distributed-accelerators
Choose, configure, validate, and troubleshoot Lightning Trainer and Fabric accelerators, devices, precision, DDP, FSDP, DeepSpeed, model parallel, cluster, and hardware backends.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- distributed-accelerators
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Choose, configure, validate, and troubleshoot Lightning Trainer and Fabric accelerators, devices, precision, DDP, FSDP, DeepSpeed, model parallel, cluster, and hardware backends.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/precision-and-devices.md6.0 KB
- 参考文档references/strategy-guide.md10.3 KB
- 参考文档references/troubleshooting.md8.2 KB
- 脚本scripts/strategy_config_check.py8.8 KB
03 SKILL.md 正文
Lightning Distributed Accelerators
Use this sub-skill when the task involves Trainer or Fabric accelerator/device selection, precision, distributed strategies, FSDP, DeepSpeed, DDP, tensor/model parallel, CUDA/GPU, MPS, TPU/XLA, multi-node clusters, or optional dependency and hardware troubleshooting.
Route Here
- Choose
accelerator,devices,num_nodes,strategy,precision, plugins, and cluster environment settings forlightning.pytorch.Trainerorlightning.Fabric. - Diagnose GPU visibility, device-count mismatches, CPU fallback, MPS/TPU/XLA limitations, NCCL/Gloo launch problems, and optional dependency import failures.
- Configure
DDPStrategy,FSDPStrategy,DeepSpeedStrategy,ModelParallelStrategy, precision plugins, or strategy registry entries. - Plan large-model training with DDP, FSDP, DeepSpeed ZeRO, activation checkpointing, CPU offload, tensor parallelism, or 2D parallelism.
Route Elsewhere
- Core
LightningModule,LightningDataModule, callbacks, loggers, checkpoint callbacks, and standardTrainer.fitsemantics:../training-core/SKILL.md. - Manual Fabric loop structure,
fabric.setup,fabric.backward, checkpoint IO, and custom loop ownership:../fabric-expert-loops/SKILL.md. LightningCLI, YAML, environment-variable, and parser syntax for these settings:../cli-configuration/SKILL.md.
Start Here
- Identify whether the user is using
Trainer,Fabric, or CLI config; route to sibling sub-skills only for the non-distributed wiring. - Pick a baseline from
references/strategy-guide.md: start withstrategy="auto"for one process,ddpfor normal multi-GPU,fsdpfor DDP out-of-memory, anddeepspeedfor DeepSpeed-specific ZeRO/offload workflows. - Cross-check accelerator/device/precision support in
references/precision-and-devices.mdbefore recommending GPU, MPS, TPU, FP16, BF16, FP8, or quantization settings. - Validate syntax without launching workers:
python scripts/strategy_config_check.py --mode trainer --accelerator cpu --devices 1 --strategy auto --precision 32-true
python scripts/strategy_config_check.py --mode fabric --accelerator auto --devices 1 --strategy auto --precision bf16-mixed
- For failures, use
references/troubleshooting.mdand report whether the evidence is import-only, syntax-only, or requires real distributed/GPU validation. This skill was drafted from CPU inspection evidence and does not claim GPU or distributed runtime was validated.
Reference Map
references/strategy-guide.md: decision table, Trainer/Fabric configuration patterns, DDP/FSDP/DeepSpeed/model-parallel examples, strategy constructor knobs, and cluster launch notes.references/precision-and-devices.md: accelerator names,devicesformats, precision modes, optional precision plugins, and hardware support boundaries.references/troubleshooting.md: install/import, optional dependencies, invalid API usage, backend limitations, launch errors, device mismatch, and workflow-specific misconfigurations.scripts/strategy_config_check.py: safe public-package validator for import availability and syntactic configuration checks; it does not spawn subprocesses or start distributed workers.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).