[ROCm][Bugfix] Fix RuntimeError in MMEncoderAttention by replacing .view() with .reshape() (vllm-project#31203)

Signed-off-by: Andreas Karatzas <akaratza@amd.com>

Author: AndreasKaratzas

tests/models/multimodal/conftest.py

@@ -19,7 +19,7 @@ def pytest_collection_modifyitems(config, items):
         return
 
     # Disable Flash/MemEfficient SDP on ROCm to avoid HF Transformers
-    # accuracy issues
+    # accuracy issues: https://github.com/vllm-project/vllm/issues/30167
     # TODO: Remove once ROCm SDP accuracy issues are resolved on HuggingFace
     torch.backends.cuda.enable_flash_sdp(False)
     torch.backends.cuda.enable_mem_efficient_sdp(False)

vllm/attention/layers/mm_encoder_attention.py

@@ -136,7 +136,7 @@ def _forward_sdpa(
             cu_seqlens=cu_seqlens,
         )
         if is_reshaped:
-            output = output.view(bsz, q_len, -1)
+            output = output.reshape(bsz, q_len, -1)
         return output
 
     def _forward_fa(
@@ -174,7 +174,7 @@ def _forward_fa(
             fa_version=self._fa_version,
         )
         if is_reshaped:
-            output = output.view(bsz, q_len, -1)
+            output = output.reshape(bsz, q_len, -1)
         return output
 
     def forward_native(