vllm

An 18 notebook course that isolates and measures each component of agentic loop engineering on real, industry standard software datasets.

ElasticMM: Elastic and Efficient MLLM Serving System

Self-hosted, OpenAI-compatible inference for the agentic era: reasoning LLMs, universal tool calling, and the Responses API alongside embeddings, speech, and image models — many models sharing your GPUs, one gateway. Powered by Ray Serve.

Local-first CLI for benchmarking LLMs on real hardware — quality, speed, reliability, and a real multi-turn agent loop.

Legacy: self-hosted LLM/RAG stack installer for DGX Spark (arm64). The work continued as AGmind Systems Lab — measured local-LLM performance evidence (Strix Halo, DGX Spark, llama.cpp, vLLM): https://agmind.ai

paper list on Video Moment Retrieval (VMR), or Temporal Video Grounding (TVG), Video Grounding (VG), or Temporal Sentence Grounding in Videos (TSGV)

useful Gentoo overlay Curated ebuilds, AI, tools & science

Serving Qwen3.8-27B-FP8 on a single DGX Spark (GB10): 7.88 to 58.5 tok/s single-stream from decode strategy alone, weights untouched. Speculative decoding and prefix caching benchmarked, plus DFlash 2 — the only Qwen3.8-27B build that can serve it under vLLM.

Native Windows vLLM 0.27.1 wheels: Python 3.13, PyTorch 2.13 + CUDA 13.0, SM 7.5-12.0 for RTX 20/30/40/50, OpenAI-compatible serving, FlashAttention/Rust, 10 KV formats, Multi-TurboQuant, and experimental CPU/RAM/NVMe prompt-KV offload - no WSL or Docker.

Serving 4-bit Qwen3.8-27B on a single DGX Spark (GB10): 75 tok/s single-stream, 246 tok/s aggregate at 8-way concurrency. NVFP4 vs MixedInt4-AutoRound vs the FP8 baseline, measured on one harness — including why the quantization advantage collapses to +0.2% by c16.

Benchmarking Open-Ended Inference Optimization by AI Agents

GLM-5.2 (744B/40B MoE) on a 4× DGX Spark / GB10 (sm_121) cluster: portable Triton sparse-MLA kernels, a data-free expert prune, MTP draft, and a one-script bootstrap.

Open-source, self-hostable social listening.

Synthetic medical VQA pipeline: 119K images annotated by frontier VLMs, cross-validated at 93% agreement, fine-tuned on 3 model families (2-3B params)

Carbon Limiting Auto Tuning for Kubernetes

Code for building self-expanding knowledge graphs with Outlines, vLLM, neo4j, and Modal.

Lightweight proxy for LLM

From-scratch C++23/CUDA inference engine for the NVIDIA RTX 5090 (sm_120a). The best single-GPU backend for agentic AI: tool calling, long-context loops, reasoning and concurrent sub-agents. Decode beats llama.cpp b9976 by 42-48% on dense GGUF (measured 2026-07-12), at-or-ahead of vLLM on NVFP4. 100% written by Claude Code.

Plug-and-play, dependency-free web command center for DGX Spark fleets (and any GPU boxes): live per-node GPU/CPU/mem/temps/power, optional RouterOS switch monitoring, and per-model decode/prefill/TTFT/KV cards. Point it at your own hardware.

GPU 性能与 AI Infra 学习项目：CUDA/Triton 算子、NCU/NSYS、vLLM/SGLang/TRT-LLM/ms-swift、PyTorch/DeepSpeed/ms-swift 训练、并行架构