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
Backendimpl. OneDeviceDescriptorper TPU chip visible to the host, driven byTPU_ACCELERATOR_TYPE+TPU_NUM_DEVICES. - drv
- PJRT plugin loader.
Functions§
- register
- Registers the TPU backend with
reg.