Multistep samplers and state caching

A classical two-step Adams–Bashforth update illustrates the dependency: it combines the current field value with a previous one under a constant-step formula. Those evaluations must belong to the intended trajectory, coordinate spacing, and field. DPM-Solver multistep methods use their own structure-specific coefficients and parameterization; the classical formula is an explanatory comparison, not their implementation. Variable step ratios, startup, and endpoints need the appropriate method rather than blindly reusing constant-step weights.

Changing conditioning, guidance, model weights, adapters, prediction conversion, or schedule can change the field whose history was stored. Some changes can be handled by a derived restart or lower-order transition, but arbitrary reuse has no automatic order guarantee. A pause/resume checkpoint must preserve the actual sample, coordinate, solver history, configuration, and any required random state together. Sharing a mutable scheduler object across interleaved requests without isolating that state can mix histories even when the denoiser itself is safely shared.

Static text-encoder embeddings can be reused when prompt tokens, encoder revision, adapters, relevant preprocessing, precision policy, and other conditioning inputs match. A denoiser output generally also depends on the current latent, noise coordinate, guidance context, and model state. Equal prompt and timestep do not imply equal latent states for two requests. Approximate feature or activation caching across steps deliberately omits some recomputation; it needs a stated error policy and quality evaluation rather than being described as exact memoization.

Diffusers v0.35.1’s DPMSolverMultistepScheduler stores previous model outputs, order-progress state, and a step index, and resets history when setting a new inference schedule. Those fields demonstrate why a request restart must use the intended initialization path. They are implementation details of the pinned version, not a promise that every scheduler is stateful in exactly the same way. Record the actual configuration and supported reset behavior, and use separate request state or a verified grouping mechanism for independent trajectories.

Build fixtures that interleave two requests, restart at a new grid, alter guidance midway, change an adapter, and cancel while cached buffers are still referenced. Compare a no-reuse reference at the same numerical settings. Track semantic validation failures alongside hit rate; a cache can achieve perfect hits by returning the wrong tensor. Memory ownership also matters: a retained history entry must not alias a scratch output overwritten by a later denoiser call. Release or recycle storage only after every scheduled user has completed.

The cache instrument measures capacity effects for its declared address stream. Reset, set both C=4 and s=4, predict four misses and thirteen hits after seventeen requests, then Step once. Separately Reset, set both C=2 and s=4, predict seventeen misses after seventeen requests, then Step once. One Step advances the initial sixteen-request stream by one. This fully associative LRU toy has no latent values, conditioning, solver history, or semantic invalidation. Its hit counts cannot establish that a real sampler cache’s keys or ownership rules are correct.