intel-arc

Open recipes, engine patches, and benchmark harnesses for LLM inference on Intel Arc Pro B60/B70 (Battlemage, Xe2). MoE 35B at 160 t/s decode / 7.5K t/s prefill single-stream, 27B at 50~ t/s decode / 1.7K t/s prefill single stream. vLLM XPU MTP unlocked. Muse Glimmer recipe added!!

Makes Intel Arc Pro B70 GPUs actually fast on Ubuntu Server. 11 llama.cpp cherry-picks that fix the big B70 bugs (MoE slot-init SEGV, Q8_0 reorder crash, OOM reorder, missing BF16 GET_ROWS, wrong Xe2 warptile, slow K-quant DMMV, etc.) + Mesa 26 + runtime env workarounds + SYCL/Vulkan backend-selection rules. 2-7x speedup on 4x B70, bench-verified.

PosterChan AI is a Nostr-powered personal cloud and a self-hosted AI powerhouse, running on one box you own.

Linux (Arch) optimization guide for Samsung Galaxy Book6 Pro (NP960XJG) — Panther Lake, Arc B390, fingerprint, BIOS analysis, idle power 7.5W→1.9W

Plug-and-play llama.cpp runtime for Intel Arc GPUs. Auto-detects your card, picks safe SYCL defaults, and exposes an OpenAI-compatible API.

videogen-b70: ComfyUI + PyTorch XPU recipes for MiniMax H3 T2V on Intel Arc Pro B70. Auto-detects one or two cards (CLIP on GPU1, UNet on GPU0). Not tensor parallel.

Per-precision XMX (matrix engine) profiling for Intel Arc GPUs via Level Zero metric streamers — observes any workload without wrapping it

Reproducible Qwen3.6-27B FP8 MTP2 runtime and benchmark kit for dual Intel BMG GPUs

Template for starting PyTorch projects with Intel GPUs (XPU) on Windows.

Layer-by-layer model training and modification for 80B+ MoE models on consumer GPUs. Abliteration, LongRoPE, LoRA merge, weight visualization. Built because nothing else could do it. https://justcalljon.pro

ComfyUI SeedVR2 video/image upscaler fork with experimental Intel XPU / PyTorch XPU support for Arc and Core Ultra iGPU

Fast local voice cloning with Qwen3-TTS on affordable Intel Arc GPUs using IPEX. No CUDA. No cloud. Zero cost.

Preregistered local-inference experiments on consumer GPUs: admission control, cache restoration, corrected results, and reproducible evidence.

A job queue and browser UI for iw3 (nunif 2D to stereo-3D video conversion), with live progress, measured ETAs and a 2-minute preview. Optional Cove extension included.

HQH-XPU: Intel Arc A770 ComfyUI optimization bundle — TINT4 (oneDNN hybrid), OmniXPU, XPU-CacheClean, ClipProj (XPU) & UniversalIO; one install auto-deploys all nodes.

ComfyUI custom node for llama-cpp-python with Intel GPU SYCL acceleration. Adapted for 0.3.39+ MTMD rewrite. Pair with allanmeng/llama-cpp-python-sycl-windows for SYCL whl builds.

Boltz-2 on Intel Arc — NVIDIA-free protein folding + binding affinity (AlphaFold3-class) via torch.xpu.

GUI manager for llama.cpp (SYCL) with OpenAI-compatible API and automatic model switching.

Serial SYCL kernel stack for Laguna-XS on Intel Arc Pro B70 — quality-gated speed