ferrotorch_gpu/
lib.rs

1//! CUDA GPU backend for ferrotorch.
2//!
3//! This crate provides device management, memory allocation, and host/device
4//! data transfers built on [`cudarc`]. It is the bridge between ferrotorch's
5//! CPU tensor world and NVIDIA GPUs.
6//!
7//! # Feature flags
8//!
9//! | Feature | Default | Description |
10//! |---------|---------|-------------|
11//! | `cuda`  | **yes** | Links against the CUDA driver API via cudarc. Disable to compile on machines without a GPU. |
12//!
13//! # Quick start
14//!
15//! ```rust,no_run
16//! use ferrotorch_gpu::{GpuDevice, cpu_to_gpu, gpu_to_cpu};
17//!
18//! let device = GpuDevice::new(0).unwrap();
19//! let host_data = vec![1.0f32, 2.0, 3.0];
20//! let gpu_buf = cpu_to_gpu(&host_data, &device).unwrap();
21//! let back = gpu_to_cpu(&gpu_buf, &device).unwrap();
22//! assert_eq!(back, host_data);
23//! ```
24
25pub mod allocator;
26pub mod backend_impl;
27pub mod blas;
28pub mod buffer;
29pub mod conv;
30pub mod device;
31pub mod error;
32pub mod flash_attention;
33pub mod graph;
34pub mod kernels;
35pub mod memory_guard;
36pub mod module_cache;
37pub mod pool;
38pub mod rng;
39pub mod stream;
40pub mod tensor_bridge;
41pub mod transfer;
42
43// Re-exports for ergonomic use.
44pub use allocator::CudaAllocator;
45pub use backend_impl::{CudaBackendImpl, get_cuda_device, init_cuda_backend};
46pub use blas::gpu_bmm_f32;
47pub use blas::{gpu_bmm_f32_into, gpu_matmul_f32_into};
48pub use blas::{gpu_matmul_f32, gpu_matmul_f64};
49pub use buffer::CudaBuffer;
50pub use conv::gpu_conv2d_f32;
51pub use device::GpuDevice;
52pub use error::{GpuError, GpuResult};
53pub use flash_attention::gpu_flash_attention_f32;
54pub use kernels::{gpu_add, gpu_mul, gpu_neg, gpu_relu, gpu_sub};
55pub use kernels::{
56    gpu_add_into, gpu_embed_lookup_into, gpu_gelu_into, gpu_layernorm_into, gpu_mul_into,
57    gpu_permute_0213_into, gpu_scale_into, gpu_slice_read_into, gpu_small_matmul_into,
58    gpu_softmax_into, gpu_transpose_2d_into,
59};
60pub use kernels::{gpu_broadcast_add, gpu_broadcast_mul, gpu_broadcast_sub};
61pub use kernels::{gpu_causal_mask_indirect, gpu_slice_write_indirect};
62pub use kernels::{
63    gpu_dropout, gpu_embed_lookup, gpu_gelu, gpu_layernorm, gpu_permute_0213, gpu_slice_read,
64    gpu_slice_write, gpu_small_bmm, gpu_small_matmul, gpu_softmax, gpu_transpose_2d,
65};
66pub use memory_guard::{
67    MemoryGuard, MemoryGuardBuilder, MemoryGuardedDevice, MemoryHook, MemoryPressureListener,
68    MemoryReservation, MemoryStats, MemoryWatchdog, OomPolicy, PressureLevel,
69};
70pub use pool::{cached_bytes, empty_cache, empty_cache_all, round_len};
71pub use rng::{CudaRngManager, PhiloxGenerator, PhiloxState, cuda_rng_manager, fork_rng, join_rng};
72pub use tensor_bridge::{GpuFloat, GpuTensor, cuda, cuda_default, tensor_to_cpu, tensor_to_gpu};
73pub use transfer::{alloc_zeros, alloc_zeros_f32, alloc_zeros_f64, cpu_to_gpu, gpu_to_cpu};
ferrotorch_gpu/lib.rs

ferrotorch_gpu/
lib.rs
