https://singularity.sg/Short notes on technology, learning, and things worth remembering.Hugoen-SGhttps://singularity.sg/2026/08/compiling-pytorch-cuda-extensions-in-docker-without-a-gpu/Mon, 17 Aug 2026 14:25:36 +0800https://singularity.sg/2026/08/compiling-pytorch-cuda-extensions-in-docker-without-a-gpu/One detail that is easy to forget when building CUDA containers is that the GPU itself is normally not available during docker build. The compiler and CUDA toolkit can be present, but PyTorch cannot inspect the target card and choose an architecture automatically. For an RTX 3090 build, the important setting is: ENV TORCH_CUDA_ARCH_LIST="8.6" Compute capability 8.6 is the correct target for the RTX 3090. This variable is specifically used by PyTorch CUDA extension builds. A plain nvcc or CMake CUDA project needs its equivalent compiler or CMAKE_CUDA_ARCHITECTURES setting instead.https://singularity.sg/2026/08/deepseek-released-the-agent-harness-too/Mon, 17 Aug 2026 09:33:52 +0800https://singularity.sg/2026/08/deepseek-released-the-agent-harness-too/DeepSeek has released DeepSeek Harness, or dsh, as an open-source agent harness under the MIT license. Model releases usually give us weights or an API and leave the actual agent runtime to tools such as OpenCode. This release includes that missing layer. What is different The main idea is that everything is a plugin. The model adapter, tool registry, session log, agent loop, filesystem, sandbox, approval policy, persistence, and telemetry can all be replaced or patched through configuration. Profiles and bundles compose those plugins into a working runtime, with both a Web UI and a headless runner.https://singularity.sg/2026/08/ubuntu-26.04-nvidia-cuda-13.3-and-rootless-docker/Sun, 16 Aug 2026 15:17:17 +0800https://singularity.sg/2026/08/ubuntu-26.04-nvidia-cuda-13.3-and-rootless-docker/GPU support: Tested on an RTX 3090. Theoretically compatible cards are listed at the end of this guide. Validated on 16 August 2026 with: Ubuntu 26.04 LTS (resolute), x86-64 Linux kernel 7.0.0-29-generic NVIDIA GeForce RTX 3090 UEFI Secure Boot enabled Canonical-signed NVIDIA 595 server-open kernel driver CUDA Toolkit 13.3 Rootless Docker Engine NVIDIA Container Toolkit 1.20.0https://singularity.sg/2026/08/warmachine-local-kali-linux-assistant/Sat, 15 Aug 2026 12:20:00 +0800https://singularity.sg/2026/08/warmachine-local-kali-linux-assistant/WarMachine is my work-in-progress model for running a Kali Linux assistant locally. The current 4B version operates more like an OpenCode-compatible security assistant than a fully autonomous pentesting agent. I run it on a local GPU inside a sandbox for Hack The Box and other authorized pentesting work. It helps construct commands, runs them through structured tool calls, and uses the output to suggest what to try next. I find it most useful for code and command completion, especially when I know the objective but do not want to stop and reconstruct the exact syntax for each tool.https://singularity.sg/2026/08/my-llama-cpp-command-for-qwen3-coder-next/Sat, 15 Aug 2026 12:10:00 +0800https://singularity.sg/2026/08/my-llama-cpp-command-for-qwen3-coder-next/I use this command to run Qwen3-Coder-Next locally with llama-server. I am keeping the complete command and an explanation of each option here so I do not have to reconstruct it later. This is the best configuration I have found so far for an RTX 4090 with 24 GB of VRAM and 64 GB of system RAM. My priority is a longer context window rather than maximum inference speed. There is still room to optimize it further, but this setup works well enough for regular use.https://singularity.sg/2026/08/running-qwen-image-edit-2511-4-bit-on-amazon-eks/Sat, 15 Aug 2026 12:00:00 +0800https://singularity.sg/2026/08/running-qwen-image-edit-2511-4-bit-on-amazon-eks/I published Qwen Image Edit 2511 4-bit as a selective NF4 quantization rather than quantizing every transformer layer indiscriminately. Some layers remain at higher precision to preserve output quality. The resulting Diffusers-compatible model can operate below 20 GB of VRAM, including on 16 GB GPUs with CPU offload. Gary Stafford used the model in his Amazon EKS deployment. His implementation packages the 17 GB model for an NVIDIA L40S, distributes it from S3 to node-local EBS, loads it into a FastAPI inference service, and reports model and GPU readiness through health checks. The accompanying deployment repository includes the containers, Kubernetes manifests, caching DaemonSet, API, and end-to-end inference tests.https://singularity.sg/2026/08/the-ssd-squeeze-tools-to-beat-nvme-shortages-and-price-hikes/Sat, 15 Aug 2026 09:57:05 +0800https://singularity.sg/2026/08/the-ssd-squeeze-tools-to-beat-nvme-shortages-and-price-hikes/Micron reported in its June 2026 SEC filing that memory and storage demand was outpacing industry supply and that its NAND average selling prices had risen by the mid-80% range quarter over quarter. Before buying another drive, it is worth seeing how much space you can recover from the SSD you already own. Use WinDirStat or Czkawka to identify and remove unnecessary files, freeing up space on your SSD. Find large and unnecessary files Disk cleanup tools like WinDirStat and Czkawka can help you visualize disk usage and identify large or unnecessary files that can be deleted to free up space. This is especially useful for NVMe SSDs, which can be expensive and in short supply. By regularly cleaning up your disk, you can maintain optimal performance and avoid running out of storage. I use them both regularly when I see my SSD filling up, and they have helped me reclaim a significant amount of space.