Support copying tensor from cpu to gpu without using copy engines (#10007)

sgl-kernel/python/sgl_kernel/__init__.py

@@ -23,6 +23,7 @@ from sgl_kernel.cutlass_moe import cutlass_w4a8_moe_mm, get_cutlass_w4a8_moe_mm_
 from sgl_kernel.elementwise import (
     FusedSetKVBufferArg,
     apply_rope_with_cos_sin_cache_inplace,
+    copy_to_gpu_no_ce,
     downcast_fp8,
     fused_add_rmsnorm,
     gelu_and_mul,

sgl-kernel/python/sgl_kernel/elementwise.py

@@ -1,5 +1,5 @@
 from dataclasses import dataclass
-from typing import Optional
+from typing import List, Optional
 
 import torch
 from sgl_kernel.utils import get_cuda_stream, is_arch_support_pdl
@@ -367,3 +367,7 @@ def downcast_fp8(
     torch.ops.sgl_kernel.downcast_fp8(
         k, v, k_out, v_out, k_scale, v_scale, loc, mult, offset, get_cuda_stream()
     )
+
+
+def copy_to_gpu_no_ce(input: List[int], output: torch.Tensor):
+    torch.ops.sgl_kernel.copy_to_gpu_no_ce(input, output)