fastgen-validate

name: fastgen-validatedescription: >-  Validate NVIDIA FastGen integrations and training quality: numerical parity against reference implementations, seeded side-by-side media, gradient and trainability checks, checkpoint round trips and repository tests. Use it whenever a claim about correctness, parity or quality needs evidence.license: Apache-2.0metadata:   author: "Julius Berner <jberner@nvidia.com>"  domain: ml

Validate FastGen behavior

Apply fastgen-runtime. This skill owns comparison and acceptance evidence; fastgen-inference generates outputs and fastgen-train executes training. Follow the new-network checklist for adapters, and the run contracts when callback logs are the evidence.

1. Manifest. Freeze source and checkpoint revisions, environment, inputs, seeds and noise tensors, preprocessing, dtype, schedule, guidance, grid/NFE and output geometry; keep validation inputs held out from training. Fix tolerances and quality criteria before looking at results.
2. Parity. Compare upstream and FastGen at widening scope: normalised inputs and encoded conditioning and latents, the same noisy state and time through the denoiser, converted predictions and sampler steps, decoded outputs. Match the actual noise and time grids rather than seeds, since equal seeds do not imply equal RNG consumption. Prefer independently constructed upstream components over shared prepared ones, check that shared weights and buffers survive preparation, and treat precision differences as a separate experiment. Find the first differing intermediate before loosening tolerances; a loosened tolerance is a separately documented comparison that leaves the failure against the manifest's criteria on record.
3. Media. Generate matched seeded examples through both implementations across prompts and conditioning modes; present side-by-side galleries for human judgment and record judgments separately from metrics. Small galleries are not benchmarks.
4. Trainability. Check finite losses and gradients before clipping, intended parameter updates and frozen modules; run generation followed by backward under the real distributed wrapper, since inference can leave lazy state behind. Inspect per-iteration media and fixed-input outputs around a short SFT run, exercise auxiliary phases, save, reload and compare outputs, and verify all saved training state when resume is claimed.
5. Distillation quality. Compare held-out teacher and student outputs at the target NFE with the actual checkpoint and sample count; use task metrics (for images the FID script), perceptual inspection and timing/memory measurements. Count network evaluations, not sampling intervals, when methods use guidance or multiple heads.
6. Tests. Reuse the patterns in tests/ (shared helpers, markers such as large_model and integration, tiny CPU models) and run make lint and the relevant tests as described in CONTRIBUTING.md before claiming a code change is validated.

Keep a compact report with commands, manifests, metrics, media, tracker links, hardware and explicit pass/fail/not-run states; keep large artifacts outside the source tree and name repeated evaluations distinctly so they cannot replace the reference. Measure and report reference nondeterminism separately.