Pinned Loading

1. agi-lab-framework agi-lab-framework Public

   From-scratch autonomous ML-research lab on a single 18GB laptop — C17/Metal training engine (no ML frameworks), 31-role LLM agent organization, 4-layer hybrid-RAG memory. ~74K LOC; ran pre-register…

   Python

2. bert bert Public

   Long-context project memory for MCP hosts, measured on Claude Code — a local MCP server with hybrid retrieval (dense + BM25 + cross-encoder rerank). Falsification-first benchmarks, honest nulls.

   Python

3. llm-inference-benchmarking llm-inference-benchmarking Public

   Latency-throughput, CUDA-graph, and FP8-vs-FP16 inference benchmarks for Qwen2.5-7B-Instruct on vLLM 0.23.0 / NVIDIA H100 NVL

   Python 1

4. moe-speed-of-light moe-speed-of-light Public

   Python

5. fla-org/flash-linear-attention fla-org/flash-linear-attention Public

   🚀 Efficient implementations for emerging model architectures

   Python 6k 811

6. NVIDIA/cudnn-frontend NVIDIA/cudnn-frontend Public

   cuDNN Frontend is NVIDIA's modern, open-source entry point to the cuDNN library and a growing collection of high-performance open-source kernels.

   Cuda 999 338