* @class nixlBlobDesc * @brief A descriptor class, with additional metadata in form of a blob * bundled with a nixlBasicDesc. */

* @class nixlBasicDesc * @brief A basic descriptor class, single contiguous memory/storage * element, alongside supporting methods */

A wrapper class to manage CUDA events, also for better overlapping convenience. Attributes: event: the CUDA event captured.

Configuration for a language model.

RAII wrapper bundling an fd + ownership flag + (optional) path string. Used by FILE_SEG backends so per-MD bookkeeping (owned/path) lives in one place

Comprehensive configuration for model deployment and benchmarking. Combines strategy, runtime, and system configurations into a single objec

Can be sharded

* @struct nixlRemoteDesc * @brief A descriptor structure for remote buffers, with remote agent name bundled with a * nixlBasicDesc. */

Represents a communication pattern between distributed processes. Attributes: matrix: Communication matrix as numpy array mem_typ

NIXL Benchmarking utility for KV cache performance testing. This class provides a configurable interface for running benchmarks on NIXL

Configuration for the GPU cluster.

ETCD-based MPI utilities - NOT PERFORMANCE OPTIMIZED (for control path only)

* @struct nixlAgentConfig * @brief Per Agent configuration information, such as if progress thread should be used. * Other configs such as as

The core expert-parallel (EP) communication buffers for Mixture of Experts (MoE) model, which supports dispatch and combine operations using NVLi

Implementation of the DeepSeek-R1 model architecture. This class represents the DeepSeek-R1 model and provides methods to access its par

Implementation of the GPT-OSS model architecture. This class represents the GPT-OSS model with Grouped Query Attention (GQA), providing

Implementation of the Llama 3.1 model architecture. This class represents the Llama 3.1 model and provides methods to access its paramet

Implementation of the Qwen3 model architecture. This class represents the Qwen3-235B-A22B model with Mixture-of-Experts (MoE) and Groupe

Extends CTPerftest to handle multiple traffic patterns sequentially. The patterns are executed in sequence, and the results are aggregated. A

Python wrapper for the C++ SharedRingBuffer class

Configuration for an inference task.

* @class nixlAgent * @brief nixlAgent forms the main transfer object class */

* @brief Section descriptor for nixl * * This class is used to store a section descriptor for nixl. * It is derived from nixlMetaDesc and contains

* @enum nixl_status_t * @brief An enumeration of status values and error codes for NIXL */

Generic IOV descriptor class independent of NIXL

Stats class for measuring benchmark metrics

Trait for types that can synchronize their state to a backend

Abstract base class defining the interface for model architectures. All model architectures should inherit from this class and implement

* @brief Binary notification with variable-length encoding and fragmentation support * * The notification payload consists of agent_name + message,

* @brief Notification header for all fragments (10 bytes) * * This is present in every fragment and contains only the essential * fields needed for

* @brief Metadata for fragment 0 only (10 bytes) * * This contains metadata that is constant across all fragments, * so we only send it once in the

Buffer chunk structure for control request pool */

Python equivalent of BufferHeader struct from cyclic_buffer.h

Connection state tracking for multi-rail connections */

Control request pool with pre-allocated buffers for SEND/RECV operations */

Lightweight data request pool for WRITE/READ operations */

Trait for testing common descriptor list operations

* @brief Process-wide shared DOCA context * * DOCA only supports one metrics context per process, so all agents share * this context. The underlyin

RAII guard for tests that mutate process-global environment variables. On construction it acquires a shared lock (serializing all `EnvGuard` users) a

* @class GMockBackendEngine * @brief A GMock implementation of nixlBackendEngine for GTest testing purposes. * * This class provides a Google Mo

Shared GUSLI device config used by utils and nixl_worker