tinker_backend

TinkerBackend - concrete :class:~evsys_sdk.training.backend.Backend implementation against tinker.ServiceClient.

Constructed via await TinkerBackend.create(model_name=..., lora_rank=...)

• the underlying create_lora_training_client_async is async, so the factory is async. Algorithm composers run the loop inside one asyncio.run(...) call so the constructor and the loop body share the same event loop:

  async def _train(self, ctx): backend = await TinkerBackend.create(...) loop = TrainingLoop(backend=backend, ...) return await loop.run(num_steps=cfg.max_steps)

  def train(self, ctx): return asyncio.run(self._train(ctx))

The wrapper is thin - every method is a one-to-one passthrough to the underlying tinker client, plus the one ergonomic affordance of returning a wrapped :class:TinkerSamplingClient from snapshot_sampling_client (so callers don't have to know about the sampler URI machinery).