atomr-accel-cuda 0.10.0

GPU acceleration via the actor model. Wraps NVIDIA CUDA libraries (cuBLAS, cuDNN, cuFFT, cuRAND, cuSOLVER, cuSPARSE, cuTENSOR, cuBLASLt, NVRTC, NCCL) as supervised atomr actors with generation-validated buffers and a uniform async surface.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93

//! # atomr-accel-cuda
//!
//! GPU acceleration via the actor model. Wraps NVIDIA CUDA libraries as
//! actors on top of [`atomr`](../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.

// Subjective clippy lints that fight the actor-message design:
// * `type_complexity` — actor messages and kernel envelopes return
//   tuples of typed `Arc<CudaSlice<T>>` keep-alives; refactoring to
//   `type` aliases would worsen the public API.
// * `too_many_arguments` — kernel-launcher fns mirror the underlying
//   CUDA library entry points (cuDNN conv, cuSPARSE SpMV) which take
//   8–10 args; collapsing to a config struct just moves the fields.
// * `arc_with_non_send_sync` — CUDA driver handles (CudaGraph,
//   cudnnHandle) are `!Send` by design and only ever shared inside
//   the producing actor.
// * `large_enum_variant` — kernel-message enums have one large
//   conv-descriptor variant; boxing it would fragment the hot path.
#![allow(
    clippy::type_complexity,
    clippy::too_many_arguments,
    clippy::arc_with_non_send_sync,
    clippy::large_enum_variant,
    // Many `unsafe` FFI shims below intentionally elide the `# Safety`
    // doc — invariants are documented at the module level alongside the
    // matching `cudarc::*::sys` types.
    clippy::missing_safety_doc,
    // Phase 0 introduced typed-dispatch BlasMsg::Gemm; `BlasMsg::Sgemm`
    // remains as a deprecated back-compat alias used by examples /
    // benches / migration tests. The crate intentionally calls its own
    // deprecated surface during the deprecation window.
    deprecated,
    // `Drop` on owned-by-Arc handles is safe; the explicit `drop()` in
    // a few places is documentation, not behaviour.
    clippy::drop_non_drop,
    // Internal-only `len`/`is_empty` symmetry isn't load-bearing for
    // dispatch traits.
    clippy::len_without_is_empty,
    clippy::vec_init_then_push,
    clippy::not_unsafe_ptr_arg_deref,
    dead_code,
    unused_macros
)]

pub mod completion;
pub mod device;
pub mod dispatcher;
pub mod dtype;
pub mod error;
pub mod event;
pub mod gpu_ref;
pub mod graph;
/// 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`.
pub mod hopper;
pub mod host;
pub mod kernel;
pub mod memory;
#[cfg(feature = "nvrtc")]
pub mod module;
#[cfg(feature = "nccl")]
pub mod multi_device;
pub mod nvrtc_cache;
#[cfg(feature = "telemetry")]
pub mod observability;
pub mod p2p;
pub mod pipeline;
pub mod placement;
pub mod prelude;
pub mod replay;
pub mod stream;
#[cfg(feature = "streams")]
pub mod streams_pipeline;
pub mod sys;