This repository is a research prototype for efficient LLM inference, specifically optimizing attention memory management via the vAttention technique (ASPLOS 2025). It contains three primary components: (1) a modified NVIDIA GPU kernel driver (nvidia-vattn-uvm-driver) that enables sub-2MB page sizes for Unified Virtual Memory (UVM), (2) a fork of the Sarathi-Serve LLM serving framework (sarathi-lean) integrated with vAttention, and (3) custom attention kernel implementations (pod_attn, vattention) using CUTLASS and FlashAttention backends. The system supports multiple attention backends (FlashAttention, FlashInfer, paged, variable-length) and includes extensive benchmarking and plotting scripts for reproducibility of research results. Architecturally, the user-space inference stack uses a scheduler (Orca or Sarathi), a cache engine (vLLM-style or vAttention), and model executors with tensor/pipeline parallelism. The vAttention cache engine replaces traditional paging with CUDA virtual memory allocation, decoupling virtual and physical memory for KV-cache. The custom NVIDIA UVM driver patch is required to support fine-grained page sizes. The repository also includes microbenchmarks for attention kernel performance and page size effects, as well as artifact scripts to reproduce figures from the associated paper. While the codebase demonstrates solid engineering for a research project (clean abstractions, registry pattern, modular design), it is explicitly not production-ready. Key limitations include strict version dependencies (PyTorch 2.3.0, CUDA 12.1), reliance on a modified kernel driver, and incomplete feature parity with production systems like vLLM. The inclusion of a fully vendored CUTLASS library and auto-generated NVIDIA driver headers adds significant size but provides reusable GPU primitives. The research scripts are functional but contain hardcoded paths and are tightly coupled to the experimental setup.
ExLlamaV2 is a high-performance inference library for running quantized large language models (LLMs) on consumer GPUs, supporting GPTQ and the custom EXL2 quantization formats. It provides a Python library with custom CUDA/C++ extensions for efficient attention, MLP, sampling, and KV cache management, along with advanced features like dynamic batching, paged attention, speculative decoding, LoRA, tensor parallelism, and multimodal (image/video) support. The library includes a WebSocket server for remote inference and a suite of example scripts demonstrating various use cases. While the codebase exhibits strong engineering with clean abstractions (e.g., module system, filter pattern, multi-backend attention) and highly optimized CUDA kernels, the repository has been archived in favor of ExLlamaV3, meaning no further updates or bug fixes are expected. This makes it a valuable reference for understanding quantized LLM inference internals but risky for new projects requiring ongoing support.