[Feat][UT] Support Deepseekv32 FULL_DECODE_ONLY mode and add unit test of sfa_v1 (#3763)

### What this PR does / why we need it?
- Add support for DeepSeek v3.2 in FULL_DECODE_ONLY mode.
- Add unit test for sfa_v1.

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?


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

---------

Signed-off-by: 1Fire4 <wangdingyi2@huawei.com>

vllm_ascend/attention/sfa_v1.py

@@ -91,7 +91,7 @@ M = TypeVar("M", bound=AscendSFAMetadata)
 class AscendSFAMetadataBuilder:
     # Does this backend/builder support ACL Graphs for attention (default: no).
     aclgraph_support: ClassVar[AttentionCGSupport] = \
-        AttentionCGSupport.NEVER
+        AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE
     """
     NOTE: Please read the comment at the top of the file before trying to
     understand this class
@@ -189,6 +189,26 @@ class AscendSFAMetadataBuilder:
             sin=sin,
             cos=cos)
 
+    def build_for_graph_capture(
+        self,
+        common_attn_metadata: AscendCommonAttentionMetadata,
+        attn_state: AscendAttentionState = AscendAttentionState.DecodeOnly,
+        model: Optional[nn.Module] = None,
+    ):
+        if attn_state == AscendAttentionState.DecodeOnly:
+            attn_metadata = self.build(
+                common_prefix_len=0,
+                common_attn_metadata=common_attn_metadata,
+                model=model,
+            )
+        else:
+            raise NotImplementedError(
+                "Currently we only support building dummy metadata for DecodeOnly state"
+            )
+
+        attn_metadata.attn_state = attn_state
+        return attn_metadata
+
 
 class AscendSFAImpl(MLAAttentionImpl):
     """

vllm_ascend/worker/model_runner_v1.py

@@ -1894,7 +1894,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
         )
 
         forward_context = get_forward_context()
-        if forward_context.cudagraph_runtime_mode == CUDAGraphMode.FULL:
+        if forward_context.cudagraph_runtime_mode == CUDAGraphMode.FULL \
+            and not self.use_sparse:
             # TODO: maybe_padded_num_tokens will be removed, use num_input_tokens instead
             if self.vllm_config.model_config.use_mla:
                 if self.pcp_size * self.dcp_size > 1:
@@ -2687,11 +2688,15 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                         [0] * dcp_world_size for _ in range(pcp_world_size)
                     ] for _ in range(num_tokens)]
                     long_seq_metadata.num_computed_tokens_of_pcp_dcp = num_computed_tokens_of_pcp_dcp
-                common_attn_metadata = AscendCommonAttentionMetadata(
-                    query_start_loc=torch.tensor(
+                if self.speculative_config:
+                    query_start_loc = torch.tensor(
                         [0] + self.actual_seq_lengths_q[:num_reqs],
                         device=self.device,
-                        dtype=torch.int32),
+                        dtype=torch.int32)
+                else:
+                    query_start_loc = self.query_start_loc[:num_reqs + 1]
+                common_attn_metadata = AscendCommonAttentionMetadata(
+                    query_start_loc=query_start_loc,
                     query_start_loc_cpu=self.query_start_loc_cpu[:num_reqs +
                                                                  1],
                     seq_lens_cpu=self.seq_lens_cpu,
@@ -2737,7 +2742,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
         forward_context = get_forward_context()
         assert forward_context is not None
         if forward_context.cudagraph_runtime_mode == CUDAGraphMode.FULL and \
-            not forward_context.capturing:
+            not forward_context.capturing and not self.use_sparse:
             if self.vllm_config.model_config.use_mla:
                 # FIXME: Try using `auto_dispatch_capture=True`
                 if self.pcp_size * self.dcp_size > 1: