[kernel slimming] Clean many useless sgl-kernel deprecated kernels (#20277)

sgl-kernel/benchmark/bench_fp4_gemm.py

@@ -5,8 +5,8 @@ import os
 import torch
 import triton
 from flashinfer import mm_fp4
-from sgl_kernel import cutlass_scaled_fp4_mm, scaled_fp4_quant
 
+from sglang.jit_kernel.nvfp4 import cutlass_scaled_fp4_mm, scaled_fp4_quant
 from sglang.srt.utils import get_device_capability, is_sm100_supported
 
 # CI environment detection

sgl-kernel/benchmark/bench_fp8_gemm.py

@@ -7,7 +7,8 @@ from typing import Optional, Tuple
 import torch
 import triton
 from sgl_kernel import fp8_scaled_mm as sgl_scaled_mm
-from sgl_kernel import sgl_per_tensor_quant_fp8
+
+from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
 
 # Optional vLLM import
 try:
@@ -97,7 +98,7 @@ def sglang_scaled_fp8_quant(
     if scale is None:
         scale = torch.zeros(1, device=input.device, dtype=torch.float32)
         is_static = False
-    sgl_per_tensor_quant_fp8(input, output, scale, is_static)
+    per_tensor_quant_fp8(input, output, scale, is_static)
 
     return output, scale
 

sgl-kernel/benchmark/bench_nvfp4_scaled_gemm.py

@@ -5,8 +5,8 @@ import os
 
 import torch
 import triton
-from sgl_kernel import cutlass_scaled_fp4_mm, scaled_fp4_quant
 
+from sglang.jit_kernel.nvfp4 import cutlass_scaled_fp4_mm, scaled_fp4_quant
 from sglang.srt.utils import get_device_capability
 
 # CI environment detection

sgl-kernel/benchmark/bench_per_tensor_quant_fp8.py

@@ -7,7 +7,8 @@ import numpy as np
 import torch
 import triton
 import triton.testing
-from sgl_kernel import sgl_per_tensor_quant_fp8
+
+from sglang.jit_kernel.per_tensor_quant_fp8 import per_tensor_quant_fp8
 
 # Optional imports
 try:
@@ -51,7 +52,7 @@ def sglang_scaled_fp8_quant(
     if scale is None:
         scale = torch.zeros(1, device=input.device, dtype=torch.float32)
         is_static = False
-    sgl_per_tensor_quant_fp8(input, output, scale, is_static)
+    per_tensor_quant_fp8(input, output, scale, is_static)
 
     return output, scale
 

sgl-kernel/benchmark/bench_rotary_embedding.py

@@ -3,9 +3,9 @@ import os
 
 import torch
 import triton
-from sgl_kernel import FusedSetKVBufferArg
 from sgl_kernel.testing.rotary_embedding import (
     FlashInferRotaryEmbedding,
+    FusedSetKVBufferArg,
     MHATokenToKVPool,
     RotaryEmbedding,
     create_inputs,