[Kernel] Migrate linear-attention, MiniMax-sparse and diffusion kernels to sglang.kernels (RFC #29630, Phase 2.5, 6/7) (#30793)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

test/registered/attention/test_gdn_prefill_cutedsl.py

@@ -23,6 +23,10 @@ if not (torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 1
         allow_module_level=True,
     )
 
+from sglang.kernels.ops.attention.linear.gdn_blackwell import (  # noqa: E402
+    chunk_gated_delta_rule_cutedsl,
+    prepare_metadata_cutedsl,
+)
 from sglang.srt.layers.attention.fla.fused_recurrent import (  # noqa: E402
     fused_recurrent_gated_delta_rule,
 )
@@ -30,10 +34,6 @@ from sglang.srt.layers.attention.fla.index import (  # noqa: E402
     prepare_chunk_indices,
     prepare_chunk_offsets,
 )
-from sglang.srt.layers.attention.linear.kernels.gdn_blackwell import (  # noqa: E402
-    chunk_gated_delta_rule_cutedsl,
-    prepare_metadata_cutedsl,
-)
 
 
 @pytest.mark.parametrize("num_seqs", [1, 5, 257])

test/registered/attention/test_kda_prefill_cutedsl.py

@@ -25,15 +25,15 @@ if not (torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 1
         allow_module_level=True,
     )
 
+from sglang.kernels.ops.attention.linear.kda_blackwell import (  # noqa: E402
+    chunk_kda_cutedsl,
+    prepare_metadata,
+)
 from sglang.srt.layers.attention.fla.index import (  # noqa: E402
     prepare_chunk_indices,
     prepare_chunk_offsets,
 )
 from sglang.srt.layers.attention.fla.kda import fused_recurrent_kda  # noqa: E402
-from sglang.srt.layers.attention.linear.kernels.kda_blackwell import (  # noqa: E402
-    chunk_kda_cutedsl,
-    prepare_metadata,
-)
 
 
 def _l2norm(x: torch.Tensor) -> torch.Tensor:

test/registered/jit/benchmark/minimax/bench_minimax_decode_topk.py

@@ -11,7 +11,7 @@ import triton
 
 from sglang.jit_kernel.benchmark import marker
 from sglang.jit_kernel.minimax_decode_topk import minimax_decode_topk
-from sglang.srt.layers.attention.minimax_sparse_ops.decode.flash_with_topk_idx import (
+from sglang.kernels.ops.attention.minimax_sparse.decode.flash_with_topk_idx import (
     _topk_index_merge_kernel,
     _topk_index_partial_kernel,
 )

test/registered/jit/minimax/test_minimax_decode_topk_page_table.py

@@ -20,7 +20,7 @@ from sglang.jit_kernel.minimax_decode_topk import (
     minimax_decode_topk,
     minimax_decode_topk_page_table,
 )
-from sglang.srt.layers.attention.minimax_sparse_ops.decode.topk_sparse import (
+from sglang.kernels.ops.attention.minimax_sparse.decode.topk_sparse import (
     flash_decode_with_gqa_share_sparse,
 )
 from sglang.test.ci.ci_register import register_cuda_ci