Load ConvRot DiT

If you own an RTX 20-series or 30-series card, ConvRot is the most important thing that happened to quantization since GGUF. fp8 is great if your card can do fp8 math - but 30-series cards have no fp8 acceleration at all, and Turing cards (sm75, your 2060/2070/2080 Ti) don't even handle BF16 well. INT8-ConvRot is the format built for exactly that hardware, and Load ConvRot DiT is how you get it into a ComfyUI graph from this pack.

The name is worth unpacking. "DiT" here is the diffusion transformer - the actual video/image generation model, as opposed to the text encoder, which gets its own loader. This node reads a ConvRot-quantized checkpoint from ComfyUI/models/diffusion_models/ and wires it in as a normal MODEL. The whole pack (which still carries the legacy name "comfyui-svdint4") is built around making modern models run on old GPUs, and this is its front door.

How it works

ConvRot weights are stored as quantized tensors with rotation applied - the README's own words are "W8A8, W4A8, and W4A4 dispatch," meaning 8-bit weights/8-bit activations, 4-bit weights/8-bit activations, and 4-bit/4-bit. The loader reads the safetensors header for ConvRot quantization metadata and hands each layer the right path, using the pack's own kernels where they exist and comfy-kitchen's INT8 operators for the rest.

Two practical things you should know before you even hit run:

• Files without supported ConvRot quantization metadata are hidden from the file list. The tooltip says exactly that. If a downloaded ConvRot file doesn't show up, it's not ConvRot (or not a format this pack supports) - it's not a bug.
• W4A8 requires comfy-kitchen. The packed W4A8 kernel is this pack's own, but the grouped-codebook paths depend on the separate comfy-kitchen CUDA backend being installed and enabled. Without it you'll get a clear runtime error, not silent garbage.

The inputs that matter

• unet_name - the ConvRot DiT file from models/diffusion_models.
• force_int8_gemm - defaults to false, which follows each layer's activation format. Flip to true to force INT8 GEMM activations everywhere; mostly a thing to try if a specific checkpoint misbehaves.
• patch_attention (optional) - w8a8 (default), sage, or sdpa. On Turing, w8a8 and sage use the bundled exact-sm75 kernels; elsewhere w8a8 uses Comfy Kitchen's INT8 attention and sage uses the installed SageAttention package. If you want sparse Sol attention, that's a separate patch node - this loader only does dense attention.

Output: a single MODEL socket that feeds straight into your sampler. Pair it with Load ConvRot CLIP for the text side and you're done.

How to install

cd ComfyUI/custom_nodes
git clone https://github.com/wjie98/comfyui-svdint4
cd comfyui-svdint4
python -m pip install -v --no-build-isolation -e ./kernel

Manager will clone the pack if you search "comfyui-svdint4," but the kernel build is a manual step either way - that pip command compiles the CUDA kernels and needs torch-with-CUDA plus ninja. You also want comfy-kitchen installed for the W4A8 paths. Restart ComfyUI afterward.

Where people get burned

The classic complaint is the empty file dropdown. Check that the checkpoint is actually ConvRot - regular GGUF or fp8 files won't appear, on purpose. Second is the missing-kernel error, which is always the manual build step skipped. And third: this loader is the dense-attention path. If you're chasing long-video sparse attention on a Turing card, you add the Sol patch node after this, not instead of it.