use sgl_per_token_group_quant_fp8 kernel (#3493)

python/pyproject.toml

@@ -25,7 +25,7 @@ runtime_common = [
 ]
 srt = [
     "sglang[runtime_common]", "cuda-python",
-    "sgl-kernel>=0.0.3.post3", "torch", "vllm>=0.6.4.post1,<=0.7.2",
+    "sgl-kernel>=0.0.3.post4", "torch", "vllm>=0.6.4.post1,<=0.7.2",
     "flashinfer_python>=0.2.0.post2", "outlines>=0.0.44,<=0.1.11"
 ]
 

python/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py

@@ -33,6 +33,10 @@ _is_rocm = torch.cuda.is_available() and torch.version.hip
 if _is_cuda:
     from sgl_kernel import gelu_and_mul, silu_and_mul
 
+    from sglang.srt.layers.quantization.fp8_kernel import (
+        sglang_per_token_group_quant_fp8,
+    )
+
 if _is_cuda or _is_rocm:
     from sgl_kernel import moe_align_block_size as sgl_moe_align_block_size
 
@@ -488,7 +492,10 @@ def invoke_fused_moe_kernel(
         else:
             assert len(block_shape) == 2
             block_n, block_k = block_shape[0], block_shape[1]
-            A, A_scale = per_token_group_quant_fp8(A, block_k)
+            if _is_cuda:
+                A, A_scale = sglang_per_token_group_quant_fp8(A, block_k)
+            else:
+                A, A_scale = per_token_group_quant_fp8(A, block_k)
             assert triton.cdiv(A.shape[-1], block_k) == A_scale.shape[-1]
             assert triton.cdiv(B.shape[-2], block_n) == B_scale.shape[-2]
             assert triton.cdiv(B.shape[-1], block_k) == B_scale.shape[-1]

python/sglang/srt/layers/quantization/fp8_kernel.py

@@ -27,6 +27,10 @@ from sglang.srt.utils import get_device_core_count, get_device_name, is_hip
 is_hip_ = is_hip()
 fp8_type_ = torch.float8_e4m3fnuz if is_hip_ else torch.float8_e4m3fn
 
+_is_cuda = torch.cuda.is_available() and torch.version.cuda
+if _is_cuda:
+    from sgl_kernel import sgl_per_token_group_quant_fp8
+
 logger = logging.getLogger(__name__)
 
 
@@ -135,6 +139,36 @@ def per_token_group_quant_fp8(
     return x_q, x_s
 
 
+def sglang_per_token_group_quant_fp8(
+    x: torch.Tensor,
+    group_size: int,
+    eps: float = 1e-10,
+    dtype: torch.dtype = fp8_type_,
+):
+    assert (
+        x.shape[-1] % group_size == 0
+    ), "the last dimension of `x` cannot be divisible by `group_size`"
+    assert x.is_contiguous(), "`x` is not contiguous"
+
+    finfo = torch.finfo(dtype)
+    fp8_max = finfo.max
+
+    fp8_min = -fp8_max
+
+    x_q = torch.empty_like(x, device=x.device, dtype=dtype)
+    M = x.numel() // group_size
+    N = group_size
+    x_s = torch.empty(
+        x.shape[:-1] + (x.shape[-1] // group_size,),
+        device=x.device,
+        dtype=torch.float32,
+    )
+
+    sgl_per_token_group_quant_fp8(x, x_q, x_s, group_size, eps, fp8_min, fp8_max)
+
+    return x_q, x_s
+
+
 @triton.jit
 def _w8a8_block_fp8_matmul(
     # Pointers to inputs and output