[Bugfix] Correctly handle the output shape in multimodal attention (#5443)

### What this PR does / why we need it?
Fix https://github.com/vllm-project/vllm-ascend/issues/5297, for
`AscendMMEncoderAttention` forward, we should keep the output shape
consistence with the input

- vLLM version: release/v0.13.0
- vLLM main:
81786c8774

---------

Signed-off-by: wangli <wangli858794774@gmail.com>

tests/e2e/conftest.py

@@ -781,6 +781,11 @@ PROMPT_CONFIGS = {
             "fps": 1,
         },
     },
+    "hunyuan-vl": {
+        "model": "Tencent-Hunyuan/HunyuanOCR",
+        "prompt_fn": hunyuan_prompt,
+        "mm_processor_kwargs": {},
+    },
 }
 
 

vllm_ascend/ops/mm_encoder_attention.py

@@ -93,6 +93,7 @@ class AscendMMEncoderAttention(MMEncoderAttention):
     ):
         bsz, q_len = query.size()[:2]
         kv_len = key.size(1)
+        is_reshaped = query.dim() == 4
 
         # q, k, v: [b, s, head, head_dim] -> [b * s, head, head_dim]
         q, k, v = self.reshape_qkv_to_3d(query, key, value, bsz, q_len, kv_len)
@@ -134,7 +135,12 @@ class AscendMMEncoderAttention(MMEncoderAttention):
         if enable_pad:
             context_layer = context_layer[..., :origin_shape]
 
+        if is_reshaped:
             context_layer = einops.rearrange(context_layer,
                                              "(b s) h d -> b s h d",
                                              b=bsz).contiguous()
+        else:
+            context_layer = einops.rearrange(context_layer,
+                                             "(b s) h d -> b s (h d)",
+                                             b=bsz).contiguous()
         return context_layer