Commit graph

2 commits

Author SHA1 Message Date
autoscriptlabs
201ec9e321 docs: clarify unified memory advantage on Grace Hopper
- Remove incorrect 'host memory staging' limitation
- Add section explaining NVLink-C2C unified memory benefits
- Zero-copy RDMA works automatically on DGX Spark
2026-01-09 14:16:48 -05:00
autoscriptlabs
031bc48953 Initial release: NCCL Mesh Plugin for direct-connect RDMA topologies
- Enables NCCL over multi-subnet mesh topologies
- 8+ GB/s bandwidth over 100Gbps RDMA
- Successfully tested with distributed LLM inference (Mistral-7B)
- Custom subnet-aware NIC selection
- Background handshake thread for deadlock-free connection setup
2026-01-09 14:09:33 -05:00