Instructions to use elef4nt/sdmlx-acceleration-patches with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use elef4nt/sdmlx-acceleration-patches with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download elef4nt/sdmlx-acceleration-patches --local-dir sdmlx-acceleration-patches
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
SDMLX Acceleration Patches
SDXL/FLUX.1 acceleration patches and exact Qwen Image2.1 adapters for SDMLX / ComfyUI on Apple Silicon.
The SDXL/FLUX.1 packages described below are not fully expanded UNet deltas. They contain SDMLX-mapped low-rank LoRA factors:
- target keys already match the SDMLX MLX UNet
- linear factors use
[out, rank]and[rank, in] - Conv2D down factors use MLX layout
[rank, kernel_h, kernel_w, in] - runtime scale is
strength * alpha / rank
Each .sdmlxpatch package contains:
patch.safetensorsmanifest.json- optional
source_metadata.json
Packages
SDXL
| Package | Recommended Steps | CFG | Sampler | Scheduler | Force No CFG |
|---|---|---|---|---|---|
dmd2_sdxl_4step_lora_fp16.sdmlxpatch |
4 | 1.0 | lcm |
normal |
true |
dmd2-lighting8step_cfg1.5.sdmlxpatch |
8 | 1.5 | lcm |
normal |
false |
sdxl_lightning_4step_lora.sdmlxpatch |
4 | 1.0 | euler |
sgm_uniform |
true |
sdxl_lightning_8step_lora.sdmlxpatch |
8 | 1.0 | lcm |
normal |
true |
lcm-lora-sdxl.sdmlxpatch |
8 | 1.0 | lcm |
normal |
true |
Hyper-SDXL-8steps-CFG-lora.sdmlxpatch |
8 | 5.0 | lcm |
normal |
false |
Hyper-SDXL-12steps-CFG-lora.sdmlxpatch |
12 | 7.5 | lcm |
normal |
false |
FLUX.1
| Package | Target | Recommended Steps | Notes |
|---|---|---|---|
Hyper-Flux.1-Dev-4-step-Lora.sdmlxpatch |
FLUX.1 Dev | 4 | Optional SDMLX FLUX.1 acceleration patch. |
The recommendations are metadata only. SDMLX should keep sampler settings explicit in the UI.
Source Licenses
These packages are transformed LoRA factors. They are provided with source attribution and remain subject to their upstream licenses:
dmd2_*: sourcetianweiy/DMD2, Hugging Face metadata listscc-by-nc-4.0; the model card text names Creative Commons Attribution-NonCommercial-ShareAlike 4.0.sdxl_lightning_*: sourceByteDance/SDXL-Lightning, licenseopenrail++.lcm-lora-sdxl: sourcelatent-consistency/lcm-lora-sdxl, licenseopenrail++.Hyper-SDXL-*: sourceByteDance/Hyper-SD; SD-related Hyper-SD models are covered by the CreativeML Open RAIL++-M license in the upstreamLICENSE.md.Hyper-Flux.1-Dev-4-step-Lora: sourceLingyuzhou/Hyper_Flux.1_Dev_4_step_Lora. No substantive upstream license grant is documented: Hugging Face metadata sayslicense: 123and the includedLICENSEfile is empty (checked 2026-08-04). Do not assume redistribution or commercial-use rights; verify the upstream terms or obtain permission. Use is also subject to applicable FLUX.1 Dev terms.
Do not publish or redistribute packages whose upstream license does not fit your use case. In particular, DMD2 is non-commercial/share-alike. OpenRAIL++ and OpenRAIL++-M packages include use-based restrictions that downstream users must follow.
The SDMLX code license does not relicense these packages. For the repository-level notice, see THIRD_PARTY_NOTICES.md in the SDMLX code repository.
Hugging Face Upload
Use Git LFS for package tensors:
git lfs track "*.safetensors"
git lfs track "*.sdmlxpatch/patch.safetensors"
git add .gitattributes README.md *.sdmlxpatch
git commit -m "Add SDMLX acceleration patches"
git push
Suggested repo type: model.
Qwen Image2.1 β 0.2.0 adapter distribution
Built with Qwen. These adapters use the separate sdmlx-qwen21-speed-patch-v1
format. Their tensor bytes are unchanged from the accepted existing Suite assets;
they are not SDXL UNet factors and cannot be used on another model family.
| Package | Fixed profile | Other profile |
|---|---|---|
Qwen-Image-2.1-viggle-v0.3-r256.sdmlxpatch |
Viggle6step / mixed9step | standard |
Qwen-Image-2.1-pruna-v0.1-8step.sdmlxpatch |
Pruna8step | standard |
Qwen-Image-2.1-qwen-turbo-r128.sdmlxpatch |
Extracted Turbo r1288step | standard |
Fixed profiles: CFG1, empty negative prompt, strength1. The SDMLX node keeps the correct per-profile curve explicit; the loader never infers sampling from names. An r128 extraction is distinct from the full official Turbo checkpoint.
All three remain subject to the Qwen Research License Agreement, including
its non-commercial-use limitation. Each folder includes the complete upstream
LICENSE, NOTICE, sources and SHA256. See individual README/source_metadata files.
This does not change licenses of the pre-existing SDXL/FLUX.1 packages.
The legacy r128 manifest field local_only is retained for compatibility;
SDMLX0.2.0 uses a separate pinned distribution registry to retrieve the exact bytes.
Quantized