Crate atomr_accel_cuda 

atomr-accel-cuda

GPU acceleration via the actor model. Wraps NVIDIA CUDA libraries as actors on top of atomr. See README.md and the architecture document under docs/ for the full design.

Foundation Phase F1 (current)

• Two-tier supervision: device::DeviceActor (stable address) ↔ device::ContextActor (owns Arc<CudaContext>, restartable).
• gpu_ref::GpuRef with generation-token validity checks.
• dispatcher::GpuDispatcher pinning actor execution to a single OS thread.
• completion::HostFnCompletion for sub-microsecond stream completion via cuLaunchHostFunc.
• stream::PerActorAllocator as the default §5.7 strategy.
• kernel::BlasActor performing cuBLAS SGEMM as the canonical demo.

Phases F2–F5 (cuDNN, cuFFT, NCCL, TensorRT, the PythonGpuBridge) and the four blueprint sub-crates are deferred.

Modules

completion

Completion strategies (§5.10).

device

DeviceActor (outer tier) + ContextActor (inner tier) — §5.11.

dispatcher

GpuDispatcher (§5.1) — pinned single-thread runtime that ensures the actor’s CUDA context stays current on the same OS thread for the actor’s whole lifetime.

dtype

CudaDtype — CUDA-side dtype mappings and capability markers.

error

Error taxonomy and the supervisor decider for context-poisoning recovery (§5.3, §5.11 of the architecture document).

event

EventActor — typed actor surface around CudaEvent.

gpu_ref

GpuRef<T> — opaque, message-friendly handle to a GPU buffer (§5.8).

graph

GraphActor — record a CUDA stream-capture once, replay many.

hopper

Phase 5: Hopper / Blackwell host-side primitives. The module surface is always compiled (the tma::TensorMapDescriptor builder and cluster::LaunchSpec types are useful even on hosts that don’t link a Hopper driver). The hopper cargo feature gates the FFI implementations of cuTensorMapEncodeTiled / cudaLaunchKernelExC. Hopper / Blackwell primitives (Phase 5).

host

Host-side support: pinned (page-locked) memory pool + PinnedBuf<T>.

kernel

Kernel-actor wrappers around CUDA library handles (§3.2).

memory

Managed (unified) memory + Phase 3 driver-API helpers.

module

ModuleActor — load prebuilt cubin/PTX from disk (or memory) and launch its kernels.

multi_device

Top-level multi-device actors that span multiple DeviceActors.

nvrtc_cache

Persistent disk cache for NVRTC-compiled CUDA kernels (Phase 0.6).

observability

Observability glue: install [atomr_telemetry::TelemetryExtension] on a host ActorSystem and expose a small set of GPU-specific probes that callers feed from kernel actors / placement actors / stream allocators.

p2p

P2P (peer-to-peer) topology + cross-device async memcpy.

pipeline

Multi-stream pipeline pattern.

placement

PlacementActor — picks the best-fit DeviceActor for each request based on a configurable PlacementPolicy.

prelude

Common imports for users of atomr-accel-cuda.

replay

Deterministic-replay harness.

stream

Stream allocation strategies (§5.7).

streams_pipeline

atomr-streams-based pipeline helpers — the F10 successor to the actor-based crate::pipeline::PipelineExecutor.

sys

sys — thin Rust wrappers over cudarc’s raw *::sys FFI for library entry points that aren’t yet exposed by the safe layer.