nemo-mbridge-perf-megatron-fsdp
Operational guide for enabling Megatron FSDP in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.
- License 许可
- Apache-2.0
- 最近更新
- 2026-11-08
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- nemo-mbridge-perf-megatron-fsdp
- license
- Apache-2.0
- description
- Operational guide for enabling Megatron FSDP in Megatron-Bridge, including config knobs, code anchors, pitfalls, and verification.
- when_to_use
- Using FSDP-based data parallelism instead of DDP, or tracing an OOM or regression to a FSDP config change; 'use_megatron_fsdp', 'data_parallel_sharding_strategy', 'sharded data parallel', 'Megatron FSDP'.
02 资源文件 5
- 参考文档BENCHMARK.md4.2 KB
- 参考文档card.yaml2.1 KB
- 参考文档evals/evals.json1.4 KB
- 参考文档skill-card.md3.7 KB
- 其他skill.oms.sig4.7 KB
03 SKILL.md 正文
Megatron FSDP Skill
For stable background and recommendation level, see:
- @docs/training/megatron-fsdp.md
- @skills/nemo-mbridge-perf-megatron-fsdp/card.yaml
Enablement
Minimal Megatron FSDP override in Bridge:
cfg.dist.use_megatron_fsdp = True
cfg.ddp.use_megatron_fsdp = True
cfg.ddp.data_parallel_sharding_strategy = "optim_grads_params"
cfg.ddp.average_in_collective = False
cfg.checkpoint.ckpt_format = "fsdp_dtensor"
Example recipe fixup:
cfg = llama3_8b_pretrain_config()
cfg.dist.use_megatron_fsdp = True
cfg.ddp.use_megatron_fsdp = True
cfg.ddp.data_parallel_sharding_strategy = "optim_grads_params"
cfg.ddp.average_in_collective = False
cfg.checkpoint.ckpt_format = "fsdp_dtensor"
cfg.checkpoint.save = "/tmp/fsdp_ckpts"
cfg.checkpoint.load = None
Performance harness note:
python scripts/performance/launch.py --use_megatron_fsdp true
Code Anchors
Bridge config definition:
use_megatron_fsdp: bool = False
"""Use Megatron's Fully Sharded Data Parallel. Cannot be used together with use_torch_fsdp2."""
use_torch_fsdp2: bool = False
"""Use the torch FSDP2 implementation. FSDP2 is not currently working with Pipeline Parallel.
It is still not in a stable release stage, and may therefore contain bugs or other
potential issues."""
Bridge validation:
if self.dist.use_megatron_fsdp and self.dist.use_torch_fsdp2:
raise ValueError(...)
...
assert not self.dist.use_tp_pp_dp_mapping, "use_tp_pp_dp_mapping is not supported with Megatron FSDP"
...
assert self.checkpoint.ckpt_format == "fsdp_dtensor", (
"Megatron FSDP only supports fsdp_dtensor checkpoint format"
)
Runtime wrapper selection:
if use_megatron_fsdp:
DP = FullyShardedDataParallel
elif use_torch_fsdp2:
DP = TorchFullyShardedDataParallel
else:
DP = DistributedDataParallel
...
DP(
config=get_model_config(model_chunk),
ddp_config=ddp_config,
module=model_chunk,
...
pg_collection=pg_collection,
)
Perf harness overrides:
recipe.ddp.use_megatron_fsdp = True
recipe.ddp.data_parallel_sharding_strategy = "optim_grads_params"
recipe.ddp.keep_fp8_transpose_cache = False
recipe.ddp.average_in_collective = False
...
recipe.checkpoint.load = None
Pitfalls
- Public recipes often expose
use_megatron_fsdpbut still default tockpt_format="torch_dist". If save/load is enabled, switch tofsdp_dtensor. use_torch_fsdp2exists, but on the validated branch Bridge still fails before training because_ddp_wrappassespg_collection.- CPU offloading is only valid when
pipeline_model_parallel_size == 1and activation recomputation is disabled. - Upstream warns that FSDP and TP/CP can want different
CUDA_DEVICE_MAX_CONNECTIONSsettings on Hopper and earlier. - Megatron FSDP and FSDP2 are mutually exclusive.
Verification
Use the existing 2-GPU functional smoke test:
CUDA_VISIBLE_DEVICES=0,1 uv run python -m torch.distributed.run --nproc_per_node=2 \
-m pytest tests/functional_tests/training/test_megatron_fsdp.py::TestMegatronFSDP::test_fsdp_pretrain_basic -v -s
Success criteria:
- Pytest reports
1 passed - The log shows finite loss at the last iteration
- The run finishes without a checkpoint format assertion
04 同仓库更多 Skills
vss-summarize-video
Use to summarize a recorded video via the LVS summarization microservice (HITL-gated) with a VLM fallback. Not for report generation or live RTSP captioning.
vss-setup-video-analytics-api
Use to deploy the vss-video-analytics-api REST service standalone (config-source, data-log bind, Elasticsearch, optional Kafka). Not for full warehouse deploy.
vss-setup-behavior-analytics
Use to deploy the vss-behavior-analytics service standalone (entrypoint, config-source, optional calibration). Not for the full warehouse deploy.
vss-search-archive
Use this skill to run top-level VSS fusion search on archived video, or to ingest video files / RTSP streams for search. Do NOT use for ad-hoc visual Q&A (use vss-ask-video), live captioning (use vss-deploy-dense-captioning), or video summarization and reports (use vss-summarize-video).
vss-query-analytics
Use this skill when reading video-analytics metrics, incidents, alerts, and sensor data via the VA-MCP server (port 9901). Not for live VLM or incident-range narrative reports.
vss-manage-video-io-storage
Use to call the VIOS REST API (sensor list, timelines, clip extraction, snapshots, add/delete sensors and streams). Not for VLM inference or search.