Nemo mbridge perf activation recompute vs Nemo mbridge perf memory tuning
Nemo mbridge perf activation recompute is an open-source agent skill for ai models and llm tooling. Nemo mbridge perf memory tuning is an open-source agent skill for knowledge and memory.
- Nemo mbridge perf activation recompute is best for fix a gpu out-of-memory error during megatron bridge pretraining.
- Nemo mbridge perf memory tuning is best for fix a training job that runs out of gpu memory on one rank.
- What it is
- Nemo mbridge perf activation recompute is an open-source agent skill for ai models and llm tooling.Nemo mbridge perf memory tuning is an open-source agent skill for knowledge and memory.
- Best for
- Fix a GPU out-of-memory error during Megatron Bridge pretrainingFix a training job that runs out of GPU memory on one rank
- Kind
- SkillSkill
- Pricing
- Open sourceOpen source
- Platforms
- Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, GooseClaude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Windsurf, Goose
- GitHub stars
- 4K4K
- Last commit
- Nov 2026Nov 2026
- License
- Apache-2.0Apache-2.0
- Open source
- Good for
- Fix a GPU out-of-memory error during Megatron Bridge pretraining; Pick which selective recompute modules to enable for an MoE model; Compare memory and step time with and without activation recomputeFix a training job that runs out of GPU memory on one rank; Reduce LoRA activation memory when sequence parallelism is enabled; Decide whether to change pipeline or tensor parallel sizes to fit a model
What Nemo mbridge perf activation recompute does
Selective module choice
Explains which recompute labels target attention, MLA, MoE, dense MLP and norm activations.
Full-layer recompute settings
Covers uniform and block placement with the number of layers to checkpoint.
Measurement protocol
Defines the matched control, per-rank memory fields and timing windows to record.
Compatibility checks
Lists restrictions involving MoE overlap, FP8 scaling, context parallelism and CUDA graphs.
Failure diagnosis table
Maps symptoms like persistent OOMs or allocator caching to likely causes and next steps.
Reference benchmark tables
Provides matched H100 runs for specific Megatron Bridge recipes as context, not rankings.
What Nemo mbridge perf memory tuning does
Allocator fragmentation fix
Explains how to enable expandable memory segments in PyTorch as the first step against out-of-memory errors.
Parallelism tradeoffs
Compares pipeline, tensor and data parallel resizing with their memory effects and measured throughput costs.
LoRA input re-gather
Describes the option that reduces retained LoRA-A activations when sequence parallelism is on.
Activation recompute guidance
Summarizes selective recompute choices and their GPU utilization cost on large dense models.
Memory estimation before launch
Shows how to call the Bridge theoretical estimator to inspect the most-loaded GPU shard.
OOM diagnosis table
Maps common symptoms such as single-rank OOM and CPU offload errors to causes and fixes.
Nemo mbridge perf activation recompute vs Nemo mbridge perf memory tuning: questions
Is Nemo mbridge perf activation recompute or Nemo mbridge perf memory tuning cheaper?
Both are open source. Compare current plans on each official site.
What is the difference between Nemo mbridge perf activation recompute and Nemo mbridge perf memory tuning?
Nemo mbridge perf activation recompute is an open-source agent skill for ai models and llm tooling. Nemo mbridge perf memory tuning is an open-source agent skill for knowledge and memory.
