Fix GPU kernel ordering and MXFP8 quantization dispatch (#37331)

Co-authored-by: Pranjal Shankhdhar <pranjalssh@meta.com>
Co-authored-by: Chengze Fan <fancz2002@gmail.com>

python/sglang/kernels/jit/include/sgl_kernel/utils.cuh

@@ -167,7 +167,10 @@ template <bool kUsePDL>
 SGL_DEVICE void PDLTriggerSecondary() {
 #if SGL_ARCH_HOPPER_OR_GREATER
   if constexpr (kUsePDL) {
-    asm volatile("griddepcontrol.launch_dependents;" :::);
+    // The "memory" clobber is load-bearing: without it the compiler may sink
+    // this kernel's stores past the trigger, and the dependent grid's
+    // griddepcontrol.wait only covers writes issued BEFORE launch_dependents.
+    asm volatile("griddepcontrol.launch_dependents;" ::: "memory");
   }
 #endif
 }

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -1293,6 +1293,12 @@ def mxfp8_group_quantize(x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
     assert x.is_contiguous(), "MXFP8 quantization requires a contiguous 2D tensor."
     _, k = x.shape
     assert k % 32 == 0, f"{k=} must be divisible by 32"
+    if _is_hip and _is_gfx95_supported:
+        from sglang.kernels.ops.quantization.mxfp8_amd_gfx95 import (
+            mxfp8_e4m3_quantize,
+        )
+
+        return mxfp8_e4m3_quantize(x)
     downcast_to_mxfp = _get_triton_mxfp8_downcast()
     q_input, scale_u8 = downcast_to_mxfp(x, torch.float8_e4m3fn, axis=1)
     return q_input.contiguous(), scale_u8.contiguous()