2805 Bowers Ave, Santa Clara, CA 95051 | 408-730-2275
research@colfax-intl.com

Research

  • FlashAttention-3 for Inference: INT8 Quantization and Query Head Packing for MQA/GQA (External)

    FlashAttention-3 for Inference: INT8 Quantization and Query Head Packing for MQA/GQA (External)

    In this blog post presented on the Character.AI research blog, we explain two techniques that are important for using FlashAttention-3 for inference: in-kernel pre-processing of tensors via warp specialization and query head packing for MQA/GQA. Go to article…

  • GPU passthrough on Proxmox VE 8.2

    GPU passthrough on Proxmox VE 8.2

    In this guide, we will walk through the steps to enable GPU passthrough and by extension PCIe passthrough on a virtual machine (VM) deployed through Proxmox. PCIe passthrough provides a path for VMs to directly access underlying PCIe hardware, in the case of this article, an Nvidia® A30 GPU. This setup is ideal for scenarios […] Go to article…

  • FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision

    FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision

    In this blogpost, we describe three main techniques that we use to speed up attention on Hopper GPUs in FlashAttention-3: exploiting asynchrony of the Tensor Cores and TMA to (1) overlap overall computation and data movement via warp-specialization and (2) interleave block-wise matmul and softmax operations, and (3) incoherent processing that leverages hardware support for FP8 low-precision. Go to article…

  • Sharing NVIDIA® GPUs at the System Level: Time-Sliced and MIG-Backed vGPUs

    Sharing NVIDIA® GPUs at the System Level: Time-Sliced and MIG-Backed vGPUs

    While some modern applications for GPUs aim to consume all GPU resources and even scale to multiple GPUs (deep learning training, for instance), other applications require only a fraction of GPU resources (like some deep learning inferencing) or don’t use GPUs all the time (for example, a developer working on an NVIDIA CUDA® application may […] Go to article…

  • Delivering 1 PFLOP/s of Performance with FP8 FlashAttention-2

    Delivering 1 PFLOP/s of Performance with FP8 FlashAttention-2

    We recently released an update to our FlashAttention-2 forward pass implementation on NVIDIA Hopper™ architecture that incorporates a number of new optimizations and improvements, including a software pipelining scheme and FP8 support. In this article, we will explain a challenge with achieving layout conformance of register fragments for WGMMA instructions that we encountered in the […] Go to article…

  • A note on the algebra of CuTe Layouts

    A note on the algebra of CuTe Layouts

    The core abstraction of NVIDIA’s CUTLASS library for high-performance linear algebra is the CuTe Layout. In this technical note, we give a rigorous, mathematical treatment of the algebra of these layouts and certain layout operations. Currently, the main goal is to lay down conditions for when the operations of complementation, composition, and logical division are […] Go to article…

  • A Case Study in CUDA Kernel Fusion: Implementing FlashAttention-2 on NVIDIA Hopper Architecture using the CUTLASS Library

    A Case Study in CUDA Kernel Fusion: Implementing FlashAttention-2 on NVIDIA Hopper Architecture using the CUTLASS Library

    We provide an optimized implementation of the forward pass of FlashAttention-2, a popular memory-aware scaled dot-product attention algorithm, as a custom fused CUDA® kernel targeting NVIDIA Hopper™ architecture and written using the open-source CUTLASS library. In doing so, we explain the challenges and techniques involved in fusing online-softmax with back-to-back GEMM kernels, utilizing the Hopper-specific […] Go to article…

  • Developing CUDA Kernels for GEMM on NVIDIA Hopper Architecture using CUTLASS

    Developing CUDA Kernels for GEMM on NVIDIA Hopper Architecture using CUTLASS

    We explain how to develop NVIDIA CUDA® kernels for optimized general matrix multiplication (GEMM) on NVIDIA Hopper™ architecture using the template collection CUTLASS and its core library CuTe. Our main contribution is to provide an implementation of a GEMM kernel that uses the Tensor Memory Accelerator (TMA) and Warp Group Matrix-Multiply-Accumulate (WGMMA) operations introduced with NVIDIA Hopper™ architecture. Go to article…