Nunchaku INT4/FP4 Kernels on RTX 3090: ComfyUI vs Wan2GP Optimization

The landscape of AI image generation workflows is constantly evolving, with new tools and optimizations appearing all the time. For users running these models on consumer hardware like an NVIDIA GeForce RTX 3090, compatibility becomes a major concern. This article explores whether Nunchaku INT4/FP4 kernels work effectively on an RTX 3090, while also analyzing the differences between ComfyUI and Wan2GP’s installation requirements to find the best setup without breaking your environment.

The core question is whether an RTX 3090 can use these Nunchaku wheels. The answer is yes, but with important distinctions between FP4 and INT4 support.

Nunchaku on the RTX 3090: INT4 vs. FP4

The RTX 3090 uses an Ampere architecture (SM 8.6), which means it falls squarely into the category of GPUs that Nunchaku explicitly supports for INT4 operations. While FP4 is primarily designed for newer architectures like Blackwell, your 3090 can definitely leverage INT4 kernels without issue. However, this conversation also highlights a broader trend in AI development: the fragmentation of software compatibility. We are seeing how different frameworks handle these optimizations differently, which leads us to look closer at the comparison between ComfyUI and Wan2GP.

ComfyUI vs. Wan2GP Installation Requirements

A major point of confusion in this space is the difference in installation requirements between these two popular tools. One noticeable detail is that Wan2GP does not list xformers as a hard requirement, whereas ComfyUI historically had a much stronger relationship with it. This isn’t just about random dependencies; it reflects their underlying philosophies. ComfyUI aims to be a generalist framework with maximum compatibility, supporting hundreds of nodes and thousands of combinations. Wan2GP, on the other hand, acts as an opinionated application that prefers stability over raw performance in its default configurations. This leads us directly into one of the most critical technical comparisons: FlashAttention wheels.

FlashAttention Wheels

The installation commands reveal a significant difference between the two tools. ComfyUI typically uses standard, explicitly labeled wheels for FlashAttention that specify exact CUDA and Torch versions. In contrast, Wan2GP utilizes a wheel from a specialized repository called `deepbeepmeep/kernels`. This suggests that Wan2GP is building its own «mini distribution» of optimized kernels rather than relying on the standard ecosystem. While ComfyUI prioritizes universal compatibility, Wan2GP seems to be optimizing for a specific, curated stack. This brings us to another key optimization area discussed in this analysis: SageAttention.

SageAttention Optimization Strategies

The versioning of SageAttention further illustrates the trade-offs between stability and performance. There is a clear distinction between installing `sageattention==1.0.6` versus using the specialized 2.2.0 wheel. Wan2GP lists the stable version 1.0.6 in its requirements, prioritizing a setup that works across a wide range of PCs without requiring users to troubleshoot complex CUDA mismatches. Meanwhile, ComfyUI and modern guides often push for the optimized 2.x versions to maximize performance. Ultimately, this analysis shows that there is no single «best» configuration. Users must decide whether they want the maximum stability provided by Wan2GP’s conservative approach or the raw, configurable performance of a ComfyUI setup.