AI Briefing
KO

ExLlamaV3 v1.0.0 Officially Released

·2026.07.15 16:17

Key point

ExLlamaV3 v1.0.0 has been released, offering improved inference speed and support for a variety of models.

Details

After more than a year of development, the official ExLlamaV3 v1.0.0 version has been released. This update focuses on inference performance optimization and dependency removal.

Key updates:

  • Dependency removal: Removed dependencies on Flash-attention-2 and xformers, simplifying installation and runtime environments.
  • Performance optimization: A new Attention kernel resolves the speed degradation issue that occurred during KV quantization, and actually improves inference speed. Additionally, GEMM/GEMV performance has been significantly improved on Ampere architecture.
  • Expanded model support: Extended Tensor-parallel support to most models, including Gemma 4, and added an MoE kernel scheduler and a new Conv1d kernel.
  • Other improvements: Added a new INT8 GEMV kernel, support for GptOssForCausalLM and NemotronHForCausalLM, along with various bug fixes and quality-of-life (QoL) improvements.
optimizationopen-sourcemodel-servingllm-inferenceexllama
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditTextCLF Releases Calibration-Free TQ Quantization Method and Open-Source Quant FactoryRedditllama.cpp PR Proposes 3-7x Faster CPU Prompt Processing via Tiled mul_mat for k-quantsPyTorchKRvLLM Introduces Hardware-Agnostic Layer Architecture to Support Diverse AcceleratorsPyTorchKRvLLM rope_scaling Override Causes 36% Model Configuration MismatchHacker NewsJevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs DetectionGeekNewsRootless Linux Sandbox 'Drop' Released for Isolating AI Agents and Third-Party Programs