fx-graph-workflows
Operate Optimum Torch FX graph transformations and version-gated tensor parallelism guidance.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- fx-graph-workflows
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Operate Optimum Torch FX graph transformations and version-gated tensor parallelism guidance.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/fx-optimization.md9.3 KB
- 参考文档references/tensor-parallelism.md7.6 KB
- 参考文档references/troubleshooting.md7.1 KB
- 脚本scripts/fx_transform_smoke.py8.8 KB
03 SKILL.md 正文
fx-graph-workflows
Use this sub-skill when a task involves Optimum's Torch FX graph workflow surface:
- CPU-compatible graph rewrites from
optimum.fx.optimization. - Custom
Transformation/ReversibleTransformationclasses,compose(...), andreverse=Truerestoration. - GraphModule lint/recompile behavior and computation-preserving validation.
- Cautious triage for optional automatic tensor parallelism from
optimum.fx.parallelization.
Do not use this sub-skill for GPTQ quantization, exporter task mapping, CLI export commands, dummy inputs, or normalized configs. Route those to the generated gptq-quantization, exporters-and-cli, or utilities-and-configs sub-skills.
Linked runtime files
- references/fx-optimization.md — transformation APIs, GraphModule expectations, built-in transformations, compose/reverse behavior, and local validation patterns.
- references/tensor-parallelism.md — optional automatic tensor-parallel API map, backend/version gates, layer replacements, and safe triage.
- references/troubleshooting.md — common failures and recoveries for FX imports, graph mutation, preservation checks, native-test downloads, Python 3.11 tensor-parallel import errors, and CUDA/NCCL requirements.
- scripts/fx_transform_smoke.py — deterministic CPU smoke test adapted from transformation behavior tests using only a tiny local
torch.nn.Module.
Operating sequence
-
Confirm the user is asking for FX graph transformations or tensor-parallel triage, not export/quantization/config utilities.
-
For optimization work, require a
torch.fx.GraphModuleor trace a localtorch.nn.Moduleinto one. Keep inputs local and deterministic; avoid Hub downloads unless the user explicitly authorizes them. -
Read references/fx-optimization.md before changing a graph. Apply transformations with default lint/recompile unless batching a compose chain.
-
Run the bundled smoke test when validating the local environment:
python scripts/fx_transform_smoke.py python scripts/fx_transform_smoke.py --check-compose -
For tensor parallelism, read references/tensor-parallelism.md first. Treat it as optional/advanced unless the environment has a compatible Python, CUDA-capable PyTorch with
torch.compile, initialized distributed/NCCL process groups, and enough GPUs for the requested world size. -
Use references/troubleshooting.md whenever import, graph lint/recompile, preservation, reverse, or backend errors appear.
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).