Skip to content

[WIP] integrate MORI for FSDP all-gather - #940

Draft
lorri-rao wants to merge 3 commits into
AMD-AGI:mainfrom
Z-Y00:mori-fsdp-all-gather
Draft

[WIP] integrate MORI for FSDP all-gather#940
lorri-rao wants to merge 3 commits into
AMD-AGI:mainfrom
Z-Y00:mori-fsdp-all-gather

Conversation

@lorri-rao

@lorri-rao lorri-rao commented Jul 31, 2026

Copy link
Copy Markdown
Contributor

Working in progress to integrate MORI for FSDP all gather

@lorri-rao
lorri-rao force-pushed the mori-fsdp-all-gather branch 3 times, most recently from a3f0347 to ae15c83 Compare August 6, 2026 23:38
@lorri-rao
lorri-rao force-pushed the mori-fsdp-all-gather branch from ae15c83 to 34e12ac Compare August 12, 2026 01:10
Wire MORI into TorchTitan and Megatron FSDP2, add runtime preflight and multi-node validation, and provide a clear custom all-gather backend selector.
@lorri-rao
lorri-rao force-pushed the mori-fsdp-all-gather branch from 34e12ac to 3b0a849 Compare August 12, 2026 02:10
Z-Y00 added 2 commits August 12, 2026 23:56
Size device-driven workspaces from observed shards and default dense-node collectives to async completion so large models avoid excess HBM while overlapping compute. Document execution modes and add comparison tooling and tests.
Use device events, automatic SHMEM sizing, and persistent backing-buffer registration to eliminate host waits and per-step IPC churn. Add standalone bandwidth and GEMM-interference reproducers with focused tests and documentation.
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.

2 participants