[Quant Kernel] refactored per token group quant fp8 to support int8 up-to 2x faster (#4396)

sgl-kernel/python/sgl_kernel/__init__.py

@@ -31,6 +31,7 @@ from sgl_kernel.gemm import (
     int8_scaled_mm,
     sgl_per_tensor_quant_fp8,
     sgl_per_token_group_quant_fp8,
+    sgl_per_token_group_quant_int8,
     sgl_per_token_quant_fp8,
 )
 from sgl_kernel.moe import moe_align_block_size, topk_softmax

sgl-kernel/python/sgl_kernel/gemm.py

@@ -96,6 +96,20 @@ def sgl_per_token_group_quant_fp8(
     )
 
 
+def sgl_per_token_group_quant_int8(
+    input: torch.Tensor,
+    output_q: torch.Tensor,
+    output_s: torch.Tensor,
+    group_size: int,
+    eps: float,
+    int8_min: float,
+    int8_max: float,
+) -> None:
+    torch.ops.sgl_kernel.sgl_per_token_group_quant_int8(
+        input, output_q, output_s, group_size, eps, int8_min, int8_max
+    )
+
+
 def sgl_per_tensor_quant_fp8(
     input: torch.Tensor,
     output_q: torch.Tensor,