A long-standing goal in both industry and academia is to develop an LLM inference platform that is portable across hardware architectures, eliminates the need for low-level hand-tuning, and still delivers best-in-class efficiency. In this work, we demonstrate that portable, efficient cross-platform LLM inference is indeed possible and share our experience. We develop a state-of-the-art paged attention kernel, the core performance-critical component of many LLM deployments, that builds exclusively on the domain-specific just-in-time compiled language Triton to achieve state-of-the-art performance on both NVIDIA and AMD GPUs. We describe our high-level approach, the key algorithmic and system-level improvements, the parameter auto-tuning required to unlock efficiency, and the integrations into a popular inference server that are necessary to bring the performance of a generic Triton attention kernel from 19.7% of the state-of-the-art to 105.9%. Our results highlight how open-source domain-specific languages can be leveraged to unlock model portability across different GPU vendors.

Multi-GPU programming traditionally requires developers to navigate complex trade-offs between performance and programmability. High-performance implementations typically rely on low-level HIP/CUDA communication libraries that demand substantial engineering effort for even basic overlap patterns, while simpler abstractions often sacrifice performance. We present Iris, a multi-GPU communication library implemented entirely in Python and Triton that eliminates this trade-off. Iris provides tile-based symmetric memory abstractions that naturally align with Triton's programming model, enabling developers to write single-source kernels that seamlessly interleave computation and communication. We demonstrate a taxonomy of compute-communication overlap patterns--from bulk-synchronous to fine-grained workgroup specialization--that can be implemented with minimal code changes in Iris, often requiring just a few additional lines within the same Triton kernel. Our evaluation shows that Iris achieves near-optimal bandwidth utilization in microbenchmarks and delivers up to 1.79x speedup over PyTorch and RCCL for GEMM+All-Scatter workloads, demonstrating that high-level implementations can match or exceed heavily-optimized libraries while dramatically simplifying multi-GPU programming.

As large language models (LLMs) continue to scale, their workloads increasingly rely on distributed execution across multiple GPUs. However, the conventional bulk synchronous parallel~(BSP) model used in such settings introduces significant performance inefficiencies. To characterize these bottlenecks, we introduce the ''Three Taxes'' (Bulk Synchronous, Inter-Kernel Data Locality, and Kernel Launch Overhead) as an analytical framework. We propose moving beyond the rigid BSP model to address key inefficiencies in distributed GPU execution. By exploiting libraries like Iris for Triton, we gain access to in-kernel communication primitives that enable the design of novel fine-grained programming patterns, offering greater flexibility and performance than traditional BSP-based approaches. These patterns systematically eliminate the three taxes by creating direct, tile-level producer-consumer pipelines and replacing global barriers with fine-grained dataflow synchronization. Applying this methodology to critical kernels, from the foundational All-Gather + general matrix multiplication operation to the complex Flash Decode algorithm, we observe a 10-20% speedup in end-to-end latency over BSP-based approaches, establishing a more programmable and efficient paradigm for distributed LLM workloads.

In the era of LLMs, dense operations such as GEMM and MHA are critical components. These operations are well-suited for parallel execution using a tilebased approach. While traditional GPU programming often relies on low level interfaces like CUDA or SYCL, Triton has emerged as a DSL that offers a more user-friendly and portable alternative by programming at a higher level. The current Triton starts at the workgroup (aka threadblock) level, and directly lowers to per-thread level. And then attempt to coalesce and amend through a series of passes, promoting information from low-level representation. We believe this is pre-mature lowering based on the below observations. 1. GPU has a hierarchical structure both physically and logically. Modern GPUs often feature SIMD units capable of directly operating on tiles on a warp or warpgroup basis, such as blocked load and blocked MMA. 2. Multi-level gradual lowering can make compiler decoupled and clean by separating considerations inter and intra a logical layer. 3. Kernel developers often need fine control to get good performance on the latest hardware. FlashAttention2 advocates explicit data partition between warps to make a performance boost. In this context, we propose ML-Triton which features multi-level compilation flow and programming interface. Our approach begins at the workgroup level and progressively lowers to the warp and intrinsic level, implementing a multilevel lowering align with the hierarchical nature of GPU. Additionally, we extend triton language to support user-set compiler hint and warp level programming, enabling researchers to get good out-of-the box performance without awaiting compiler updates. Experimental results demonstrate that our approach achieves performance above 95% of expert-written kernels on Intel GPU, as measured by the geometric mean.

Large language models (LLMs) are remarked by their substantial computational requirements. To mitigate the cost, researchers develop specialized CUDA kernels, which often fuse several tensor operations to maximize the utilization of GPUs as much as possible. However, those specialized kernels may still leave performance on the table as CUDA assembly experts show that manual optimization of GPU SASS schedules can lead to better performance, and trial-and-error is largely employed to manually find the best GPU SASS schedules. In this work, we employ an automatic approach to optimize GPU SASS schedules, which thus can be integrated into existing compiler frameworks. The key to automatic optimization is training an RL agent to mimic how human experts perform manual scheduling. To this end, we formulate an assembly game, where RL agents can play to find the best GPU SASS schedules. The assembly game starts from a \textit{-O3} optimized SASS schedule, and the RL agents can iteratively apply actions to mutate the current schedules. Positive rewards are generated if the mutated schedules get higher throughput by executing on GPUs. Experiments show that CuAsmRL can further improve the performance of existing specialized CUDA kernels transparently by up to $26\%$, and on average $9\%$. Moreover, it is used as a tool to reveal potential optimization moves learned automatically.

The Australian government has said it will buy a new drone and upgrade existing patrol and response aircraft in a 1.5 billion Australian dollar ($964.88m) The military will buy a fourth MQ-4C Triton drone and upgrade the air force's existing fleet of 14 P-8A Poseidon maritime patrol aircraft, Pat Conroy, minister for defence industry, said in a statement on Tuesday. The Triton will be delivered in 2024 and be based in northern Australia. The aircraft upgrades will provide enhancements to anti-submarine warfare, maritime strike and intelligence collection capabilities, the statement said. The first Poseidon will enter the upgrade programme in 2026, with the final aircraft to be completed in 2030.

Unpaired image translation algorithms can be used for sim2real tasks, but many fail to generate temporally consistent results. We present a new approach that combines differentiable rendering with image translation to achieve temporal consistency over indefinite timescales, using surface consistency losses and \emph{neural neural textures}. We call this algorithm TRITON (Texture Recovering Image Translation Network): an unsupervised, end-to-end, stateless sim2real algorithm that leverages the underlying 3D geometry of input scenes by generating realistic-looking learnable neural textures. By settling on a particular texture for the objects in a scene, we ensure consistency between frames statelessly. Unlike previous algorithms, TRITON is not limited to camera movements -- it can handle the movement of objects as well, making it useful for downstream tasks such as robotic manipulation.

This is the development repository of Triton, a language and compiler for writing highly efficient custom Deep-Learning primitives. The aim of Triton is to provide an open-source environment to write fast code at higher productivity than CUDA, but also with higher flexibility than other existing DSLs. The foundations of this project are described in the following MAPL2019 publication: Triton: An Intermediate Language and Compiler for Tiled Neural Network Computations. Please consider citing this work if you use Triton! The official documentation contains installation instructions and tutorials.

Working as an aerospace engineer in Malaysia, Chee How Lim dreamed of building a startup that could really take off. Today his company, Tapway, is riding a wave of computer vision and AI adoption in Southeast Asia. A call for help in 2019 with video analytics led to the Kuala Lumpur-based company's biggest project to date. Malaysia's largest operator of toll highways, PLUS, wanted to reduce congestion for its more than 1.5 million daily travelers. A national plan called for enabling car, taxi, bus and truck traffic to flow freely across multiple lanes -- but that posed several big challenges.

After meeting at an entrepreneur matchmaking event, Ulrik Hansen and Eric Landau teamed up to parlay their experience in financial trading systems into a platform for faster data labeling. In 2020, the pair of finance industry veterans founded Encord to adapt micromodels typical in finance to automated data annotation. Micromodels are neural networks that require less time to deploy because they're trained on less data and used for specific tasks. Encord's NVIDIA GPU-driven service promises to automate as much as 99 percent of businesses' manual data labeling with its micromodels. "Instead of building one big model that does everything, we're just combining a lot of smaller models together, and that's very similar to how a lot of these trading systems work," said Landau.