Seeding¶

How experiment seeds reach each model back-end; see Models.

rupsycho.seeding ¶

Reproducible sampling: map an experiment seed onto each model back-end.

Every model call of an experiment carries a seed (parameters.seeds). How that seed reaches the model differs per back-end, and silently ignoring it would defeat the purpose of a seeded experiment:

Local Hugging Face pipelines have no seed argument. Their sampling is driven by global random number generators, so the seed is applied with transformers.set_seed right before every call.
API / server back-ends with a seed field (OpenAI, DeepSeek, Ollama, ...) receive the seed through a copy of the model with that field set; chat models around a Hugging Face endpoint receive it as a request parameter.
Back-ends without any seed support (e.g. Google Gemini) cannot be seeded. A warning is emitted once per model type; repetitions are then independent samples.

Custom model types can be registered with :func:register_seeder.

Seeder `module-attribute` ¶

Seeder = Callable[[Any, int], Runnable]

A function (model, seed) -> seeded runnable.

register_seeder ¶

register_seeder(model_type: type, seeder: Seeder) -> None

Register how models of model_type are seeded.

The most recent registration wins and takes precedence over the built-in strategies.

PARAMETER	DESCRIPTION
`model_type`	The model class (subclasses match too). TYPE: `type`
`seeder`	Function `(model, seed)` returning a runnable that behaves like `model` but samples reproducibly for `seed`. TYPE: `Seeder`

Example

from rupsycho.seeding import register_seeder

register_seeder(MyModel, lambda model, seed: model.model_copy(update={"rng_seed": seed}))

is_thread_safe ¶

is_thread_safe(model: Any) -> bool

Whether calls to model may run concurrently in several threads.

Local pipelines rely on process-global state (random generators, accelerator memory) and are therefore not thread safe; API clients are.

supports_seeding ¶

supports_seeding(model: Any) -> bool

Whether :func:seed_model can make model reproducible.

seed_model ¶

seed_model(model: Any, seed: int) -> Runnable

Return a runnable that samples from model reproducibly for seed.

PARAMETER	DESCRIPTION
`model`	The model (any LangChain runnable). TYPE: `Any`
`seed`	The seed of the current experiment run. TYPE: `int`

RETURNS	DESCRIPTION
`Runnable`	A runnable to use in place of `model`. For models that cannot be seeded the model
`Runnable`	itself is returned and a warning is emitted (once per model type).