support prefill cache mode use fia op

Signed-off-by: shiyuan680 <[email protected]>

Author: shiyuan680

vllm_ascend/attention/attention_v1.py

@@ -373,18 +373,43 @@ def _forward_prefill_cache_hit(
         batch_size = attn_metadata.query_lens.shape[0]
         block_table = attn_metadata.block_tables[:batch_size, :]
 
-        torch_npu._npu_flash_attention_qlens(
-            query=query,
-            key_cache=self.key_cache,
-            value_cache=self.value_cache,
-            block_table=block_table,
-            mask=compress_mask,
-            seq_len=attn_metadata.query_lens,
-            context_lens=attn_metadata.seq_lens,
-            num_kv_heads=self.num_kv_heads,
-            num_heads=self.num_heads,
-            scale_value=self.scale,
-            out=output)
+        if torch.version.cann.startswith("8.3"):
+            # TODO:The npu_fused_infer_attention_score op is planned to
+            # be utilized in a wider range in upcoming versions.
+            num_block, block_size, _, _ = self.key_cache.shape  # type: ignore
+            key = self.key_cache.view(  # type: ignore
+                num_block, block_size, -1)
+            value = self.value_cache.view(  # type: ignore
+                num_block, block_size, -1)
+
+            output, _ = torch_npu.npu_fused_infer_attention_score(
+                query=query,
+                key=key,
+                value=value,
+                atten_mask=compress_mask,
+                block_table=block_table,
+                input_layout="TND",
+                block_size=block_size,
+                actual_seq_lengths=attn_metadata.actual_seq_lengths_q,
+                actual_seq_lengths_kv=attn_metadata.actual_seq_lengths_q,
+                num_key_value_heads=self.num_kv_heads,
+                num_heads=self.num_heads,
+                scale=self.scale,
+                sparse_mode=3,
+            )
+        else:
+            torch_npu._npu_flash_attention_qlens(
+                query=query,
+                key_cache=self.key_cache,
+                value_cache=self.value_cache,
+                block_table=block_table,
+                mask=compress_mask,
+                seq_len=attn_metadata.query_lens,
+                context_lens=attn_metadata.seq_lens,
+                num_kv_heads=self.num_kv_heads,
+                num_heads=self.num_heads,
+                scale_value=self.scale,
+                out=output)
         return output
 
     def _forward_decode_only(

vllm_ascend/worker/model_runner_v1.py

@@ -907,8 +907,12 @@ def _make_attention_mask(self, seq_lens, position,
                 max_seq_len, self.dtype, self.device)
         # Prefill with cache hit.
         elif attn_state == AscendAttentionState.PrefillCacheHit:
-            return self.attn_mask_builder.get_attn_mask(
-                128, self.dtype, self.device)
+            if torch.version.cann.startswith("8.3"):
+                return ~torch.tril(
+                    torch.ones((2048, 2048), dtype=torch.bool, device=self.device))
+            else:
+                return self.attn_mask_builder.get_attn_mask(
+                    128, self.dtype, self.device)
         # Decode-only situation.
         else:
             return None