diff --git a/python/sglang/kernels/ops/diffusion/__init__.py b/python/sglang/kernels/ops/diffusion/__init__.py
index 6faa255f28..ea70f58a21 100644
--- a/python/sglang/kernels/ops/diffusion/__init__.py
+++ b/python/sglang/kernels/ops/diffusion/__init__.py
@@ -287,6 +287,13 @@ _SPECS: tuple[tuple[str, KernelBackend, str, frozenset, str], ...] = (
         _CUDA,
         "Paired in-place Helios transposed Q/K RoPE.",
     ),
+    (
+        "diffusion.complex_rope",
+        KernelBackend.TRITON,
+        "rope.complex_rope_triton:fused_complex_rope",
+        _CUDA,
+        "Paired RoPE preserving PyTorch complex64 multiplication rounding.",
+    ),
     (
         "diffusion.hunyuan_qkv_rope_pack",
         KernelBackend.TRITON,
@@ -294,6 +301,13 @@ _SPECS: tuple[tuple[str, KernelBackend, str, frozenset, str], ...] = (
         _CUDA,
         "HunyuanVideo QKV pack + RoPE.",
     ),
+    (
+        "diffusion.rmsnorm_preserve_reduction",
+        KernelBackend.TRITON,
+        "norm.rmsnorm_preserve_reduction:rmsnorm_preserve_reduction",
+        _CUDA,
+        "Cast-before-weight RMSNorm preserving the native FP32 mean reduction.",
+    ),
     (
         "diffusion.silu_mul",
         KernelBackend.TRITON,
@@ -543,6 +557,8 @@ _EXPORTS: dict[str, str] = {
     "try_fused_bias_mul_add": "sglang.kernels.kda_kernels.norm_scale_shift_jit",
     "try_fused_bias_scale_residual_norm_scale_shift": "sglang.kernels.kda_kernels.norm_scale_shift_jit",
     "triton_one_pass_rms_norm": "norm.rmsnorm_onepass_triton",
+    "can_use_rmsnorm_preserve_reduction": "norm.rmsnorm_preserve_reduction",
+    "rmsnorm_preserve_reduction": "norm.rmsnorm_preserve_reduction",
     "can_use_fused_rmsnorm_scale_shift": "norm.rmsnorm_scale_shift_bitexact",
     "can_use_fused_scale_residual_rmsnorm_scale_shift": "norm.rmsnorm_scale_shift_bitexact",
     "fused_rmsnorm_scale_shift_bitexact": "norm.rmsnorm_scale_shift_bitexact",
@@ -598,6 +614,8 @@ _EXPORTS: dict[str, str] = {
     "can_use_helios_qk_rope": "rope.helios_qk_rope_jit",
     "fused_inplace_helios_qk_rope": "rope.helios_qk_rope_jit",
     "apply_rotary_embedding": "rope.rotary_triton",
+    "can_use_fused_complex_rope": "rope.complex_rope_triton",
+    "fused_complex_rope": "rope.complex_rope_triton",
     # Tensor layout transformations fused with downstream quantization
     "try_flux2_token_cat_fp8": "sglang.kernels.kda_kernels.flux2_token_cat_fp8_triton",
     # Activation-function fusions