Skip to content

feat: add IPv6 support for network utilities and distributed training - #2

Open
narutolhy wants to merge 1 commit into
mainfrom
ipv6-support
Open

feat: add IPv6 support for network utilities and distributed training#2
narutolhy wants to merge 1 commit into
mainfrom
ipv6-support

Conversation

@narutolhy

Copy link
Copy Markdown
Collaborator

Summary

  • Add is_ipv6_address() and format_addr() helpers for proper IPv6 URL formatting (RFC 2732)
  • Extend gethostip() with IPv6 fallback for IPv6-only environments
  • Use format_addr() across infra (sglang_server, rpc_server, rollout_controller, remote_inf_engine)
  • Fix NCCL tcp:// init_method to bracket IPv6 addresses in fsdp_engine
  • Use Gloo cpu_group for object collectives in fsdp_engine to avoid cross-node hang when IB is unavailable
  • Add per-worker port_range partitioning to avoid TOCTOU port collision races
  • Bind SGLang to :: on IPv6 hosts while registering actual host IP

Test plan

  • Verify IPv6 address detection and formatting
  • Test multi-node training on IPv6-only cluster
  • Confirm NCCL weight update group initializes correctly with bracketed IPv6

🤖 Generated with Claude Code

- Add is_ipv6_address() and format_addr() helpers for proper IPv6 URL formatting
- Extend gethostip() with IPv6 fallback for IPv6-only environments
- Use format_addr() across infra (sglang_server, rpc_server, rollout_controller,
  remote_inf_engine) for proper IPv6 bracket wrapping in host:port strings
- Fix NCCL tcp:// init_method to bracket IPv6 addresses in fsdp_engine
- Use Gloo cpu_group for object collectives in fsdp_engine to avoid NCCL
  cross-node hang when IB is unavailable
- Add per-worker port_range partitioning to avoid TOCTOU port collision races
- Bind SGLang to "::" on IPv6 hosts while registering actual host IP

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant