CUDA error: out of memory - (Minimax only)

A user reports that the MiniMax H3 video generation model fails with a CUDA out-of-memory error while Wan 2.2 runs normally, citing insufficient GPU memory during VAE encoding. The error occurs in the MiniMax H3 pipeline when adding image conditions, despite memory management attempts to pin model data to RAM.

Hi i only get the following error message with minimax Wan 2.2 is running normal without any memory error - To create a public link, set share=True in launch . Loading Model ‘ckpts\MiniMax-H3-FL2VA-pruned rank8 int8 convrot.safetensors’ … Loading Text Encoder ‘ckpts\Qwen3-VL-32B-Instruct\Qwen3-VL-32B-Instruct-layer50 quanto bf16 int8.safetensors’ … Loading MiniMax H3 Video VAE ‘ckpts\MiniMax-H3-video vae fp16.safetensors’… Memory Management for the GPU Poor mmgp 3.7.12 by DeepBeepMeep Switching to partial pinning since full requirements for pinned models is 51105.5 MB while estimated available reservable RAM is 26184.6 MB. You may increase the value of parameter ‘perc reserved mem max’ to a value higher than 0.40 to force full pinnning. Partial pinning of data of ‘transformer’ to reserved RAM Found 200 tied weights for a total of 0.00 MB, last : blocks.49.attn.q proj.output scale ↔ blocks.49.attn.v proj.output scale The model was partially pinned to reserved RAM: 100 large blocks spread across 18547.98 MB Hooked to model ‘transformer’ MiniMaxH3Model Async loading plan for model ‘transformer’ : base size of 58.22 MB will be preloaded with a 370.96 MB async circular shuttle Partial pinning of data of ‘text encoder’ to reserved RAM Unable to pin more tensors for this model as the maximum reservable memory has been reached 5331.80 . The model was partially pinned to reserved RAM: 24 large blocks spread across 5331.80 MB Hooked to model ‘text encoder’ Qwen3VLTextModel Async loading plan for model ‘text encoder’ : base size of 1483.75 MB will be preloaded with a 465.16 MB async circular shuttle Unable to pin data of ‘vision encoder’ to reserved RAM as there is no reserved RAM left. Transfer speed from RAM to VRAM may be slower. Hooked to model ‘vision encoder’ Qwen3VLVisionModel Unable to pin data of ‘vae’ to reserved RAM as there is no reserved RAM left. Transfer speed from RAM to VRAM may be slower. Hooked to model ‘vae’ ModuleDict Unable to pin data of ‘video encoder’ to reserved RAM as there is no reserved RAM left. Transfer speed from RAM to VRAM may be slower. Hooked to model ‘video encoder’ ModuleDict Unable to pin data of ‘audio vae’ to reserved RAM as there is no reserved RAM left. Transfer speed from RAM to VRAM may be slower. Hooked to model ‘audio vae’ MiniMaxH3AudioVAE Traceback most recent call last : File “C:\pinokio\api\wan.git\app\wgp.py”, line 7589, in generate media samples = wan model.generate ^^^^^^^^^^^^^^^^^^^ File “C:\pinokio\api\wan.git\app\models\minimax h3\pipeline.py”, line 31, in wrapped return method self, args, kwargs ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File “C:\pinokio\api\wan.git\app\venv\Lib\site-packages\torch\utils contextlib.py”, line 124, in decorate context return func args, kwargs ^^^^^^^^^^^^^^^^^^^^^ File “C:\pinokio\api\wan.git\app\models\minimax h3\pipeline.py”, line 407, in generate self. add image condition image start, 0, presentation, visual latents, keyframes File “C:\pinokio\api\wan.git\app\models\minimax h3\pipeline.py”, line 267, in add image condition latent = self. encode video video ^^^^^^^^^^^^^^^^^^^^^^^^^ File “C:\pinokio\api\wan.git\app\models\minimax h3\pipeline.py”, line 231, in encode video return self.vae.encode condition video.unsqueeze 0 .to device=self.device, dtype=self.vae. model dtype , keep all latents=keep all latents .cpu ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File “C:\pinokio\api\wan.git\app\venv\Lib\site-packages\torch\utils device.py”, line 109, in torch function return func args, kwargs ^^^^^^^^^^^^^^^^^^^^^ torch.AcceleratorError: CUDA error: out of memory Search for cudaErrorMemoryAllocation' in https://docs.nvidia.com/cuda/cuda-runtime-api/group CUDART TYPES.html for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA LAUNCH BLOCKING=1 Compile with TORCH USE CUDA DSA to enable device-side assertions. Error Queue autosaved successfully to error queue.zip