1# minimax-h3-image-gb10: the same pinned image as the kromoss-dmd2-gb10 capsule2# (nvcr.io/nvidia/pytorch:25.09-py3 + diffusers 0.41.0), upgraded to diffusers 0.43.0, the first3# release with MiniMax-H3-Image, plus bitsandbytes for the 4-bit text encoder.4#5# Base: nvcr.io/nvidia/pytorch:25.09-py3 (the image gx10's start_moss_image.sh runs on),6# pinned by digest. The pins below are exactly what that script's7# `pip install --upgrade-strategy only-if-needed 'diffusers>=0.38' 'transformers>=5.3' ...`8# resolved to in the working kromoss-dmd2 container (read off `docker diff`).9#10# torchao is removed for the recipe's documented reason: diffusers reaches torchao eagerly11# while importing AutoencoderKLFlux2, and NGC 25.09's torchao 0.13 lacks the FqnToConfig12# newer diffusers expects. Only plain bf16/fp32 weights are loaded here.13FROM nvcr.io/nvidia/pytorch@sha256:f3b4a33fd60d5e0358287bb44d57e91eda45e8f1681fef1bfd6c96937141713014RUN pip install --no-cache-dir \15      diffusers==0.41.0 transformers==5.20.2 tokenizers==0.23.3 accelerate==1.15.0 \16      safetensors==0.8.0 huggingface_hub==1.33.0 hf_xet==1.7.0 pillow==13.0.0 \17      fastapi==0.161.0 starlette==1.9.0 uvicorn==0.54.0 \18 && pip uninstall -y torchao \19 && python3 -c "from diffusers import AutoencoderKLFlux2; import transformers; print('ok', transformers.__version__)"2021# diffusers 0.43.0 (MiniMax-H3-Image modular pipeline) with --no-deps: everything it needs is22# already pinned above. bitsandbytes 0.50.2 has an aarch64 CUDA wheel.23RUN pip install --no-cache-dir --no-deps diffusers==0.43.0 bitsandbytes==0.50.2 \24 && python3 -c "from diffusers import ModularPipeline, MiniMaxH3ImageTransformer2DModel, AutoencoderKLMiniMaxH3Image, MiniMaxH3Scheduler; import bitsandbytes; print('ok')"