Higher Performance with Lower SM Occupancy through Zero-Copy and TMA Offloading#453

Open

monethuang1 wants to merge 26 commits intodeepseek-ai:mainfrom

monethuang1:trmt-zero-copy

Commits on Oct 10, 2025

Init zero-copy internode dispatch/combine

committed
Use 1 SM per channel for cached_notify
Zhehao Lin
committed
add combine tma alighment and fix rdma receivers num

Xiaojie Huang
authored and
Zhehao Lin
committed
Reduce SMEM usage
Zhehao Lin
committed
Support multiple buffers for zero-copy
Zhehao Lin
committed
Add assertion against too large recv token count
Zhehao Lin
committed
Use separate sizes for Dispatch/Combine input buffers
Zhehao Lin
committed
Ensure NVLink fused buffers are properly aligned by reordering NVL buffer segments
Zhehao Lin
committed
Cleanup TMA-related code (1)
Zhehao Lin
committed
Cleanup TMA-related code (2)
Zhehao Lin
committed
Don't change NVSHMEM_CUMEM_GRANULARITY
Zhehao Lin
committed
Cleanup TMA-related code (3)
Zhehao Lin
committed
Use elect_one_sync()
Zhehao Lin
committed
Add assertion against unsupported cases
Zhehao Lin
committed
Fix src_meta size
Zhehao Lin
committed
Increase combine SMEM size for 8K hidden_size
Zhehao Lin
committed
Error when hidden size is too large
Zhehao Lin
committed
Improve comments
Zhehao Lin
committed
Fix NVL buffer size hint
Zhehao Lin
committed
Fix NVL buffer size hint
Zhehao Lin
committed
Restore .gitignore
Zhehao Lin
committed
Minor format
Zhehao Lin
committed
More minor
Zhehao Lin
committed
Minor (comment)
Zhehao Lin
committed
Fix build for 32-bit topk idx
Zhehao Lin
committed

Commits on Oct 11, 2025

Fix calc_diff for combine
Zhehao Lin
committed