Skip to content

Environment Variables

torch_ha reads two user-facing environment variables directly, listed below. For variables shared across the SDK, such as those for logging and backend selection, see the documentation for the lower layers.


Memory

Variable Type Default When applied Description
HA_FILL_KERNEL Integer On First fill When 0, fill_ and zero_ do not use the device memset kernel and fill everything on the host path
HA_FILL_KERNEL_MIN_BYTES Bytes 1048576 (1 MiB) First fill Fills smaller than this use the host path. When 0, every fill the kernel can handle goes to the kernel. A negative or non-numeric value falls back to the default

Variables of Lower Layers

Layer Variables
Runtime HA_* — see hart environment variables
Collective communication LCCL_* — see LCCL environment variables
Distributed launcher MASTER_ADDR, MASTER_PORT, RANK, WORLD_SIZE — standard PyTorch

What Is Not Controlled by Environment Variables

Setting How to set it
Device selection Set through the API with torch.accelerator.set_device_index. Distributed §4.2
Weight layout Applied automatically when an nn.Linear or nn.Embedding module is moved to lpu
Eager compute kernels (add, sub, mul, mm) Always enabled. No environment variable disables or changes them. Operators §3.1

See Also