diff --git a/python/sglang/kernels/ops/diffusion/__init__.py b/python/sglang/kernels/ops/diffusion/__init__.py index 6501af6e7d..86db6b3142 100644 --- a/python/sglang/kernels/ops/diffusion/__init__.py +++ b/python/sglang/kernels/ops/diffusion/__init__.py @@ -350,7 +350,6 @@ _SPECS: tuple[tuple[str, KernelBackend, str, frozenset], ...] = ( KernelBackend.TRITON, "rope.flux3_video_qknorm_rope_triton:flux3_video_qknorm_rope", _CUDA, - "Bit-exact FLUX 3 video VAE QK RMSNorm + rotate-half RoPE, head dim 64.", ), ( "diffusion.interleaved_rope_fp64", diff --git a/python/sglang/kernels/ops/diffusion/rope/flux3_video_qknorm_rope_triton.py b/python/sglang/kernels/ops/diffusion/rope/flux3_video_qknorm_rope_triton.py index 8645f50101..d5a1165d54 100644 --- a/python/sglang/kernels/ops/diffusion/rope/flux3_video_qknorm_rope_triton.py +++ b/python/sglang/kernels/ops/diffusion/rope/flux3_video_qknorm_rope_triton.py @@ -13,7 +13,7 @@ import torch import triton import triton.language as tl -from sglang.kernels.ops.diffusion.common.numerics import cuda_rsqrtf, round_bf16_to_fp32 +from sglang.kernels.numerics import cuda_rsqrtf, round_bf16_to_fp32 from sglang.srt.utils.custom_op import register_custom_op