axlearn
Routes AXLearn training, language-model, vision, audio/ASR, and GCP CLI workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- axlearn
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Routes AXLearn training, language-model, vision, audio/ASR, and GCP CLI workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/repo-provenance.md2.1 KB
- 参考文档references/repo-routing-metadata.json530 B
- 参考文档references/troubleshooting.md2.5 KB
- 脚本scripts/check_install.py1.1 KB
03 SKILL.md 正文
AXLearn
AXLearn is a JAX-based deep learning library with a config system, trainer runtime, vision/audio experiment catalogs, and a GCP launcher/ops CLI.
Use this skill when the user asks about:
axlearn.commonconfigs, modules, trainers, inputs, learners, checkpointers, or launchers.axlearnCLI commands such asgcp config,gcp bundle,gcp launch,gcp vm,gcp bastion,gcp dataflow,gcp logs, orgcp auth.- Vision workflows such as ImageNet, ResNet, CLIP, or other image-classification configs.
- Audio/ASR workflows such as Conformer, LibriSpeech, feature extraction, or WER evaluation.
- GPT / language-model trainer catalogs, tokenizers, MoE configs, or flash-attention paths.
Start here
- Read
references/repo-provenance.mdif you need to check whether this skill is current for the checkout. - Read
references/troubleshooting.mdwhen installation, import, or optional dependency checks fail. - Use
scripts/check_install.pyfor a safe import/version smoke check. - Route to the matching sub-skill:
sub-skills/training-core/for trainer configs, fake-data smoke checks, and tokenizer setup.sub-skills/language-models/for GPT, MoE, flash-attention, and tokenizer catalog workflows.sub-skills/cli-cloud/for GCP config, bundle, launch, VM, bastion, Dataflow, logs, and auth.sub-skills/vision-workflows/for ResNet/ImageNet and other vision model recipes.sub-skills/audio-asr/for Conformer, LibriSpeech, and ASR workflows.
Installation and smoke check
For local inspection, install the editable package with the extras that match the workflow. Start with the base package, then add only the extras you need:
python -m pip install -e .
python -m pip install -e .[core,dev]
Common add-ons:
audiofor ASR workflows.gcpfor cloud CLI workflows.orbaxwhen checkpoint utilities are needed.devonly if you plan to run repo tests.
Minimal smoke checks:
python -I -c "import axlearn; print(axlearn.__file__)"
axlearn --help
If you are using the cloud CLI routes, also check:
axlearn gcp --help
Routing guidance
- Use
training-corefor local config construction,SpmdTrainer,launch_trainer_main, fake inputs, and short tutorial-style probes. - Use
language-modelswhen the task names Fuji, Gala, Honeycrisp, Qwen, C4, Pajama, MoE, or flash attention. - Use
cli-cloudwhen the task names GCP activation, bundling, launching, bastion, Dataflow, logs, or auth. - Use
vision-workflowswhen the task names ImageNet, ResNet, image classification, or CLIP-like vision recipes. - Use
audio-asrwhen the task names LibriSpeech, Conformer, speech features, ASR, or WER.
If the task spans trainer config mechanics plus a domain family, start in training-core and then jump to the domain sub-skill.
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).