Every output whose length only the call knows, embeddings, states, token vectors and probabilities, lands in an Arena you pass, and the verb answers a view over it without copying. Arena(executor) places it where that executor computes, decided once, and Arena(library) makes host pages for the calling thread.
Each call rewinds its arena, so the previous call’s views die and a steady loop allocates once; reading a stale one raises Status.INVALID_ARGUMENT.
To keep a result past the next call, pass that call another arena, and give every deferred call its own.
An arena from a VRAM executor holds device memory, which host accessors refuse with Status.DEVICE, and serves only that executor’s calls, keeping the executor alive until the arena goes.
The engine classifies every input pointer through the driver, so TapeView.of takes a host span or a host or device pointer, such as buffer.unsafe_ptr() of a MAX DeviceBuffer, with its offsets in host memory.
ArenaDescriptionuform_arena_description_t: where an arena's outputs live.struct ArenaDescriptionresidency | Residency | RAM for host pages, VRAM for one device's memory. |
|---|---|---|
device | Int32 | The device ordinal, as Executor.for_vram takes it, or -1 for host pages. |
Copyable, Deinitable, ImplicitlyCopyable, MovableArenaThe memory every output whose length scales with a call lands in, freed when it goes out of scope.3struct ArenaMade where an executor computes, host pages or the memory of its GPU, decided once; nothing is
allocated until a call writes into it. Every call handed it rewinds and refills it, which
invalidates the previous call's outputs, so a steady loop allocates once. One arena serves one
call at a time; a second concurrent use is refused by the engine as CONTENDED, which leaves
the holder's results readable.
A result refuses every read with INVALID_ARGUMENT once a later call rewinds its arena, and a
result over a device arena refuses every host read with DEVICE. A device arena serves only
calls on the executor that made it, and keeps that executor alive until it is freed.
Deinitable, Movable__init__Host pages for calls on the calling thread. Calls uform_arena_init.def __init__(out self, library: Library)def __init__(out self, executor: Executor)Memory where executor computes, serving the calls of its residency. Calls uform_arena_init.
executorExecutor | The executor whose calls it serves; host pages may outlive it, and a device arena keeps it alive. |
|---|
__deinit__Frees the arena and every output in it, then lets a device arena's executor go. Calls uform_arena_free.def __deinit__(deinit self)The executor and the loader reference are surrendered only after the call, for the reason
_ExecutorHandle records.
describeWhere its outputs live, as the executor that made it decided. Calls uform_arena_describe.def describe(self) -> ArenaDescription