Kernels

AIAI-generated summary Engineering at Meta built Triton Low-Level Extensions to give engineers explicit control over GPU execution while preserving Triton’s productive programming model. Meta says the extension powers production workloads across Meta from AI training to large-scale LLM pretraining on NVIDIA and AMD hardware.

AIAI-generated summary This paper presents a systematic analysis of deep learning compilers, covering their architectural evolution, compilation workflows, optimization techniques, and emerging research directions. Experiments using PyTorch and TorchInductor on an NVIDIA RTX A6000 demonstrate speedups of up to 8.21x for fusion-friendly computation graphs, while other workloads exhibit limited improvements.

AIAI-generated summary A report evaluating the Locality Domain feature introduced in CUDA 13.4 examines multi-die scaling, data placement, and cross-die communication on NVIDIA Blackwell GPUs. The study demonstrates that die-aware adaptations achieve significant speedups for tools including NCCL 2.32.3, nvbandwidth, and DeepGEMM, while also identifying memory allocation overheads.

AIAI-generated summary DeepSeek V4.1 Flash speeds on GB300 NVL72 increased by up to 1.5x tokens per second per agent over the past month on AA-AgentPerf. The improvements stem from a newer vLLM version, new vLLM kernels, FlashInfer autotuning, and an updated setup.

AIAI-generated summary ROCm Blogs presents a guide detailing how to profile a DeepSeek-V4-Flash training proxy in Primus using madengine and rocprofv3. The workload uses an eight-layer BF16 training proxy configured with eight AMD Instinct MI350X GPUs on a single node.

AIAI-generated summary DeepSeek V4.1 Flash 1017 matches GLM-5.3 intelligence at about one third the cost and four times the speed according to Baseten. Baseten details three optimizations used to make Flash models faster including kernel fusion Programmatic Dependent Launch and host loop minimization.

AIAI-generated summary Octomind states that Flash Attention computes attention results in tiles without keeping the full score matrix in GPU memory. Octomind notes that the method does not reduce model weights or the persistent KV cache.

AIAI-generated summary FlyDSL allows developers to write GPU programs in Python and compile specialized kernels on demand for LLM serving. ATOM integrates FlyDSL mono decode for models including MiniMax-M3, GLM-5.2, Kimi-K3, and DeepSeek-V4.1.

AIAI-generated summary Aditya Kumar released an overview detailing how to build XLA from source and teach it a new pass. The instructions explain how to add a custom pass to rewrite expressions within the compiler pipeline.

AIAI-generated summary NVIDIA releases NVIDIA CUDA 13.5 to deliver general availability for NVIDIA Vera Rubin support across major new platforms. NVIDIA says the update provides innovations across libraries, compilers, NVIDIA CUDA Tile, and developer tools to improve scheduling and performance.

AIAI-generated summary ROCm Blogs has made available GLM-5.2 AgentX on AMD Instinct MI355X to boost throughput and interactivity. ROCm Blogs states that cache efficiency, faster decode kernels, careful AgentX validation, and upstream collaboration raised peak throughput 1.72 times and peak P90 interactivity 2.46 times.

AIAI-generated summary Together AI released a compiler that automates GPU tuning for newly released open source models using AI agents to write faster code. Together AI claims the tool has outperformed hand-tuned megakernels by achieving speeds up to 1.3x faster on Llama 1B.