Kijai

/

MiniMax-H3_comfy

like 178

In my case, the process errors out and freezes completely. I had to reinstall ComfyUI from scratch; installing the nodes and dependencies for W4A8 messed everything up. I hope they manage to implement this soon, that would HELP A LOT with speed and RAM/VRAM bottlenecks.

In my case, the process errors out and freezes completely. I had to reinstall ComfyUI from scratch; installing the nodes and dependencies for W4A8 messed everything up. I hope they manage to implement this soon, that would HELP A LOT with speed and RAM/VRAM bottlenecks.

https://github.com/Comfy-Org/ComfyUI/commit/344b43989e8c56b5bb4a66cf028c834192ab59dd
This PR has already been merged into the main branch.

For me its just very slow.

I am also having a very slow experience in comparison to the normal int8 model.

Same for me. Very slow and corrupted output when used together with the experimental W4A8 checkpoint. However, no error messages in the GUI or in the console. I'm just curious whether it's a known limitation and to be expected, or a bug?

Same for me. Very slow and corrupted output when used together with the experimental W4A8 checkpoint. However, no error messages in the GUI or in the console. I'm just curious whether it's a known limitation and to be expected, or a bug?

With a LoRA? That had some issues that were only fixed in the past few hours.

Hi Kijai, first of thanks for all these optimizations QOL improvements you make with your node. Personally I have experienced that the w4a8 model is multitudes slower than the int8 model. I had earlier used the w4a8 conv models of this person: https://huggingface.co/AX1Y2JP/MiniMax-H3-W4A8-ConvRot. And these models are fast, just a bit slower than the int8 models. But these same models don't seem to work with your code implementation . They had made a custom wheel that is based of your code and somehow that seems to work well for their models and are much faster.

Same for me. Very slow and corrupted output when used together with the experimental W4A8 checkpoint. However, no error messages in the GUI or in the console. I'm just curious whether it's a known limitation and to be expected, or a bug?

With a LoRA? That had some issues that were only fixed in the past few hours.

EDIT:
Damn, I misread your question as "Which LoRA", and now my response makes less sense. I still think I have the correct LoRAs, but I will re-download them and test again.

ORIGINAL:
Both converted 4-Step LightX2V LoRAs from this repository (minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy and minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy_resized_avg_rank_21_bf16). When I try to use either one together with minimax_h3_fl2va_pruned_w4a8_mixed, the generation takes a very long time and the resulting video is corrupted. Without the LoRAs, the checkpoint works perfectly fine. Also, using the LoRAs with the minimax_h3_ref2va_pruned_int8_convrot causes no issues at all.

I just checked, and I've downloaded both LoRAs in the last 4 hours, so I believe I have the latest versions?

Btw, thank you for all the work you're doing.

Same for me. Very slow and corrupted output when used together with the experimental W4A8 checkpoint. However, no error messages in the GUI or in the console. I'm just curious whether it's a known limitation and to be expected, or a bug?

With a LoRA? That had some issues that were only fixed in the past few hours.

Using only w4a8 results in fast generation speeds and high quality.
Using int8 with LoRA yields fast generation speeds and moderate quality, with slight, barely noticeable artifacts.
Using w4a8 with LoRA results in slow generation speeds and poor quality, with artifacts accumulating over time.
When using the dedicated LoRA loader and sampler from https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo with w4a8 and LoRA, generation is extremely fast but quality is very poor, with artifacts pervading every part of the image.

My issue turned to be an outdated ComfyUI installation. And by "outdated", I mean by something like 5 hours... The speed of development is incredible, Kijai & Co. must be pulling all-nighters. Really appreciate that...

Using w4a8 with LoRA results in slow generation speeds and poor quality, with artifacts accumulating over time.

Update ComfyUI. It works fine now!

My issue turned to be an outdated ComfyUI installation. And by "outdated", I mean by something like 5 hours... The speed of development is incredible, Kijai & Co. must be pulling all-nighters. Really appreciate that...

Using w4a8 with LoRA results in slow generation speeds and poor quality, with artifacts accumulating over time.

Update ComfyUI. It works fine now!

Shit, just finished the INT8 acceleration benchmark using ComfyUI-SolAttn_triton and ComfyUI-MiniMaxH3-AdaptiveCache.

I can confirm that the lora slowdown problem is indeed fixed, thanks again Kijai.

I used https://comfy.org/workflows/e8099b642c9f-e8099b642c9f/ for testing.

The latest version of Comfy resolves the W4A8 compatibility issues encountered when using Turbo; the destructive artifacts seen in the previous version are virtually gone, leaving only slight, barely perceptible, and infrequent flickering artifacts. However, the image quality is somewhat lackluster, and using Sol-Attn in conjunction with Turbo leads to a noticeable drop in instruction-following performance.

Sol-Attn is currently the most stable and fundamental performance optimization.

AdaptiveCache is a secondary optimization layer; it becomes more effective as the step count increases and shows clear synergy with Sol.

EasyCache is effective at 20 steps but generally inferior to AdaptiveCache; under current parameters, Turbo4/8 offers little practical value.

Turbo4 should be used with Sol only, whereas Turbo8 works best with Sol + AdaptiveCache.

On the 5090, the actual speeds of W4A8 and INT8 are nearly identical; however, since W4A8 compresses the DiT weights from approximately 20GB to around 12GB, it offers significantly higher overall value.

Once optimization reaches a certain level, the bottleneck begins to shift from the DiT to post-processing components like the VAE or Audio VAE.

https://github.com/FFFFFFpy/ComfyUI-MiniMaxH3-AdaptiveCache
https://github.com/kijai/ComfyUI-SolAttn_triton