nemo-mbridge-perf-megatron-fsdp

nemo-mbridge-perf-megatron-fsdp is an operational reference skill from NVIDIA that documents how to enable Megatron FSDP (Fully Sharded Data Parallel) inside Megatron-Bridge. It addresses a concrete engineering problem: developers moving from DDP to FSDP-based data parallelism, or tracing an out-of-memory error or throughput regression back to an FSDP configuration change, need to know exactly which config knobs to set and which pitfalls to avoid. The skill provides the minimal config overrides (use_megatron_fsdp, data_parallel_sharding_strategy, average_in_collective, and the required fsdp_dtensor checkpoint format) plus an example recipe fixup.

Beyond enablement snippets, it supplies code anchors pointing at the relevant Bridge source lines for config definitions, validation assertions, runtime wrapper selection, and the performance-harness overrides, so an agent can verify guidance against the actual codebase. A pitfalls section covers common mistakes such as recipes defaulting to torch_dist checkpoint format, the mutual exclusivity of Megatron FSDP and FSDP2, CPU-offload constraints, and CUDA_DEVICE_MAX_CONNECTIONS differences on Hopper. A verification section gives an exact 2-GPU pytest smoke-test command and explicit success criteria (1 passed, finite loss, no checkpoint-format assertion).

It is aimed at ML infrastructure and performance engineers working with Megatron-Bridge large-model training. The skill is pure documentation plus code/command references; its outputs are configuration instructions and shell/pytest commands, and NVSkills-Eval reported a 100% Security score for it. There are no destructive operations, credential handling, or network exfiltration involved.