What problem does it solve? Slow save/load checkpoint tests in mflux download real model weights and generate images, making them expensive to run. This Skill guides writing a fast twin test that exercises the real ModelSaver/WeightLoader/WeightApplier code path on real component classes built at toy dimensions, completing in about a second with no downloads. ## Core Features & Use Cases - Checkpoint roundtrip coverage: Proves a model's quantized checkpoint survives a save/load roundtrip byte-for-byte using the shared TinyCheckpointRoundtrip harness. - Dimension constraint guidance: Derives shrink knobs from each component's real constructor while respecting interlocking constraints like rope_axes_dim summing to head_dim and quantization_group_size multiples. - Known pitfall handling: Covers component dict keying by name, dual seeds, skip_quantization gaps, and tensors_per_shard tuning for multi-shard coverage. - Use Case: After porting a new model to mflux, ask for a tiny test to get cheap, fast checkpoint coverage that mirrors the slow save/load test without touching src/. ## Quick Start Ask the AI to make a tiny test for a specific mflux model, such as "make tiny test for qwen", and it will produce the test file under tests/model_saving/.
