[Refactor] Reducing code duplication across FP8 CUDA quantization kernels (#4163)

sgl-kernel/benchmark/bench_per_token_group_quant_fp8.py

@@ -1,13 +1,12 @@
 import itertools
-import math
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Tuple
 
 import torch
 import triton
 import triton.language as tl
 from sgl_kernel import sgl_per_token_group_quant_fp8
 
-from sglang.srt.utils import get_device_core_count, get_device_name, is_hip
+from sglang.srt.utils import is_hip
 
 is_hip_ = is_hip()
 fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn

sgl-kernel/benchmark/bench_per_token_quant_fp8.py

@@ -40,9 +40,6 @@ def calculate_diff(batch_size: int, seq_len: int):
     scale_diff = torch.abs(vllm_scale - sglang_scale).mean().item()
     output_diff = torch.abs(vllm_out.float() - sglang_out.float()).mean().item()
 
-    print(f"Scale difference: {scale_diff}")
-    print(f"Output difference: {output_diff}")
-
     if torch.allclose(
         vllm_out.to(torch.float32), sglang_out.to(torch.float32), rtol=1e-3, atol=1e-5
     ) and torch.allclose(vllm_scale, sglang_scale, rtol=1e-3, atol=1e-5):