Seeded generation reproducibility boundaries

Initial latent generation and stochastic sampler steps can consume random values. Reusing a generator object after one request starts from its advanced state unless it is explicitly reset or reconstructed. Diffusers v0.35.1 recommends passing generator state and explains this consumption behavior. A deterministic ODE-style sampler may use no new stochastic draws after initialization, but its output still depends on the initial latent. A stochastic sampler can require an entire per-step noise stream in addition to that initial state.

A global generator shared across requests can assign different draws when batch order, step count, cancellation, or retries change. Give each request an explicit stream policy, with stable logical identifiers and a declared retry rule. Separate per-request generators or a supported counter-based mapping can reduce scheduling coupling, but neither is implemented merely by attaching a request ID to a log. If a model uses multiple random sources, include every relevant source and the transformation from random state to tensors. Preserve actual initial latents or noise tensors when that is the needed comparison boundary.

Record model and autoencoder revisions, tokenizer and conditioning artifacts, adapters, scheduler configuration, realized grid, guidance policy, prediction type, precision, initial latent or generator state, and relevant software/hardware settings. Floating-point reduction order, kernel selection, nondeterministic operations, and batch shape can change results even with the same random draws. CPU-generated initialization can help align one source of randomness across devices but does not force the subsequent denoiser and decoder arithmetic to match bitwise.

PyTorch and the pinned Diffusers reproducibility documentation explicitly limit guarantees across releases and platforms. Deterministic-algorithm settings can improve repeatability in a supported fixed environment while changing performance or rejecting unsupported operations. Define the oracle: exact latent or pixel equality, bounded numerical error, perceptual similarity, or a distributional test. Passing a tolerance for one generated image is not proof that two samplers have the same output law, and bitwise equality on one machine is not a universal portability guarantee.

Replay a fixed request alone, in different compatible batch orders, after another request consumes random values, and after a neighboring request is canceled. Verify the intended stream assignment and compare outputs using the declared oracle. Test retries separately: the service may promise the same candidate again or intentionally create a new sample, but that policy should not depend on accidental process-global RNG progress. For resumable generation, preserve the latent, solver history, coordinate, and random state at one consistent boundary rather than only the original seed.

The modular instrument is a transparent sequence-position analogy, not a random generator. Reset, set both m=16 and a=5, predict residue fifteen after three additions, then Step three times. Separately Reset, set both m=16 and a=3, predict residue nine after three additions, then Step three times. Both sequence rule and position affect subsequent values. This short exact periodic orbit has no Gaussian sampling, statistical randomness, device kernels, or model outputs. It helps distinguish a fully specified stream position from the fact that two processes once started at zero.