wtdcode/vllm-backport
Dogfooding vLLM backport for older GPUs.
01 / MACHINE INTELLIGENCE
Models you can run. Agents you can inspect. Tools you can build on.
← All topicsDogfooding vLLM backport for older GPUs.
Open-source (Apache-2.0) PDF takeoff for construction & flooring — the first engine an AI agent drives natively over MCP, not bolted on. One-click room detection, materials + quantities, built for preconstruction. Runs entirely in your browser.
A Command-Line Interface (CLI) and Skill to interact with Papers with Code. For agents and humans.
GLM-5.3 Flash EXL3 for 2x DGX Sparks.
A Simple, Lightweight, and Extensible Serving Framework for X-AnyLabeling
Stop That Shit(别再造史了)|面向 Codex/GPT 场景的多平台 Hook + Skill Guard:拦截 AI coding agent 无需求的哈希、校验和与任务范围膨胀。 A multi-platform Hook + Skill Guard for AI coding agents in Codex/GPT workflows: stop unrequested hashes, checksums, and task-scope creep.
Modular cross-runtime WebGPU library for shaders, 3D scenes, GPU tensors, neural networks, and math viz
Run any model in Grok Bot — one-command setup, model picker UI, evidence-based provider wire maps, and an update-proof doctor. Not farming you, arming you.
Claude Opus Pro.
ZKE(Z Kubernetes Engine):AI 原生的 Kubernetes 云操作环境,桌面式多集群控制台加受控 AIOps Agent,基于 Server + Agent 与 QUIC/mTLS,适用于私有云、混合云及边缘环境
A small desktop robot that brings an AI coding assistant to life, physically acting out when it reads files, thinks, writes code, and finishes work.
AI Stupid Meter - API Server.
AI Stupid Meter - Web Frontend.
A local bridge for bidirectional collaboration between Claude Code and Codex. 连接 Claude Code 与 Codex 的本地实时协作桥接工具。
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.
Open the project to explore the idea and the code.
Training models inside world models.
Open the project to explore the idea and the code.
~2M live jobs from 25 ATS, enriched with LLM-extracted fields + embeddings, released CC0. Includes agent tooling to search them (hull -> learn -> rank).
OpenFlow — a visual builder for multi-agent AI workflows, built on the opencode engine
Low-level agentic design patterns. Turning old engineering wisdom into reflexes your agent reaches for on its own—on any agent.
Gemma 4 E2B inference in pure C.
TTPO: Test-Time Policy Optimization.
Production-grade recipe for DeepSeek-V4-Flash-0731 (284B MoE) on 2x NVIDIA DGX Spark: self-healing 2-node vLLM cluster, reboot-verified, tuned DSpark speculative decoding (~75 tok/s), full benchmarks, OpenAI Codex CLI integration