Mage-Flow-Turbo-CPU

Mage-Flow-Turbo-Native-Inference is a portable native inference and deployment stack for Mage-Flow-Turbo. The repository does not train or modify model weights. Python provides configuration, model/runtime identity verification, CLI/REST orchestration, lifecycle control, telemetry and evidence collection; the actual model execution path is the native stable-diffusion.cpp sd-cli runtime. v1.0.0 supports a canonical Q8 GGUF profile and a BF16 SafeTensors profile, both through the same native runtime. The project does not provide a Hugging Face Transformers inference backend and does not run a PyTorch/Transformers inference loop.

The project freezes model and runtime identities. Exact SHA-256 identities are verified before real inference. The Git repository contains no model weights. The canonical Q8 profile uses GGUF Q8_0 diffusion, Qwen3-VL-4B GGUF Q4_K_M text encoder, SafeTensors VAE, and pinned stable-diffusion.cpp sd-cli. The BF16 SafeTensors profile uses BF16 SafeTensors diffusion, the same Qwen GGUF text encoder, SafeTensors VAE, and the same pinned native runtime.

Retained strict 2×2 v1.0.0 measurements cover Kaggle CPU and Kaggle T4/T4x2 cuda0 for both profiles. CPU policy uses Kaggle Accelerator=None, backend exactly cpu, prebuilt CPU sd-cli only, no CUDA fallback, and host-memory and process-RSS telemetry. T4/T4x2 policy requires a T4 or T4x2 host, uses only physical GPU slot 0, sets CUDA_DEVICE_ORDER=PCI_BUS_ID and CUDA_VISIBLE_DEVICES=0, and uses effective inference backend cuda0 with no cuda1, multi-GPU split, auto-fit, or CPU inference fallback. The reference service binds to 127.0.0.1 by default and exposes REST endpoints including GET /healthz, GET /readyz, GET /v1/info, POST /v1/images/generate, and GET /v1/artifacts/<png>. Kaggle integration uses prebuilt CPU/CUDA runtimes attached as Kaggle Datasets and Mage-Flow, Qwen, and VAE inputs as Kaggle Models.