Quality-first local MiniMax H3 video-series generation and loopback API for dual RTX 4090 workstations, with native audio, references, P8/P9 continuity, and preserved artifacts.
-
Updated
Oct 15, 2026 - Python
Quality-first local MiniMax H3 video-series generation and loopback API for dual RTX 4090 workstations, with native audio, references, P8/P9 continuity, and preserved artifacts.
Research preview for reproducible MoE routing and memory oversubscription on consumer GPUs. Not a production inference engine.
DualDeadline adds separate gate/up and down-projection transfer deadlines to exact MoE offloading, with H200 validation and Triton-optimized predictors.
Deploying personal-grade production models on a consumer-grade PC
Flash-backed mixture-of-experts inference on Apple devices with Swift/MLX.
Isaac-lite for weak laptops: YAML scenes, lidar/camera/IMU sensors, synthetic datasets, plus layer streaming for models too big for your GPU
A research toolbox for running and studying LLMs on constrained hardware, with memory, placement, inference, and benchmarking methods.
Local AI on Windows: Qwen 27B GGUF + Qwen-Image via llama.cpp and ComfyUI. 12GB VRAM / 16GB RAM, serial inference, automatic model switching, text-to-image and reference-image editing. 小显存本地大模型与生图工作流。
LatchMoE (GPU):面向受限显存的地址稳定、图重放兼容 MoE 专家缓存卸载框架(CUDA + vLLM PIECEWISE Graph 实现,与昇腾 NPU 主仓对应)
Research artifact for Memory-Sovereign Inference: Output-Exact Execution Beyond Full Residency
Stream PyTorch model blocks from NVMe or pinned CPU memory with a bounded GPU residency budget. LoRA-finetune models far larger than host RAM on one GPU - Qwen3-VL-32B and gpt-oss-120b train in under 6 GB of VRAM.
Small-scale per-layer embeddings (PLE) experiment inspired by Gemma 3n: TinyStories GPT, PLE-table offload benchmarks (GPU/CPU/mmap) and a token explorer.
To associate your repository with the model-offloading topic, visit your repo's landing page and select "manage topics."