[Kernel] RFC #29630 finale: retire sglang.jit_kernel into sglang.kernels (#32072)

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

sgl-kernel/benchmark/bench_fp8_gemm.py

@@ -8,7 +8,9 @@ import torch
 import triton
 from sgl_kernel import fp8_scaled_mm as sgl_scaled_mm
 
-from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
+from sglang.kernels.ops.quantization._jit_per_tensor_quant_fp8 import (
+    per_tensor_quant_fp8,
+)
 from sglang.utils import is_in_ci
 
 # Optional vLLM import

sgl-kernel/benchmark/bench_fp8_gemm_swap_ab.py

@@ -18,7 +18,9 @@ import torch
 import triton
 from sgl_kernel import fp8_scaled_mm as sgl_scaled_mm
 
-from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
+from sglang.kernels.ops.quantization._jit_per_tensor_quant_fp8 import (
+    per_tensor_quant_fp8,
+)
 from sglang.utils import is_in_ci
 
 IS_CI = is_in_ci()

sgl-kernel/benchmark/bench_per_tensor_quant_fp8.py

@@ -8,7 +8,9 @@ import torch
 import triton
 import triton.testing
 
-from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
+from sglang.kernels.ops.quantization._jit_per_tensor_quant_fp8 import (
+    per_tensor_quant_fp8,
+)
 from sglang.utils import is_in_ci
 
 # Optional imports

sgl-kernel/csrc/common_extension.cc

@@ -174,7 +174,7 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
 
   // moe_fused_gate / kimi_k2_moe_fused_gate (AOT) retired: the CUDA gate/topk path
   // now routes through the unified Triton router
-  // (python/sglang/jit_kernel/moe_fused_gate.py).
+  // (python/sglang/kernels/ops/moe/moe_fused_gate.py).
 
   m.def(
       "fp8_blockwise_scaled_grouped_mm(Tensor output, Tensor a_ptrs, Tensor b_ptrs, Tensor out_ptrs, Tensor "

sgl-kernel/csrc/elementwise/dsv4_norm_rope.cu

@@ -14,7 +14,7 @@ limitations under the License.
 ==============================================================================*/
 
 // DeepSeek-V4 fused norm + RoPE kernels, ported from JIT kernel
-// python/sglang/jit_kernel/csrc/deepseek_v4/main_norm_rope.cuh
+// python/sglang/kernels/jit/csrc/deepseek_v4/main_norm_rope.cuh
 // to sgl-kernel AOT compilation with CUDA + HIP (ROCm) support.
 
 #ifndef USE_ROCM

sgl-kernel/python/sgl_kernel/moe.py

@@ -106,7 +106,7 @@ def moe_sum(
 
 # moe_fused_gate / kimi_k2_moe_fused_gate (AOT gate kernels) retired — the gate/topk
 # path is consolidated onto the unified Triton router in
-# python/sglang/jit_kernel/moe_fused_gate.py (sglang issue #26771).
+# python/sglang/kernels/ops/moe/moe_fused_gate.py (sglang issue #26771).
 
 
 def fp8_blockwise_scaled_grouped_mm(

sgl-kernel/python/sgl_kernel/testing/rotary_embedding.py

@@ -3,8 +3,10 @@ from typing import Optional, Tuple, Union
 
 import torch
 
-from sglang.jit_kernel.rope import FusedSetKVBufferArg as _JitFusedSetKVBufferArg
-from sglang.jit_kernel.rope import (
+from sglang.kernels.ops.attention.rope import (
+    FusedSetKVBufferArg as _JitFusedSetKVBufferArg,
+)
+from sglang.kernels.ops.attention.rope import (
     apply_rope_with_cos_sin_cache_inplace as _jit_apply_rope_with_cos_sin_cache_inplace,
 )
 

sgl-kernel/tests/test_cutlass_w4a8_moe_mm.py

@@ -5,7 +5,9 @@ import torch
 from sgl_kernel import cutlass_w4a8_moe_mm
 from utils import is_hopper
 
-from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
+from sglang.kernels.ops.quantization._jit_per_tensor_quant_fp8 import (
+    per_tensor_quant_fp8,
+)
 
 
 def pack_int4_values_to_int8(int4_values_interleaved: torch.Tensor) -> torch.Tensor: