Tag
The paper introduces adaptive representations for functional gradient descent, provably ensuring convergence to global minimizers and outperforming neural networks by orders of magnitude in various settings.
This pull request fuses a chain of operations in the Vulkan backend for the qwen4exp model to optimize inference performance in llama.cpp.
This paper introduces EoupCT, a novel framework that estimates and orthogonalizes unknown pre-training gradients to mitigate catastrophic forgetting in continual fine-tuning of large language models.
The author proposes optimizing llama.cpp for specific models by stripping unnecessary code and using AI to automate the process, aiming for 2x+ performance improvements.
FurE is an efficient strand-based method for 3D animal fur reconstruction from multi-view images, using a PCA decoder learned from human-hair data to overcome dataset scarcity and achieve 10x speedup over current methods.
AI agents, primarily using Opus 5.5, have significantly accelerated a 27B model's inference speed from 66 to 580 tokens per second on a Mac by rewriting its inference engine in just three days.
The article demonstrates how to solve a physical corn puzzle using Google's OR-Tools CP-SAT constraint satisfaction solver, highlighting its efficiency over naive backtracking methods.
MoEspresso 3, an inference engine, achieves 12-15 decode tokens per second for the Qwen3.8-Flash-Next model on a 2021 M1 Max with 32GB RAM, using techniques like Cache-Prior to optimize performance and outperforming some hosted solutions in benchmarks.
The tweet discusses when and how to finetune a specialized classifier model that is better than Opus, faster than Jev, and runs on user devices.
The article explains the implementation of a software occlusion culling technique in a block-based game to improve rendering performance on weak hardware, achieving significant chunk culling and frame rate gains.
KernelZero introduces a co-evolution framework with Proposer and Coder models to enhance GPU kernel generation, achieving superior performance on CUDA and Triton benchmarks.
A user built an open-source inference engine called Strata to optimize running Qwen3.8-Flash-Next on modest hardware, achieving up to 65.1 tok/s from an initial 15 tok/s.
The author introduces a learning track on large model inference optimization, covering topics such as KV Cache, Continuous Batching, PagedAttention, and a comparison between vLLM and SGLang, highlighting it as cutting-edge in AI deployment.
The article surveys the state of SIMD support in Rust in 2026, discussing advancements in SIMD libraries and implementation details.
This paper presents a reinforcement learning agent trained to generate provably optimal quadrilateral block decompositions for planar domains, demonstrating superior performance over existing tools like Gmsh in terms of usability and optimality.
A team announced they achieved first place on NVIDIA's SOL-ExecBench benchmark across all tracks, beating 583 teams with real production kernels on B200 GPUs.
The article discusses the problem of exploding variance in estimating log-sum-exp functions in machine learning and proposes using least-squares regression to mitigate statistical disadvantages.
ThinkingCap is a finetuned model based on Qwen 3.6 27B that reduces reasoning tokens by about 50% while maintaining performance, leading to significant efficiency gains in inference.
The paper proposes a formal functional architecture for AI-supported electricity trading in European markets, integrating regulatory compliance and operational constraints.
A GitHub repository collabosm provides an optimized setup for running the Qwen3.8-Flash-Next model on a Google Colab A100, achieving inference speeds faster than commercial APIs with detailed performance metrics and instructions.