Possible to add RTX5000 support? Cuda 12.8 Pytorch 2.7.0?

[holycowdude]

Hi is it possible for this to support the latest RTX5000 series GPU's please?

Cuda 12.8

PyTorch 2.7.0
https://github.com/pytorch/pytorch/releases/tag/v2.7.0
https://pytorch.org/get-started/locally/

[maybleMyers]

I have added requirementsTorch27.txt.

try:
pip install typing-extensions
pip install torch==2.7.0+cu128 torchvision==0.22.0+cu128 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirementsTorch27.txt

Let me know if it works, thanks!

[holycowdude]

It works! thank you 😊

There is an error on startup but it doesnt appear to cause any issues:

H1111\h1111.py:4301: SyntaxWarning: invalid escape sequence '\d'
demo.load(None, None, None, js="""
INFO: Could not find files for the given pattern(s).

and when I generate a framepack or hunyuan t2v it mentions this, but it still works:

SUBPROCESS: WARNING:bitsandbytes.cextension:Could not find the bitsandbytes CUDA binary at WindowsPath('K:/H1111/H1111/env/Lib/site-packages/bitsandbytes/libbitsandbytes_cuda128.dll')
SUBPROCESS: WARNING:bitsandbytes.cextension:The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit multiplication, and GPU quantization are unavailable.

[maybleMyers]

Cool. What gpu do you have? You can ignore all of those warnings.

[holycowdude]

Thanks
Its an RTX5090

[holycowdude]

@maybleMyers
the latest Xformers can be installed (for CUDA128) and it speeds up Framepack:
pip install -U xformers --index-url https://download.pytorch.org/whl/cu128

Triton-Windows also seems to install and work ok but i've not seen any speed improvements with Framepack (not tested anything else yet so may need more testing)
pip install triton-windows
pip install --upgrade bitsandbytes

Also Framepack F1 seems to ignore the 'end frame image, but standard Framepack works with the end frame

[maybleMyers]

Thanks for the feedback. will fix the endframe tomorrow. I have been trying to get a fun control style v2v setup working for framepack so have only been testing that lately.

[holycowdude]

Thanks for the software its really good!

Is there a way to keep the framepack models in memory please?
It seems to need to load them every time I generate a new Framepack video, so if they could be cached it would speed up multiple runs?

[maybleMyers]

I encountered a problem with bleeding from the last gen when I went that route, like each subsequent gen would be worse and worse. Might be worth revisitting for framepack tho.

[maybleMyers]

In f1 it kinda glitches bad when you switch images, I guess that is why kohya left it out. I will need to do more testing but I have added it on this branch https://github.com/maybleMyers/H1111/tree/f1_end

[maybleMyers]

It is just the model probably. Some lora make it alot worse also. It seems to be kind of random on what pictures make it really bad. Sometimes increasing latent window size and steps help. If you are doing an extension increase vae batch size and context frames also.

For modifying the save function, you could mess with save_videos_grid in base_hav_generate_video.py for the normal framepack tab.

    stream.bit_rate = 4000000  # 4Mbit/s
#start mod
    stream.options = {
        "colorspace": "bt709",      # Sets the colorspace tag (e.g., Rec.709)
        "color_primaries": "bt709", # Describes the chromaticity of R, G, B primaries.
        "color_trc": "bt709",       # Describes the transfer characteristic (gamma).

        # "color_range": "tv",

    }
    logger.info(f"Set video stream color options: {stream.options}")
#end mod
    for frame_array in outputs:

Maybe something like that.

[holycowdude]

@maybleMyers
Please can you check the Hunyuan T2V tab?
I have the latest version but the Lora's dont seem to be working at all (no effect)?

I'm also getting this error during clip generation, unsure if its related or not, i'm using requirementsTorch27.txt with a fresh / clean install

INFO:hunyuan_model.vae:Loading 3D VAE model (884-16c-hy) from: hunyuan/pytorch_model.pt
WARNING:bitsandbytes.cextension:Could not find the bitsandbytes CUDA binary at WindowsPath('K:/H1111/H1111/env/Lib/site-packages/bitsandbytes/libbitsandbytes_cuda128.dll')
WARNING:bitsandbytes.cextension:The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit multiplication, and GPU quantization are unavailable.
INFO:hunyuan_model.vae:VAE to dtype: torch.float16
INFO:main:Set chunk_size to 32 for CausalConv3d
INFO:main:Sample save to: outputs/20250523-000452_0_2441038712.mp4
INFO:main:Done!

Thanks

[holycowdude]

@maybleMyers
I'm also getting errors with Frampack video extension. i've tried both 'force original video resolution' and changing the 'resolution' (max dimensions) but it always fails and I get this error. the videos I tried are 16:9 portrait do they need to be a different aspect ratio?

FPE_SUBPROCESS: Initializing VideoReader...
FPE_SUBPROCESS: Video loaded: 134 frames, FPS: 24.0
FPE_SUBPROCESS: Truncating video from 134 to 132 frames for latent size compatibility
FPE_SUBPROCESS: Reading video frames...
FPE_SUBPROCESS: Frames read: (132, 1280, 720, 3)
FPE_SUBPROCESS: Native video resolution: 720x1280
FPE_SUBPROCESS: Error in video_encode: 720
FPE_SUBPROCESS: Traceback (most recent call last):
FPE_SUBPROCESS: File "K:\H1111\H1111\f1_video_cli_local.py", line 113, in video_encode
FPE_SUBPROCESS: actual_target_height, actual_target_width = find_nearest_bucket(target_h_arg, target_w_arg, resolution=resolution)
FPE_SUBPROCESS: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FPE_SUBPROCESS: File "K:\H1111\H1111\frame_pack\bucket_tools.py", line 24, in find_nearest_bucket
FPE_SUBPROCESS: for (bucket_h, bucket_w) in bucket_options[resolution]:
FPE_SUBPROCESS: ~~~~~~~~~~~~~~^^^^^^^^^^^^
FPE_SUBPROCESS: KeyError: 720
FPE_SUBPROCESS: Traceback (most recent call last):
FPE_SUBPROCESS: File "K:\H1111\H1111\f1_video_cli_local.py", line 243, in do_extension_work
FPE_SUBPROCESS: start_latent_cpu, input_image_np_for_clip, video_latents_history_cpu, fps, height, width, _ = video_encode(
FPE_SUBPROCESS: ^^^^^^^^^^^^^
FPE_SUBPROCESS: File "K:\H1111\H1111\env\Lib\site-packages\torch\utils_contextlib.py", line 116, in decorate_context
FPE_SUBPROCESS: return func(*args, **kwargs)
FPE_SUBPROCESS: ^^^^^^^^^^^^^^^^^^^^^
FPE_SUBPROCESS: File "K:\H1111\H1111\f1_video_cli_local.py", line 113, in video_encode
FPE_SUBPROCESS: actual_target_height, actual_target_width = find_nearest_bucket(target_h_arg, target_w_arg, resolution=resolution)
FPE_SUBPROCESS: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
FPE_SUBPROCESS: File "K:\H1111\H1111\frame_pack\bucket_tools.py", line 24, in find_nearest_bucket
FPE_SUBPROCESS: for (bucket_h, bucket_w) in bucket_options[resolution]:
FPE_SUBPROCESS: ~~~~~~~~~~~~~~^^^^^^^^^^^^
FPE_SUBPROCESS: KeyError: 720
FPE_SUBPROCESS: Error during extension generation: 720
FPE_SUBPROCESS: Unloaded CLIPTextModel as complete.
FPE_SUBPROCESS: Unloaded DynamicSwap_LlamaModel as complete.
FPE_SUBPROCESS: Unloaded CLIPTextModel as complete.
FPE_SUBPROCESS: Unloaded SiglipVisionModel as complete.
FPE_SUBPROCESS: Unloaded AutoencoderKLCausal3D as complete.
FPE_SUBPROCESS: Unloaded DynamicSwap_HunyuanVideoTransformer3DModelPacked as complete.
FPE_SUBPROCESS: --- Extension work cycle finished. ---
FPE_SUBPROCESS: Video extension process completed.

[maybleMyers]

I added increased bucket sizes for framepack so you should just be able to specify a higher bucket. It will round it to the nearest multiple of 32 then try to find a bucket using a scaled version of the original framepack buckets and your images aspect ratio. Framepack extension is still a wip, gonna be working on it alot this weekend.

For normal hunyuan make sure your lora is converted to the default format and check the logs, you should have something like this, specifically the 320 modules or something like that:
INFO:networks.lora:create LoRA network from weights
INFO:networks.lora:create LoRA for U-Net/DiT: 320 modules.
INFO:main:Merging LoRA weights to DiT model
INFO:networks.lora:weights are merged

If it says 0 modules it means lora is incompatible/ wrong format. bitsandbytes is for linux and that is just a warning, ignore it.

[holycowdude]

@maybleMyers is it possible to include the following two projects with H1111 please?

https://github.com/Tencent-Hunyuan/HunyuanVideo-Avatar

Jenga for HunyuanVideo
https://github.com/dvlab-research/Jenga?tab=readme-ov-file

[maybleMyers]

I could. I am going to try avatar after I finish implementing phantom. Jenga seems to have very limited model support so far but will be worth a look once they finish wan 14b support.