Veda sparse attention on Strix Halo: 1.89x faster MiniMax H3 video gen (691s -> 366s)
Sharing a ComfyUI node I've been working on, since video gen on Strix Halo is where a lot of us keep hitting the wall.
Credit first: Veda is a learned sparse-attention method for video diffusion models (ICML 2026). A small distilled predictor says in advance which tiles of the attention map actually carry the result; only the top ~10% get computed. The original project is by veda-sparse:
- Project page: https://veda-sparse.github.io/
- Original repo: https://github.com/veda-sparse/Veda-on-ComfyUI
- Predictor model file (275 MB): https://huggingface.co/Veda-Sparse/Minimax-H3-T2VA-Veda-8NFE-600Step-Preview — goes in ComfyUI/models/veda/ (opening a Veda template offers it in the missing-model dialog)
Upstream supports NVIDIA (SM80+) and Apple silicon, and lists ROCm as "no kernel, the model runs its own attention". So I ported it to ROCm and tuned it for Strix Halo. What that took:
- Backend layer: report the real gfx family (gfx1151) and unified memory on HIP, and admit ROCm devices to the Triton INT8 backend (kernels still self-test on the GPU before use, same as upstream)
- Triton's ROCm backend compiles the INT8 kernel for gfx1151: Q and K are block-quantized to int8, QK^T runs on the hardware int8 WMMA instructions (v_wmma_i32_16x16x16_iu8), P*V stays fp16 WMMA
- The launch config upstream shipped (4 warps / 3 stages) was tuned on an RTX 5070 and is wrong on this chip: measured on the real portrait grid, 8 warps / 2 stages is 2x on the attention kernel (35.3 -> 17.6 ms per call)
- A few more gfx1151-specific kernel tweaks, e.g. the quantize pass runs 23% faster with 2 warps, bit-identical output
- Accuracy verified on this hardware: kernel vs the INT8 reference is ~0.14% relative L2, below the quantization noise that's already there
Numbers on my machine, MiniMax H3, 11-second clip at 480x864:
- dense attention sampler: 691s
- sparse (Veda): 366s
- 1.89x on sampling, 2.15x on the full prompt (1016s -> 473s)
- only ~27% of the attention is computed, output looks the same to me
My ROCm port lives in rocm-ninodes (ComfyUI Manager, or comfy node install rocm-ninodes); the node is VedaSparseAttention, drop it on the MODEL wire last before the sampler:
- https://github.com/iGavroche/rocm-ninodes
If you try it on your Halo I'd be curious about numbers on other resolutions/lengths. Happy to answer questions about the sparse selection or the ROCm port.
tl;dr: Veda Sparse Attention on ROCm (optimized for Strix Halo)