profile-model-performance

Installation
SKILL.md

Profile model performance

Use this skill before changing runtime behavior. The goal is to produce a defensible bottleneck map and a short list of candidate optimizations, not to guess from code shape alone.

Workflow

  1. Scope the executed path.
    • Find the user-facing entry point: runner, CLI, interactive server, batch script, notebook, or downstream adapter.
    • Trace one generation step through input preparation, encode/context setup, model or denoise loop, cache update/finalize, decode, transfer, encode, and presentation.
    • Read flashdreams-integrations before changing framework boundaries or config contracts. Keep this skill focused on measurement and diagnosis.
    • Prefer no-GPU inspection first: config resolution, --help, --no-instantiate, static runner wiring, and small CPU tests.
Installs
1
GitHub Stars
403
First Seen
12 days ago
profile-model-performance — nvidia/flashdreams