HunyuanVideo w. BitsAndBytes (local): Expected all tensors to be on the same device

[tin2tin]

Describe the bug

Errors in the HunyuanVideo examples here:
hunyuan_video

Reproduction

Run this code from the link:

import torch
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, HunyuanVideoTransformer3DModel, HunyuanVideoPipeline
from diffusers.utils import export_to_video

quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
transformer_8bit = HunyuanVideoTransformer3DModel.from_pretrained(
 "tencent/HunyuanVideo",
 subfolder="transformer",
 quantization_config=quant_config,
 torch_dtype=torch.float16,
)

pipeline = HunyuanVideoPipeline.from_pretrained(
 "tencent/HunyuanVideo",
 transformer=transformer_8bit,
 torch_dtype=torch.float16,
 device_map="balanced",
)

prompt = "A cat walks on the grass, realistic style."
video = pipeline(prompt=prompt, num_frames=61, num_inference_steps=30).frames[0]
export_to_video(video, "cat.mp4", fps=15)

Gives this error:
HTTPError: 404 Client Error: Not Found for url: https://huggingface.co/tencent/HunyuanVideo/resolve/main/transformer/config.json

Changing the path to: hunyuanvideo-community/HunyuanVideo

Gives this error:
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

And the other example crashes on a RTX 4090 due to OOM.

(I wanted to check if FastHunyuan-diffusers would be more vram friendly, but I couldn't due to those errors)

Logs

Logs inserted above.

System Info

Win 11

Who can help?

@DN6 @a-r-r-o-w

[SahilCarterr]

Add revision='refs/pr/18' to from_pretrained until weights are merged.

import torch
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, HunyuanVideoTransformer3DModel, HunyuanVideoPipeline
from diffusers.utils import export_to_video

quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
transformer_8bit = HunyuanVideoTransformer3DModel.from_pretrained(
 "tencent/HunyuanVideo",
 subfolder="transformer",
 quantization_config=quant_config,
 torch_dtype=torch.float16,
 revision='refs/pr/18'
)

pipeline = HunyuanVideoPipeline.from_pretrained(
 "tencent/HunyuanVideo",
 transformer=transformer_8bit,
 torch_dtype=torch.float16,
 device_map="balanced",
 revision='refs/pr/18'
)

prompt = "A cat walks on the grass, realistic style."
video = pipeline(prompt=prompt, num_frames=61, num_inference_steps=30).frames[0]
export_to_video(video, "cat.mp4", fps=15)

@tin2tin

[tin2tin]

@SahilCarterr
Thank you. With your code/weights, I still get this error (running it local):
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

[a-r-r-o-w]

@tin2tin, this seems to run perfectly fine for me without errors:

import torch
from diffusers import BitsAndBytesConfig as DiffusersBitsAndBytesConfig, HunyuanVideoTransformer3DModel, HunyuanVideoPipeline
from diffusers.utils import export_to_video

quant_config = DiffusersBitsAndBytesConfig(load_in_8bit=True)
transformer_8bit = HunyuanVideoTransformer3DModel.from_pretrained(
 "hunyuanvideo-community/HunyuanVideo",
 subfolder="transformer",
 quantization_config=quant_config,
 torch_dtype=torch.bfloat16,
)

pipeline = HunyuanVideoPipeline.from_pretrained(
 "hunyuanvideo-community/HunyuanVideo",
 transformer=transformer_8bit,
 torch_dtype=torch.float16,
 device_map="balanced",
)
print(pipeline.text_encoder.device)
print(pipeline.transformer.device)
print(pipeline.vae.device)

prompt = "A cat walks on the grass, realistic style."
video = pipeline(prompt=prompt, num_frames=61, num_inference_steps=30).frames[0]
export_to_video(video, "cat.mp4", fps=15)

The OOM comes from the default height and width set to 720 and 1280. This requires roughly about 64gb with 61 frames. BnB quantization may not be of help here because the transformer must be in bfloat16 for decent results. To enable more consumer friendly inference, I recommend trying out:

• enable_model_cpu_offload
• torchao quantization on transformer, because it does support bf16
• or, instead of quantization, we can do layerwise fp8 upcasting: [core] Layerwise Upcasting #10347

Note that you might still OOM despite these changes due to the resolution being used. Even if it doesn't, you need sufficient CPU RAM to be able to hold the models during offloading. If it OOMs on CPU, you could load the transformer directly in float8_e4m3fn and then enable layerwise upcasting - I believe this is similar to what makes it runnable in low vram in UIs

[tin2tin]

It's properly caused by the text encoder is on cpu and the rest is on cuda?

text_encoder: cpu
transformer: cuda
vae: cuda

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select

Fetching 6 files: 100%|██████████████████████████████████████████████████████████████████████████| 6/6 [00:00<?, ?it/s]
Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████| 4/4 [00:01<00:00, 3.80it/s]
Loading pipeline components...: 100%|████████████████████████████████████████████████████| 7/7 [00:03<00:00, 1.77it/s]
cpu
cuda:0
cuda:0
Error: Python: Traceback (most recent call last):
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\torch\utils\_contextlib.py", line 115, in decorate_context
 return func(*args, **kwargs)
 ^^^^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\diffusers\pipelines\hunyuan_video\pipeline_hunyuan_video.py", line 589, in __call__
 prompt_embeds, pooled_prompt_embeds, prompt_attention_mask = self.encode_prompt(
 ^^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\diffusers\pipelines\hunyuan_video\pipeline_hunyuan_video.py", line 314, in encode_prompt
 prompt_embeds, prompt_attention_mask = self._get_llama_prompt_embeds(
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\diffusers\pipelines\hunyuan_video\pipeline_hunyuan_video.py", line 241, in _get_llama_prompt_embeds
 prompt_embeds = self.text_encoder(
 ^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\torch\nn\modules\module.py", line 1532, in _wrapped_call_impl
 return self._call_impl(*args, **kwargs)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\torch\nn\modules\module.py", line 1541, in _call_impl
 return forward_call(*args, **kwargs)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\transformers\models\llama\modeling_llama.py", line 891, in forward
 inputs_embeds = self.embed_tokens(input_ids)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\torch\nn\modules\module.py", line 1532, in _wrapped_call_impl
 return self._call_impl(*args, **kwargs)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\torch\nn\modules\module.py", line 1541, in _call_impl
 return forward_call(*args, **kwargs)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\torch\nn\modules\sparse.py", line 163, in forward
 return F.embedding(
 ^^^^^^^^^^^^
 File "C:\Users\peter\Documents\blender-4.4.0\4.4\python\Lib\site-packages\torch\nn\functional.py", line 2264, in embedding
 return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0! (when checking argument for argument index in method wrapper_CUDA__index_select)

[a-r-r-o-w]

cc @sayakpaul for device_map

Oh my bad, this is because everything was on cuda for me due to testing on A100 😅 Text encoder should've been automatically moved to CUDA when its forward method is called in your case. Will try on lower VRAM GPU and try to replicate

[sayakpaul]

Well, the text encoder is being placed on a CPU because with the "balanced" device_map there was nothing else available.

So, you could first compute the text embeddings and completely delete the text encoders to free up space. https://huggingface.co/docs/diffusers/main/en/api/pipelines/mochi#reproducing-the-results-from-the-genmo-mochi-repo shows an example of how encode_prompt() can be leveraged in this case.

[tin2tin]

(The Moshi code is causing OOM crash on RTX 4090)

[sayakpaul]

I provided that example as a reference for you to adapt in your use case.

shows an example of how encode_prompt() can be leveraged in this case.

[tin2tin]

@sayakpaul @a-r-r-o-w I'm sorry for tagging you on this, but I spend a full day trying to figure out how to run HunyuanVideo on a RTX 4090 through Diffusers, and I've tried to merge the Mochi script methods with the HunyuanVideo bitsandbytes script, but everything seems to dead-end for me. Is this not meant to be, or how can HunyuanVideo be run on a 4090?