huggingface / huggingface/diffusers
MiniMax-H3 VAE decoder loads in fp32 but CUDA decode runs it in fp16
- Dominant language
- Python
- Stars
- 34.5k
- Forks
- 7.3k
- Avg merge
- 3d 3h
- Merged PRs (30d)
- 91
Description
### Describe the bug
The VAE keeps everything in fp32 no matter what dtype you pass, but the decode block wraps vae.decode in a fp16 autocast, so the decoder gets downcast to fp16 anyway on every call. On CUDA this just costs VRAM, RAM and disk space.
Decoding 192 frames at 1344x768:
| |weights|peak|
|---|---|---|
|main|9.70 GiB|16.69 GiB|
|decode in fp16|5.19 GiB|7.67 GiB|
As seen on the table, this prevents to use this model on 16GB GPUs for that resolution and duration.
### Reproduction
Just normal inference on a cuda GPU will reproduce it
```py
import torch
from diffusers import ComponentsManager, ModularPipeline
manager = ComponentsManager()
manager.enable_auto_cpu_offload(device="cuda")
pipe = ModularPipeline.from_pretrained("MiniMaxAI/MiniMax-H3", components_manager=manager)
pipe.load_components(workflow="t2va", dtype=torch.bfloat16)
results = pipe(
prompt="An astronaut hiking through the mountains, humming a tune",
num_frames=192,
output=["videos", "audio", "sampling_rate"],
)
```
code to test a PoC fix:
```py
import torch
from diffusers import AutoencoderKLMiniMaxH3
class Fp16DecoderVae(AutoencoderKLMiniMaxH3):
_keep_in_fp32_modules = [
"encoder",
"quant_conv",
"post_quant_conv",
"norm1",
"norm2",
"norm_out",
"scale1",
"scale2",
]
@classmethod
def from_pretrained(cls, *args, **kwargs):
kwargs.pop("torch_dtype", None)
kwargs["dtype"] = torch.float16
return super().from_pretrained(*args, **kwargs)
vae = Fp16DecoderVae.from_pretrained("MiniMaxAI/MiniMax-H3", subfolder="vae", dtype=torch.bfloat16)
```
### main
https://github.com/user-attachments/assets/5bfbfedb-cfed-4645-87b7-58d9c8e8fa9f
### PoC fix
https://github.com/user-attachments/assets/72429905-03b6-4590-bee0-f4fd2e4a39d4
### System Info
diffusers from main
any cuda platform
### Who can help?
Contributor guide
Research direction
Start at the AutoencoderKLMiniMaxH3 implementation and trace the decode block's autocast behavior during the provided CUDA inference reproduction. Compare it with the Fp16DecoderVae PoC, then verify that the decoder uses the intended dtype and that the 192-frame, 1344x768 case no longer incurs the reported fp32 memory usage.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 55/100