# Tinker AI Lab (tinker-ai.com)

> Tinker AI Lab is a bilingual (English & Simplified Chinese) Reproducible AI Engineering Lab ("可复现 AI 工程与实测工坊") dedicated to helping developers, HomeLab builders, and makers run AI on their own hardware and workflows. Every guide ships with tested hardware specs, pinned software versions, verification dates, downloadable parameterized configs, offline capability boundaries, and known limits.

## Core Positioning & Verification Standard

- **Entity Name**: Tinker AI Lab / Tinker AI 实测工坊
- **Canonical Domain**: https://tinker-ai.com (English) | https://tinker-ai.com/zh/ (Simplified Chinese)
- **Content Pillars (maintenance-weighted as of 2026-10-27)**: Local AI / HomeLab (active) · API & Deployment Troubleshooting (active) · No-Code AI Workflows (maintain existing only) · AI Maker / STEM (existing guides kept; new posts frozen). Start-here hardware paths: Apple Silicon → mac-local-llm-benchmark; NVIDIA → nvidia-pc-vram-diagnostics; Pi/Edge → pi5-offline-voice-vision-box. Not an AI writing SaaS or coding-tool review directory.
- **Full Context Index**: [llms-full.txt](https://tinker-ai.com/llms-full.txt)

## Interactive Engineering Workbenches

- [Local LLM VRAM & Context Budget Calculator](https://tinker-ai.com/tools/vram-calculator/): Separates GGUF quantized weight size (28 model presets including DeepSeek-R1-671B, Qwen3-235B-A22B, Qwen3-Next-80B-A3B, Qwen3-32B/30B-A3B, Qwen2.5-VL-32B, and Llama-3.2-3B), dynamic KV Cache (FP16 / Q8_0 / Q4_0), and compute buffer overhead across 50 hardware profiles (Apple Silicon up to Mac Studio M5 Ultra 512GB / M3 Ultra 256GB/512GB, Dual-Mac TB5 RDMA JACCL 256GB, MacBook Pro M5 Max 128GB & MacBook Pro M6 16/24/32GB, Qualcomm Snapdragon X2 Elite Extreme 128GB, Intel Core Ultra X9 388H 64GB, NVIDIA RTX Spark N1X 128GB, AMD Ryzen AI Max+ PRO 495 192GB, AMD EPYC 12-Ch DDR5 Hybrid 512GB, NVIDIA RTX PRO 6000 96GB / PRO 5000 72GB / PRO 4500 32GB / PRO 4000 SFF 24GB, Intel Arc Pro B50 16GB, Tenstorrent Blackhole p150a 32GB, RTX 5080 SUPER 24GB / 5070 SUPER 18GB, single/dual DGX Spark GB10 128GB/256GB, Jetson AGX Thor 128GB, AMD Radeon AI PRO R9700 32GB, Strix Halo 128GB, RTX 5090 / 5080 / 5070 Ti / 5060 Ti, RX 9070 XT 16GB, Intel Arc Pro B60 24GB, Jetson Orin Nano Super, and Raspberry Pi 5 AI HAT+ 2 Hailo-10H 8GB), with one-click Modelfile export. ([中文版](https://tinker-ai.com/zh/tools/vram-calculator/))
- [Interactive AI Deployment & Streaming API Troubleshooter](https://tinker-ai.com/tools/troubleshooter/): 13 step-by-step diagnostic decision trees for container-to-host Ollama networking, silent CPU fallback, post-HTTP-200 SSE stream errors, n8n workflow timeouts, HTTP 403 CORS, DeepSeek-R1 loops, Pi 5 power/PCIe/audio, RAG vector dimension mismatches, AMD Strix Halo 15.5GB GTT / Arc B580 / DGX Spark UMA traps, n8n + MCP tool-calling loops, Blackwell sm_120 / RDNA 4 gfx1201 compute errors, Multi-GPU PCIe P2P / 200GbE ConnectX-7 RoCEv2 cluster traps, and Hybrid CPU+GPU MoE NUMA / Edge NPU (Hailo-10H & RK3588 CMA) errors. ([中文版](https://tinker-ai.com/zh/tools/troubleshooter/))
- [Reproducible Artifact & Config Workbench](https://tinker-ai.com/artifacts/): In-browser customization and download of 42 verified Docker Compose stacks, n8n + MCP workflow JSONs, macOS Thunderbolt 5 RDMA JACCL & 70W SFF RTX PRO 4000 / Arc Pro B50 / Tenstorrent Blackhole scripts, llama.cpp --n-cpu-moe & EPYC 12-Ch DDR5 Hybrid MoE scripts, Blackwell RTX PRO 5000/6000 scripts, DGX Spark 200GbE RoCEv2, Apple M5 Max Metal 4, M3 Ultra 512GB & M6 memory-tier scripts, Snapdragon X2 Elite Extreme / Intel Panther Lake / RTX Spark / Gorgon Halo unified-memory benchmarks, Jetson AGX Thor / Pi 5 Hailo-10H 8GB / RK3588 edge scripts, Qwen2.5-VL OCR runners, and resilient SSE clients. ([中文版](https://tinker-ai.com/zh/artifacts/))

## Pillar: Local AI / HomeLab / 本地 AI / HomeLab

Go beyond copy-pasted install commands: tested memory budgets, token speeds under fixed tasks, and offline verification. (不止步于翻译安装命令：提供固定任务下的显存/上下文实测、备份恢复演练与离线边界核查。)

- [Apple Silicon Mac Local Model Benchmark: Fixed Tasks, Quantization, Context & Speed](https://tinker-ai.com/guides/mac-local-llm-benchmark/) ([中文: 一台现有 Mac 的本地模型实测：固定任务、模型版本、量化、上下文和速度](https://tinker-ai.com/zh/guides/mac-local-llm-benchmark/)): Measured prompt eval and token generation speeds across Q4_K_M, Q5_K_M, and Q8_0 on unified memory—showing exactly when raising num_ctx triggers SSD swap degradation. | Tested on: MacBook Pro 14" (M3 Pro · 18GB & 36GB Unified Memory tested) · macOS 15.3 | Verified: 2026-10-23
- [NVIDIA PC Model Loading & VRAM Diagnostics: Spotting PCIe Shared-Memory Spillover](https://tinker-ai.com/guides/nvidia-pc-vram-diagnostics/) ([中文: 一台现有 NVIDIA PC 的模型加载与显存排查：定位 PCIe 共享内存回退与 Docker GPU 透传](https://tinker-ai.com/zh/guides/nvidia-pc-vram-diagnostics/)): Why a 12GB RTX GPU silently drops from 48 tok/s to 4 tok/s without throwing an OOM error—and how to cross-check nvidia-smi against ollama ps. | Tested on: NVIDIA RTX 4070 12GB VRAM · 32GB DDR5 System RAM · Ubuntu 24.04 & Windows 11 WSL2 | Verified: 2026-10-23
- [Ollama + Open WebUI Operations: Installation, Persistence, Backup, Upgrade & Safe Restore](https://tinker-ai.com/guides/ollama-openwebui-persistence-backup/) ([中文: Ollama + Open WebUI：安装、持久化、备份、升级和恢复实战](https://tinker-ai.com/zh/guides/ollama-openwebui-persistence-backup/)): Pin exact container image tags, map host bind mounts for SQLite & vector stores, run a 10-second pre-upgrade snapshot, and practice full rollback. | Tested on: Linux HomeLab (Ubuntu 24.04 · Docker Compose v2.29) & macOS Host + Docker WebUI | Verified: 2026-10-23
- [Local Document Q&A (RAG): Downloadable Sample Corpus, Citation Check & Failure Cases](https://tinker-ai.com/guides/local-rag-citation-and-failure-cases/) ([中文: 本地文档问答 (RAG)：可下载测试样例、引用检查和三大典型失败案例](https://tinker-ai.com/zh/guides/local-rag-citation-and-failure-cases/)): Test your local RAG pipeline with our downloadable ground-truth policy document—exposing silent num_ctx truncation, two-column PDF bleed, and hallucinated citations. | Tested on: Mac M3 Pro 36GB & RTX 4070 12GB · Local Ollama + Open WebUI Hybrid RAG | Verified: 2026-10-23
- [DeepSeek-R1 & Qwen2.5-Coder Local IDE Setup: <think> Token Budgets, FIM Autocomplete & Loop Fixes](https://tinker-ai.com/guides/deepseek-r1-local-coding-and-think-budget/) ([中文: DeepSeek-R1 与 Qwen2.5-Coder 本地编程实战：<think> 思考预算、FIM 补全与死循环排障](https://tinker-ai.com/zh/guides/deepseek-r1-local-coding-and-think-budget/)): Why using DeepSeek-R1 for tab autocomplete ruins IDE latency—and how to pair R1-14B (for architectural reasoning with 4K <think> budget) alongside Qwen2.5-Coder-7B (for sub-250ms FIM edits). | Tested on: MacBook Pro M3 Max (36GB Unified) & NVIDIA RTX 4070 Ti Super (16GB VRAM) | Verified: 2026-10-23
- [Concurrent Local LLM Serving: Ollama NUM_PARALLEL vs vLLM Continuous Batching Benchmark](https://tinker-ai.com/guides/vllm-vs-ollama-concurrent-homelab/) ([中文: HomeLab 多并发本地模型实测：Ollama 多槽位并发 vs vLLM 连续批处理吞吐量对比](https://tinker-ai.com/zh/guides/vllm-vs-ollama-concurrent-homelab/)): Measured P95 Time-to-First-Token (TTFT), aggregate throughput, and VRAM pre-allocation on an RTX 4090 (24GB) under 1, 4, and 8 parallel workflow workers. | Tested on: NVIDIA RTX 4090 (24GB VRAM) · 64GB DDR5 · Ubuntu 24.04 · Docker Engine 27.x | Verified: 2026-10-23
- [Local Vision OCR Benchmark: Qwen2.5-VL-7B vs Traditional OCR on Scanned PDFs, Merged Tables & Formulas](https://tinker-ai.com/guides/local-vision-ocr-qwen25vl-pdf-tables/) ([中文: 本地视觉模型 OCR 实测：Qwen2.5-VL-7B 转换扫描件 PDF、跨列合并表格与数学公式实战](https://tinker-ai.com/zh/guides/local-vision-ocr-qwen25vl-pdf-tables/)): Why standard PDF text extractors return 0 characters on scanned contracts—and how to convert multi-page scanned PDFs into clean Markdown tables locally using pdftoppm (200 DPI) + Qwen2.5-VL-7B. | Tested on: MacBook Pro M3 Max (36GB Unified) & NVIDIA RTX 4070 Ti Super (16GB VRAM) | Verified: 2026-10-26
- [AMD Ryzen AI Max+ 395 (Strix Halo 128GB) Local LLM Benchmark: 96GB GTT Pool, Vulkan vs ROCm & Qwen3-30B-A3B MoE](https://tinker-ai.com/guides/amd-strix-halo-128gb-unified-memory-local-llm/) ([中文: AMD Ryzen AI Max+ 395 (Strix Halo 128GB 统一内存) 本地大模型实测：96GB 显存池解锁、Vulkan vs ROCm 与 Qwen3-30B-A3B MoE 吞吐](https://tinker-ai.com/zh/guides/amd-strix-halo-128gb-unified-memory-local-llm/)): Tested on a 128GB LPDDR5X-8000 Strix Halo Mini PC (Radeon 8060S gfx1151): unlocking the 96GB unified VRAM pool on Linux 6.16+, tuning ROCm hipBLASLt vs Vulkan RADV, and why Qwen3-30B-A3B MoE hits 78.4 tok/s while dense 70B models cap at 6.2 tok/s. | Tested on: GMKtec EVO-X2 / Framework Desktop (AMD Ryzen AI Max+ 395 · Radeon 8060S 40CU gfx1151 · 128GB LPDDR5X-8000 256 GB/s) · Ubuntu 24.04.2 (Kernel 6.16.9) | Verified: 2026-10-27
- [Apple Mac Studio M3 Ultra (512GB / 256GB · 819 GB/s) Benchmark: Running Full DeepSeek-R1-671B & Qwen3-235B-A22B with MLX vs llama.cpp](https://tinker-ai.com/guides/mac-studio-m3-ultra-512gb-deepseek-r1-671b-qwen3-235b-mlx-benchmark/) ([中文: Apple Mac Studio M3 Ultra (512GB / 256GB · 819 GB/s) 本地大模型实测：满血 DeepSeek-R1-671B 与 Qwen3-235B-A22B 在 MLX 与 llama.cpp 下的吞吐对决](https://tinker-ai.com/zh/guides/mac-studio-m3-ultra-512gb-deepseek-r1-671b-qwen3-235b-mlx-benchmark/)): Unlock 475GB of Metal wired memory on a 512GB Mac Studio M3 Ultra to run full 671B DeepSeek-R1/V3 (4-bit) at 14.8 tok/s under 192W wall power, compare MLX vs llama.cpp on Qwen3-235B-A22B MoE (32.4 tok/s), and link two Macs via 80 Gbps Thunderbolt 5 RDMA/Bridge0. | Tested on: Mac Studio M3 Ultra (32C CPU / 80C GPU, 512GB Unified Memory @ 819 GB/s) + Mac Studio M3 Ultra (256GB @ 819 GB/s) · macOS 15.5 Sequoia | Verified: 2026-10-28
- [NVIDIA Blackwell Local LLM Benchmark: RTX 5090 (32GB) vs RTX 5060 Ti (16GB), Native FP4/MXFP4, CUDA 12.8 sm_120 & 450W Power Tuning](https://tinker-ai.com/guides/rtx-5090-vs-rtx-5060ti-16gb-blackwell-fp4-sm120-local-llm-benchmark/) ([中文: NVIDIA Blackwell 本地大模型实测：RTX 5090 (32GB) 对决 RTX 5060 Ti (16GB)、原生 FP4 加速、CUDA 12.8 (sm_120) 与 450W 甜点锁功耗](https://tinker-ai.com/zh/guides/rtx-5090-vs-rtx-5060ti-16gb-blackwell-fp4-sm120-local-llm-benchmark/)): Side-by-side GDDR7 benchmarks of RTX 5090 (32GB · 1,792 GB/s) and RTX 5060 Ti (16GB · 448 GB/s) across Qwen3-14B, Qwen3-30B-A3B MoE, and Qwen2.5-32B, testing Blackwell FP4 Tensor Cores and proving why capping RTX 5090 to 450W saves 125W while retaining 96.4% tok/s. | Tested on: NVIDIA GeForce RTX 5090 (32GB GDDR7 · 1,792 GB/s) & RTX 5060 Ti (16GB GDDR7 · 448 GB/s) · AMD Ryzen 9 9950X · 64GB DDR5 · Ubuntu 24.04.2 LTS | Verified: 2026-10-28
- [Budget High-VRAM Homelab Showdown: Intel Arc Pro B60 (24GB / Dual 48GB) vs AMD Radeon RX 9070 XT (16GB RDNA 4) on Vulkan & ROCm 6.3.4](https://tinker-ai.com/guides/intel-arc-pro-b60-24gb-vs-radeon-rx-9070-xt-rdna4-rocm-vulkan-benchmark/) ([中文: 高性价比大显存 HomeLab 对决：Intel Arc Pro B60 (24GB / 双卡 48GB) vs AMD RX 9070 XT (16GB RDNA 4) 的 Vulkan 与 ROCm 6.3.4 实测](https://tinker-ai.com/zh/guides/intel-arc-pro-b60-24gb-vs-radeon-rx-9070-xt-rdna4-rocm-vulkan-benchmark/)): Benchmarking Intel’s 24GB Arc Pro B60 workstation GPU (single 24GB and dual-card 48GB Battlematrix) against AMD’s 640 GB/s Radeon RX 9070 XT (16GB gfx1201) across Qwen3-14B, Qwen3-30B-A3B MoE, and Qwen2.5-32B. | Tested on: 2x Intel Arc Pro B60 (24GB GDDR6 · 456 GB/s · 160 XMX each) & AMD Radeon RX 9070 XT (16GB GDDR6 · 640 GB/s · gfx1201) · Linux Kernel 6.13.5 · Ubuntu 24.04.2 LTS | Verified: 2026-10-28
- [NVIDIA DGX Spark (GB10 · 128GB / Dual-Node 256GB) Local LLM Benchmark: Single-Box NVFP4 vs 200GbE ConnectX-7 RoCEv2 Distributed vLLM & llama.cpp](https://tinker-ai.com/guides/nvidia-dgx-spark-gb10-single-and-dual-node-200gbe-vllm-llama-cpp-benchmark/) ([中文: NVIDIA DGX Spark (GB10 · 128GB / 双机 256GB 互联) 本地大模型实测：单机 NVFP4 对决 ConnectX-7 200GbE RoCEv2 双节点 vLLM 与 llama.cpp](https://tinker-ai.com/zh/guides/nvidia-dgx-spark-gb10-single-and-dual-node-200gbe-vllm-llama-cpp-benchmark/)): Hands-on benchmark of the NVIDIA DGX Spark (Grace Blackwell GB10 Superchip, 128GB LPDDR5X @ 273 GB/s, sm_121) in single-node and switchless 2-node 200GbE QSFP112 ConnectX-7 RoCEv2 cluster mode across Qwen3-30B-A3B, Llama-3.3-70B, Qwen3-Next-80B-A3B, and Qwen3-235B-A22B. | Tested on: 2x NVIDIA DGX Spark (GB10 Grace Blackwell: 20-core Arm CPU + sm_121 GPU, 128GB coherent LPDDR5X @ 273 GB/s per node, linked via 200GbE QSFP112 DAC cable) | Verified: 2026-11-02
- [Workstation Local LLM Showdown: NVIDIA RTX PRO 6000 Blackwell (96GB GDDR7 · 300W Max-Q vs 600W) vs AMD Radeon AI PRO R9700 (32GB / Dual 64GB RDNA 4)](https://tinker-ai.com/guides/rtx-pro-6000-blackwell-96gb-vs-radeon-ai-pro-r9700-32gb-workstation-llm-benchmark/) ([中文: 工作站大显存显卡实测对决：NVIDIA RTX PRO 6000 Blackwell (96GB GDDR7 · 300W Max-Q vs 600W) vs AMD Radeon AI PRO R9700 (32GB / 双卡 64GB RDNA 4)](https://tinker-ai.com/zh/guides/rtx-pro-6000-blackwell-96gb-vs-radeon-ai-pro-r9700-32gb-workstation-llm-benchmark/)): Benchmarking the single-slot-friendly 96GB GDDR7 ECC NVIDIA RTX PRO 6000 Blackwell (at 600W stock and 300W Max-Q power profile) against 1x/2x AMD Radeon AI PRO R9700 (32GB/64GB RDNA 4 gfx1201) across vLLM and llama.cpp (Vulkan RADV vs ROCm). | Tested on: Host A: NVIDIA RTX PRO 6000 Blackwell (96GB GDDR7 ECC · 1,792 GB/s · sm_120) | Host B: 2x AMD Radeon AI PRO R9700 (32GB GDDR6 each = 64GB · 640 GB/s · RDNA 4 gfx1201) | Verified: 2026-11-02
- [Apple M5 Max (128GB · 614 GB/s) vs M4 Max (128GB · 546 GB/s) Local LLM Benchmark: Metal 4 GPU Neural Accelerators, 3.4x Faster Prefill & MLX vs llama.cpp](https://tinker-ai.com/guides/apple-m5-max-vs-m4-max-metal4-neural-accelerators-mlx-llama-cpp-benchmark/) ([中文: Apple M5 Max (128GB · 614 GB/s) 对决 M4 Max (128GB · 546 GB/s) 本地大模型实测：Metal 4 GPU 神经网络加速器、3.4 倍 Prefill 暴涨与 MLX 调优](https://tinker-ai.com/zh/guides/apple-m5-max-vs-m4-max-metal4-neural-accelerators-mlx-llama-cpp-benchmark/)): Side-by-side local LLM benchmark of the 16-inch MacBook Pro M5 Max (128GB unified memory @ 614 GB/s) vs M4 Max (128GB @ 546 GB/s), measuring how Apple’s new per-core GPU Neural Accelerators (Metal 4 TensorOps) slash 32K–64K prompt prefill TTFT by 3.4x in MLX and llama.cpp. | Tested on: MacBook Pro 16" Apple M5 Max (40-core GPU w/ Neural Accelerators · 128GB Unified @ 614 GB/s) vs MacBook Pro 16" Apple M4 Max (40-core GPU · 128GB Unified @ 546 GB/s) | Verified: 2026-11-02
- [AMD EPYC 12-Channel DDR5 + Single GPU Hybrid MoE Benchmark: Running DeepSeek-R1-671B & Qwen3-235B-A22B via KTransformers & ik_llama.cpp (-ot)](https://tinker-ai.com/guides/epyc-ddr5-hybrid-cpu-gpu-ktransformers-ik-llama-cpp-deepseek-r1-671b-qwen3-235b-benchmark/) ([中文: AMD EPYC 12 通道 DDR5 + 单张独显异构大模型实测：用 KTransformers 与 ik_llama.cpp (-ot) 低成本跑通满血 DeepSeek-R1-671B 与 Qwen3-235B](https://tinker-ai.com/zh/guides/epyc-ddr5-hybrid-cpu-gpu-ktransformers-ik-llama-cpp-deepseek-r1-671b-qwen3-235b-benchmark/)): How to run full 671B and 235B sparse MoE models at 14–27 tok/s without a $10,000 Mac Studio or multi-GPU cluster: pinning MLA Attention & KV Cache onto a single 24GB/32GB GPU while streaming routed experts across 12-channel DDR5 EPYC RAM (~460 GB/s) using ik_llama.cpp (-ot) and KTransformers. | Tested on: AMD EPYC 9374F / 9554P (32C/64C Zen 4 · 12x 48GB DDR5-4800 ECC = 576GB @ ~460 GB/s) + 1x NVIDIA RTX 5090 (32GB) or RTX 4090 (24GB) | Verified: 2026-11-04
- [NVIDIA Blackwell 16GB–72GB Local LLM Matrix: RTX PRO 5000 (72GB ECC) vs RTX PRO 4500 (32GB) & RTX 5080 / 5070 Ti (16GB) + Speculative Decoding](https://tinker-ai.com/guides/rtx-pro-5000-72gb-vs-rtx-pro-4500-32gb-and-rtx-5080-5070ti-16gb-blackwell-llm-benchmark/) ([中文: NVIDIA Blackwell 16GB–72GB 显卡全系实测：RTX PRO 5000 (72GB ECC) 对决 PRO 4500 (32GB)、RTX 5080 / 5070 Ti (16GB) 与投机解码翻倍加速](https://tinker-ai.com/zh/guides/rtx-pro-5000-72gb-vs-rtx-pro-4500-32gb-and-rtx-5080-5070ti-16gb-blackwell-llm-benchmark/)): Completing the Blackwell discrete GPU benchmark matrix: why the 300W RTX PRO 5000 (72GB GDDR7 ECC · 1,344 GB/s) is the single-card sweet spot for 70B FP8 & 80B MoE, why the 256-bit RTX 5070 Ti (896 GB/s) hits 93% of RTX 5080 speed, and how Draft Speculative Decoding boosts coding throughput by 1.85x. | Tested on: NVIDIA RTX PRO 5000 Blackwell (72GB GDDR7 ECC · 1,344 GB/s · 300W) | RTX PRO 4500 Blackwell (32GB ECC · 896 GB/s · 200W) | RTX 5080 (16GB · 960 GB/s) & RTX 5070 Ti (16GB · 896 GB/s) | Verified: 2026-11-04
- [128GB vs 192GB Unified-Memory AI PCs (Fall 2026): NVIDIA RTX Spark (N1X · Windows on ARM) vs AMD Ryzen AI Max+ PRO 495 (Gorgon Halo) — Model Fit, Decode Ceilings & a Reproduction Protocol](https://tinker-ai.com/guides/nvidia-rtx-spark-n1x-128gb-vs-ryzen-ai-max-pro-495-192gb-unified-memory-local-llm/) ([中文: 2026 秋季 128GB 与 192GB 统一内存 AI PC 选型：NVIDIA RTX Spark（N1X · Windows on ARM）对决 AMD Ryzen AI Max+ PRO 495（Gorgon Halo）——模型装载边界、生成速度理论上限与复现协议](https://tinker-ai.com/zh/guides/nvidia-rtx-spark-n1x-128gb-vs-ryzen-ai-max-pro-495-192gb-unified-memory-local-llm/)): RTX Spark laptops/mini desktops (shipping since mid-October 2026) and Ryzen AI Max+ 400 "Gorgon Halo" mini PCs both promise "run 100B+ models locally". The real difference is not TOPS: it is ~300 GB/s vs ~273 GB/s memory bandwidth (decode speed), CUDA Tensor Cores vs RDNA 3.5 (prefill speed), and 128GB vs 160GB GPU-addressable memory (whether Qwen3-235B-A22B Q4 fits at all). This guide derives decode ceilings from published specs, marks every number as an estimate, and ships a script so you can replace our estimates with your own measurements. | Tested on: Spec basis (not lab-measured): NVIDIA RTX Spark N1X top bin (20-core Grace ARM CPU + Blackwell GPU 6,144 CUDA cores · up to 128GB LPDDR5X 256-bit · ~300 GB/s reported) vs AMD Ryzen AI Max+ PRO 495 (16C Zen 5 + Radeon 8065S 40CU RDNA 3.5 gfx1151 · up to 192GB LPDDR5X-8533 256-bit · ~273 GB/s · up to 160GB GPU-allocatable) | Verified: 2026-11-05
- [Apple M6 MacBook Pro (16GB / 24GB / 32GB) for Local LLMs: Which Memory Tier, Which Models, and Why the Dual Neural Engine Does Not Speed Up llama.cpp](https://tinker-ai.com/guides/apple-m6-macbook-pro-16gb-24gb-32gb-local-llm-sizing-neural-engine-vs-gpu/) ([中文: Apple M6 MacBook Pro（16GB / 24GB / 32GB）跑本地大模型：选哪档内存、跑哪些模型，以及为什么双 Neural Engine 不会让 llama.cpp 变快](https://tinker-ai.com/zh/guides/apple-m6-macbook-pro-16gb-24gb-32gb-local-llm-sizing-neural-engine-vs-gpu/)): The new 2nm M6 brings per-core GPU Neural Accelerators, a dual 16-core Neural Engine, and 153 GB/s (16GB) or 170 GB/s (24/32GB) of memory bandwidth. For local LLMs the decisive spec is the memory tier: 16GB is an 8B-class machine, 24GB is a 14B machine, and 32GB is the only tier that comfortably runs Qwen3-30B-A3B. The Neural Engine is not used by llama.cpp, Ollama, or MLX — buy for memory, not for NPU TOPS. | Tested on: Spec basis (not lab-measured): 14-inch MacBook Pro with base Apple M6 (TSMC N2 · 12-core CPU: 2 super + 4 performance + 6 efficiency · 12-core GPU with per-core Neural Accelerators · dual 16-core Neural Engine · 16GB @ 153 GB/s, 24GB/32GB @ 170 GB/s, as reported at launch) | Verified: 2026-11-05
- [Thin-and-Light AI Laptops for Local LLMs (Fall 2026): Snapdragon X2 Elite Extreme (192-bit · 228 GB/s · 128GB) vs Intel Core Ultra X9 388H "Panther Lake" (Arc B390 12 Xe3 · 153.6 GB/s) — Bandwidth Ceilings, NPU Myths & Reproduction Protocol](https://tinker-ai.com/guides/snapdragon-x2-elite-extreme-128gb-vs-intel-core-ultra-x9-388h-panther-lake-xe3-local-llm/) ([中文: 2026 轻薄本跑本地大模型选型：高通骁龙 X2 Elite Extreme（192-bit · 228 GB/s · 128GB）对决 Intel 酷睿 Ultra X9 388H「Panther Lake」（Arc B390 12 Xe3 · 153.6 GB/s）——带宽上限、NPU 误区与复现协议](https://tinker-ai.com/zh/guides/snapdragon-x2-elite-extreme-128gb-vs-intel-core-ultra-x9-388h-panther-lake-xe3-local-llm/)): Qualcomm Snapdragon X2 Elite Extreme (80 TOPS Hexagon NPU · 18-core Oryon) and Intel Core Ultra Series 3 "Panther Lake" (50 TOPS NPU 5 · Arc B390 12 Xe3 cores on Intel 18A) dominate 2026 thin-and-light Windows AI laptops. For local LLMs, ignore the NPU TOPS sticker: standard llama.cpp and Ollama run on the integrated GPU, where X2 Elite Extreme’s 192-bit LPDDR5X-9533 bus (228 GB/s, up to 128GB) beats Panther Lake’s 128-bit LPDDR5X-9600 (153.6 GB/s, up to 64/96GB) by ~48% on token decode—while Intel’s XMX matrix engines fight back on prompt prefill and x86_64 compatibility. | Tested on: Spec basis (not lab-measured): Qualcomm Snapdragon X2 Elite Extreme X2E-96-100 (TSMC 3nm · 18-core 3rd-gen Oryon up to 5.0 GHz · Adreno X2-90 GPU · 80 TOPS Hexagon NPU · 192-bit LPDDR5X-9533 @ 228 GB/s · up to 128GB) vs Intel Core Ultra X9 388H "Panther Lake" (Intel 18A · 16C: 4P+8E+4LPE · Arc B390 12 Xe3 cores + XMX · 50 TOPS NPU 5 · 128-bit LPDDR5X-9600 @ 153.6 GB/s · up to 64GB/96GB) | Verified: 2026-11-06
- [Running 30B & 80B MoE Models on a 12GB–24GB Consumer GPU + Dual-Channel DDR5 PC: Native llama.cpp --n-cpu-moe Benchmark & RTX 5070 SUPER (18GB) / 5080 SUPER (24GB) Sizing](https://tinker-ai.com/guides/llama-cpp-n-cpu-moe-hybrid-offload-16gb-24gb-vram-ddr5-vs-rtx-5080-super-24gb-5070-super-18gb/) ([中文: 单张 12GB–24GB 消费级显卡 + 双通道 DDR5 跑通 30B 与 80B MoE 大模型：llama.cpp 原生 --n-cpu-moe 专家卸载实测与 RTX 5070 SUPER (18GB) / 5080 SUPER (24GB) 显存选型](https://tinker-ai.com/zh/guides/llama-cpp-n-cpu-moe-hybrid-offload-16gb-24gb-vram-ddr5-vs-rtx-5080-super-24gb-5070-super-18gb/)): You do not need an EPYC server or a $4,000 workstation card to run Qwen3-30B-A3B (18.6GB) or Qwen3-Next-80B-A3B (48.2GB) on a desktop PC. Upstream llama.cpp’s native `--n-cpu-moe N` (`-ncmoe N`) flag keeps 100% of attention layers and the Q8_0 KV cache inside your 12GB/16GB/24GB GPU while offloading only sparse MoE experts for N layers to dual-channel DDR5 RAM—delivering 46–55 tok/s on Qwen3-30B-A3B and 26–33 tok/s on Qwen3-Next-80B-A3B. We also model how 3GB-module GDDR7 refreshes (RTX 5070 SUPER 18GB & RTX 5080 SUPER 24GB, spec-estimate) shift the exact `--n-cpu-moe` threshold. | Tested on: Measured on: AMD Ryzen 9 9900X + 64GB Dual-Channel DDR5-6400 CL30 (~84 GB/s read) + NVIDIA RTX 4070 (12GB) / RTX 5060 Ti (16GB) / RTX 5070 Ti (16GB) / RTX 4090 (24GB) | Spec-estimate rows: NVIDIA RTX 5070 SUPER (18GB GDDR7 192-bit · ~672 GB/s) & RTX 5080 SUPER (24GB GDDR7 256-bit 32 Gbps · ~1,024 GB/s) | Verified: 2026-11-06
- [Multi-Mac Thunderbolt 5 RDMA Cluster for Local LLMs: rdma_ctl, MLX JACCL (--backend jaccl) vs exo vs llama.cpp RPC & Mac Studio M5 Ultra (512GB · 1.2 TB/s) Sizing](https://tinker-ai.com/guides/thunderbolt-5-rdma-mlx-jaccl-exo-llama-cpp-rpc-multi-mac-studio-cluster-benchmark/) ([中文: 多台 Mac 雷电 5 RDMA 分布式大模型集群实战：rdma_ctl 解锁、MLX JACCL (--backend jaccl) 对决 exo 与 llama.cpp RPC 及 Mac Studio M5 Ultra (512GB · 1.2 TB/s) 推算](https://tinker-ai.com/zh/guides/thunderbolt-5-rdma-mlx-jaccl-exo-llama-cpp-rpc-multi-mac-studio-cluster-benchmark/)): Apple’s native RDMA over Thunderbolt 5 (`rdma_ctl enable` in macOS 26.2+) and MLX’s JACCL backend (`mlx.launch --backend jaccl`) slash Mac-to-Mac latency from ~180 µs (TCP over Thunderbolt Bridge) to ~3 µs over 80 Gbps links—finally making Tensor Parallelism (TP) across 2 to 4 Apple Silicon Macs faster than single-node execution. We compare MLX JACCL, `exo`, and `llama.cpp` `rpc-server` on a 2× M5 Max (256GB) & 2× M3 Ultra cluster, document the 6-cable 4-node full-mesh rule, and model 2× Mac Studio M5 Ultra (512GB · 1.2 TB/s each, spec-estimate). | Tested on: Measured on: 2× MacBook Pro M5 Max (128GB · 614 GB/s each = 256GB pool) & 2× Mac Studio M3 Ultra (256GB · 819 GB/s each = 512GB pool) via Belkin 1m active Thunderbolt 5 (80 Gbps) cables | Spec-estimate row: Mac Studio M5 Ultra (512GB · ~1,200 GB/s) | Verified: 2026-11-09
- [70W Slot-Powered SFF Workstation GPUs & Open RISC-V PCIe Cards for HomeLab LLMs: NVIDIA RTX PRO 4000 Blackwell SFF (24GB GDDR7) vs Intel Arc Pro B50 (16GB) vs Tenstorrent Blackhole p150a (32GB)](https://tinker-ai.com/guides/rtx-pro-4000-blackwell-sff-24gb-70w-vs-intel-arc-pro-b50-16gb-and-tenstorrent-blackhole-p150a-homelab-llm/) ([中文: 70W 免外接供电半高工作站显卡与开源 RISC-V 加速卡实测对决：NVIDIA RTX PRO 4000 Blackwell SFF (24GB GDDR7) vs Intel Arc Pro B50 (16GB) vs Tenstorrent Blackhole p150a (32GB)](https://tinker-ai.com/zh/guides/rtx-pro-4000-blackwell-sff-24gb-70w-vs-intel-arc-pro-b50-16gb-and-tenstorrent-blackhole-p150a-homelab-llm/)): Compact SFF HomeLab servers (Minisforum MS-01, Dell Precision/OptiPlex SFF, 1U/2U shallow racks) cannot fit 300W triple-slot GPUs or auxiliary power cables. We benchmark the new wave of 70W slot-powered half-height workstation cards—NVIDIA RTX PRO 4000 Blackwell SFF (24GB GDDR7 ECC · 432 GB/s · 70W) and Intel Arc Pro B50 (16GB GDDR6 · 224 GB/s · 70W)—alongside Tenstorrent’s open-source RISC-V Tensix PCIe card, Blackhole p150a (32GB GDDR6 · 512 GB/s · 4× 800G QSFP-DD), comparing tok/s, tokens-per-watt (tok/J), ECC VRAM tax, and software compatibility (`llama.cpp` vs `tt-metal` / `vLLM`). | Tested on: NVIDIA RTX PRO 4000 Blackwell SFF (24GB GDDR7 ECC · 192-bit · 432 GB/s · 70W slot-powered, GB203GL) vs Intel Arc Pro B50 (16GB GDDR6 · 128-bit · 224 GB/s · 70W slot-powered, BMG-G21) vs Tenstorrent Blackhole p150a (32GB GDDR6 · 512 GB/s · 120 Tensix cores · 300W) & p300c (64GB Dual-ASIC · 1,024 GB/s [Spec-Est.]) | Verified: 2026-11-09

## Pillar: No-Code AI Workflows / 无代码工作流

Production-hardened workflows connected to local or hybrid models with deduplication, schema guards, and dead-letter queues. (聚焦经过验证的实用场景流程，内置去重、JSON Schema 校验与超时降级分支。)

- [n8n + Local LLM RSS Classifier: Importable Workflow, Deduplication & Retry Branches](https://tinker-ai.com/guides/n8n-local-llm-rss-pipeline/) ([中文: n8n + 本地模型的 RSS 分类：导入文件、测试输入、去重和重试](https://tinker-ai.com/zh/guides/n8n-local-llm-rss-pipeline/)): A production-ready n8n workflow that fetches RSS feeds, deduplicates URLs, classifies items via local Ollama with strict JSON output, and handles queue timeouts. | Tested on: Self-hosted n8n (Docker) + Local Ollama (Qwen2.5-7B-Instruct Q4_K_M) | Verified: 2026-10-23
- [Automated Webhook Ticket Triage: Schema Guard, Dead-Letter Fallback, Cost & Permissions](https://tinker-ai.com/guides/n8n-webhook-ticket-triage/) ([中文: 自动化工单分流工作流：字段校验、死信降级、API 成本预算与最小权限说明](https://tinker-ai.com/zh/guides/n8n-webhook-ticket-triage/)): A practical customer support / issue triage pipeline comparing local Ollama vs cloud API cost, strict input truncation, and least-privilege credential scoping. | Tested on: Docker n8n 1.76+ · Compatible with Local Ollama or Any OpenAI-Compatible Endpoint | Verified: 2026-10-23
- [n8n + Local LLM Invoice & Contract Extractor: JSON Schema Guard & Deterministic Math Reconciliation](https://tinker-ai.com/guides/n8n-pdf-invoice-structured-extractor/) ([中文: n8n + 本地模型发票与合同提取工作流：JSON 强约束与金额恒等式自动核验](https://tinker-ai.com/zh/guides/n8n-pdf-invoice-structured-extractor/)): Never trust an LLM to do accounting math alone: extract structured invoice fields locally with Qwen2.5-14B, then verify `subtotal + tax === total` in a deterministic n8n Code node. | Tested on: Docker n8n 1.76+ · Local Ollama (Qwen2.5-14B-Instruct Q4_K_M / Q5_K_M) | Verified: 2026-10-23
- [Self-Hosted Git PR Diff Reviewer with n8n & Qwen2.5-Coder: Secret Redaction & Token Caps](https://tinker-ai.com/guides/n8n-github-pr-security-review-bot/) ([中文: 基于 n8n 与本地 Qwen2.5-Coder 的代码 PR 自动审查：密钥脱敏、Diff 截断与最小权限](https://tinker-ai.com/zh/guides/n8n-github-pr-security-review-bot/)): Automate pull request bug and security reviews on your own GPU using n8n and Qwen2.5-Coder-14B—complete with lockfile filtering, regex secret scrubbing, and scoped webhook tokens. | Tested on: Self-hosted n8n + Local Ollama (`qwen2.5-coder:14b` on 16GB+ GPU / 24GB+ Mac) | Verified: 2026-10-23
- [n8n + Local LLM Read-Only Text-to-SQL Analyst: AST Guardrails, Row Limits & Zero-Write Security](https://tinker-ai.com/guides/n8n-sql-readonly-database-analyst/) ([中文: n8n + 本地模型 Text-to-SQL 数据分析工作流：只读事务沙箱、AST 语法拦截与行数熔断](https://tinker-ai.com/zh/guides/n8n-sql-readonly-database-analyst/)): Let team members query internal PostgreSQL/MySQL databases in plain language using Qwen2.5-Coder-14B—backed by triple defense-in-depth guards so an LLM can never DROP, UPDATE, or dump your database. | Tested on: Docker n8n 1.76+ · PostgreSQL 16 / MySQL 8 · Local Ollama (`qwen2.5-coder:14b`) | Verified: 2026-10-26
- [n8n + Model Context Protocol (MCP) + Local Qwen3 / Ollama Tool Calling: Sandboxed Filesystem, SQLite & Loop Guards](https://tinker-ai.com/guides/n8n-mcp-server-client-local-llm-tool-calling/) ([中文: n8n + MCP 协议 (Model Context Protocol) + 本地 Qwen3 工具调用工作流：沙箱读写、工具白名单与死循环熔断](https://tinker-ai.com/zh/guides/n8n-mcp-server-client-local-llm-tool-calling/)): Connect n8n’s MCP Client Tool to local MCP Servers using Qwen3-8B / Qwen3-30B-A3B on Ollama: why exposing >5 MCP tools crashes 8B model accuracy, how `/no_think` cuts tool-routing latency by 84%, and how to enforce read-only allowlists and `maxIterations=5`. | Tested on: Mac mini M4 Pro (24GB Unified Memory · 273 GB/s) & Ubuntu 24.04 Docker Host (RTX 4070 12GB VRAM) | Verified: 2026-10-27

## Pillar: AI Maker / STEM Projects / AI 创客 / STEM

Hands-on hardware + AI builds with exact part numbers, power/thermal checks, and honest setup duration expectations. (软硬结合的实验项目，附带完整物料表、供电与温控自检脚本及真实搭建耗时记录。)

- [Raspberry Pi 5 Offline AI Lab Box: Complete BOM, Power/Thermal Check & Step-by-Step Build](https://tinker-ai.com/guides/pi5-offline-voice-vision-box/) ([中文: 树莓派 5 离线语音与视觉 AI 实验箱：完整物料清单 (BOM)、供电排坑与分步验证](https://tinker-ai.com/zh/guides/pi5-offline-voice-vision-box/)): Build a self-contained offline AI station on Raspberry Pi 5 (8GB) with exact BOM specifications, power throttling diagnostics, and realistic setup time expectations. | Tested on: Raspberry Pi 5 (8GB RAM) · Official 27W (5V/5A) USB-C PD Power Supply · Active Cooler · 64GB A2 microSD / NVMe HAT | Verified: 2026-10-23
- [STEM Project: Raspberry Pi Local Sensor & HomeLab Log Explainer with Zero Cloud Dependency](https://tinker-ai.com/guides/pi5-local-sensor-log-guardian/) ([中文: 创客与教学项目：树莓派本地传感器与家庭服务器日志 AI 诊断盒](https://tinker-ai.com/zh/guides/pi5-local-sensor-log-guardian/)): A structured maker project teaching students how to combine deterministic threshold filtering with a local 3B model for plain-language hardware alerts. | Tested on: Raspberry Pi 5 (8GB) · BME280 I2C Temp/Humidity Sensor (or journalctl system logs) | Verified: 2026-10-23
- [Raspberry Pi 5 AI Kit (Hailo-8L NPU) + Local LLM Build: PCIe Gen 3 Vision & Split-Compute Benchmark](https://tinker-ai.com/guides/pi5-hailo8l-npu-vs-cpu-vision-benchmark/) ([中文: 树莓派 5 AI 套件 (Hailo-8L NPU) 视觉与本地大模型协同实战：PCIe 3.0 实测与功耗拆解](https://tinker-ai.com/zh/guides/pi5-hailo8l-npu-vs-cpu-vision-benchmark/)): Understand what the 13 TOPS Hailo-8L NPU accelerates (28 FPS YOLOv8 vision at <2W) and what it does NOT (GGUF LLMs)—and build a zero-contention NPU Vision + CPU 3B LLM pipeline. | Tested on: Raspberry Pi 5 (8GB) · Official Raspberry Pi AI Kit (M.2 HAT+ with Hailo-8L 13 TOPS NPU) · Pi Camera Module 3 | Verified: 2026-10-23
- [Raspberry Pi 5 100% Offline Voice Assistant Loop: Whisper.cpp STT + Ollama 3B + Piper Neural TTS](https://tinker-ai.com/guides/pi5-whisper-piper-offline-voice-assistant/) ([中文: 树莓派 5 全离线语音助手实战：Whisper.cpp 语音识别 + Ollama 3B + Piper 神经网络语音合成闭环](https://tinker-ai.com/zh/guides/pi5-whisper-piper-offline-voice-assistant/)): Build a sub-2.5s latency air-gapped voice assistant on Raspberry Pi 5 (8GB): exact ALSA audio card bindings, ARM64 NEON Whisper model sizing, and sentence-streaming Piper TTS. | Tested on: Raspberry Pi 5 (8GB) · Official 27W PSU · USB Microphone (UAC) · 3.5mm / USB Speaker | Verified: 2026-10-26
- [NVIDIA Jetson Orin Nano Super (67 TOPS) vs Raspberry Pi 5 AI HAT+ (26 TOPS) & IMX500 AI Camera: Edge VLM & Vision Benchmark](https://tinker-ai.com/guides/jetson-orin-nano-super-vs-pi5-ai-hat-26tops-vision-llm/) ([中文: NVIDIA Jetson Orin Nano Super (67 TOPS) 对决树莓派 5 AI HAT+ (26 TOPS) 与 IMX500 智能相机：边缘视觉与小模型实测](https://tinker-ai.com/zh/guides/jetson-orin-nano-super-vs-pi5-ai-hat-26tops-vision-llm/)): Side-by-side edge hardware benchmark: when to use Raspberry Pi 5 + Sony IMX500 on-sensor NPU / Hailo-8 26 TOPS (4.8W–9.6W ultra-low power vision) versus Jetson Orin Nano Super 25W mode (102 GB/s LPDDR5 running Qwen3-4B & Qwen2.5-VL at 31.4 tok/s). | Tested on: NVIDIA Jetson Orin Nano Super 8GB (JetPack 6.2 · 25W MAXN_SUPER · 102 GB/s) vs Raspberry Pi 5 8GB + AI HAT+ 26 TOPS (Hailo-8) + Raspberry Pi AI Camera (Sony IMX500) | Verified: 2026-10-27
- [Next-Gen Edge AI & Robotics Benchmark: NVIDIA Jetson AGX Thor T5000 (128GB · 2070 TFLOPS) vs Raspberry Pi 5 AI HAT+ 2 (Hailo-10H 8GB) & RK3588 NPU](https://tinker-ai.com/guides/jetson-agx-thor-t5000-128gb-vs-pi5-ai-hat-plus-2-hailo10h-8gb-and-rk3588-edge-llm-benchmark/) ([中文: 新一代边缘 AI 与具身智能硬件实测：NVIDIA Jetson AGX Thor (128GB · 2070 TFLOPS) 对决树莓派 5 AI HAT+ 2 (Hailo-10H 独立 8GB) 与 RK3588 NPU](https://tinker-ai.com/zh/guides/jetson-agx-thor-t5000-128gb-vs-pi5-ai-hat-plus-2-hailo10h-8gb-and-rk3588-edge-llm-benchmark/)): Benchmarking 2026 edge generative AI hardware across three power envelopes: NVIDIA Jetson AGX Thor T5000 (128GB Blackwell · 40W–130W), Raspberry Pi 5 + AI HAT+ 2 (Hailo-10H 40 TOPS with dedicated 8GB onboard LPDDR4X via hailo-ollama at ~3.2W NPU power), and Rockchip RK3588 6 TOPS NPU (rknn-llm). | Tested on: NVIDIA Jetson AGX Thor DevKit (T5000 · 128GB LPDDR5X @ 273 GB/s · 2070 FP4 TFLOPS) | Raspberry Pi 5 (8GB) + AI HAT+ 2 (Hailo-10H 40 TOPS · 8GB Dedicated NPU RAM) | Orange Pi 5 Plus / Rock 5B (RK3588 16GB · 6 TOPS NPU) | Verified: 2026-11-04

## Pillar: API & Deployment Troubleshooting / API 与部署排障

Root-cause diagnostics for errors that official status-code tables miss: mid-stream SSE failures, bridge firewalls, and KV cache OOMs. (覆盖简单状态码列表无法解决的运行期故障：SSE 流内报错、Docker 网桥隔离与 KV Cache 溢出。)

- [Container Cannot Connect to Host Model: Minimal Reproduction & Diagnostic Steps](https://tinker-ai.com/guides/container-to-host-llm-networking/) ([中文: 容器连不上宿主机模型：最小复现和诊断步骤](https://tinker-ai.com/zh/guides/container-to-host-llm-networking/)): Solve the #1 self-hosted AI failure (`ECONNREFUSED` / `ETIMEDOUT` calling port 11434 from Open WebUI, n8n, or Dify) in three deterministic verification commands. | Tested on: Verified on Ubuntu 24.04 (Docker Engine 27.x + UFW) & macOS 15 (Docker Desktop 4.36) | Verified: 2026-10-23
- [Post-HTTP-200 SSE Stream Errors: Failure Conditions & Recovery Boundaries](https://tinker-ai.com/guides/sse-http200-stream-errors/) ([中文: LLM API 在 HTTP 200 后流式报错：错误发生条件与恢复边界](https://tinker-ai.com/zh/guides/sse-http200-stream-errors/)): Why simple HTTP status-code tables fail in streaming LLM apps: handling mid-stream SSE error events, proxy idle cuts, and max_tokens JSON truncation. | Tested on: Node.js 22.13+ SSE Client · Applicable to Claude API, OpenAI API, and Local Ollama / vLLM | Verified: 2026-10-23
- [Fixing Local LLM HTTP 403 CORS, Reverse Proxy Host Headers & LAN Bearer Authentication](https://tinker-ai.com/guides/cors-reverse-proxy-auth-local-llm/) ([中文: 本地模型 403 CORS 跨域拒绝、Nginx/Caddy 反代报错与局域网鉴权排障指南](https://tinker-ai.com/zh/guides/cors-reverse-proxy-auth-local-llm/)): Why browser extensions and LAN web clients get HTTP 403 Forbidden from Ollama—and how to configure OLLAMA_ORIGINS, Host header rewrites, and Bearer token protection safely. | Tested on: Nginx 1.26 / Caddy 2.8 · Ollama 0.5.7 · Chrome / Firefox Extensions & LAN Web Clients | Verified: 2026-10-23
- [Troubleshooting RAG Embedding Dimension Mismatch (1024 vs 768) & Zombie VRAM Allocation](https://tinker-ai.com/guides/embedding-dimension-mismatch-and-zombie-vram/) ([中文: 本地 RAG 向量维度冲突 (1024 vs 768) 报错与显存“幽灵占用”深度排障指南](https://tinker-ai.com/zh/guides/embedding-dimension-mismatch-and-zombie-vram/)): Fix the two most common post-setup local AI failures: Chroma/PGVector dimension mismatch errors after changing embedding models, and GPU VRAM staying 95% full when ollama ps shows zero models. | Tested on: Linux / macOS / Windows WSL2 · Ollama 0.5.7 · Open WebUI / Dify / Continue (ChromaDB, Qdrant, PGVector) | Verified: 2026-10-26
- [New Hardware GPU & Unified Memory Troubleshooting: Intel Arc B580 Docker Fallback, AMD Strix Halo 15.5GB Cap & DGX Spark Page Cache OOM](https://tinker-ai.com/guides/intel-arc-b580-and-amd-uma-gtt-vram-allocation-troubleshooting/) ([中文: 新一代硬件显存与加速排障指南：修复 Intel Arc B580 容器退化 CPU、AMD Strix Halo 15.5GB 显存锁死与 DGX Spark 页缓存 OOM](https://tinker-ai.com/zh/guides/intel-arc-b580-and-amd-uma-gtt-vram-allocation-troubleshooting/)): Root-cause diagnostics and one-line verification scripts for three high-frequency new-hardware failures: Intel Arc B580 (12GB Xe2) silent CPU fallback in Docker, AMD Strix Halo deprecated `amdgpu.gttsize` caps, and NVIDIA DGX Spark (GB10 128GB) unified memory page-cache OOMs. | Tested on: Intel Arc B580 (12GB GDDR6 Xe2) · AMD Ryzen AI Max+ 395 (128GB UMA) · NVIDIA DGX Spark (GB10 128GB Coherent Unified Memory) | Verified: 2026-10-27
- [Blackwell (sm_120) & RDNA 4 (gfx1201) GPU Troubleshooting: Fixing "No Kernel Image Is Available", PCIe 5.0 Riser Xid 79 & ROCm ISA Crashes](https://tinker-ai.com/guides/blackwell-sm120-no-kernel-image-and-rdna4-gfx1201-rocm-troubleshooting/) ([中文: Blackwell (sm_120) 与 RDNA 4 (gfx1201) 深度排障指南：解决 no kernel image 报错、PCIe 5.0 延长线 Xid 79 掉卡与 ROCm 算子崩溃](https://tinker-ai.com/zh/guides/blackwell-sm120-no-kernel-image-and-rdna4-gfx1201-rocm-troubleshooting/)): Root-cause diagnostics and copy-paste fixes for the four most common failures on RTX 5090 / 5080 / 5070 Ti / 5060 Ti (`sm_120`) and AMD RX 9070 XT (`gfx1201`): CUDA <=12.6 `no kernel image` errors, `FORCE_CUBLAS` 5x slowdown, PCIe Gen5 riser `Xid 79` bus drops, and RDNA 4 `INVALID_ISA` faults. | Tested on: NVIDIA RTX 5090 32GB / RTX 5060 Ti 16GB (`sm_120`) & AMD Radeon RX 9070 XT 16GB (`gfx1201`) · Ubuntu 24.04.2 LTS (Kernel 6.13.5) | Verified: 2026-10-28
- [Multi-GPU PCIe P2P & 200GbE RoCEv2 Troubleshooting: Fixing NCCL P2P Deadlocks (iommu=pt / ACS), DGX Spark ConnectX-7 TCP Fallback, and Multi-Card R9700 Hangs](https://tinker-ai.com/guides/multi-gpu-pcie-p2p-iommu-deadlock-and-connectx7-rocev2-tcp-fallback-troubleshooting/) ([中文: 多卡 PCIe P2P 与 200GbE RoCEv2 互联排障指南：解决 NCCL 死锁卡死 (iommu=pt / ACS)、DGX Spark ConnectX-7 退化 TCP 慢速与多卡 R9700 假死](https://tinker-ai.com/zh/guides/multi-gpu-pcie-p2p-iommu-deadlock-and-connectx7-rocev2-tcp-fallback-troubleshooting/)): Step-by-step root-cause diagnostics and verified fixes for multi-GPU vLLM Tensor Parallel hangs (`iommu=pt`, PCIe ACS BAR1 P2P blocks), 2-node DGX Spark ConnectX-7 200GbE silent fallback to 1GbE TCP (`NET/Socket`), RoCEv2 MTU/GID errors (`ibv_modify_qp`), and multi-card AMD Radeon AI PRO R9700 RCCL deadlocks. | Tested on: 2x NVIDIA DGX Spark (ConnectX-7 200GbE QSFP112) | Dual RTX 5090 / RTX PRO 6000 on AMD Threadripper / Ryzen X870E | Dual AMD Radeon AI PRO R9700 (64GB) | Verified: 2026-11-02
- [Hybrid CPU+GPU MoE (KTransformers / ik_llama.cpp) & Edge NPU (Hailo-10H / RK3588) Troubleshooting: Fixing Auto-NUMA Thrashing, -ot Regex OOMs & CMA Errors](https://tinker-ai.com/guides/hybrid-moe-numa-thrashing-ktransformers-ik-llama-cpp-and-hailo10h-rkllm-npu-troubleshooting/) ([中文: 异构 CPU+GPU 大模型 (KTransformers / ik_llama.cpp) 与边缘 NPU (Hailo-10H / RK3588) 排障指南：解决跨 NUMA 掉速、-ot 正则爆显存与 CMA 报错](https://tinker-ai.com/zh/guides/hybrid-moe-numa-thrashing-ktransformers-ik-llama-cpp-and-hailo10h-rkllm-npu-troubleshooting/)): Root-cause diagnostics and verified fixes for 50–65% Linux Auto-NUMA slowdown (`kernel.numa_balancing=0`) on EPYC/Xeon MoE servers, `ik_llama.cpp` `-ot` regex CUDA OOMs on 24GB/32GB GPUs, Raspberry Pi 5 AI HAT+ 2 (`Hailo-10H`) firmware conflicts, and RK3588 `rknn-llm` CMA allocation failures. | Tested on: AMD EPYC 9004/9005 / Intel Xeon DDR5 + Single RTX GPU | Raspberry Pi 5 + AI HAT+ 2 (Hailo-10H 8GB) | Rockchip RK3588 (16GB/32GB) | Verified: 2026-11-04

## Parameterized Config Artifacts

- [run-mac-benchmark.sh — Apple Silicon Mac Fixed-Task Speed & Context Benchmark Script](https://tinker-ai.com/artifacts/mac-benchmark-runner/) ([中文: Mac 本地模型固定任务、上下文与生成速度复现脚本](https://tinker-ai.com/zh/artifacts/mac-benchmark-runner/)): Measures prompt eval rate, token generation speed (tok/s), and peak unified memory footprint under a fixed deterministic task. (在固定测试任务（seed=42）与指定上下文窗口下，精确采集预填充速度、生成速率 (tok/s) 与统一内存驻留情况。)
- [check-nvidia-vram-spill.sh — NVIDIA PC Model Loading & PCIe Shared-Memory Spillover Probe](https://tinker-ai.com/artifacts/nvidia-vram-probe/) ([中文: NVIDIA PC 模型加载与显存溢出 / PCIe 共享内存回退排查脚本](https://tinker-ai.com/zh/artifacts/nvidia-vram-probe/)): Compares physical VRAM usage in nvidia-smi against ollama ps processor split to detect silent CPU/PCIe fallback. (对比 nvidia-smi 物理显存占用与 ollama ps 的 GPU/CPU 层卸载比例，精准揪出因共享内存回退导致的 10 倍掉速。)
- [docker-compose.ollama-webui.yml — Ollama + Open WebUI Persistent Stack with Backup & Rollback Safety](https://tinker-ai.com/artifacts/ollama-openwebui-compose/) ([中文: Ollama + Open WebUI 持久化、备份与安全回滚栈 (Docker Compose)](https://tinker-ai.com/zh/artifacts/ollama-openwebui-compose/)): Pinned versions, explicit bind-mounted volumes for 30-second backup/restore, telemetry disabled, and KV cache optimization. (锁定镜像版本、显式挂载宿主机持久化目录（支持 30 秒快照备份与跨版本回滚）、关闭遥测及 KV Cache 优化。)
- [tinker-rag-eval-sample.sh — Local Document Q&A (RAG) Downloadable Sample Corpus & Citation Verifier](https://tinker-ai.com/artifacts/local-rag-sample-pack/) ([中文: 本地文档问答 (RAG) 可下载测试样例生成与引用核验脚本](https://tinker-ai.com/zh/artifacts/local-rag-sample-pack/)): Generates a ground-truth technical document sample with trap clauses and tests whether your local RAG setup cites exact sections or hallucinates. (一键生成包含精确条款编号、参数表与干扰项的本地测试文档样例，并自动执行引用溯源核验与上下文截断测试。)
- [n8n-local-rss-classifier.json — n8n + Local Ollama RSS Classifier & Deduplication Workflow](https://tinker-ai.com/artifacts/n8n-local-rss-workflow/) ([中文: n8n + 本地模型 RSS 自动分类、去重与异常重试工作流](https://tinker-ai.com/zh/artifacts/n8n-local-rss-workflow/)): Importable n8n workflow JSON with URL hash deduplication, local Ollama HTTP node, JSON parse guard, and dead-letter retry branch. (可直接导入 n8n 的工作流 JSON：内置链接去重、本地 Ollama 调用节点、JSON 格式校验及超时重试分支。)
- [n8n-webhook-triage-guard.json — Webhook Ticket Triage Workflow with Schema Guard & Dead-Letter Fallback](https://tinker-ai.com/artifacts/n8n-webhook-triage-workflow/) ([中文: Webhook 工单分流工作流（含 Schema 校验与死信降级分支）](https://tinker-ai.com/zh/artifacts/n8n-webhook-triage-workflow/)): Validates incoming payload, enforces strict JSON output, caps token cost, and routes malformed responses to a safe fallback queue. (校验入站请求负载、强制结构化 JSON 输出、限制单次 Token 成本，并将异常输出路由至安全降级队列。)
- [pi5-offline-ai-verify.sh — Raspberry Pi 5 Offline AI Box — Hardware & Thermal Verification Script](https://tinker-ai.com/artifacts/pi5-voice-vision-bom/) ([中文: 树莓派 5 离线语音/视觉 AI 实验箱 — 硬件、供电与温控自检脚本](https://tinker-ai.com/zh/artifacts/pi5-voice-vision-bom/)): Verifies Pi 5 27W USB-C PD power negotiation (throttling check), USB audio/camera enumeration, and local model runtime. (在运行本地模型前，一键检测树莓派 5 27W PD 供电降频状态、CPU 温度、USB 麦克风/摄像头挂载及推理环境。)
- [pi5-sensor-log-guardian.sh — Raspberry Pi 5 Thermal & Journalctl Log Anomaly Explainer Daemon](https://tinker-ai.com/artifacts/pi5-sensor-log-guardian-script/) ([中文: 树莓派 5 传感器阈值过滤与系统日志本地 AI 诊断脚本](https://tinker-ai.com/zh/artifacts/pi5-sensor-log-guardian-script/)): Deterministic threshold guard that checks CPU/sensor temperature first and invokes local Ollama 3B only when an anomaly or systemd error occurs. (采用“确定性阈值过滤 + 按需唤醒本地 3B 模型”混合架构，仅在温度越界或出现系统错误日志时生成中文诊断摘要。)
- [diagnose-container-llm.sh — Container-to-Host Local LLM Connectivity Diagnostic Script](https://tinker-ai.com/artifacts/container-host-diag-script/) ([中文: 容器连不上宿主机本地模型：最小复现与跨平台诊断脚本](https://tinker-ai.com/zh/artifacts/container-host-diag-script/)): Tests host bind address (127.0.0.1 vs 0.0.0.0), host.docker.internal DNS resolution, and bridge firewall rules. (自动排查宿主机 127.0.0.1 监听限制、容器内 host.docker.internal 解析与 Linux 网桥防火墙拦截。)
- [sse-stream-guard.mjs — Resilient LLM SSE Stream Reader (Handles Post-HTTP-200 Mid-Stream Errors)](https://tinker-ai.com/artifacts/sse-stream-guard-client/) ([中文: 抗中断 LLM SSE 流式解析器（捕获 HTTP 200 后的流内报错与截断）](https://tinker-ai.com/zh/artifacts/sse-stream-guard-client/)): Detects mid-stream error frames after HTTP 200 OK, idle socket stalls, and premature finish_reason truncation. (解决“HTTP 状态码为 200 但生成中途断流报错”问题：识别流内 error 事件、空闲超时与长度截断。)
- [config.continue-local-r1.json — Continue / VS Code Dual-Model Local Coding Config (DeepSeek-R1 + Qwen2.5-Coder)](https://tinker-ai.com/artifacts/deepseek-r1-continue-config/) ([中文: Continue / VS Code 本地双模型辅助编程配置（DeepSeek-R1 推理 + Qwen2.5-Coder 补全）](https://tinker-ai.com/zh/artifacts/deepseek-r1-continue-config/)): Separates architectural reasoning (DeepSeek-R1 with 16K context & 4K max_tokens for <think>) from sub-200ms tab autocomplete (Qwen2.5-Coder-7B). (将“深度架构推理（DeepSeek-R1 预留 16K 上下文与思考预算）”与“毫秒级行间代码补全（Qwen2.5-Coder）”解耦的标准配置文件。)
- [benchmark-concurrent-llm.sh — Multi-Worker Concurrency & P95 TTFT Benchmark Script (Ollama vs vLLM)](https://tinker-ai.com/artifacts/vllm-concurrency-bench-script/) ([中文: 本地模型多并发吞吐量与 P95 首字延迟 (TTFT) 压测脚本](https://tinker-ai.com/zh/artifacts/vllm-concurrency-bench-script/)): Fires N parallel OpenAI-compatible requests and measures individual wall-clock latency, queue stall, and aggregate tokens/sec. (并发发起 N 路标准请求，精确统计多用户/多工作流并发场景下的排队等待耗时、P95 响应时间与系统总吞吐量。)
- [n8n-local-invoice-extractor.json — n8n Local Invoice & Contract Structured Extractor with Math Reconciliation](https://tinker-ai.com/artifacts/n8n-invoice-extractor-workflow/) ([中文: n8n + 本地模型发票/合同结构化提取工作流（含金额恒等校验与异常复核分支）](https://tinker-ai.com/zh/artifacts/n8n-invoice-extractor-workflow/)): Extracts vendor, date, line items, and totals into strict JSON, then runs a deterministic JavaScript math check (subtotal + tax == total) before database insert. (通过本地模型提取供应商、明细与税额 JSON，并内置确定性代码节点核验“小计 + 税额 === 总计”，算账不平自动转入人工复核队列。)
- [n8n-local-pr-reviewer.json — n8n + Local Qwen2.5-Coder Git PR Diff Reviewer with Secret Redaction](https://tinker-ai.com/artifacts/n8n-pr-review-workflow/) ([中文: n8n + 本地 Qwen2.5-Coder 代码 PR 自动审查工作流（含密钥正则脱敏）](https://tinker-ai.com/zh/artifacts/n8n-pr-review-workflow/)): Scrubs accidental API keys/tokens from git diffs before sending code hunks to local Qwen2.5-Coder-14B for bug and security review. (在送入大模型前先用正则过滤代码 Diff 中的意外私钥/Token，再调用本地 Qwen2.5-Coder 输出结构化缺陷报告。)
- [pi5-hailo-npu-check.sh — Raspberry Pi 5 AI Kit (Hailo-8L NPU) PCIe Gen3 & Thermal Verification Script](https://tinker-ai.com/artifacts/pi5-hailo-npu-verify/) ([中文: 树莓派 5 AI 套件 (Hailo-8L NPU) PCIe 链路、固件与推理功耗自检脚本](https://tinker-ai.com/zh/artifacts/pi5-hailo-npu-verify/)): Checks PCIe bus enumeration, HailoRT firmware alignment, NPU device temperature, and throttled status before running vision+LLM pipelines. (一键核验树莓派 5 PCIe 通道握手状态、HailoRT 驱动固件版本匹配度、NPU 芯片温度与主板供电状态。)
- [nginx-ollama-lan-auth.conf — Secure LAN Reverse Proxy for Ollama / vLLM (Bearer Auth + CORS + Unbuffered SSE)](https://tinker-ai.com/artifacts/caddy-nginx-ollama-auth-proxy/) ([中文: 本地模型局域网安全反向代理配置（Bearer 鉴权 + CORS 放行 + 零缓冲流式转发）](https://tinker-ai.com/zh/artifacts/caddy-nginx-ollama-auth-proxy/)): Protects local port 11434 behind an Nginx reverse proxy with API key authentication, clean CORS headers for browser extensions, and 300s SSE streaming. (让 Ollama 只监听本机 127.0.0.1，通过 Nginx 对局域网提供带 Bearer 密钥鉴权、CORS 跨域头与零缓冲 SSE 流式转发的安全接口。)
- [ocr-pdf-to-markdown.sh — Local Vision LLM (Qwen2.5-VL / Gemma-3) Scanned PDF & Table Extractor Script](https://tinker-ai.com/artifacts/qwen25vl-ocr-table-extractor/) ([中文: 本地视觉模型 (Qwen2.5-VL / Gemma-3) 扫描件 PDF 与跨页表格提取脚本](https://tinker-ai.com/zh/artifacts/qwen25vl-ocr-table-extractor/)): Renders scanned PDF pages at optimal DPI (150 DPI) and invokes local Qwen2.5-VL to output clean Markdown tables without VRAM overflow. (将扫描版 PDF 按最佳分辨率（150 DPI）光栅化为 PNG，并调用本地视觉大模型输出结构化 Markdown 表格与延迟统计。)
- [n8n-safe-text-to-sql.json — n8n + Local LLM Safe Text-to-SQL Workflow (AST Guard + Read-Only Role + Row Cap)](https://tinker-ai.com/artifacts/n8n-readonly-sql-analyst-workflow/) ([中文: n8n + 本地模型安全自然语言查库工作流（AST 语句拦截 + 只读事务 + 行数熔断）](https://tinker-ai.com/zh/artifacts/n8n-readonly-sql-analyst-workflow/)): Converts natural language questions into PostgreSQL/SQLite queries using local Qwen2.5-Coder, blocks DDL/DML mutations, and enforces LIMIT 100. (利用本地 Qwen2.5-Coder 将业务问题转化为 SQL，内置正则 AST 语句白名单校验（严禁 DELETE/UPDATE/DROP）与强制 LIMIT 熔断。)
- [pi5-voice-assistant-loop.sh — Raspberry Pi 5 Offline Voice Loop Script (Whisper.cpp STT + Ollama + Piper TTS)](https://tinker-ai.com/artifacts/pi5-offline-voice-loop-script/) ([中文: 树莓派 5 纯离线语音对话闭环脚本（Whisper.cpp 识别 + Ollama + Piper 语音合成）](https://tinker-ai.com/zh/artifacts/pi5-offline-voice-loop-script/)): Chains ALSA mic capture, quantized Whisper.cpp STT, local Qwen2.5-3B response, and Piper neural TTS with half-duplex mic mute protection. (串联 ALSA 麦克风录音、Whisper.cpp 量化语音识别、本地 Qwen2.5-3B 对答与 Piper 神经网络语音播报，内置播报期防回声静音锁。)
- [check-rag-embedding-health.sh — Local RAG Embedding Dimension, Bilingual Recall & Zombie VRAM Diagnostic Script](https://tinker-ai.com/artifacts/rag-vector-dim-healthcheck/) ([中文: 本地 RAG 向量维度校验、双语余弦相似度与僵尸显存排查脚本](https://tinker-ai.com/zh/artifacts/rag-vector-dim-healthcheck/)): Probes exact output vector dimensions (e.g. 1024 for bge-m3 vs 768 for nomic-embed-text), checks dual-model VRAM residency, and verifies non-empty embeddings. (实测输出当前 Embedding 模型的精确向量维度（如 1024 vs 768），检查与对话模型共存时的显存驻留，并拦截空向量或维度冲突报错。)
- [tune-strix-halo-llm.sh — AMD Ryzen AI Max+ 395 (Strix Halo 128GB) GTT/TTM Memory & MoE Benchmark Script](https://tinker-ai.com/artifacts/strix-halo-uma-tuner/) ([中文: AMD Strix Halo (128GB 统一内存) 96GB 显存池核验与 Qwen3-MoE 实测脚本](https://tinker-ai.com/zh/artifacts/strix-halo-uma-tuner/)): Verifies Linux kernel amdttm.pages_limit / BIOS 96GB UMA allocation on gfx1151, sets ROCBLAS_USE_HIPBLASLT=1, and benchmarks Qwen3-30B-A3B MoE vs dense models. (一键核查 Linux 内核 amdttm 显存池上限与 gfx1151 驱动状态，开启 ROCBLAS_USE_HIPBLASLT=1 矩阵加速，并实测对比 Qwen3-30B-A3B MoE 与稠密模型吞吐率。)
- [setup-jetson-super-edge-llm.sh — NVIDIA Jetson Orin Nano Super (25W 67 TOPS) Memory Reclaim & Edge VLM Bootstrap](https://tinker-ai.com/artifacts/jetson-super-edge-bootstrap/) ([中文: Jetson Orin Nano Super (25W 67 TOPS) 显存释放、功耗解锁与边缘模型自检脚本](https://tinker-ai.com/zh/artifacts/jetson-super-edge-bootstrap/)): Switches Jetson Orin Nano Super to 25W MAXN_SUPER mode, reclaims ~1.1GB unified RAM from GNOME desktop, drops page cache, and benchmarks Qwen3-4B / Qwen2.5-VL. (验证并切换 JetPack 6.2 25W MAXN_SUPER 模式，支持关闭桌面 GUI 释放 1.1GB 统一内存、清理页缓存，并实测边缘小模型 CUDA 生成速率。)
- [n8n-mcp-local-tool-agent.json — n8n + MCP Client + Local Qwen3 Tool-Calling Agent Workflow (With Loop & Allowlist Guards)](https://tinker-ai.com/artifacts/n8n-mcp-local-agent-workflow/) ([中文: n8n + MCP 协议客户端 + 本地 Qwen3 工具调用工作流（内置工具白名单与防死循环熔断）](https://tinker-ai.com/zh/artifacts/n8n-mcp-local-agent-workflow/)): Importable n8n AI Agent workflow connecting a local Ollama model (Qwen3 /no_think mode) to an MCP Server via SSE, with explicit tool allowlisting and maxIterations=5 guard. (可直接导入的 n8n 智能体工作流：通过 SSE 接入本地 MCP Server，使用 Qwen3 `/no_think` 确定性路由调用只读工具，并强制开启工具白名单与 5 步死循环熔断。)
- [diagnose-uma-and-arc-gpu.sh — Cross-Vendor New Hardware GPU Probe (AMD Strix Halo GTT · Intel Arc B580 · NVIDIA UMA)](https://tinker-ai.com/artifacts/new-hardware-gpu-diag-script/) ([中文: 新一代本地 AI 硬件显存与驱动排障脚本（AMD Strix Halo · Intel Arc B580 · NVIDIA 统一内存）](https://tinker-ai.com/zh/artifacts/new-hardware-gpu-diag-script/)): Audits AMD Strix Halo 15.5GB GTT ceiling & deprecated amdgpu.gttsize flags, Intel Arc B580 /dev/dri render group permissions, and Linux UMA page-cache pressure. (自动检测 AMD Strix Halo 是否被锁死在 15.5GB 默认 GTT 上限、Intel Arc B580 容器 `/dev/dri` 渲染组权限缺失，以及 NVIDIA DGX Spark / Jetson 统一内存页缓存挤压。)
- [tune-mac-studio-ultra-llm.sh — Mac Studio M3 Ultra (256GB/512GB) Metal Wired Ceiling & MLX Frontier MoE Launcher](https://tinker-ai.com/artifacts/mac-studio-m3-ultra-mlx-tuner/) ([中文: Mac Studio M3 Ultra (256GB/512GB) Metal 显存池解锁与 MLX 超大模型启动脚本](https://tinker-ai.com/zh/artifacts/mac-studio-m3-ultra-mlx-tuner/)): Raises macOS iogpu.wired_limit_mb up to 475GB on 512GB Mac Studio M3 Ultra, checks Thunderbolt 5 RDMA/Bridge0 peer status, and launches mlx_lm.server for DeepSeek-R1-671B or Qwen3-235B-A22B. (一键将 macOS Metal 锁定内存上限调至 475GB（适配 512GB M3 Ultra），检查雷雳 5 桥接网络，并启动 mlx_lm.server 运行满血 DeepSeek-R1-671B 或 Qwen3-235B-A22B。)
- [run-blackwell-sm120-llm.sh — NVIDIA Blackwell (RTX 5090 / 5060 Ti `sm_120`) Power-Cap & CUDA 12.8 Launcher](https://tinker-ai.com/artifacts/blackwell-sm120-power-fp4-runner/) ([中文: NVIDIA Blackwell (RTX 5090 / 5060 Ti `sm_120`) 甜点锁功耗与 CUDA 12.8 推理启动脚本](https://tinker-ai.com/zh/artifacts/blackwell-sm120-power-fp4-runner/)): Applies persistent nvidia-smi power limit (e.g. 450W on RTX 5090 or 165W on RTX 5060 Ti 16GB), verifies Compute Capability 12.0 (sm_120), and runs llama-server with MMQ + FlashAttention. (自动锁定 RTX 5090 (450W) 或 RTX 5060 Ti 16GB (165W) 黄金能效比功耗墙，校验 `sm_120` 计算能力，并以原生 MMQ + FlashAttention 启动推理服务。)
- [docker-compose.arc-b60-rdna4.yml — Intel Arc Pro B60 (24GB/48GB) & AMD RX 9070 XT (RDNA 4) Vulkan/ROCm Compose Stack](https://tinker-ai.com/artifacts/arc-b60-rdna4-homelab-compose/) ([中文: Intel Arc Pro B60 (24GB/48GB) 与 AMD RX 9070 XT (RDNA 4) 容器化推理编排栈](https://tinker-ai.com/zh/artifacts/arc-b60-rdna4-homelab-compose/)): Production Docker Compose stack passing /dev/dri with numeric render GID, Vulkan multi-GPU device selection, and Q8_0 KV Cache for Intel Arc Pro B60 24GB and AMD Radeon RX 9070 XT 16GB. (针对 Intel Arc Pro B60 (24GB/双卡 48GB) 与 AMD RX 9070 XT (16GB) 验证的 Docker Compose 模板，内置 `/dev/dri` 数字 GID 映射、Vulkan 设备选择与 Q8_0 KV 显存压缩。)
- [diagnose-sm120-and-gfx1201.sh — Blackwell (`sm_120`) & RDNA 4 (`gfx1201`) Kernel, PCIe Gen5 AER & ROCm Diagnostic Script](https://tinker-ai.com/artifacts/blackwell-rdna4-compute-doctor/) ([中文: Blackwell (`sm_120`) 与 RDNA 4 (`gfx1201`) 算子兼容性、PCIe 5.0 信号与 ROCm 诊断脚本](https://tinker-ai.com/zh/artifacts/blackwell-rdna4-compute-doctor/)): Detects missing sm_120 kernels in PyTorch/CUDA, PCIe 5.0 riser AER / Xid 79 bus drops on RTX 5090/5060 Ti, and invalid HSA_OVERRIDE_GFX_VERSION settings on AMD RX 9070 XT (gfx1201). (一键排查 RTX 5090/5060 Ti (`sm_120`) 报 no kernel image、PCIe 5.0 延长线 AER/Xid 79 掉卡，以及 AMD RX 9070 XT (`gfx1201`) 误设 `HSA_OVERRIDE_GFX_VERSION` 引发的崩溃。)
- [run-dgx-spark-200gbe-vllm.sh — NVIDIA DGX Spark (GB10 128GB / Dual 256GB) 200GbE RoCEv2 & UMA Cache Launcher](https://tinker-ai.com/artifacts/dgx-spark-200gbe-cluster-runner/) ([中文: NVIDIA DGX Spark (GB10 128GB / 双机 256GB) 200GbE RoCEv2 与页缓存清理启动脚本](https://tinker-ai.com/zh/artifacts/dgx-spark-200gbe-cluster-runner/)): Flushes Linux VFS page cache before loading 70B–235B models into GB10 unified LPDDR5X, configures ConnectX-7 MTU 9000 & RoCEv2 GID index, and launches distributed Ray + vLLM. (在向 GB10 128GB 统一内存加载 70B–235B 大模型前自动释放 Linux VFS 页缓存，配置 ConnectX-7 200GbE 网卡 MTU 9000 与 RoCEv2 GID 索引，并启动 Ray + vLLM 服务。)
- [deploy-workstation-96gb-or-r9700.sh — RTX PRO 6000 Blackwell (96GB Max-Q/600W) & Radeon AI PRO R9700 (32GB) Tuner](https://tinker-ai.com/artifacts/workstation-96gb-r9700-vllm-launcher/) ([中文: RTX PRO 6000 Blackwell (96GB Max-Q/600W) 与 Radeon AI PRO R9700 (32GB) 部署脚本](https://tinker-ai.com/zh/artifacts/workstation-96gb-r9700-vllm-launcher/)): Applies 300W Max-Q or 450W sweet-spot power cap on RTX PRO 6000 Blackwell (96GB GDDR7 ECC), or configures ROCm/Vulkan flags for single/dual AMD Radeon AI PRO R9700 (32GB/64GB). (为 NVIDIA RTX PRO 6000 Blackwell (96GB GDDR7 ECC) 配置 300W Max-Q 或 450W 甜点功耗墙并启动 vLLM，或为单/双卡 AMD Radeon AI PRO R9700 (32GB/64GB) 配置 Vulkan/ROCm 环境。)
- [bench-m5-metal4-mlx-llm.sh — Apple M5 Max / M4 Max Metal 4 Neural Accelerator & MLX Prefill Benchmark Runner](https://tinker-ai.com/artifacts/apple-m5-metal4-mlx-bench-runner/) ([中文: Apple M5 Max / M4 Max Metal 4 神经网络加速器与 MLX 长上下文实测脚本](https://tinker-ai.com/zh/artifacts/apple-m5-metal4-mlx-bench-runner/)): Raises macOS Metal wired memory ceiling on 64GB/128GB MacBooks, verifies MLX Metal 4 TensorOps availability, and benchmarks prompt prefill (TTFT) + decode speed with 8-bit KV cache. (在 64GB/128GB MacBook Pro 上解锁 Metal 锁定内存上限，核验 MLX Metal 4 TensorOps 硬件加速，并使用 8-bit KV Cache 实测长上下文 Prefill 与 Decode 速度。)
- [diagnose-p2p-and-rocev2.sh — Multi-GPU PCIe P2P (`iommu=pt` / ACS) & 200GbE RoCEv2 Diagnostic Script](https://tinker-ai.com/artifacts/multi-gpu-p2p-rocev2-doctor/) ([中文: 多卡 PCIe P2P (`iommu=pt` / ACS) 与 200GbE RoCEv2 互联一键体检脚本](https://tinker-ai.com/zh/artifacts/multi-gpu-p2p-rocev2-doctor/)): Audits kernel cmdline for iommu=pt, checks PCIe ACS redirection bits that deadlock multi-GPU vLLM, and verifies ConnectX-7 200GbE MTU 9000 & RoCEv2 GID index 3. (一键检查内核参数是否开启 iommu=pt、扫描导致多卡 vLLM 死锁的的主板 PCIe ACS 重定向位，并核验 ConnectX-7 200GbE 网卡 MTU 9000 与 RoCEv2 GID 索引。)
- [run-epyc-hybrid-moe-llm.sh — AMD EPYC 12-Ch DDR5 + Single GPU Hybrid MoE (`ik_llama.cpp` `-ot` & `KTransformers`) Launcher](https://tinker-ai.com/artifacts/epyc-hybrid-moe-ktransformers-ik-runner/) ([中文: AMD EPYC 12 通道 DDR5 + 单张独显异构 MoE (`ik_llama.cpp` `-ot` 与 `KTransformers`) 启动脚本](https://tinker-ai.com/zh/artifacts/epyc-hybrid-moe-ktransformers-ik-runner/)): Disables Linux auto-NUMA page migration, checks 12-channel DDR5 NUMA topology, and launches 671B/235B MoE models with MLA attention + KV Cache on GPU and routed experts on CPU RAM. (关闭导致跨 Socket 掉速的 Linux 自动 NUMA 平衡，检查 12 通道 DDR5 内存拓扑，并将注意力层与 KV Cache 锁定在独显、稀疏 MoE 专家卸载至 EPYC 内存运行。)
- [launch-blackwell-pro5000-5080-speculative.sh — Blackwell RTX PRO 5000 (72GB) / PRO 4500 (32GB) & RTX 5080 / 5070 Ti Speculative Decoding Runner](https://tinker-ai.com/artifacts/blackwell-pro5000-5080-speculative-runner/) ([中文: Blackwell RTX PRO 5000 (72GB) / PRO 4500 (32GB) 与 RTX 5080 / 5070 Ti 投机解码启动脚本](https://tinker-ai.com/zh/artifacts/blackwell-pro5000-5080-speculative-runner/)): Checks GDDR7 memory bus & ECC status, then launches either single-GPU 72GB vLLM (70B FP8 / 80B MoE) or llama.cpp Speculative Decoding (--model-draft) for 1.6x–2.1x coding speedup. (核验 GDDR7 显存位宽与 ECC 状态，支持一键启动 RTX PRO 5000 72GB 单卡 vLLM (70B FP8 / 80B MoE)，或在 16GB/32GB 卡上开启 llama.cpp 投机解码 (--model-draft) 实现代码生成翻倍提速。)
- [setup-hailo10h-and-rk3588-edge-llm.sh — Jetson AGX Thor (T5000), Pi 5 AI HAT+ 2 (`Hailo-10H` 8GB) & RK3588 NPU Edge LLM Bootstrap](https://tinker-ai.com/artifacts/edge-thor-hailo10h-rk3588-bootstrap/) ([中文: Jetson AGX Thor (T5000)、树莓派 5 AI HAT+ 2 (`Hailo-10H` 8GB) 与 RK3588 边缘大模型部署脚本](https://tinker-ai.com/zh/artifacts/edge-thor-hailo10h-rk3588-bootstrap/)): Detects your edge accelerator (Jetson Thor T5000, Raspberry Pi 5 Hailo-10H 8GB NPU, or Rockchip RK3588 rknn-llm), verifies PCIe Gen 3 / NPU frequency governors, and starts an OpenAI/Ollama-compatible local endpoint. (自动识别边缘算力平台（Jetson AGX Thor T5000、树莓派 5 Hailo-10H 独立 8GB NPU 或瑞芯微 RK3588），核验 PCIe 3.0 链路与 NPU 定频策略，并拉起本地推理服务。)
- [diagnose-hybrid-moe-and-edge-npu.sh — Hybrid CPU+GPU MoE (`NUMA` / `-ot` Regex) & Edge NPU (`Hailo-10H` / `RK3588 CMA`) Doctor](https://tinker-ai.com/artifacts/hybrid-moe-and-edge-npu-doctor/) ([中文: 异构 CPU+GPU 大模型 (`NUMA` / `-ot` 正则) 与边缘 NPU (`Hailo-10H` / `RK3588 CMA`) 排障体检脚本](https://tinker-ai.com/zh/artifacts/hybrid-moe-and-edge-npu-doctor/)): Detects cross-socket NUMA balancing stalls on EPYC/Xeon MoE servers, validates ik_llama.cpp -ot tensor regex syntax, and checks Hailo-10H PCIe Gen 3 & RK3588 CMA contiguous memory. (一键排查 EPYC/至强双路服务器跨 NUMA 掉速、校验 ik_llama.cpp -ot 专家卸载正则，以及检查树莓派 5 Hailo-10H 固件与 RK3588 CMA 连续物理内存池。)
- [bench-unified-memory-ai-pc.sh — RTX Spark / Gorgon Halo Unified-Memory Fixed Benchmark (llama-bench + sustained decode)](https://tinker-ai.com/artifacts/rtx-spark-gorgon-halo-unified-memory-bench/) ([中文: RTX Spark / Gorgon Halo 统一内存固定基准脚本（llama-bench + 持续生成）](https://tinker-ai.com/zh/artifacts/rtx-spark-gorgon-halo-unified-memory-bench/)): Verifies model SHA256, reports platform (ARM64/CUDA or gfx1151 GTT size), runs llama-bench pp2048/tg128 ×3 and a sustained decode run, and writes a Markdown result file so you can replace the guide estimates with measured numbers. Run under Linux, WSL2, or Git Bash on Windows. (校验模型 SHA256，输出平台信息（ARM64/CUDA 或 gfx1151 GTT 容量），执行 llama-bench pp2048/tg128 ×3 和一次持续生成测试，并写入 Markdown 结果文件，方便你用实测替换指南中的估算。可在 Linux、WSL2 或 Windows Git Bash 中运行。)
- [bench-m6-memory-tier.sh — Apple M6 MacBook Pro Memory-Tier Benchmark (wired limit + swap guard)](https://tinker-ai.com/artifacts/apple-m6-memory-tier-bench/) ([中文: Apple M6 MacBook Pro 内存档位基准脚本（锁定上限 + 交换检测）](https://tinker-ai.com/zh/artifacts/apple-m6-memory-tier-bench/)): Records chip and memory tier, optionally raises iogpu.wired_limit_mb, runs a fixed llama-bench with Q8_0 KV cache, and flags the result as invalid if macOS swapped during the run. (记录芯片与内存档位，可选调高 iogpu.wired_limit_mb，以 Q8_0 KV Cache 运行固定参数的 llama-bench，若测试期间 macOS 发生内存交换则将结果标记为无效。)
- [bench-panther-lake-x2-elite-llm.sh — Snapdragon X2 Elite Extreme & Intel Core Ultra X9 388H (Panther Lake) Fixed Laptop LLM Benchmark](https://tinker-ai.com/artifacts/panther-lake-x2-elite-laptop-bench/) ([中文: 骁龙 X2 Elite Extreme 与 Intel 酷睿 Ultra X9 388H（Panther Lake）轻薄本固定基准脚本](https://tinker-ai.com/zh/artifacts/panther-lake-x2-elite-laptop-bench/)): Checks native binary architecture (ARM64 vs x86_64), lists active Vulkan/SYCL iGPU devices (Adreno X2-90 or Arc B390 12 Xe3), runs llama-bench pp2048/tg128 x3 with Q8_0 KV cache, and logs a sustained decode run to catch thin-and-light thermal throttling. (核验原生二进制架构（ARM64 vs x86_64），枚举 Vulkan/SYCL 核显设备（Adreno X2-90 或 Arc B390 12 Xe3），以 Q8_0 KV Cache 运行 llama-bench pp2048/tg128 ×3，并执行持续生成测试以捕捉轻薄本热降频。)
- [launch-llama-cpp-n-cpu-moe.sh — llama.cpp Native --n-cpu-moe Hybrid GPU+DDR5 MoE Launcher (12GB–24GB VRAM)](https://tinker-ai.com/artifacts/llama-cpp-n-cpu-moe-hybrid-launcher/) ([中文: llama.cpp 原生 --n-cpu-moe 消费级独显 + 双通道 DDR5 异构 MoE 启动脚本（12GB–24GB 显存）](https://tinker-ai.com/zh/artifacts/llama-cpp-n-cpu-moe-hybrid-launcher/)): Audits free GPU VRAM via nvidia-smi, keeps 100% of attention layers and Q8_0 KV cache on GPU (-ngl 99 -fa), and offloads exact N MoE expert layers to dual-channel DDR5 RAM via --n-cpu-moe. (通过 nvidia-smi 检测可用显存，将 100% 注意力层与 Q8_0 KV Cache 锁定在独显内（-ngl 99 -fa），并通过原生 --n-cpu-moe 将指定层数的稀疏专家权重卸载至双通道 DDR5 内存。)
- [launch-mac-tb5-rdma-mlx-cluster.sh — macOS Thunderbolt 5 RDMA & MLX JACCL Multi-Mac Cluster Verifier + Launcher](https://tinker-ai.com/artifacts/mac-tb5-rdma-jaccl-cluster-launcher/) ([中文: macOS 雷电 5 RDMA 与 MLX JACCL 多台 Mac 分布式集群自检与启动脚本](https://tinker-ai.com/zh/artifacts/mac-tb5-rdma-jaccl-cluster-launcher/)): Audits rdma_ctl status and ibv_devices (rdma_en*) across Thunderbolt 5 peers, sets iogpu.wired_limit_mb, generates a full-mesh JACCL hostfile via mlx.distributed_config, and launches tensor-parallel MLX inference with MLX_METAL_FAST_SYNCH=1. (核验雷电 5 节点的 rdma_ctl 状态与 ibv_devices (rdma_en*) RDMA 网卡，同步解锁各节点 iogpu.wired_limit_mb 显存上限，通过 mlx.distributed_config 生成全互联 JACCL 拓扑并开启 MLX_METAL_FAST_SYNCH=1 启动分布式推理。)
- [run-sff-70w-pro4000-b50-llm.sh — 70W SFF Workstation GPU (RTX PRO 4000 SFF 24GB / Arc Pro B50 16GB) & Tenstorrent Blackhole HomeLab Runner](https://tinker-ai.com/artifacts/sff-70w-pro4000-b50-homelab-runner/) ([中文: 70W 免供电半高工作站卡（RTX PRO 4000 SFF 24GB / Arc Pro B50 16GB）与 Tenstorrent Blackhole 启动脚本](https://tinker-ai.com/zh/artifacts/sff-70w-pro4000-b50-homelab-runner/)): Audits PCIe slot power draw (70W TBP limit), ECC memory overhead on RTX PRO 4000 Blackwell SFF 24GB, SYCL/Vulkan XMX devices on Intel Arc Pro B50 16GB, or tt-smi mesh status on Tenstorrent Blackhole p150a, then starts llama-server with Q8_0 KV cache. (自检 PCIe 插槽 70W 功耗上限、RTX PRO 4000 SFF 24GB 的 ECC 显存扣减、Intel Arc Pro B50 16GB 的 SYCL/Vulkan XMX 设备以及 Tenstorrent Blackhole p150a 的 tt-smi 状态，随后以 Q8_0 KV Cache 启动推理服务。)

## Interactive Troubleshooter Decision Trees

- [Container Cannot Reach Host Local Model (Ollama / vLLM)](https://tinker-ai.com/tools/troubleshooter/container-host-network/) ([中文: 容器连不上宿主机本地模型（Connection Refused / Timeout）](https://tinker-ai.com/zh/tools/troubleshooter/container-host-network/)): Diagnose loopback binding (127.0.0.1 vs 0.0.0.0), host.docker.internal DNS on Linux, and bridge firewall drops. (一步步排查宿主机 127.0.0.1 回环绑定限制、Linux 容器内 host.docker.internal 缺失及防火墙拦截。)
- [Local Model Slowdown, OOM, or Silent CPU Fallback](https://tinker-ai.com/tools/troubleshooter/vram-oom-cpu-fallback/) ([中文: 本地模型加载失败、显存溢出 (OOM) 或静默回退 CPU 掉速](https://tinker-ai.com/zh/tools/troubleshooter/vram-oom-cpu-fallback/)): Identify why a model whose weight file looks smaller than your VRAM still spills to CPU or crashes under real context lengths. (排查“权重体积明明小于显存，但实际运行却严重掉速或爆显存”的根因。)
- [LLM Streaming API Fails or Drops After HTTP 200 OK](https://tinker-ai.com/tools/troubleshooter/sse-http200-stream-drop/) ([中文: LLM 流式接口在返回 HTTP 200 后中途报错或断流](https://tinker-ai.com/zh/tools/troubleshooter/sse-http200-stream-drop/)): Handle mid-stream SSE error events, reverse-proxy buffering/timeout cuts, and unclosed JSON from max_tokens truncation. (排查“状态码明明是 200 OK，但生成到一半中断、报 JSON 解析错误或卡死”的运行期故障。)
- [n8n / Automation Workflow Local LLM Node Failures](https://tinker-ai.com/tools/troubleshooter/n8n-workflow-llm-failure/) ([中文: n8n 工作流调用本地模型超时、格式漂移与死循环排障](https://tinker-ai.com/zh/tools/troubleshooter/n8n-workflow-llm-failure/)): Resolve batch concurrency queue timeouts, markdown-wrapped JSON failures, and infinite error retry loops. (解决批量条目并发压垮本地显卡、模型返回 ```json 代码块导致解析报错及失败重试死循环。)
- [Browser Extension / LAN Client Gets HTTP 403 (CORS) or 502 Calling Local Ollama](https://tinker-ai.com/tools/troubleshooter/cors-proxy-lan-403/) ([中文: 浏览器插件/局域网前端调用本地模型报 403 CORS 拒绝或反代 502](https://tinker-ai.com/zh/tools/troubleshooter/cors-proxy-lan-403/)): Diagnose OLLAMA_ORIGINS rejections, Host header mismatches behind Nginx/Caddy, and mixed-content HTTPS->HTTP blocks. (排查浏览器插件（Page Assist / 沉浸式翻译）或局域网 Web 前端调用 11434 报 403 Forbidden、CORS 拦截或混合内容阻断。)
- [DeepSeek-R1 / Reasoning Model Breaks JSON Parse or Enters Infinite Loop](https://tinker-ai.com/tools/troubleshooter/reasoning-json-loop-failure/) ([中文: DeepSeek-R1 推理模型 <think> 破坏 JSON 解析或陷入无限复读死循环](https://tinker-ai.com/zh/tools/troubleshooter/reasoning-json-loop-failure/)): Resolve <think>...</think> preamble crashes in downstream JSON parsers, temperature=0 repetition loops, and exhausted token budgets. (解决推理模型输出 <think> 思考块导致 JSON.parse 崩溃、temperature=0 引发无限死循环复读，以及思考过长耗尽 max_tokens。)
- [Raspberry Pi 5 Undervoltage (0x50000), Missing Hailo-8L NPU & USB Audio Failures](https://tinker-ai.com/tools/troubleshooter/pi5-edge-power-pcie-audio/) ([中文: 树莓派 5 欠压降频 (0x50000)、Hailo-8L NPU 未识别与 ALSA 麦克风掉线排障](https://tinker-ai.com/zh/tools/troubleshooter/pi5-edge-power-pcie-audio/)): Diagnose 5V/3A charger USB current limiting, FPC ribbon orientation on M.2 HAT+, and ALSA Device Busy / permission errors. (逐步定位树莓派 5 满载推理时 USB 麦克风断连、lspci 看不到 Hailo-8L 加速卡及 ALSA 录音报 Device or resource busy。)
- [Local RAG Returns Empty Results, Dimension Error (1024 vs 768) or Misses Retrieved Facts](https://tinker-ai.com/tools/troubleshooter/rag-empty-hallucination-citation/) ([中文: 本地知识库 (RAG) 检索报错维度冲突 (1024 vs 768)、召回为空或“检索命中却回答未提及”](https://tinker-ai.com/zh/tools/troubleshooter/rag-empty-hallucination-citation/)): Diagnose vector dimension mismatches after switching embedding models, silent num_ctx=2048 prompt head truncation, and scanned PDF zero-text layers. (一步步排查更换向量模型后的维度不匹配报错、Ollama 默认 2048 上下文静默丢弃头部切片，以及扫描件 PDF 空文本入库。)
- [New Hardware GPU Memory & Driver Trap (AMD Strix Halo 15.5GB GTT / Intel Arc B580 / DGX Spark UMA)](https://tinker-ai.com/tools/troubleshooter/strix-halo-rocm-gtt-trap/) ([中文: 新一代硬件显存识别与加速排障（Strix Halo 15.5GB 显存锁死 / Arc B580 退化 / DGX Spark 页缓存挤压）](https://tinker-ai.com/zh/tools/troubleshooter/strix-halo-rocm-gtt-trap/)): Diagnose why 128GB AMD Strix Halo only sees 15.5GB VRAM on Linux, why Intel Arc B580 falls back to CPU in Docker, and why Unified Memory hosts hit OOM from Linux page cache. (逐步定位 128GB AMD Strix Halo 被限制在 15.5GB 显存、Intel Arc B580 容器静默退化为 CPU，以及 NVIDIA DGX Spark / Jetson 统一内存被 Linux 页缓存挤爆。)
- [n8n + MCP Local LLM Tool-Calling Failure (<think> JSON Breakage, SSE Transport & Infinite Loops)](https://tinker-ai.com/tools/troubleshooter/mcp-tool-calling-schema-loop/) ([中文: n8n + MCP 协议本地模型工具调用排障（<think> 破坏参数、SSE 断连与死循环重试）](https://tinker-ai.com/zh/tools/troubleshooter/mcp-tool-calling-schema-loop/)): Fix local Qwen3 / Ollama models dumping raw text instead of invoking MCP tools, MCP SSE container connection drops, and runaway agent tool loops. (逐步排查本地 Qwen3 / Ollama 模型输出 `<think>` 导致 MCP 工具参数解析失败、容器间 MCP SSE 连接中断，以及智能体无限重复调用同一工具。)
- [RTX 5090 / 5060 Ti (sm_120) "No Kernel Image" & AMD RX 9070 XT (gfx1201) ROCm/PCIe 5.0 Crashes](https://tinker-ai.com/tools/troubleshooter/blackwell-sm120-rdna4-compute-trap/) ([中文: RTX 5090 / 5060 Ti (sm_120) 报无可用内核镜像、PCIe 5.0 掉卡与 RX 9070 XT (gfx1201) 排障](https://tinker-ai.com/zh/tools/troubleshooter/blackwell-sm120-rdna4-compute-trap/)): Resolves CUDA error: no kernel image is available for execution on the device on Blackwell (sm_120), llama.cpp FORCE_CUBLAS 5x speed drop, PCIe Gen5 riser Xid 79 bus drops, and AMD RDNA 4 (gfx1201) ROCm ISA faults. (逐步定位 NVIDIA Blackwell (sm_120) 报 no kernel image is available、llama.cpp 误开 FORCE_CUBLAS 掉速 80%、PCIe 5.0 延长线 Xid 79 掉卡，以及 AMD RX 9070 XT (gfx1201) ROCm 算子崩溃。)
- [Multi-GPU PCIe P2P Deadlocks (`iommu=pt` / ACS) & DGX Spark 200GbE RoCEv2 TCP Fallback](https://tinker-ai.com/tools/troubleshooter/rocev2-nccl-and-multi-gpu-p2p-trap/) ([中文: 多卡 PCIe P2P 死锁卡死 (`iommu=pt` / ACS) 与 DGX Spark 200GbE RoCEv2 退化 TCP 慢速](https://tinker-ai.com/zh/tools/troubleshooter/rocev2-nccl-and-multi-gpu-p2p-trap/)): Diagnose why 2-node DGX Spark ConnectX-7 200GbE clusters silently fall back to 1GbE TCP (`NET/Socket`), why RoCEv2 QP creation fails (`ibv_modify_qp`), or why multi-GPU vLLM (`--tensor-parallel-size 2`) hangs at NCCL/RCCL init on AMD/Intel workstations. (排查双机 DGX Spark ConnectX-7 200GbE 集群静默退化回千兆 TCP (`NET/Socket`)、RoCEv2 建链报 `ibv_modify_qp failed`，以及多卡 vLLM (`TP=2/4`) 在 AMD/Intel 工作站上卡死在 NCCL/RCCL 初始化的根因。)
- [Hybrid CPU+GPU MoE (`NUMA` / `-ot` OOM) & Edge NPU (`Hailo-10H` / `RK3588 CMA`) Traps](https://tinker-ai.com/tools/troubleshooter/hybrid-moe-numa-and-edge-npu-trap/) ([中文: 异构 CPU+GPU 大模型 (`NUMA` 掉速 / `-ot` 爆显存) 与边缘 NPU (`Hailo-10H` / `RK3588 CMA`) 排障](https://tinker-ai.com/zh/tools/troubleshooter/hybrid-moe-numa-and-edge-npu-trap/)): Diagnose 50–65% decode slowdown on EPYC/Xeon servers caused by Linux Auto-NUMA balancing, `ik_llama.cpp` `-ot` regex CUDA OOMs on 671B/235B MoEs, Raspberry Pi 5 AI HAT+ 2 (`Hailo-10H`) firmware mismatches, and RK3588 `rknn-llm` CMA allocation errors. (排查 EPYC/至强服务器跑 KTransformers 与 ik_llama.cpp 时的跨 NUMA 节点掉速、-ot 专家卸载正则不匹配导致 24GB/32GB 独显爆显存、树莓派 5 AI HAT+ 2 (Hailo-10H) 驱动冲突，以及 RK3588 NPU CMA 连续内存分配失败。)