MPS Float8 support issue

You’re running **Wan2.2 T2V** models in **ComfyUI on a Mac**, and the error: > **Trying to convert Float8_e4m3fn to the MPS backend but it does not have support for that dtype.** …means exactly what it says: **Apple MPS does NOT support Float8 (e4m3fn) tensor types.** Wan2.2 text-to-video *requires* Float8 support → which currently only works on **CUDA GPUs** (NVIDIA). So on **Mac**, you cannot load or run these Float8 quantized models directly. This is why the pipeline fails at the **KSamplerAdvanced** step. --- # ✅ Why This Happens MPS supports: - float16 (fp16) - float32 (fp32) - some int8 - some float32 composite ops But **MPS does NOT support float8** (`e4m3fn` or `e5m2`). Wan2.2 fp8_scaled models include: - `wan2.2_t2v_high-noise_q4_k_m` - `wan2.2_t2v_low-noise_q4_k_m` - `umt5_xxl_fp8_e4m3fn_scaled` - etc. All of these assume: - CUDA backend - float8 kernels - torch 2.4+ with NVIDIA support --- # ❌ Not fixable with a setting There is **no ComfyUI setting** or **patch** that makes float8 work on MPS. The limitation is at the **hardware + PyTorch backend** level. --- # ✅ Options You *Can* Do ## **1. Use the CPU backend (very slow, experimental)** You *can* force CPU, but it will be **extremely slow** and may still fail on float8 ops. In ComfyUI → **Settings → System** → ✔ enable **Force CPU** But T2V models are too large for practical CPU use. So: **not recommended**. --- ## **2. Run Wan2.2 on a cloud GPU / remote machine** If you want Wan2.2, you must run on **NVIDIA GPU**: Cheap options: - RunPod (cheapest) - Vast.ai - Lambda GPU Cloud - Paperspace - Colab Pro+ This works perfectly because CUDA supports float8 and q4_k_m quantization. --- ## **3. Switch to models that support MPS** For Mac T2V / T2I compatible with MPS: ### Works on Mac (MPS) - Stable Diffusion 1.5 - Stable Diffusion XL - SD3.5-turbo (some builds) - LTX-Video - Mochi Video - ZeroScope v2 - LaVie (some configs) - HunyuanVideo (fp16 only) ### Does **NOT** work on Mac - Wan2.2 (requires float8) - UMT5 (float8) - Many q4_k_m quantized diffusion models --- # 🧩 The *Real* Fix for Your Workflow If you want to continue on Mac: ### ✔ Replace all Float8 / fp8 models with **fp16** versions You need fp16 versions of: - Wan2.2 (currently not released publicly in fp16) - UMT5 encoder (might have fp16, check) But **Wan2.2 fp16 T2V is not released**. Only the float8/q4 CUDA-optimized version is available. So: > **Wan2.2 T2V cannot be run on Apple Silicon today.** --- # 🚀 If you want, I can - Convert your current ComfyUI workflow into a **Mac-compatible version** - Build a **replacement T2V pipeline** (LTX Video or Mochi) for MPS - Create a **GPU deployment checklist** for running Wan2.2 on RunPod/Vast - Rewrite your existing workflow so it loads models conditionally depending on device (MPS vs CUDA) Just tell me which direction you want.