Initial commit

2026-03-04 00:08:42 +00:00
commit 4d88dc0584
4449 changed files with 1556714 additions and 0 deletions
@@ -0,0 +1,50 @@
+//! RDMA Transport Layer for High-Performance Distributed Training
+//!
+//! This module provides RDMA (Remote Direct Memory Access) support for
+//! high-bandwidth, low-latency communication between nodes in a distributed
+//! training cluster.
+//!
+//! # Features
+//! - InfiniBand verbs integration via libibverbs
+//! - GPU Direct RDMA support for direct GPU-to-GPU transfers
+//! - Zero-copy data transfer
+//! - Reliable connected (RC) queue pairs
+//! - Completion queue polling for async operations
+//! - Collective communication primitives (AllReduce, Broadcast, etc.)
+//!
+//! # Requirements
+//! - libibverbs development libraries
+//! - RDMA-capable network adapter (InfiniBand, RoCE, iWARP)
+//! - For GPU Direct: NVIDIA GPUs with GPUDirect RDMA support
+//!
+//! # Example
+//! ```rust,ignore
+//! use rtx_distributed::rdma_transport::{RdmaTransport, RdmaConfig};
+//!
+//! let config = RdmaConfig::default();
+//! let transport = RdmaTransport::new(config)?;
+//! transport.connect(peer_addr).await?;
+//! transport.send(&data).await?;
+//! ```
+
+pub mod collectives;
+pub mod config;
+pub mod discovery;
+pub mod ffi;
+pub mod gpu_direct;
+pub mod transport;
+pub mod types;
+
+#[cfg(test)]
+mod tests;
+
+// Re-export all public types
+pub use collectives::{RdmaCollectiveStats, RdmaCollectives};
+pub use config::RdmaConfig;
+pub use discovery::{discover_rdma_devices, is_rdma_available};
+pub use ffi::*;
+pub use gpu_direct::{GpuDirectConfig, GpuDirectRdmaTransport};
+pub use transport::{
+    RdmaCompletionQueue, RdmaSge, RdmaTransport, RdmaWorkCompletion, RdmaWorkRequest,
+};
+pub use types::*;