Sleeping Robots

A complete guide to running Qwen 3.8 Flash Next (125B MoE, 6B active) on the AMD Strix Halo 128 GB platform.

Following up on the hipCUB fix post, I built and benchmarked EngramHalo.cpp, a llama.cpp fork with RDNA 3.5 kernel patches and MTP speculative decoding.

Qwen 3.8 27B hits 18 t/s via MTP speculative decoding on Strix Halo, up from 9.6 t/s baseline.

Hy3's 299B MoE runs at 10 tok/s on 128 GB Strix Halo in baseline mode, but the MTP speculative head that gives CUDA users a +27-37% boost can't allocate on Vulkan unified memory.

Colibri streams GLM-5.2's 19,456 experts from disk to run a 744-billion-parameter MoE model on consumer hardware.

RedNote's 2B continuous autoregressive TTS hits RTF 0.35 on the NVIDIA 4070 Super with the MeanFlow-distilled checkpoint, putting it among the fastest voice-cloning-capable models I've tested locally.

BeeLlama.cpp's DFlash speculative decoding nearly triples dense model throughput on AMD Strix Halo, but in a strict head-to-head against my existing MTP setups, MTP still wins by 23-67% depending on the configuration.

Benchmarking OpenMOSS's 8B llama.cpp GGUF backend and 100M ONNX Nano model on AMD's Ryzen AI MAX+ 395, with thread scaling analysis and a surprising Nano result that beats everything else I've tested.

The AI discourse is stuck between two poles that refuse to talk to each other.

A community PR optimizing CUDA kernels for GFX1151 delivers +24% prefill throughput on MoE models, but combining those same kernel changes with MTP speculative decoding makes inference slower.

AMD released ROCm 7.13 with Strix Halo optimizations.

Getting Resemble AI's expressive TTS model running on AMD Strix Halo with no NVIDIA hardware.

Getting Unsloth Studio's full training pipeline running on AMD Strix Halo (gfx1151) using pip-packaged ROCm nightlies, no /opt/rocm required.

Google's official Gemma 4 MTP assistant heads bring speculative decoding to MoE models that couldn't benefit before, and nearly quadruple dense model throughput on AMD Strix Halo's bandwidth-limited unified memory.

Eight optimization attempts on Echo-TTS CPU inference, the five that worked, quality evaluation with voice cloning, and how the optimized CPU path ended up faster than the GPU hybrid.

Multi-Token Prediction turns Qwen 3.6 27B from 6 t/s to 30 t/s on AMD Strix Halo, succeeding where draft models and ngram decoding failed, by using prediction heads baked into the model itself.

Running a diffusion-based TTS model on AMD's Strix Halo, patching CUDA-only code for CPU, discovering a bf16 GPU hang on gfx1151, and a hybrid GPU/CPU trick that beats every other TTS model I've tested.

Porting Tencent's CUDA-only 3D world model to AMD's Radeon 8060S via ROCm Docker, flash-attention CK kernels, a fully compiled gsplat with wave32 patches, and complete 3D reconstruction output including Gaussian splats.

Ollama gained traction by being the first easy llama.cpp wrapper, then spent years dodging attribution, misleading users, and pivoting to cloud, all while riding VC money earned on someone else's engine.

Benchmarking speculative decoding with Gemma 4 E2B as a draft model for Gemma 4 31B on AMD Strix Halo, a bandwidth-bound setup where the optimal draft-max differs from discrete GPUs.

A full-stack web interface that puts the Pi coding agent in the browser, with system-level access, session history, and model switching through a local LiteLLM proxy.

How LiteLLM, llama-swap, and Lemonade Server compose into a unified local inference platform, routing dozens of models across GPU and NPU through a single API endpoint, accessible anywhere via Tailscale and a local reverse proxy.

Running a 2B parameter tokenizer-free TTS model in both Python and C++ on AMD's integrated GPU, near-real-time speech synthesis on CPU, and the Vulkan crash that stopped GPU acceleration in its tracks.

Running a 600+ language zero-shot TTS model on an AMD integrated GPU, voice cloning benchmarks, ROCm compatibility adventures, and the container workaround that actually worked.

Setting up AMD's Lemonade Server on Strix Halo to run LLM and Whisper inference on the XDNA 2 NPU, driver builds, architecture decisions, and benchmarks against the integrated GPU.

A Telegram bot backed by the Pi coding agent with autonomous scheduling, persistent memory, cross-session search, and a web dashboard.

Running Fish Audio's 4B parameter S2-Pro text-to-speech model locally on an AMD Strix Halo integrated GPU via ROCm and Podman.

A browser-based AI music generation tool powered by ACE-Step, ported to Linux for local generation on AMD Strix Halo hardware.

A full-stack quiz platform that turns markdown files and YouTube transcripts into mixed-format quizzes with AI grading, contextual chat, and performance analytics.

A modular collection of services for building a personal AI agent, tool use, memory, browser automation, TTS, and multi-platform chat interfaces.

Snap a photo of a handwritten list, OCR it with a local vision model, and print a formatted checklist on a thermal receipt printer.

A Dockerfile and docker-compose setup for running llama.cpp with its Python bindings in a container, because finding a working one shouldn't be this hard.

An all-in-one Docker container that bundles LLMs, embeddings, vision, and TTS into a single unified inference server.

Generating weekly Spotify Wrapped-style reports from browser history using local models and Browser Recall data.

Building a personal browser history search engine with full-text recall, inspired by Microsoft's Recall and rewind.ai.

Building a custom speech-to-text solution for Linux Wayland users using NVIDIA's Canary model, Silero VAD, and ydotool.

Building a natural language voice controller for the Linux desktop using Qt6, a tiny 1.7B LLM, and a clever vector embedding trick for tool calling.

Building a private, browser-based voice assistant using WebAssembly, Moonshine STT, Piper TTS, and local LLMs.

A text-driven simulation where you interact with a guy named John through natural language, and an LLM determines how his world responds.