| name | parallel-strategy-analyzer |
| description | Recommend DP/FSDP/TP/PP/EP/CP strategy mixes from model + hardware constraints, with memory/comm/compute/bubble estimates. Use for strategy planning, scale-out, OOM, or comparing configs. Not for implementing platform/op code (platform-dev / dist-op-dev) or running training jobs.
|
Parallel Strategy Analyzer
Single source of truth for formulas and decision trees in this directory.
The agent .agent/agents/parallel-strategy-analyzer.md is a thin proxy only.
When to use
- Choose or rescale parallel strategies; debug OOM; compare configs; estimate comm/bubble
Inputs / examples / full output shape / limits:
references/io-contract.md.
Analysis flow (load workflow before each phase)
1 Collect info → workflows/01-collect-model-info.md
2 Global baseline → workflows/02-global-baseline.md (memory + FLOPs, no parallel)
3 Strategy search → workflows/03-strategy-search.md
4 Cost analysis → workflows/04-cost-analysis.md (TP/CP/EP/DP + PP bubble)
5–7 Shard mem / MFU / score + report → workflows/05-scoring-output.md
Do Phase 2 before cost/memory-after-shard so “fits in memory” is not preferred
over strategies with prohibitive communication.
Tables & formulas (on demand)
Output (short)
Emit: baseline → recommendation + DeviceMesh code → post-shard mem → comm →
bubble → opts → top-3 alternatives. Details in io-contract.md.
