resolve_lagllama_device_string

lagllama_device.resolve_lagllama_device_string

Lag-Llama device selection. models/lagllama/settings.yaml uses device: cpu for CPU-only workers and local runs. Explicit cuda / cuda:0 values are preserved. Requested cpu is never upgraded based on job tier so multivariate / covariate fan-out stays on the configured device (notably CPU-only Batch jobs).

resolve_lagllama_device_string(requested, *, worker_compute_tier, cuda_available) -> str

Return the device string passed to torch.device for Lag-Llama.

Parameter	Type	Default	Description
requested	str	-	Value from model settings (e.g. "cpu", "cuda").
worker_compute_tier	str	-	Raw WORKER_COMPUTE_TIER env (e.g. "gpu"). Ignored for device selection; kept for API stability with existing call sites.
cuda_available	bool	-	Whether CUDA is available at runtime. Ignored unless requested is a CUDA device string (callers may use it in the future for validation).