[MoE] Retire the AOT moe_fused_gate / kimi_k2_moe_fused_gate gate kernels (#26771) (#29997)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

sgl-kernel/python/sgl_kernel/__init__.py

@@ -93,9 +93,7 @@ else:
         apply_shuffle_mul_sum,
         fp8_blockwise_scaled_grouped_mm,
         fused_qk_norm_rope,
-        kimi_k2_moe_fused_gate,
         moe_align_block_size,
-        moe_fused_gate,
         moe_sum,
         moe_sum_reduce,
         prepare_moe_input,
@@ -181,10 +179,8 @@ else:
         "gptq_gemm",
         "gptq_shuffle",
         "int8_scaled_mm",
-        "kimi_k2_moe_fused_gate",
         "merge_state_v2",
         "moe_align_block_size",
-        "moe_fused_gate",
         "moe_sum",
         "moe_sum_reduce",
         "prepare_moe_input",

sgl-kernel/python/sgl_kernel/moe.py

@@ -102,74 +102,9 @@ def moe_sum(
     )
 
 
-def moe_fused_gate(
-    input_tensor,
-    bias,
-    num_expert_group,
-    topk_group,
-    topk,
-    num_fused_shared_experts=0,
-    routed_scaling_factor=0,
-    apply_routed_scaling_factor_on_output=False,
-):
-    # This fused kernel function is used to select topk expert in a hierarchical 2-layer fashion
-    # it split group of expert into num_expert_group, and use top2 expert weight sum in each group
-    # as the group weight to select expert groups and then select topk experts within the selected groups
-    # the #experts is decided by the input tensor shape and we currently only support power of 2 #experts
-    # and #experts should be divisible by num_expert_group. #expert/num_expert_group <= 32 is limited for now.
-    # for non-supported case, we suggest to use the biased_grouped_topk func in sglang.srt.layers.moe.topk
-    # num_fused_shared_experts: if > 0, the last several experts will be
-    #   replaced with shared experts. the shared experts will be divided by the
-    #   routed_scaling_factor - this is intended to cancel out later when routed+shared
-    #   output is scaled so that shared experts are not scaled.
-    # routed_scaling_factor: if > 0, the experts will be scaled by this factor
-    # apply_routed_scaling_factor_on_output: if true, output will be
-    #   scaled by the routed_scaling_factor
-    return torch.ops.sgl_kernel.moe_fused_gate.default(
-        input_tensor,
-        bias,
-        num_expert_group,
-        topk_group,
-        topk,
-        num_fused_shared_experts,
-        routed_scaling_factor,
-        apply_routed_scaling_factor_on_output,
-    )
-
-
-def kimi_k2_moe_fused_gate(
-    input_tensor,
-    bias,
-    topk,
-    renormalize=True,
-    routed_scaling_factor=1.0,
-    apply_routed_scaling_factor_on_output=False,
-):
-    """
-    Simplified fused kernel for Kimi K2 model (num_expert_group=1).
-    This kernel removes the grouped topk logic since all experts belong to a single group.
-
-    Args:
-        input_tensor: Gating output tensor [num_tokens, num_experts]
-        bias: Correction bias tensor [num_experts]
-        topk: Number of experts to select per token
-        renormalize: Whether to renormalize the topk weights
-        routed_scaling_factor: Scaling factor for expert weights
-        apply_routed_scaling_factor_on_output: If true, apply scaling factor to output
-
-    Returns:
-        Tuple of (topk_weights, topk_ids)
-        - topk_weights: [num_tokens, topk] float32 tensor
-        - topk_ids: [num_tokens, topk] int32 tensor
-    """
-    return torch.ops.sgl_kernel.kimi_k2_moe_fused_gate.default(
-        input_tensor,
-        bias,
-        topk,
-        renormalize,
-        routed_scaling_factor,
-        apply_routed_scaling_factor_on_output,
-    )
+# moe_fused_gate / kimi_k2_moe_fused_gate (AOT gate kernels) retired — the gate/topk
+# path is consolidated onto the unified Triton router in
+# python/sglang/jit_kernel/moe_fused_gate.py (sglang issue #26771).
 
 
 def fp8_blockwise_scaled_grouped_mm(