1、kvcache支持fp8的scale

vllm/v1/attention/backends/flash_attn.py

@@ -700,8 +700,9 @@ class FlashAttentionImpl(AttentionImpl):
                 prefix_scheduler_metadata=attn_metadata.prefix_scheduler_metadata,
                 suffix_scheduler_metadata=attn_metadata.scheduler_metadata,
                 # q_descale=layer._q_scale,
-                # k_descale=layer._k_scale,
-                # v_descale=layer._v_scale,
+                q_descale=None,
+                k_descale=layer._k_scale,
+                v_descale=layer._v_scale,
             )
         return output
 
@@ -779,6 +780,9 @@ class FlashAttentionImpl(AttentionImpl):
                 # q_descale=layer._q_scale.expand(descale_shape),
                 # k_descale=layer._k_scale.expand(descale_shape),
                 # v_descale=layer._v_scale.expand(descale_shape),
+                q_descale=None,
+                k_descale=layer._k_scale,
+                v_descale=layer._v_scale,
                 is_prefix_cache=False,
             )
 
@@ -932,12 +936,12 @@ def cascade_attention(
             return_softmax_lse=True,
             scheduler_metadata=prefix_scheduler_metadata,
             # fa_version=fa_version,
-            # q_descale=q_descale.expand(descale_shape)
-            # if q_descale is not None else None,
-            # k_descale=k_descale.expand(descale_shape)
-            # if k_descale is not None else None,
-            # v_descale=v_descale.expand(descale_shape)
-            # if v_descale is not None else None,
+            q_descale=q_descale.expand(descale_shape)
+            if q_descale is not None else None,
+            k_descale=k_descale.expand(descale_shape)
+            if k_descale is not None else None,
+            v_descale=v_descale.expand(descale_shape)
+            if v_descale is not None else None,
             is_prefix_cache=True,
         )
 
@@ -985,12 +989,12 @@ def cascade_attention(
             return_softmax_lse=True,
             scheduler_metadata=suffix_scheduler_metadata,
             # fa_version=fa_version,
-            # q_descale=q_descale.expand(descale_shape)
-            # if q_descale is not None else None,
-            # k_descale=k_descale.expand(descale_shape)
-            # if k_descale is not None else None,
-            # v_descale=v_descale.expand(descale_shape)
-            # if v_descale is not None else None,
+            q_descale=q_descale.expand(descale_shape)
+            if q_descale is not None else None,
+            k_descale=k_descale.expand(descale_shape)
+            if k_descale is not None else None,
+            v_descale=v_descale.expand(descale_shape)
+            if v_descale is not None else None,
             is_prefix_cache=True,
         )