AI Briefing#fp8
The latest AI and developer news about #fp8, with the original source and a short summary.
The latest AI and developer news about #fp8, with the original source and a short summary.
The release removes deprecated APIs and adds FP8 support, YaRN truncation controls, and performance improvements across CUDA, Metal, Vulkan, and other backends.
The 321B parameter model based on GLM-5.3-Flash defaults to max reasoning, requiring token usage checks before comparison.
StepFun released the weights for its 600B MoE model Step 5 Preview and confirmed it can run on 8 H200s.
PyTorch 2.15 improves Mac GPU MoE decoding by 5-6.5x and supports FP8.
The update introduces native FP8 support and optimizes Metal kernels, reducing driver memory peaks from 270.135 GiB to 8.108 GiB for video decoding.
The open-source project replicates the 71-block Swin/ViT architecture of DLSS-NR build 310.8.0 using FP8 tensor cores, achieving byte-for-byte intermediate matches with the original.
In PCIe-based VLA training, FP8 transmission reduced bytes by 48.4%, but step time increased because post-merger serialization placed it on the critical path.
Poolside updated the FP8 and NVFP4 checkpoints for Laguna S 2.1.
A look at efficient training methodologies and the latest technology trends for foundation models, which serve as the basis of LLMs.
DeepLearning.AI has released a course on optimizing efficient LLM inference using vLLM.
This analyzes the FP8 standard mismatch and kernel optimization issues that arise when running DeepSeek-V4-Flash on AMD MI300X.
RedHatAI released the Qwen3.6-27B FP8 model on Hugging Face.
SGLang has been found to have FP8 KV cache corruption and a memory leak in image requests.
Qwen 3.6 27B was remodeled with Opus CoT and Hermes tool-calling SFT and released in FP8.
In an apples-to-apples FP8 comparison, the advantage between Ironwood and B200 reversed depending on the model.
Meta has open-sourced RCCLX, an AMD extension of RCCL, boosting performance with DDA and low-precision collectives.
Furiosa SDK 2026.1 strengthens batching, caching, and Kubernetes support.
Hugging Face and AMD collaborated to release custom kernels specifically for MI300X that maximize the performance of Llama 3.1 405B.
This introduces a method for efficiently fine-tuning the FLUX.1-dev model with under 10GB of VRAM using the diffusers library and QLoRA.
Stability AI partnered with NVIDIA to release TensorRT-optimized Stable Diffusion 3.5 models, boosting performance by up to 2.3x.
Hugging Face's distributed training framework Accelerate has announced the release of version 1.0.0.
A Hugging Face guide walks through deploying Llama 3.1 405B on Google Cloud Vertex AI.
Hugging Face has launched the Optimum-NVIDIA library, which speeds up LLM inference by up to 28x on NVIDIA platforms.