training-config
Integrate DeepSpeed training into PyTorch, author and validate config JSON/HJSON, plan launcher resources, choose ZeRO/offload settings, and handle checkpoint save/load/export safely.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training-config
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Integrate DeepSpeed training into PyTorch, author and validate config JSON/HJSON, plan launcher resources, choose ZeRO/offload settings, and handle checkpoint save/load/export safely.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/api-reference.md3.6 KB
- 参考文档references/configuration.md3.9 KB
- 参考文档references/launcher-and-checkpointing.md4.5 KB
- 参考文档references/troubleshooting.md4.0 KB
- 脚本scripts/launcher_resource_preview.py5.7 KB
- 脚本scripts/validate_ds_config.py7.8 KB
03 SKILL.md 正文
DeepSpeed Training Config
Use this sub-skill when a future agent needs to wire DeepSpeed into a PyTorch training loop, create or audit ds_config.json, build deepspeed launcher commands, select ZeRO stages/offload, or plan training checkpoint save/load/export. This sub-skill covers training runtime setup only; route inference-only kernel injection to inference-injection, pipeline/MoE/sequence parallel details to parallelism-moe, and build diagnostics to ops-tooling.
Fast Path
- Integrate the model with
deepspeed.initialize(...); calldeepspeed.init_distributed()only when the script needs a distributed process group beforeinitialize. - Author config through
references/configuration.md, then runscripts/validate_ds_config.py --world-size <ranks> ds_config.jsonbefore launching. - Build resource filters and
--no_sshcommands throughreferences/launcher-and-checkpointing.md; preview filters withscripts/launcher_resource_preview.py. - Choose ZeRO settings with
references/configuration.md, keeping deprecated aliases out of new configs. - Plan checkpoint calls so all ranks participate; for ZeRO-3 exports, enable
stage3_gather_16bit_weights_on_model_saveor use the generatedzero_to_fp32.py/utility path after checkpointing.
Bundled References
references/api-reference.md:deepspeed.initialize,init_distributed, engine training calls,DeepSpeedConfig, andDeepSpeedZeroConfigfield facts.references/configuration.md: batch-size contract, JSON/HJSON authoring, precision, optimizer/scheduler, ZeRO/offload choices, and safe validation workflow.references/launcher-and-checkpointing.md: launcher hostfile/include/exclude/no-SSH syntax plus checkpoint save/load/export guardrails.references/troubleshooting.md: common failures and fixes for config parsing, launcher filters, checkpoint hangs, ZeRO-3 export, and CUDA op probing.
Bundled Scripts
scripts/validate_ds_config.py: Parses JSON or HJSON whenhjsonis installed, detects duplicate JSON keys, checks batch-size consistency, and warns about selected ZeRO pitfalls without importing or launching DeepSpeed.scripts/launcher_resource_preview.py: Parses hostfile text plus--include/--excludestyle filters and reports the selected resources or conflicts without launching training.
Native Candidates
Safe native candidates to consult during verification are the config parser/dict tests and launcher parser tests. Broad GPU/distributed checkpoint matrices should be skipped by default unless the verifier has explicit safe hardware and time approval.
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).