[CP&SP] Integrate FIA operator in mla_cp._forward_decode (#5641)

### What this PR does / why we need it?
Replace the npu_multi_head_latent_attention with FIA operator in
mla_cp.py _forward_decode.
Adjust mla_attn_dpc_pcp in acl_graph.py

### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?

- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef

---------

Signed-off-by: 白永斌 <baiyongbin3@h-partners.com>
Signed-off-by: Bai Yongbin <845473182@qq.com>
Signed-off-by: tongyuzhou <t00886357@china.huawei.com>
Co-authored-by: 白永斌 <baiyongbin3@h-partners.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: tongyuzhou <t00886357@china.huawei.com>

tests/ut/attention/test_mla_cp.py

@@ -450,11 +450,11 @@ class TestAscendMLAImpl(TestBase):
         self.assertEqual(result.shape[2], self.impl.kv_lora_rank + 1)
 
     @patch('vllm_ascend.attention.context_parallel.mla_cp.get_forward_context')
-    @patch("torch_npu.atb.npu_multi_head_latent_attention")
+    @patch("torch_npu.npu_fused_infer_attention_score")
     @patch('torch_npu.npu_attention_update')
     @patch_distributed_groups(dcp_size=2, pcp_size=2, needs_mocks=False)
     def test_forward_decode_pcp_dcp(self, mock_npu_attention_update,
-                                    mock_npu_multi_head_latent_attention,
+                                    mock_npu_fused_infer_attention_score,
                                     mock_get_forward_context):
         self.impl.dcp_size = 2
         self.impl.pcp_size = 2
@@ -470,8 +470,8 @@ class TestAscendMLAImpl(TestBase):
 
         q_nope = torch.randn(B, N, self.impl.qk_nope_head_dim)
         q_pe = torch.randn(B, N, self.impl.qk_rope_head_dim)
-        k_nope = torch.randn(NB, BS, 1, self.impl.kv_lora_rank)
+        k_nope = torch.randn(NB, 1, BS, self.impl.kv_lora_rank)
-        k_pe = torch.randn(NB, BS, 1, self.impl.qk_rope_head_dim)
+        k_pe = torch.randn(NB, 1, BS, self.impl.qk_rope_head_dim)
 
         attn_metadata = MagicMock()
         attn_metadata.attn_state = AscendAttentionState.SpecDecoding
@@ -485,7 +485,7 @@ class TestAscendMLAImpl(TestBase):
 
         mock_npu_attention_update.return_value = (torch.randn(
             B, self.impl.num_heads, self.impl.kv_lora_rank), None)
-        mock_npu_multi_head_latent_attention.return_value = [
+        mock_npu_fused_infer_attention_score.return_value = [
             torch.randn(B, N, self.impl.kv_lora_rank),
             torch.randn(B, N, 1)
         ]

tests/ut/compilation/test_acl_graph.py

@@ -754,7 +754,7 @@ class TestPCPDCPGraphParams(TestBase):
 
     @patch('torch.npu.graph_task_update_end', )
     @patch('torch.npu.graph_task_update_begin', MagicMock())
-    @patch('torch_npu.atb.npu_multi_head_latent_attention', MagicMock())
+    @patch('torch_npu.npu_fused_infer_attention_score.out', MagicMock())
     def test_update_mla_dcp_pcp_params(self, _mock_graph_task_end):
         input_positions = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8])
         block_table = torch.zeros(2, 5, dtype=torch.long)
@@ -793,16 +793,20 @@ class TestPCPDCPGraphParams(TestBase):
         qk_rope_head_dim = 32
         qk_nope_head_dim = 64
         query = torch.randn(4, num_heads, qk_head_dim)
-        q_pe = query[..., qk_nope_head_dim:]
+
         q_nope = query[..., :qk_nope_head_dim]
+        q_pe = query[..., qk_rope_head_dim:]
         k_nope = torch.randn(4, num_heads, qk_nope_head_dim)
         k_pe = torch.randn(4, num_heads, qk_rope_head_dim)
+        input_layout = "BNSD"
+        actual_seq_lengths_kv = [1, 1]
         out = torch.randn(2, 16, 128)
         lse = torch.randn(2, 16, 8)
         self.graph_params.attn_params[4] = []
         self.graph_params.attn_params[4].append(
-            (q_nope, q_pe, k_nope, k_pe, block_table, seq_lens, num_heads,
+            (q_nope, k_nope, q_pe, k_pe, num_heads, num_kv_heads, input_layout,
-             scale, num_kv_heads, out, lse))
+             None, 0, scale, block_table, 128, None, actual_seq_lengths_kv,
+             out, lse))
 
         with patch("torch_npu._C._npu_setStream", return_value=None):
             update_mla_attn_dcp_pcp_params(self.update_stream, forward_context,

vllm_ascend/attention/context_parallel/mla_cp.py

@@ -14,6 +14,8 @@ from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.utils.math_utils import cdiv
 from vllm.v1.kv_cache_interface import AttentionSpec, MLAAttentionSpec
 
+from vllm_ascend.attention.attention_v1 import AscendAttentionState
+
 # isort: off
 from vllm_ascend.attention.mla_v1 import (
     AscendMLADecodeMetadata,
@@ -244,8 +246,12 @@ class AscendMlaCPMetadataBuilder(AscendMLAMetadataBuilder):
         self.batch_seq_mask_buf[: batch_seq_mask.shape[0]].copy_(batch_seq_mask, non_blocking=True)
         batch_seq_mask = self.batch_seq_mask_buf[: batch_seq_mask.shape[0]]
         cp_seq_len = torch.where(cp_seq_len == 0, 1, cp_seq_len)
-        decode_metadata.cp_seq_len = cp_seq_len
+        decode_metadata.cp_seq_len = cp_seq_len.tolist()
         decode_metadata.batch_seq_mask = batch_seq_mask
+
+        actual_seq_lengths_q = torch.arange(self.num_decodes_flatten) + 1
+        decode_metadata.actual_seq_lengths_q = actual_seq_lengths_q
+
         return decode_metadata
 
 
@@ -535,18 +541,53 @@ class AscendMlaCPImpl(AscendMLAImpl):
             num_heads = self.num_heads * self.dcp_size
         else:
             num_heads = self.num_heads
-
-        k_nope = k_nope.view(-1, block_size, self.num_kv_heads, self.kv_lora_rank)
-        k_pe = k_pe.view(-1, block_size, self.num_kv_heads, self.qk_rope_head_dim)
-        q_nope = q_nope.view(num_tokens, num_heads, -1)
-        q_pe = q_pe.view(num_tokens, num_heads, -1)
         # use pcp & dcp split computed token nums from scheduler to compute actual seq_len and seq_mask
-        seq_len = decode_meta.cp_seq_len
+        k_nope = k_nope.view(-1, self.num_kv_heads, block_size, self.kv_lora_rank)
+        k_pe = k_pe.view(-1, self.num_kv_heads, block_size, self.qk_rope_head_dim)
+
+        actual_seq_lengths = None
+        input_layout = "BNSD"
+
+        if (
+            attn_metadata.attn_state
+            in [
+                AscendAttentionState.SpecDecoding,
+                AscendAttentionState.ChunkedPrefill,
+                AscendAttentionState.DecodeOnly,
+            ]
+            and self.speculative_config is not None
+        ):
+            input_layout = "TND"
+            # TODO: If the driver is upgraded later, the contiguous function can be deleted.
+            q_nope = q_nope.view(num_tokens, num_heads, -1).contiguous()
+            q_pe = q_pe.view(num_tokens, num_heads, -1)
+            sparse_mode = 3
+            spec_attn_mask = attn_metadata.decode.attn_mask  # type:ignore
+            actual_seq_lengths = decode_meta.actual_seq_lengths_q
+        else:
+            q_nope = q_nope.view(num_tokens, num_heads, 1, -1).contiguous()
+            q_pe = q_pe.view(num_tokens, num_heads, 1, -1)
+            sparse_mode = 0
+            spec_attn_mask = None
 
         common_kwargs = {
-            "return_lse": True,
+            "query_rope": q_pe,
-            "calc_type": "calc_type_ring",
+            "key_rope": k_pe,
+            "num_heads": num_heads,
+            "num_key_value_heads": self.num_kv_heads,
+            "input_layout": input_layout,
+            "atten_mask": spec_attn_mask,
+            "sparse_mode": sparse_mode,
+            "scale": self.scale,
+            "antiquant_mode": 0,
+            "antiquant_scale": None,
+            "block_table": decode_meta.block_table,
+            "block_size": block_size,
+            "actual_seq_lengths": actual_seq_lengths,
+            "actual_seq_lengths_kv": decode_meta.cp_seq_len,
+            "softmax_lse_flag": True,
         }
+
         forward_context: ForwardContext = get_forward_context()
         if forward_context.is_draft_model:
             graph_params = get_draft_graph_params()
@@ -560,72 +601,58 @@ class AscendMlaCPImpl(AscendMLAImpl):
             graph_params.events[num_tokens].append(event)
             workspace = graph_params.workspaces.get(num_tokens)
             if workspace is None:
-                workspace = torch_npu.atb._npu_multi_head_latent_attention_get_workspace(
+                workspace = torch_npu._npu_fused_infer_attention_score_get_max_workspace(
                     q_nope,
-                    q_pe,
                     k_nope,
-                    k_pe,
+                    k_nope,
-                    decode_meta.block_table,
-                    seq_len,
-                    num_heads,
-                    self.scale,
-                    self.num_kv_heads,
                     **common_kwargs,
                 )
                 update_graph_params_workspaces(num_tokens, workspace)
             attn_output = torch.empty_like(q_nope)
-            softmax_lse = torch.empty((num_tokens, num_heads, 1), dtype=q_nope.dtype, device=q_nope.device)
+            if input_layout == "BNSD":
+                softmax_lse = torch.empty((num_tokens, num_heads, 1, 1), dtype=torch.float, device=q_nope.device)
+            else:
+                softmax_lse = torch.empty((num_tokens, num_heads, 1), dtype=torch.float, device=q_nope.device)
+
             graph_params.attn_params[num_tokens].append(
                 (
                     weak_ref_tensors(q_nope),
-                    weak_ref_tensors(q_pe),
                     weak_ref_tensors(k_nope),
+                    weak_ref_tensors(q_pe),
                     weak_ref_tensors(k_pe),
-                    decode_meta.block_table,
-                    seq_len,
                     num_heads,
-                    self.scale,
                     self.num_kv_heads,
+                    input_layout,
+                    weak_ref_tensors(spec_attn_mask) if spec_attn_mask is not None else None,
+                    sparse_mode,
+                    self.scale,
+                    weak_ref_tensors(decode_meta.block_table),
+                    block_size,
+                    actual_seq_lengths,
+                    decode_meta.cp_seq_len,
                     weak_ref_tensors(attn_output),
                     weak_ref_tensors(softmax_lse),
                 )
             )
             torch.npu.graph_task_group_begin(stream)
-            torch_npu.atb.npu_multi_head_latent_attention(
+            torch_npu.npu_fused_infer_attention_score.out(
-                q_nope,
+                q_nope, k_nope, k_nope, **common_kwargs, workspace=workspace, out=[attn_output, softmax_lse]
-                q_pe,
-                k_nope,
-                k_pe,
-                decode_meta.block_table,
-                seq_len,
-                num_heads,
-                self.scale,
-                self.num_kv_heads,
-                **common_kwargs,
-                workspace=workspace,
-                output=attn_output,
-                lse=softmax_lse,
             )
             handle = torch.npu.graph_task_group_end(stream)
             graph_params.handles[num_tokens].append(handle)
         else:
-            attn_output = torch.empty_like(q_nope)
+            attn_output, softmax_lse = torch_npu.npu_fused_infer_attention_score(
-            softmax_lse = torch.empty((num_tokens, num_heads, 1), dtype=q_nope.dtype, device=q_nope.device)
-            torch_npu.atb.npu_multi_head_latent_attention(
                 q_nope,
-                q_pe,
                 k_nope,
-                k_pe,
+                k_nope,
-                decode_meta.block_table,
+                **common_kwargs,
-                seq_len,
-                num_heads,
-                self.scale,
-                self.num_kv_heads,
-                return_lse=True,
-                calc_type="calc_type_ring",
-                output=attn_output,
-                lse=softmax_lse,
             )
+        if input_layout == "BNSD":
+            B_attn, N_attn, S, D = attn_output.shape
+            B_lse, N_lse, Q_S, _ = softmax_lse.shape
+
+            attn_output = attn_output.permute(0, 2, 1, 3).reshape(B_attn * S, N_attn, D)
+            softmax_lse = softmax_lse.permute(0, 2, 1, 3).reshape(B_lse * Q_S, N_lse, 1)
 
         # Update out&lse
         attn_out_lse = _process_attn_out_lse(attn_output, softmax_lse, decode_meta.batch_seq_mask)

vllm_ascend/compilation/acl_graph.py

@@ -468,45 +468,56 @@ def update_mla_attn_dcp_pcp_params(update_stream, forward_context, runtime_shape
         ):
             (
                 q_nope,
-                q_pe,
                 k_nope,
+                q_pe,
                 k_pe,
-                block_table,
-                seq_len,
                 num_heads,
-                scale,
                 num_kv_heads,
+                input_layout,
+                spec_attn_mask,
+                sparse_mode,
+                scale,
+                block_table,
+                block_size,
+                actual_seq_lengths,
+                actual_seq_lengths_kv,
                 attn_output,
                 softmax_lse,
             ) = param
 
             decode_meta = forward_context.attn_metadata[key].decode
             seq_len = decode_meta.cp_seq_len
+            if isinstance(seq_len, torch.Tensor):
+                seq_len = seq_len.tolist()
+            actual_seq_lengths_kv = seq_len
 
-            # For pcp + spec decode, we flatten seq_lens
+            pad_length = runtime_shape - len(actual_seq_lengths_kv)
-            # to avoid irregular attn_mask shape,
+            if pad_length > 0:
-            # so there's no need to divide runtime_shape by spec_multiple
+                actual_seq_lengths_kv = actual_seq_lengths_kv + [0] * (runtime_shape - len(actual_seq_lengths_kv))
-            pad_length = runtime_shape - len(seq_len)
-            pad_tensor = torch.zeros(pad_length, dtype=seq_len.dtype, device=seq_len.device)
-            seq_len = torch.cat([seq_len, pad_tensor], dim=0)
 
             torch.npu.graph_task_update_begin(update_stream, handle)
 
-            torch_npu.atb.npu_multi_head_latent_attention(
+            torch_npu.npu_fused_infer_attention_score.out(
                 q_nope,
-                q_pe,
                 k_nope,
-                k_pe,
+                k_nope,
-                block_table,
+                query_rope=q_pe,
-                seq_len,
+                key_rope=k_pe,
-                num_heads,
+                num_heads=num_heads,
-                scale,
+                num_key_value_heads=num_kv_heads,
-                num_kv_heads,
+                input_layout=input_layout,
-                return_lse=True,
+                atten_mask=spec_attn_mask,
-                calc_type="calc_type_ring",
+                sparse_mode=sparse_mode,
+                scale=scale,
+                antiquant_mode=0,
+                antiquant_scale=None,
+                softmax_lse_flag=True,
+                block_table=block_table,
+                block_size=block_size,
+                actual_seq_lengths_kv=actual_seq_lengths_kv,
+                actual_seq_lengths=actual_seq_lengths,
                 workspace=graph_params.workspaces.get(runtime_shape),
-                output=attn_output,
+                out=[attn_output, softmax_lse],
-                lse=softmax_lse,
             )
             torch.npu.graph_task_update_end(update_stream)