GitHub - Luce-Org/lucebox-hub: Lucebox optimization hub: hand-tuned LLM inference, built for specific consumer hardware.
Lucebox optimization hub: hand-tuned LLM inference, built for specific consumer hardware. - Luce-Org/lucebox-hub
Curated Resource List
Summary
A curated, structured resource list for learning GPU performance engineering, progressing from CUDA fundamentals through kernel optimization, inference engines, and distributed serving. Sources are limited to original papers, official specifications, and direct implementation work.
Highlights
- Organized from a single inference request to single GPU, optimized kernels, inference engines, and distributed systems
- Strict source policy: only original papers, official specifications, creator repositories, and direct implementation work
- Covers hardware-specific content for NVIDIA, AMD, Google TPU, and AWS Trainium
- Includes a 'Start here' section providing a minimum mental model for newcomers to the field
- Spans from fundamentals like CUDA programming model and roofline to production topics like MoE serving and disaggregation
auto-generated
Context
Audience
ML engineers, systems engineers, and researchers working on GPU-accelerated inference, kernel optimization, and production serving
DomainGPU Performance Engineering
Formatcurated resource list
Accessfree online (GitHub)
Source Overview
A curated resource list for GPU performance engineering, organized from CUDA fundamentals through kernel optimization, inference engines, and distributed serving.
Sources are limited to original papers, official specifications, and direct implementation work. Covers hardware-specific content for NVIDIA, AMD, TPU, and Trainium.
Topics
Related
CUDA Programming GuideTritonCUTLASSvLLMTensorRT-LLMRoofline Model
