revert TND modify when dcp pcp (#3948)

### What this PR does / why we need it?
1、revert TND modify when dcp pcp, which is introduced by
f57bdb09fc
2、deal aclgraph pad border issue

- vLLM version: v0.11.0
- vLLM main:
83f478bb19

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>

vllm_ascend/attention/attention_v1.py

@@ -869,6 +869,7 @@ class AscendAttentionBackendImpl(AttentionImpl):
         else:
             num_heads = self.num_heads
 
+        q_nope = query.view(query.shape[0], 1, query.shape[1], query.shape[2])
         k_nope = self.key_cache.view(self.key_cache.shape[0],
                                      self.key_cache.shape[1], -1)
         value = self.value_cache.view(self.key_cache.shape[0],
@@ -879,7 +880,7 @@ class AscendAttentionBackendImpl(AttentionImpl):
             'num_key_value_heads':
             self.num_kv_heads,
             'input_layout':
-            "TND",
+            "BSND",
             'atten_mask':
             None,
             'scale':
@@ -895,14 +896,12 @@ class AscendAttentionBackendImpl(AttentionImpl):
             'block_size':
             self.key_cache.shape[1],
             'actual_seq_lengths_kv':
-            attn_metadata.seq_lens_list[:attn_metadata.num_decode_tokens],
-            'actual_seq_lengths':
-            attn_metadata.actual_seq_lengths_q[:attn_metadata.
-                                               num_decode_tokens]
+            attn_metadata.decode_meta.
+            num_computed_tokens_of_pcp_dcp[:, self.pcp_rank, self.dcp_rank],
         }
         graph_params = get_graph_params()
         forward_context: ForwardContext = get_forward_context()
-        num_tokens = query.shape[0]
+        num_tokens = q_nope.shape[0]
         if forward_context.capturing:
             stream = torch_npu.npu.current_stream()
 
@@ -914,16 +913,16 @@ class AscendAttentionBackendImpl(AttentionImpl):
             workspace = graph_params.workspaces.get(num_tokens)
             if workspace is None:
                 workspace = torch_npu._npu_fused_infer_attention_score_get_max_workspace(
-                    query, k_nope, value, **common_kwargs)
+                    q_nope, k_nope, value, **common_kwargs)
                 update_graph_params_workspaces(num_tokens,
                                                weak_ref_tensors(workspace))
-            attn_out = torch.empty_like(query)
+            attn_out = torch.empty_like(q_nope)
             attn_lse = torch.empty((num_tokens, num_heads, 1, 1),
                                    dtype=torch.float,
-                                   device=query.device)
+                                   device=q_nope.device)
 
             graph_params.attn_params[num_tokens].append(
-                (weak_ref_tensors(query), weak_ref_tensors(k_nope),
+                (weak_ref_tensors(q_nope), weak_ref_tensors(k_nope),
                  weak_ref_tensors(value), self.num_heads, self.num_kv_heads,
                  self.scale, attn_metadata.block_tables,
                  self.key_cache.shape[1], attn_metadata.decode_meta.
@@ -933,7 +932,7 @@ class AscendAttentionBackendImpl(AttentionImpl):
                  self.pcp_rank, self.dcp_rank, self.dcp_size))
             torch.npu.graph_task_group_begin(stream)
             torch_npu.npu_fused_infer_attention_score.out(
-                query,
+                q_nope,
                 k_nope,
                 value,
                 **common_kwargs,
@@ -943,7 +942,11 @@ class AscendAttentionBackendImpl(AttentionImpl):
             graph_params.handles[num_tokens].append(handle)
         else:
             attn_out, attn_lse = torch_npu.npu_fused_infer_attention_score(
-                query, k_nope, value, **common_kwargs)
+                q_nope, k_nope, value, **common_kwargs)
+
+        attn_out = attn_out.view(attn_out.shape[0], attn_out.shape[2],
+                                 attn_out.shape[3])
+        attn_lse = attn_lse.view(attn_lse.shape[0], attn_lse.shape[1], 1)
 
         attn_out_lse_list = []
         # Concat out&lse: [bs,num_heads,v_head_dim] + [bs,num_heads,1] -> [bs,num_heads,v_head_dim+1]
@@ -1017,7 +1020,8 @@ class AscendAttentionBackendImpl(AttentionImpl):
                     prefill_query, key, value, attn_metadata,
                     output[num_decode_tokens:], prefill_query.shape[0])
                 attn_metadata.seq_lens = seq_lens_back
-            output[num_decode_tokens:] = output_prefill
+            output[num_decode_tokens:output_prefill.shape[0] +
+                   num_decode_tokens] = output_prefill
         return output
 
     def forward(
@@ -1089,7 +1093,9 @@ class AscendAttentionBackendImpl(AttentionImpl):
             if has_prefill:
                 if self.pcp_size > 1:
                     kv = torch.cat([key, value], dim=-1)
-                    all_kv = get_pcp_group().all_gather(kv, dim=0)
+                    num_actual_tokens_pcp_padded = attn_metadata.num_actual_tokens_pcp_padded // self.pcp_size
+                    all_kv = get_pcp_group().all_gather(
+                        kv[:num_actual_tokens_pcp_padded].contiguous(), dim=0)
                     pcp_allgather_restore_idx = attn_metadata.prefill.pcp_allgather_restore_idx if attn_metadata.prefill else None
                     all_kv = torch.index_select(all_kv, 0,
                                                 pcp_allgather_restore_idx)

vllm_ascend/compilation/acl_graph.py

@@ -301,9 +301,9 @@ def update_attn_dcp_pcp_params(update_stream, forward_context, runtime_shape):
         ):
             (q_nope, k_nope, value, num_heads, num_kv_heads, scale,
              block_table, block_size, actual_seq_lengths_kv, attn_output,
-             softmax_lse, cp_rank, dcp_rank, dcp_size) = param
+             softmax_lse, pcp_rank, dcp_rank, dcp_size) = param
             actual_seq_lengths_kv = forward_context.attn_metadata[
-                key].decode_meta.num_computed_tokens_of_pcp_dcp[:, cp_rank,
+                key].decode_meta.num_computed_tokens_of_pcp_dcp[:, pcp_rank,
                                                                 dcp_rank]
             pad_length = runtime_shape - len(actual_seq_lengths_kv)
             pad_tensor = np.zeros(pad_length,

vllm_ascend/worker/model_runner_v1.py

@@ -476,6 +476,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
         self.pcp_padded_slot_mapping = torch.zeros(self.max_num_tokens,
                                                    dtype=torch.int32,
                                                    device=self.device)
+        self.num_actual_tokens_pcp_padded = 0
         if self.speculative_config and self.pcp_size > 1:
             self.input_ids_pcp_full = torch.zeros(self.max_num_tokens,
                                                   dtype=torch.int32,
@@ -1915,7 +1916,9 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                 hidden_states = hidden_states[:-pad_size, :]
 
         if self.pcp_size > 1:
-            hidden_states = get_pcp_group().all_gather(hidden_states, 0)
+            hidden_states = get_pcp_group().all_gather(
+                hidden_states[:self.num_actual_tokens_pcp_padded //
+                              self.pcp_size], 0)
             hidden_states = torch.index_select(
                 hidden_states, 0,
                 self.pcp_allgather_restore_idx[:hidden_states.shape[0]])
@@ -4304,6 +4307,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
         num_decodes = sum(self.input_batch.num_computed_tokens_cpu[:num_reqs]
                           >= self.input_batch.num_prompt_tokens[:num_reqs])
         num_actual_tokens_pcp_padded = total_num_scheduled_tokens * self.pcp_size
+        self.num_actual_tokens_pcp_padded = num_actual_tokens_pcp_padded
         long_seq_metadata = None
         if self.pcp_size * self.dcp_size > 1:
             long_seq_metadata = AscendPrefillContextParallelMetadata(