---

name: speedster-optimization description: "Guides Speedster inference optimization workflows, compiler/backend selection, save/load, and troubleshooting for optimized models." disable-model-invocation: true metadata: disco-role: operating license: NO_LICENSE

Speedster optimization

Use this sub-skill when the user wants to accelerate an existing model, choose optimize_model parameters, save or load an optimized learner, or understand why a compiler/backend path was selected or skipped.

Triggers

• Optimize a PyTorch, TensorFlow, ONNX, Hugging Face, or Diffusers model.
• Choose metric_drop_ths, optimization_time, dynamic_info, ignore_compilers, ignore_compressors, or device.
• Save or reload an optimized model.
• Interpret telemetry, latency summaries, or backend-specific warnings.

Read next

• references/api-reference.md for verified function signatures, accepted inputs, and return behavior.
• references/workflows.md for end-to-end recipes and parameter choices.
• references/framework-recipes.md for framework-specific notes and compiler/backend selection.
• references/troubleshooting.md for install/import/backend and data-format failures.
• ../nebullvm-backends/SKILL.md when the question is really about DataManager, device parsing, compiler lists, or optional dependency probes.
• scripts/speedster_quick_probe.py when you want a safe import/signature/backend smoke probe.

What to include

• optimize_model, save_model, and load_model.
• Data input shapes and dynamic-shape guidance.
• Device selection and backend selection.
• Compatible compiler names and optional dependencies.
• Telemetry opt-out and save/load behavior.

What to exclude

• Full compiler installation procedures that belong to the NebullVM backend support sub-skill.
• Long notebook or source-repo example paths.
• Heavy benchmark runs and large optimization sweeps.

Quick decision rule

If the user asks “how do I make this model faster?”, start here. If the user asks “which backend/compiler package should I install?”, read the NebullVM backend sub-skill next.