transformer-engine
Route Transformer Engine installation/build, PyTorch, JAX, and shared precision/compatibility workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- transformer-engine
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Route Transformer Engine installation/build, PyTorch, JAX, and shared precision/compatibility workflows.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/precision-recipes.md3.8 KB
- 参考文档references/repo-provenance.md1.5 KB
- 参考文档references/repo-routing-metadata.json451 B
- 参考文档references/troubleshooting.md3.7 KB
- 脚本scripts/inspect_transformer_engine_runtime.py6.9 KB
03 SKILL.md 正文
Transformer Engine
Use this skill for NVIDIA Transformer Engine questions. Start here when the user mentions Transformer Engine, TE, the transformer_engine package, or a current-source checkout that must be installed, inspected, or explained.
Route map
| User need | Go to |
|---|---|
| Install, build, or verify the package or a source checkout | install-build |
| Use Transformer Engine with PyTorch | pytorch |
| Use Transformer Engine with JAX/Flax | jax |
| Choose a recipe, format, or hardware gate | precision recipes |
| Diagnose loader, version, CUDA, or recipe failures | troubleshooting |
Fast start
- If the runtime stack is unclear, run the bundled runtime inspector in
scripts/inspect_transformer_engine_runtime.py. - If the task is about installation or source builds, read the install/build sub-skill first.
- If the task is framework-specific, route directly to the PyTorch or JAX sub-skill rather than trying to infer from common package names alone.
- If the task is about FP8, MXFP8, NVFP4, BF16, or hardware compatibility, read the shared precision reference before choosing a recipe.
What this skill does not do
- It does not tell users to open original repo docs, examples, or tests as the runtime answer.
- It does not make FP8/MXFP8/NVFP4 claims without a support probe or hardware gate.
- It does not hide install or import problems behind generic advice when a bundled script or reference can be used instead.
Architecture notes
- A100/SM80 is BF16/FP16-capable but not an FP8/MXFP8/NVFP4 runtime target.
- FP8 requires compute capability 8.9 or newer.
- NCCL EP is a Hopper-or-newer feature.
- Mixed PyTorch/JAX environments are normal for this repo; use the sub-skill that matches the framework the user actually wants.
Shared references
- references/precision-recipes.md for recipe classes, support checks, and hardware selection.
- references/troubleshooting.md for cross-cutting loader, CUDA, cuBLAS, cuDNN, and version failures.
- scripts/inspect_transformer_engine_runtime.py for a quick import/support probe that does not depend on the source checkout.
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).