Fix torch.OutOfMemoryError: Allocation on device 0 would exceed allowed memory in ComfyUI (FLUX & Wan 2.1)

Root Cause & Memory Allocation Mechanics

Modern flow-matching Diffusion Transformers have three distinct memory phases inside a ComfyUI execution graph: (1) Text Conditioning (`T5-XXL` at `9.8 GB` FP16 or `UMT5-XXL` at `11.4 GB` FP16), (2) Iterative Latent Denoising (`12B–14B DiT` weights plus cross-attention activation maps), and (3) Spatial/Temporal VAE Decoding. On Windows 11, if ComfyUI's smart memory manager miscalculates free VRAM because Desktop Window Manager (DWM) or a web browser is holding 1.2 GB, either Windows silently spills tensors into slow shared system RAM (dropping speed by 10x) or `--cuda-malloc` throws `Allocation on device 0 would exceed allowed memory` during the final VAE upscaling convolution.

Exact CLI Remediation Commands

# Unload any background Ollama LLMs stealing VRAM before running ComfyUI!
ollama stop $(ollama ps | awk 'NR>1 {print $1}')

# Launch ComfyUI with explicit VRAM reserve and PyTorch expandable segments
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
python main.py --reserve-vram 1.5 --use-sage-attention

# For 8GB–12GB GPUs, force lowvram weight streaming:
python main.py --lowvram --reserve-vram 1.0

# In ComfyUI workflow nodes:
# 1. Use 'Load Diffusion Model (weight_dtype: fp8_e4m3fn)' or 'Unet Loader (GGUF)'
# 2. Load 't5xxl_fp8_e4m3fn.safetensors' in DualCLIPLoader
# 3. Replace 'VAE Decode' with 'VAE Decode (Tiled)' (tile_size=256 or 512)

Frequently Asked Questions