[Perf] Add new npu_fused_infer_attention_score op to improve perfomance in splitfuse cases and resolve long-seq mask problems (#2962)

### What this PR does / why we need it?
Add new npu_fused_infer_attention_score op to improve perfomance in
splitfuse cases and resolve long-seq mask problems .

1. The original op's performance is suboptimal in certain scenarios,
necessitating optimization through the _new op_
(npu_fused_infer_attention_score)。
2. For ultra-long sequences (128k), the original operator will allocate
a large attn_mask, which consumes excessive CPU memory. In contrast, the
_new op_ supports a fixed-size compressed mask, effectively resolving
this issue.

NOTE1: The current PR retains the original logic and uses a version
check of the CANN package to determine whether the _new op_ can be
enabled. This ensures no impact on existing users. In future versions,
this version check and the original logic will be deprecated, and the
_new op_ scheduling will be uniformly adopted.
NOTE2: This pr relies on future CANN version, which is not available
now.
NOTE3: To enable the new op in chunked prefill, the parameter
additional_config should be set like `--additional-config
'{"ascend_scheduler_config":
{"enabled":true,"enable_chunked_prefill":true}}' \` at least.

### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
CI passed




- vLLM version: v0.10.2
- vLLM main:
6c5f82e5aa

---------

Signed-off-by: tangtianyi <tangtianyi4@huawei.com>
Signed-off-by: Angazenn <supperccell@163.com>
Co-authored-by: Angazenn <supperccell@163.com>

vllm_ascend/attention/attention_mask.py

@@ -39,11 +39,22 @@ class AttentionMaskBuilder:
         self,
         max_seq_len: int,
         dtype: torch.dtype,
+        device: torch.device = None,
     ):
+        # NOTE: The device argument specifies the target NPU
+        # to be used for the newly added FIA operator.
+        # Only pass this parameter when using the new FIA operator.
+
         attn_mask = _generate_attn_mask(max_seq_len, dtype)
 
         self._seq_len_cached = attn_mask.shape[0]
         self.attn_mask_cache = attn_mask
+        self.device = device
+        if torch.version.cann.startswith("8.3"):
+            assigned_mask_dim = 2048
+            self.chunked_prefill_attn_mask = torch.triu(
+                torch.ones(assigned_mask_dim, assigned_mask_dim),
+                diagonal=1).to(torch.int8).to(device)
 
     @staticmethod
     def get_mask_scale_factor(dtype: torch.dtype = torch.float16):
@@ -66,24 +77,28 @@ class AttentionMaskBuilder:
 
     def get_splitfuse_attn_mask(
         self,
-        seq_lens: torch.Tensor,
+        seq_lens: torch.Tensor = None,
-        position: torch.Tensor,
+        position: torch.Tensor = None,
-        dtype: torch.dtype,
+        dtype: torch.dtype = None,
-        device: torch.device,
+        device: torch.device = None,
     ) -> torch.Tensor:
-        if dtype not in [torch.float16, torch.bfloat16]:
+        if torch.version.cann.startswith("8.3"):
-            raise ValueError(
+            return self.chunked_prefill_attn_mask
-                "splitfuse_attn_mask now only supports bf16 and fp16")
+        else:
-        max_seq_len = max(seq_lens, default=0)
+            if dtype not in [torch.float16, torch.bfloat16]:
-        self._update_attn_cache(max_seq_len, dtype)
+                raise ValueError(
-        # FIXME: Currently the mask value of chunked-prefill situation and Prefill-Only situation
+                    "splitfuse_attn_mask now only supports bf16 and fp16")
-        # is not the same. Fix this in the future when kernel is ready.
+            max_seq_len = max(seq_lens, default=0)
-        mask_scale_factor = AttentionMaskBuilder.get_mask_scale_factor(dtype)
+            self._update_attn_cache(max_seq_len, dtype)
-        attn_mask = torch.index_select(self.attn_mask_cache,
+            # FIXME: Currently the mask value of chunked-prefill situation and Prefill-Only situation
-                                       dim=0,
+            # is not the same. Fix this in the future when kernel is ready.
-                                       index=position)[:, :max_seq_len]
+            mask_scale_factor = AttentionMaskBuilder.get_mask_scale_factor(
-        attn_mask *= mask_scale_factor
+                dtype)
-        return attn_mask.contiguous().to(device, non_blocking=True)
+            attn_mask = torch.index_select(self.attn_mask_cache,
+                                           dim=0,
+                                           index=position)[:, :max_seq_len]
+            attn_mask *= mask_scale_factor
+            return attn_mask.contiguous().to(device, non_blocking=True)
 
     def _update_attn_cache(self, seqlen: int, dtype: torch.dtype):
         if seqlen > self._seq_len_cached:

vllm_ascend/attention/attention_v1.py

@@ -456,18 +456,43 @@ class AscendAttentionBackendImpl(AttentionImpl):
             attn_metadata.seq_lens = \
                 attn_metadata.seq_lens.to(device=query.device)
 
-        torch_npu._npu_paged_attention_splitfuse(
+        if torch.version.cann.startswith("8.3"):
-            query=query,
+            # TODO:The npu_fused_infer_attention_score op is planned to
-            key_cache=self.key_cache,
+            # be utilized in a wider range in upcoming versions.
-            value_cache=self.value_cache,
+            num_block, block_size, _, _ = self.key_cache.shape  # type: ignore
-            mask=attn_metadata.attn_mask,
+            key = self.key_cache.view(  # type: ignore
-            block_table=attn_metadata.block_tables,
+                num_block, block_size, -1)
-            seq_len=attn_metadata.query_lens,
+            value = self.value_cache.view(  # type: ignore
-            context_lens=attn_metadata.seq_lens,
+                num_block, block_size, -1)
-            num_kv_heads=self.num_kv_heads,
+
-            num_heads=self.num_heads,
+            output, _ = torch_npu.npu_fused_infer_attention_score(
-            scale_value=self.scale,
+                query=query,
-            out=output)
+                key=key,
+                value=value,
+                atten_mask=attn_metadata.attn_mask,
+                block_table=attn_metadata.block_tables,
+                input_layout="TND",
+                block_size=block_size,
+                actual_seq_lengths=attn_metadata.query_start_loc[1:],
+                actual_seq_lengths_kv=attn_metadata.seq_lens,
+                num_key_value_heads=self.num_kv_heads,
+                num_heads=self.num_heads,
+                scale=self.scale,
+                sparse_mode=3,
+            )
+        else:
+            torch_npu._npu_paged_attention_splitfuse(
+                query=query,
+                key_cache=self.key_cache,
+                value_cache=self.value_cache,
+                mask=attn_metadata.attn_mask,
+                block_table=attn_metadata.block_tables,
+                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(
@@ -561,12 +586,18 @@ class AscendAttentionBackendImpl(AttentionImpl):
                                                    output)
             # Normal V1 situation.
             else:
+                if torch.version.cann.startswith("8.3"):
+                    # npu_fused_infer_attention_score does not support cases
+                    # where query.shape[0] != attn_metadata.query_start_loc[-1].
+                    # Thus we need unpad it here.
+                    num_tokens = attn_metadata.query_start_loc[-1]
+                    query = query[:num_tokens]
                 output = self._forward_v1_style(query, attn_metadata, output)
 
         # to make in-place change to the output tensor
         if hasattr(layer, 'quant_method') and use_kv_cache_int8:
             output = output.view(num_tokens, self.num_heads, self.head_size)
-        ori_output[:, :, :] = output[:num_tokens, :, :]
+        ori_output[:num_tokens, :, :] = output[:num_tokens, :, :]
         return output.view(num_tokens, self.hidden_size)
 
 

vllm_ascend/worker/model_runner_v1.py

@@ -301,8 +301,13 @@ class NPUModelRunner(LoRAModelRunnerMixin):
             use_mla=self.model_config.use_mla,
         )
 
-        self.attn_mask_builder = AttentionMaskBuilder(
+        if torch.version.cann.startswith("8.3"):
-            self.model_config.max_model_len, self.dtype)
+            self.attn_mask_builder = AttentionMaskBuilder(
+                self.scheduler_config.max_num_batched_tokens, self.dtype,
+                self.device)
+        else:
+            self.attn_mask_builder = AttentionMaskBuilder(
+                self.model_config.max_model_len, self.dtype)
 
         # Set up speculative decoding.
         self.spec_attn_mask = None
@@ -860,8 +865,11 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                              attn_state) -> torch.Tensor:
         # Chunk Prefill situation.
         if attn_state == AscendAttentionState.ChunkedPrefill and not self.vllm_config.model_config.use_mla:
-            return self.attn_mask_builder.get_splitfuse_attn_mask(
+            if torch.version.cann.startswith("8.3"):
-                seq_lens, position, self.dtype, self.device)
+                return self.attn_mask_builder.get_splitfuse_attn_mask()
+            else:
+                return self.attn_mask_builder.get_splitfuse_attn_mask(
+                    seq_lens, position, self.dtype, self.device)
         # Prefill without cache situation.
         elif attn_state == AscendAttentionState.PrefillNoCache:
             max_seq_len = max(seq_lens, default=0)