Skip to main content

Crate ironaccelerator_tpu

Crate ironaccelerator_tpu 

Source
Expand description

§ironaccelerator-tpu

Google TPU backend via the PJRT (Pretty Just Runtime) C plugin interface. PJRT is the stable cross-framework TPU entry point — the same ABI JAX and PyTorch/XLA use. The backend dynamically loads pjrt_c_api_tpu_plugin.so (or libtpu.so on older images), calls GetPjrtApi, creates a PJRT_Client, and enumerates the attached TPU chips as devices.

Enumeration is all we need at the backend-trait level; compilation and execution sit on top of a StableHLO graph that the higher layers build and hand back through PJRT. That machinery lives in a follow-on crate and is not part of 1.1.

Hosts without the PJRT plugin — i.e. every machine that is not a Cloud TPU VM or a libtpu-provisioned GKE pod — get an unavailable backend and an empty device list.

Re-exports§

pub use backend::TpuBackend;
pub use backend::TPU_BACKEND;

Modules§

backend
TPU Backend impl. One DeviceDescriptor per TPU chip visible to the host, driven by TPU_ACCELERATOR_TYPE + TPU_NUM_DEVICES.
drv
PJRT plugin loader.

Functions§

register
Registers the TPU backend with reg.