Delete sgl-kernel AOT router GEMM and fused A GEMM (#30280)

Co-authored-by: Brayden Zhong <brayden@radixark.ai>
Co-authored-by: root <root@sgl-b300-inference.datacrunch.io>

sgl-kernel/python/sgl_kernel/__init__.py

@@ -55,7 +55,6 @@ else:
     )
     from sgl_kernel.gemm import (
         awq_dequantize,
-        dsv3_fused_a_gemm,
         fp8_scaled_mm,
         gptq_gemm,
         gptq_shuffle,
@@ -160,8 +159,6 @@ else:
         "copy_to_gpu_no_ce",
         "cutlass_mla_decode",
         "cutlass_mla_get_workspace_size",
-        "dsv3_fused_a_gemm",
-        "dsv3_router_gemm",
         "dsv4_fused_k_norm_rope_flashmla",
         "dsv4_fused_q_indexer_rope_hadamard_quant",
         "dsv4_fused_q_norm_rope",

sgl-kernel/python/sgl_kernel/gemm.py

@@ -31,21 +31,6 @@ def fp8_scaled_mm(mat_a, mat_b, scales_a, scales_b, out_dtype, bias=None):
     )
 
 
-def dsv3_fused_a_gemm(
-    mat_a: torch.Tensor,
-    mat_b: torch.Tensor,
-    output: Optional[torch.Tensor] = None,
-) -> torch.Tensor:
-    if output is None:
-        output = torch.empty(
-            (mat_a.shape[0], mat_b.shape[1]),
-            device=mat_a.device,
-            dtype=mat_a.dtype,
-        )
-    torch.ops.sgl_kernel.dsv3_fused_a_gemm.default(output, mat_a, mat_b)
-    return output
-
-
 def sgl_per_token_group_quant_8bit(
     input: torch.Tensor,
     output_q: torch.Tensor,