Merge int8_convrot models ?
Could you upload a single-file format version that can be used in ComfyUI,
like this:
fl2va_int8_convrot.safetensors
ref2va_int8_convrot.safetensors
Thanks~
I agree, single int8 models would be much better.
+1!
I made a single file, but it cannot used by comfy, the comfy loaders do not support the structure yet and there is no PR to support it... 😑
I made a single file, but it cannot used by comfy, the comfy loaders do not support the structure yet and there is no PR to support it... 😑
@darksidewalker Try this node? https://github.com/Saganaki22/ComfyUI-Hyperflow
I made a single file, but it cannot used by comfy, the comfy loaders do not support the structure yet and there is no PR to support it... 😑
@darksidewalker Try this node? https://github.com/Saganaki22/ComfyUI-Hyperflow
That is a lora loader. Not working with the full checkpoint.
I made a single file, but it cannot used by comfy, the comfy loaders do not support the structure yet and there is no PR to support it... 😑
can you give that file ? or link
Did it. Single-file INT8 for ComfyUI, both fl2va and ref2va:
https://huggingface.co/BennyDaBall/HyperFlow-H3-Flash-INT8-ComfyUI
@darksidewalker is right that the stock loaders can't run it as is. HyperFlow feeds the model a second time value and some blocks are low-rank, so it needs one small loader node:
https://github.com/BennyDaBall930/comfyui_hyperflow_int8_nodes
It's also in ComfyUI Manager as HyperFlow INT8 Nodes. You need the official ComfyUI-HyperFlow-Flash pack installed too.
It's Video Rebirth's own INT8 weights converted bit for bit to ComfyUI's layout, nothing requantized, running on ComfyUI's native INT8 kernels. Workflows for t2v, i2v, flf2v and ref2v are in the repo.
On my 5090 at 1344x768, a 5 s clip (124 frames) with audio takes 46 s, vs 66 s for the Q4_K_M GGUF at the same settings. About 38 s with ComfyUI's sparse attention node (that workflow is included).
@VideoRebirthAI if you want this in your official ComfyUI pack, I've got the loader ready to PR. It's small and reuses your model_patch.py as is, so your users would only need your pack plus these weights.
Unofficial conversion, weights stay under the MiniMax H3 Community License.
We really need an official release that provides models quantized and packaged according to the standard INT8 format supported by ComfyUI. The current method—using the INT8 weight-conversion approach provided by HyperFlow—is problematic: it doesn't support the (down, up) projection structure and frequently leads to compatibility issues with other H3 LoRAs or plugins. Furthermore, attempting to convert these into the standard ComfyUI-supported INT8 format would require expanding the low-rank linear layers back into full-rank matrices before quantization, which compromises precision. The only viable solution is to perform the quantization and packaging directly from the base model.