domains / families / science / nemo-mbridge-perf-megatron-fsdp
nemo-mbridge-perf-megatron-fsdp
separate
Operational guide for enabling Megatron FSDP in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.
When to use
- enabling Megatron FSDP
- using FSDP-based data parallelism instead of DDP
- tracing an OOM or regression to a FSDP config change
- configuring use_megatron_fsdp
- configuring data_parallel_sharding_strategy
- sharded data parallel setup
- Megatron FSDP verification
When not to use
- using torch FSDP2
- using use_tp_pp_dp_mapping
- using torch_dist checkpoint format
Skill page
Source
| Repository | Path |
|---|
| NVIDIA/skills | skills/nemo-mbridge-perf-megatron-fsdp/SKILL.md |