[1/2] Optimizations and refactors about quant kernel (#9534)

python/sglang/srt/layers/quantization/int8_kernel.py

@@ -12,7 +12,13 @@ from sglang.srt.utils import get_device_name, is_cuda
 
 _is_cuda = is_cuda()
 if _is_cuda:
-    from sgl_kernel import sgl_per_token_group_quant_int8
+    # Temporary
+    try:
+        from sgl_kernel import sgl_per_token_group_quant_8bit
+    except ImportError:
+        from sgl_kernel import (
+            sgl_per_token_group_quant_int8 as sgl_per_token_group_quant_8bit,
+        )
 
 logger = logging.getLogger(__name__)
 
@@ -204,7 +210,7 @@ def sglang_per_token_group_quant_int8(
         dtype=torch.float32,
     )
 
-    sgl_per_token_group_quant_int8(x, x_q, x_s, group_size, eps, int8_min, int8_max)
+    sgl_per_token_group_quant_8bit(x, x_q, x_s, group_size, eps, int8_min, int8_max)
 
     return x_q, x_s