[Kernel] Introduce sglang.kernels namespace and migrate scattered triton_ops kernels (RFC #29630, Phase 2) (#30044)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

benchmark/bench_attention_sink/bench_attention_sink_triton.py

@@ -3,10 +3,10 @@ import argparse
 import torch
 import triton
 
-from sglang.srt.layers.attention.triton_ops.decode_attention import (
+from sglang.kernels.ops.attention.decode_attention import (
     decode_attention_fwd_grouped,
 )
-from sglang.srt.layers.attention.triton_ops.extend_attention import extend_attention_fwd
+from sglang.kernels.ops.attention.extend_attention import extend_attention_fwd
 
 # gpt oss
 head_num = 64

benchmark/kernels/decoding_attention_triton/triton_flashinfer_cudnn.py

@@ -6,8 +6,8 @@ import torch
 import torch.utils.benchmark as benchmark
 from flashinfer import BatchDecodeWithPagedKVCacheWrapper
 
+from sglang.kernels.ops.attention.decode_attention import decode_attention_fwd
 from sglang.srt.layers.attention.flashinfer_backend import should_use_tensor_core
-from sglang.srt.layers.attention.triton_ops.decode_attention import decode_attention_fwd
 
 
 def benchmark_forward(

benchmark/kernels/lora_csgmv/tune_lora_csgmv.py

@@ -58,9 +58,9 @@ from typing import Any, Dict, List, Optional
 import torch
 import triton
 
-from sglang.srt.lora.triton_ops.chunked_sgmv_expand import _chunked_lora_expand_kernel
-from sglang.srt.lora.triton_ops.chunked_sgmv_shrink import _chunked_lora_shrink_kernel
-from sglang.srt.lora.triton_ops.lora_tuning_config import (
+from sglang.kernels.ops.gemm.chunked_sgmv_expand import _chunked_lora_expand_kernel
+from sglang.kernels.ops.gemm.chunked_sgmv_shrink import _chunked_lora_shrink_kernel
+from sglang.kernels.ops.gemm.lora_tuning_config import (
     DEFAULT_EXPAND_CONFIG,
     DEFAULT_SHRINK_CONFIG,
     get_lora_config_file_name,
@@ -362,9 +362,9 @@ def save_config(
         "..",
         "python",
         "sglang",
-        "srt",
-        "lora",
-        "triton_ops",
+        "kernels",
+        "ops",
+        "gemm",
         "csgmv_configs",
         version_dir,
     )

benchmark/kernels/sliding_window_attention_triton/bench_triton_swa_kernel.py

@@ -5,7 +5,7 @@ import torch.nn.functional as F
 import triton.testing as tt
 
 from sglang.benchmark.bench_utils import run_bench
-from sglang.srt.layers.attention.triton_ops.extend_attention import extend_attention_fwd
+from sglang.kernels.ops.attention.extend_attention import extend_attention_fwd
 
 
 def extend_attention_fwd_torch(

benchmark/kernels/verify_splitkv_triton/bench_verify_splitkv.py

@@ -19,10 +19,10 @@ import argparse
 import torch
 import triton
 
-from sglang.srt.layers.attention.triton_ops.extend_attention import (
+from sglang.kernels.ops.attention.extend_attention import (
     extend_attention_fwd,
 )
-from sglang.srt.layers.attention.triton_ops.verify_splitkv import verify_splitkv_fwd
+from sglang.kernels.ops.attention.verify_splitkv import verify_splitkv_fwd
 from sglang.srt.utils import is_gfx95_supported