AGENT SKILL

deepspeed

Use DeepSpeed for distributed training, inference acceleration, ZeRO configuration, parallelism/MoE design, profiling, autotuning, and operational diagnostics.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
deepspeed
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Use DeepSpeed for distributed training, inference acceleration, ZeRO configuration, parallelism/MoE design, profiling, autotuning, and operational diagnostics.
disable-model-invocation
true

02 资源文件 6

03 SKILL.md 正文

DeepSpeed Repo Skill

Use this skill when a user asks to work with DeepSpeed APIs, configuration files, launch commands, inference acceleration, model parallelism, MoE, profiling, autotuning, or operational tooling. DeepSpeed is a PyTorch-focused optimization library with many optional GPU, distributed, and compiled-op paths, so route first by the user's workflow and then check the relevant troubleshooting reference before running heavy commands.

First Checks

  1. Confirm the user wants DeepSpeed itself, not a downstream framework integration that merely accepts a DeepSpeed config.
  2. Read references/install-and-environment.md when installing, importing, or debugging optional CUDA/ROCm/vendor accelerator behavior.
  3. Run scripts/check_deepspeed_env.py for a read-only environment check, then scripts/inspect_deepspeed_api.py when API fields or signatures may have drifted.
  4. Check references/repo-provenance.md before refreshing this skill against a source checkout.
  5. Do not run native examples, distributed tests, autotuning, NVMe/GDS tools, builds, downloads, or remote SSH commands without explicit safety confirmation.

Route By Task

  • Use sub-skills/training-config/ for PyTorch training integration, ds_config.json authoring, launcher resource filters, ZeRO/offload choices, checkpoint save/load/export, and config validation.
  • Use sub-skills/inference-injection/ for deepspeed.init_inference, kernel or manual injection policies, inference tensor parallelism, inference quantization, checkpoint reshaping, v2/FastGen routing, and hybrid-engine boundaries.
  • Use sub-skills/parallelism-moe/ for PipelineModule, pipeline schedules, MoE layers and optimizer groups, expert tensor parallelism, sequence parallel APIs, AutoSP, and activation checkpointing.
  • Use sub-skills/ops-tooling/ for ds_report, install/build diagnostics, JIT/prebuilt op flags, autotuning, FLOPS profiling, monitor backends, DeepNVMe/AIO/GDS tools, compression APIs, and repo policy checks.

Shared References

  • references/install-and-environment.md: package installation, PyTorch-first requirement, optional extras, compiled ops, and backend safety tiers.
  • references/troubleshooting.md: cross-cutting import, accelerator, config, launcher, and optional-tool failure modes.
  • references/repo-provenance.md: source snapshot and refresh baseline.

Shared Scripts

  • scripts/check_deepspeed_env.py: read-only import, package metadata, PyTorch, CUDA visibility, and installed CLI discovery.
  • scripts/inspect_deepspeed_api.py: read-only signature and config-field inspection for the main DeepSpeed APIs covered by this skill.

Safety Defaults

  • Prefer read-only commands first: imports, --help, config parsing, and tiny CPU checks.
  • Treat deepspeed launcher commands, checkpoint tests, CUDA op builds, autotuning, ds_io, ds_nvme_tune, GDS/AIO code, and ds_ssh as real workload or infrastructure commands.
  • Do not promise GPU, compiled-op, NVMe, GDS, monitor, or model-download behavior unless the current environment has been explicitly verified for that path.
  • Keep generated examples self-contained. Do not tell future agents to open or execute original repo docs, tests, scripts, examples, or notebooks as runtime dependencies.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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).

★ 351