W4A8
PulseAugur coverage of W4A8 — every cluster mentioning W4A8 across labs, papers, and developer communities, ranked by signal.
1 day(s) with sentiment data
-
New QMapAnything framework optimizes 3D Transformers via structure-aware quantization
Researchers have developed QMapAnything, a novel Post-Training Quantization (PTQ) framework designed to optimize feed-forward 3D models that utilize Transformers. This new approach addresses the challenge of heterogeneo…
-
llama.cpp b12122 release optimizes CUDA integration for MMQ operations
The llama.cpp project has released version b12122, which includes several updates focused on CUDA integration. These updates involve fusing NVFP4 gate/up matmuls and GLU operations within the MMQ (Mixed-Quantized Matrix…
-
ComfyUI Video Upscaler Adds W4A8 Quantization Support
A new version of the ComfyUI-SeedVR2-VideoUpscaler-with-TensorRT tool, v1.6.6, has been released, introducing support for W4A8 quantization. This update aims to improve image quality over NVFP4 while maintaining similar…
-
Qwen3.8 27B model hits 280 tok/s with new MXFP4 optimization
A developer has achieved significant performance gains with the Qwen3.8 27B model by implementing MXFP4 kernels on dual R9700 GPUs. This optimization, which utilizes W4A8 quantization, has reportedly surpassed FP8 perfo…
-
New research explores quantization techniques for efficient AI model deployment
Two new research papers explore methods for optimizing large language models (LLMs) and edge vision models for deployment on resource-constrained hardware. The first paper, a survey on Quantization-Aware Training (QAT),…
-
Quantization trade-offs studied for machine translation models
Researchers have investigated the impact of quantization techniques on the inference efficiency and translation quality of machine translation models. Their study focused on two model families, EuroLLM and Hy-MT2, acros…
-
INT4 Weight-Only Quantization: Decode Speedup, Prefill Stagnation Explained
Weight-only INT4 quantization, while effective for reducing memory traffic and speeding up the decoding phase of LLM inference, does not improve the prefill phase. This is because prefill is compute-bound, meaning it is…