SECourses Musubi Tuner - 1-Click to Install App for LoRA Training and Full Fine Tuning Qwen Image, Qwen Image Edit, Wan 2.1 and Wan 2.2, FLUX Klein, FLUX 2, Z Image Base and Turbo Models with Musubi Tuner with Ready Presets

App Installer Download Link : https://www.patreon.com/SECourses/posts/secourses-musubi-137551634

• Latest Zip File : SECourses_Musubi_Trainer_v28.zip

• Main training tutorial to learn training and how to use this app (mandatory to watch) : https://youtu.be/DPX3eBTuO_Y

• Wan 2.2 training tutorial : https://youtu.be/ocEkhAsPOs4

• SwarmUI : https://www.patreon.com/posts/114517862

• ComfyUI : https://www.patreon.com/posts/105023709

• Currently 1-click to install on Windows, RunPod and Massed Compute with uv installation - ultra fast

• We have model auto downloader that supports below models (run Windows_Download_Training_Model_Files.bat)

• Qwen Image Training Models (Old and new 2512) with Torch Compile

  • Qwen Image Edit Plus (2509 and 2511) Training Models
  • Wan 2.1 Text to Video Training Models with Torch Compile
  • Wan 2.2 Text to Video Training Models with Torch Compile
  • Wan 2.2 Image to Video Training Models with Torch Compile
  • Z-Image Turbo LoRA training with Torch Compile
  • Z-Image Base Fine Tuning with Torch Compile
  • Z-Image Base LoRA Training with Torch Compile
  • FLUX 2 Dev LoRA Training with Torch Compile
  • FLUX Klein 9B and 4B LoRA Training with Torch Compile

• Please use model downloader to not have any issues because your selected models may be wrong
  - Moreover, the Musubi Tuner automatically does FP8 and FP8 scaled conversion while loading BF16 model into RAM so we always use BF16 models for training
  - I specifically developed this model downaloder with like UGET method so that it is both ultra fast and ultra robust - not using Hugging Face downloader anymore
  - Moreover the downloader script verifies SHA 256 hash of the models and prevents any possibly corrupted model downloads

• This is an interface app that is based on famous Kohya Musubi Tuner

  • It has all the features of Kohya Musubi Tuner + extra features with huge amount of details and very easy to use
  • Musubi Tuner Official repo : https://github.com/kohya-ss/musubi-tuner

• The installer will install with Torch 2.8 and CUDA 12.9 and pre-compiled xFormers, Triton, Flash Attention and Sage Attention libraries

  • We literally support all of the GPUs like RTX 3000, 4000, 5000 series or, A40, L40, A100, H100, B200 etc
  • Flash Attention and Sage Attention may not work on RTX 2000 or 1000 series but I have plan for them hopefully soon with newest libraries compiling myself

• Currently it fully supports Qwen Image Models training, Wan 2.2 models training and Wan 2.1 models training

  • Also has automatic Qwen VL captioning

Torch Compile Toolchain

When the Inductor backend is selected through Enable torch.compile or Accelerate Dynamo, the trainer prepares and validates the native compiler environment before loading the model. Non-codegen backends such as eager remain available without a host compiler.

• Windows discovery covers existing developer shells, Visual Studio and Build Tools installations, vswhere, registry and custom-drive installs, legacy layouts, and every installed x64 MSVC toolset. A real C++17/OpenMP build and an nvcc host-compiler probe must pass.
• Linux discovery covers configured, versioned, Conda, GCC, Clang, NVIDIA HPC, Intel, and ARM C++ compilers. Candidates must compile and link C++17/OpenMP code and, when available, pass an nvcc -ccbin probe.
• The RunPod and Massed Compute installers add build-essential and ninja-build when the host does not already provide them.
• CUDA selection follows torch.version.cuda, Ninja is discovered automatically, and stable Inductor/Triton caches are created under .cache/torch_compile.
• Successful native toolchain probes are fingerprint-cached inside the current installation and inherited by training workers, avoiding repeated temporary probe.exe runs until the compiler, CUDA, SDK, Python, PyTorch, or requirements change. A fresh installation starts with a fresh validation cache.
• If toolchain validation or the initial torch.compile attempt fails, training logs the reason and continues in eager mode by default.
• Direct GUI runs, worker subprocesses, direct training CLI runs, and Accelerate Dynamo all use the same backend setup.
• If the first real compiled block fails for a graph-specific reason, training transparently retries in eager mode while preserving compiled state-dict compatibility. Set MUSUBI_TORCH_COMPILE_FALLBACK=0 to make that failure strict.

Advanced validation switches are MUSUBI_TORCH_COMPILE_REQUIRE_CUDA, MUSUBI_TORCH_COMPILE_REQUIRE_NINJA, and MUSUBI_TORCH_COMPILE_REQUIRE_OPENMP (enabled by default). Set a switch to 0 only when intentionally relaxing that check.

Run the full local diagnostic on Windows:

.\venv\Scripts\python.exe -m musubi_tuner.torch_compile_toolchain --require-openmp --smoke-test

On Linux:

./venv/bin/python -m musubi_tuner.torch_compile_toolchain --require-openmp --smoke-test

Windows Attention Performance

Every supported trainer keeps the SDPA setting portable across operating systems. PyTorch's native fused FlashAttention SDPA remains the first choice. When that backend is unavailable, the trainer can use the installed external FlashAttention implementation only after a CUDA forward/backward training probe succeeds. Import, GPU capability, probe, or runtime tensor incompatibilities retain the working PyTorch SDPA path.

The Training Settings panel in every training tab provides Use Legacy PyTorch SDPA beside Split Attention. It is off by default, so the guarded automatic strategy is used. Enable it to force PyTorch's older, usually slower SDPA path for compatibility testing. The selection is saved with GUI presets and emitted as --use_legacy_sdpa in runtime configs and generated workflow scripts.

Set MUSUBI_DISABLE_EXTERNAL_FLASH_SDPA=1 before launching a trainer to disable the compatibility path for debugging or strict backend comparisons. This environment switch and the GUI checkbox never affect explicit FlashAttention, SageAttention, or xFormers selections.

Updates

- Click images to see their full sizes

31 July 2026 Krea 2 ConvRot INT8 Training

• Integrated upstream musubi-tuner pull request #1008: ConvRot INT8 base-weight quantization for Krea 2 LoRA training (arXiv:2512.03673).
• The same ConvRot scheme the Model Quantizer already uses for inference checkpoints is now available during training: a block-diagonal Hadamard rotation (group size 256) smooths activation outliers, then the frozen DiT base weights are quantized per-channel to INT8 and the forward runs a fused Triton INT8 GEMM.
• New ConvRot INT8 checkbox and ConvRot INT8 Backward dropdown (bf16 / int8) in the Krea 2 LoRA tab's Model Settings. It is an alternative to FP8 Base + Scaled FP8, not an addition — the GUI rejects the combination.
• Same weight VRAM as FP8 (1 byte per base weight), but the Linear forward is roughly 2.5x faster than BF16 and 2.7x faster than Scaled FP8, and measured output error against the BF16 reference is lower than block-wise Scaled FP8 (~1.4e-2 vs ~2.5e-2).
• The biggest win is on GPUs with no FP8 hardware (RTX 30 series and older), where the FP8 path has to dequantize to BF16 on every forward pass.
• triton (Linux) / triton-windows (Windows) are already in this application's requirements, so the fused kernels work out of the box on both platforms. Without triton it still runs, via a slower dequantized BF16 fallback.
• Composes with block swap, gradient checkpointing, and torch.compile. Not supported with Krea 2 Turbo sample generation or full fine-tuning.

13 July 2026 Image Full-DiT Fine-Tuning

• Integrated upstream musubi-tuner pull request #997 on top of the maintained performance and torch.compile changes.
• Extended the guarded Windows SDPA acceleration and legacy compatibility control to every supported model trainer and GUI training tab.
• Added GUI training-mode selection for FLUX.2 dev, FLUX.2 Klein 4B/9B, Ideogram 4, and Krea 2.
• Full-DiT runtime configs automatically remove LoRA-only and FP8-base options and enforce supported precision, optimizer, distributed, and block-swap combinations.
• Added ready demo presets for FLUX.2, FLUX.2 Klein, Krea 2, Ideogram 4, Z-Image, and supported Wan 2.1 LoRA fine-tuning in Demo_Training_Configs_FLUX-2_Z-Image_FLUX-Klein_WAN-21_Krea2_Ideogram4.
• Ideogram 4 full fine-tuning requires a plain FP32/FP16/BF16 conditional DiT; the official FP8/NVFP4 model remains LoRA-only.

11 July 2026 Quantizer Preset Review

• INT8 ConvRot with fixed group size 256 is now the balanced default for validated model families.
• Current ComfyUI native FP8, tensor/row INT8, ConvRot, MXFP8, and NVFP4 paths are distinguished from QuantOps-only INT8 blockwise.
• Qwen Image, Z-Image, Anima, Boogu, and models without public ConvRot validation keep safer model-specific defaults.
• Model filters, Qwen3.5/Gemma 4 protections, FLUX.1/Klein, ERNIE Image, LTX aliases, and re-quantization safety were updated.
• See the model-by-model preset reference for the current decisions.

25 June 2026 V28.6 Update

• Quantize model backend code fully updated
• Quantize presets fully updated and vastly improved
• Queue system added so set your settings and click start conversion it will auto queue
• Following presets fully added and default quality set to Normal - pretty good quality

• Full new screenshot of the quantize page as below
• I am currently testing FP8 Krea 2 Tensor vs Blockwise vs INT8 Blockwise
• Just run Windows_Install_and_Update.bat

---

---

---

---

---

---

---

Stage 1 Training Published Config Results Examples

30+ examples shared here : https://medium.com/@furkangozukara/qwen-image-lora-trainings-stage-1-results-and-pre-made-configs-published-as-low-as-training-with-ba0d41d76a05

---

---

---

---

---

App Screenshots

---

---

---

---

---

---

---

---

---

---

---

---

---