Anduin Xue
Anduin Xue

Anduin's Tech Blog

All Posts

27
MAY

Setup DGX Spark with Qwen 3.6 27B Dense

Deploying large language models on edge devices often feels like a high-stakes game of memory Tetris, where every megabyte counts and stability hangs by a thread. The NVIDIA DGX Spark, with its unified 128GB memory architecture, presents a unique playground for running dense models like Qwen 3.6 27B in full BF16 precision without quantization. This approach challenges the common assumption that edge deployment requires compromising on quality for speed or size. By leveraging the raw power of the GB10 Grace-Blackwell chip, it becomes possible to host a model that demands nearly 90GB of memory while leaving ample room for system overhead and long-context caching. The key lies not just in the hardware, but in the meticulous orchestration of software layers. Stripping away the graphical interface to eliminate GPU memory leaks is just the beginning; the real magic happens in the configuration of vLLM. Parameters like chunked prefill and prefix caching transform raw bandwidth into responsive...--AI Generated

DGX Spark Qwen3.6 vLLM Blackwell BF16 LLM Deployment
11
APR

Ultimate Fix: Rudder (Yaw) Axis Conflict for X-Plane 12 on Linux

Controlling a high-fidelity simulator like X-Plane 12 on Linux with an Xbox controller often feels like trying to drive a race car with a broken steering wheel, especially when the rudder axis stubbornly locks in place. The core of this frustration lies in a fundamental mismatch: modern triggers report as two independent absolute axes, while flight simulators demand a single, centered axis that returns to zero. Traditional workarounds using legacy drivers frequently fail with new hardware, leaving pilots stranded with a rudder that refuses to center. A more elegant solution involves intercepting these signals at the kernel level to synthesize a virtual joystick that perfectly mimics a professional flight yoke. By mathematically merging the left and right trigger inputs into a single, high-resolution value, the system transforms the chaotic 0-100% trigger data into a precise -1 to +1 command that advanced aircraft like the ZIBO 737 can instantly recognize and trust. This approach does n...--AI Generated

Linux Python X Plane 12 Xbox Series X evdev Anduin
09
MAR

Highly Reliable Data Disaster Recovery Architecture: Asynchronous Snapshot Backup Solution for Heterogeneous File Systems

In high-performance computing environments, the relentless pursuit of raw I/O throughput often forces a compromise on data safety, creating a paradox where the most efficient file systems lack the inherent ability to recover from logical errors. Traditional backup strategies frequently fail to bridge this gap; tools like rsync risk propagating accidental deletions to the backup target if synchronization is too frequent, while infrequent runs leave a dangerous window of vulnerability. Conversely, relying entirely on snapshot-capable file systems like ZFS for primary storage can introduce unacceptable performance bottlenecks due to write amplification and memory overhead, especially when dealing with heterogeneous environments that might include NTFS or other non-snapshot-aware formats. This exploration challenges the conventional wisdom that one must choose between speed and security by proposing a decoupled architecture where the production layer prioritizes performance using Ext4 with...--AI Generated

ext4 ZFS rsync Disaster Recovery Snapshot Data Backup
26
FEB

Rescuing a Stuck Nextcloud Sync: Debugging Flatpak Clients and Server-Side Ghost Files

A stuck Nextcloud sync isn’t always a client problem—it’s often a silent war between the database and the filesystem, where a phantom file haunts the sync process like a digital ghost. When a Flatpak client endlessly retries downloading a file that no longer exists, the real culprit lies buried in the server’s database, where a deleted file from a Group Folder still holds a tenuous, broken link in the sharing system. The error isn’t in the code, but in the logic: a file ID points to a trash path with a corrupted, unreadable name, causing the server to return a 503 error every time the client asks for it. This creates a perfect storm—sync halts, logs fill with fatal errors, and users are left stranded. But the solution isn’t a restart or a reinstall; it’s a forensic dive into the `oc_filecache` and `oc_share` tables, where the ghost is identified by its unique file ID, then surgically erased. The real insight? Nextcloud’s strength—its relational tracking of files, shares, and permission...--AI Generated

Database Nextcloud Docker Flatpak Sync Error File Cache
24
FEB

Anduin的CQE定理：为什么大型组织的采购永远在买电子垃圾

What happens when compliance becomes the only law in corporate procurement? A deep dive into the absurd machinery behind large-scale IT purchases reveals a system that systematically rewards mediocrity, punishes innovation, and turns technology into a theater of bureaucratic theater. At the heart of this dysfunction lies the CQE Theorem: in any large organization, you can have two out of three—compliance, quality, efficiency—but never all three. The result? A world where the best products are buried under layers of fake certifications, rigged bidding games, and endless compliance theater, while the real winners are not the most capable, but the most skilled at playing the game. Imagine a procurement process where the most expensive, least functional system wins—not because it’s good, but because it follows every rule perfectly. Where a 100 million dollar project ends up with a patchwork of incompatible tools, each chosen not for performance but for how easily it fits a checklist. Where...--AI Generated

AI Generated Enterprise IT Bidding Scandal Compliance Trap Product vs Process Technical Integrity
22
FEB

Run Human Fall flat on AnduinOS

What if your dream of playing Human: Fall Flat on Linux turns into a nightmare of 2 FPS and instant crashes—despite owning a top-tier gaming rig? This deep dive reveals why the game’s native Linux version, long abandoned since 2018, is a ticking time bomb of hardware conflicts and memory exploits, silently sabotaging your experience through ancient code that clashes with modern CPUs and security protections. But the real twist isn’t just about avoiding the native build—it’s about uncovering a hidden dependency trap: even Proton, the savior for many Linux gamers, fails silently without 32-bit Vulkan and OpenGL drivers, leaving you stranded in a world of soft-rendered lag and broken compatibility. The solution isn’t a patch or a workaround—it’s a full reset, a clean slate, and a precise chain of steps that rebuild the game’s environment from the ground up. Yet beyond the technical fix lies a deeper question: in an era where software dies silently, how do we preserve the joy of games that...--AI Generated

Linux Gaming Steam Proton Human Fall Flat 32 bit Drivers Proton 9 Game Crashes
21
FEB

告别丑陋的 Tag 后缀：使用 Docker Buildx 构建真正的多架构镜像

Imagine a world where your Docker image seamlessly runs on both x86 and ARM devices—no more messy tag suffixes, no more fragmented builds. With Docker Buildx and QEMU user-mode emulation, this isn’t just possible; it’s effortless. The magic lies in how a standard x86 machine can compile ARM-native code by tricking the kernel into running foreign binaries through real-time instruction translation. But beyond the convenience, there’s a deeper truth: true multi-architecture support isn’t about packing two systems into one image—it’s about creating a smart manifest list that acts like a menu, serving the right architecture based on the host. This means leaner downloads, faster pulls, and no wasted bandwidth. Yet, the system has its boundaries: local Docker daemons can’t natively store these manifest lists, and a single architecture-specific command in your Dockerfile can break the entire build. So how do you write a Dockerfile that’s truly universal? What if your app depends on a binary th...--AI Generated

bash Linux Docker Container Docker Buildx Multi Arch ARM64
10
JAN

给Caddy上Cloudflare以清洗流量

The system employs a modular Caddy configuration strategy to securely route traffic through Cloudflare with mutual TLS authentication and strict proxy trust. Caddy is configured to only accept requests from Cloudflare by validating mTLS certificates using Cloudflare's public CA, while also presenting its own certificate signed by Cloudflare to ensure authenticity. The configuration is built dynamically by combining three components: a Cloudflare IP trust definition that whitelists Cloudflare’s edge network IPs, a baseline configuration setting up logging, HTTP redirects, TLS, and HSTS, and individual service configurations for domains like grafana.anduinos.com and download.anduinos.com, each importing the necessary security and routing rules. These components are stitched together at build time via a script that aggregates all *.conf files from multiple service directories into a single Caddyfile, which is then embedded into the final Docker image. During runtime, the real Cloudflare c...--AI Generated

Reverse Proxy ACME Caddy Docker Cloudflare MTLS
06
JAN

AI 时代的普通人究竟应该学习什么？

In an era where AI can generate code, text, and solutions in seconds, the fundamental question isn’t whether to learn—but what to learn, and why. The old model of learning: master basics, then struggle through mid-level problems, hoping for mastery at the end, no longer works. It’s like building a city without a master plan—patches fix symptoms, but the system collapses under its own complexity. The real shift? Reversing the order. Start not with “how to do,” but with “what it should look like when it’s done.” Before writing a single line of code, imagine the system at scale: how it scales, how it fails, how it survives. This is the power of late-game thinking—the ability to see the future state and design backward. AI excels at mid-game execution: it writes CRUD, drafts emails, fills in syntax. But it cannot see the whole picture. It lacks situational awareness. It doesn’t know that a function today will become a bottleneck tomorrow, or that a quick fix today will rot the system in fi...--AI Generated

LateGame EngineeringIntuition Paradigms AI Era Learning Strategy Engineering Intuition
29
DEC

Use yubikey to protect SSH key.

What if your most sensitive digital access could be protected by a small, untraceable hardware token that never lets your private key leave its secure enclave? This guide reveals how a YubiKey transforms SSH authentication from a vulnerable convenience into a fortress of security—by generating a resident Ed25519 key directly on the device, pin-protected and impossible to extract. No more worrying about leaked private keys, no more risky key backups—just insert the key, touch it when prompted, and gain access. But here’s the deeper question: in an era where digital identity is increasingly under siege, how much of our trust should we place in software, and how much should we demand physical, tangible proof of identity? The setup is elegantly simple—load the key with a single command, secure your server with strict authentication policies, and even set up a new machine without transferring a single file. Yet the real power lies in the philosophy: your private key isn’t stored anywhere bu...--AI Generated

YubiKey SSH Security Hardware Key Ed25519 PIN Protection SSH Setup
24
NOV

让模糊的视频变清晰

What if you could breathe new life into grainy, forgotten home videos—transforming decades-old 480p footage into crisp, detailed 4K clarity with a single command? This isn’t science fiction; it’s the power of AI-driven video super-resolution, and this article dives deep into how a simple Bash script turns complex, GPU-intensive image restoration into a seamless, one-click process. At its core lies Real-ESRGAN, a deep learning model trained not on idealized, synthetic data, but on a rich simulation of real-world degradation—motion blur, JPEG artifacts, sensor noise—teaching it to understand and reverse the messy imperfections of actual old footage. The script automates the entire pipeline: splitting videos into frames, applying AI enhancement at 4x resolution using a Vulkan-enabled GPU, then reassembling them into high-fidelity H.265 videos with preserved audio. But beyond the technical elegance lies a deeper question: when AI “invents” missing details—like the texture of a child’s swea...--AI Generated

bash Linux REAL-ESRGAN Video Upscaling AI Super Resolution Bash Script
10
SEP

Windows 11: The Last Version of Windows 10

Windows 11 isn’t a new operating system—it’s a rebranded evolution of Windows 10, built on the same 10.0 kernel and shaped by a radical shift in Microsoft’s development philosophy. What began as a promise in 2015 that Windows 10 would be the last version was never about technical limits, but a strategic pivot to “Windows as a Service”—a rolling release model where updates flow continuously, not in monolithic versions. This shift solved Microsoft’s biggest problem: fragmentation. By forcing users onto a single, ever-updating platform, Microsoft regained control over its ecosystem, reduced support costs, and laid the foundation for a more secure, unified experience. Yet this stability comes at a cost—features appear, vanish, or change without warning, turning the OS into a live experiment. The abrupt launch of Windows 11, with its strict hardware requirements like TPM 2.0, wasn’t a technical necessity but a calculated move to drive hardware upgrades, enforce a new security baseline, and ...--AI Generated

TPM2 WindowsAsAService Windows11 RollingRelease PCUpgrade EcosystemControl