the validated payload on Prefill
• require Prefill max_image_pixels and disable_image_resize to match PD Master
• use torch.int64 for model input_ids so virtual multimodal token IDs fit safely
• probabilistically isolate virtual multimodal token-ID ranges across Prefill nodes when no Config Server is available
• document that probabilistic ranges reduce, but do not eliminate, cross-node collision risk

Deployment scripts and documentation

• update PD deployment examples for the new transfer-mode argument
• add test/start_scripts/qwen35/qwen35_pd_2p1d.sh
  • two TP2 Prefill nodes
  • one TP4 Decode node
  • CPU cache enabled only on Prefill
  • NCCL KV transfer and Qwen3.5 MTP configuration
  • PD Master starts after all P/D launchers

Validation

• targeted unit tests: 25 passed
• Qwen3.5-0.8B NCCL 2P/1D deployment reached healthy and ready state
• PD Master, both Prefill nodes, and Decode node returned HTTP 200 from /health
• three OpenAI-compatible multimodal chat requests returned HTTP 200 and correctly identified their images
• Decode logs confirmed MTP execution; aggregate draft-token acceptance was approximately 38.24% for the three requests
• graceful shutdown left no LightLLM processes or listening service ports behind

Review follow-ups before merge

• avoid blocking forever when a submodule exits before sending its initialization result
• strictly validate fixed pd_master_mode strings instead of accepting malformed suffixes such as 2p1x
• initialize model_created for PD Master so /v1/models does not pass created=None to ModelCard

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.