[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>

test/manual/lora/test_lora_tuning_config.py

@@ -6,7 +6,7 @@ import tempfile
 import unittest
 from unittest.mock import patch
 
-from sglang.srt.lora.triton_ops.lora_tuning_config import (
+from sglang.kernels.ops.gemm.lora_tuning_config import (
     DEFAULT_EXPAND_CONFIG,
     DEFAULT_SHRINK_CONFIG,
     get_lora_config_file_name,
@@ -15,7 +15,7 @@ from sglang.srt.lora.triton_ops.lora_tuning_config import (
     get_lora_shrink_config,
 )
 
-_MODULE = "sglang.srt.lora.triton_ops.lora_tuning_config"
+_MODULE = "sglang.kernels.ops.gemm.lora_tuning_config"
 
 # Shared fixture
 _TUNED_CONFIGS = {
@@ -88,7 +88,7 @@ class TestConfigSelection(unittest.TestCase):
 
     def setUp(self):
         get_lora_configs.cache_clear()
-        from sglang.srt.lora.triton_ops import lora_tuning_config
+        from sglang.kernels.ops.gemm import lora_tuning_config
 
         lora_tuning_config._logged_configs.clear()
 

test/manual/test_triton_attention_rocm_mla.py

@@ -3,10 +3,10 @@ import unittest
 
 import torch
 
-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.rocm_mla_decode_rope import (
+from sglang.kernels.ops.attention.rocm_mla_decode_rope import (
     decode_attention_fwd_grouped_rope,
 )
 from sglang.srt.layers.rotary_embedding import DeepseekScalingRotaryEmbedding

test/manual/test_trtllm_fp8_kv_kernel.py

@@ -6,7 +6,7 @@ import unittest
 
 import torch
 
-from sglang.srt.layers.attention.triton_ops.trtllm_fp8_kv_kernel import (
+from sglang.kernels.ops.kvcache.trtllm_fp8_kv_kernel import (
     fused_fp8_set_kv_buffer,
 )
 from sglang.test.test_utils import CustomTestCase