[AMD] Accelerate AITER unified-attention decode with scaled FP8 Q (#31856)

Co-authored-by: jacky.cheng <yichiche@amd.com>
Co-authored-by: YC Yen-Ching Tseng <yctseng@amd.com>

python/sglang/srt/layers/attention/aiter_backend.py

@@ -65,7 +65,10 @@ from sglang.kernels.ops.attention.utils import (
     launch_reshape_and_cache_flash,
     pad_sequence_with_mask,
 )
-from sglang.kernels.ops.quantization.fp8_kernel import fp8_dtype
+from sglang.kernels.ops.quantization.fp8_kernel import (
+    fp8_dtype,
+    scaled_fp8_quant,
+)
 from sglang.srt.configs.model_config import AttentionArch
 from sglang.srt.layers.attention.aiter_utils import (
     forward_decode_vectorized_5d,
@@ -2605,6 +2608,12 @@ class AiterAttnBackend(AttentionBackend):
                         page_table = self.forward_metadata.swa_page_table
 
                 max_kv_len = page_table.shape[1] * self.page_size
+                q_descale = None
+                if self.kv_cache_dtype == fp8_dtype:
+                    q_descale = (
+                        layer.k_scale if layer.k_scale is not None else self.k_scale
+                    )
+                    q, _ = scaled_fp8_quant(q, q_descale)
 
                 unified_attention(
                     q=q.view(-1, layer.tp_q_head_num, layer.qk_head_dim),
@@ -2624,7 +2633,7 @@ class AiterAttnBackend(AttentionBackend):
                     window_size=window_size,
                     block_table=page_table,
                     softcap=0,
-                    q_descale=None,
+                    q_descale=q_descale,
                     k_descale=k_descale,
                     v_descale=v_descale,
                     sinks=sinks,