qwen2vl and qwen2.5vl contiguous on rocm and torch.sdpa

Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
Signed-off-by: JartX <sagformas@epdcenter.es>

Author: JartX

vllm/model_executor/models/qwen2_5_vl.py

@@ -423,6 +423,11 @@ def forward(
             ).contiguous()
         elif self.attn_backend == _Backend.TORCH_SDPA:
             # Execute attention entry by entry for speed & less VRAM.
+            from vllm.platforms import current_platform
+            if current_platform.is_rocm():
+                q = q.contiguous()
+                k = k.contiguous()
+                v = v.contiguous()
             outputs = []
             for i in range(1, len(cu_seqlens)):
                 start_idx = cu_seqlens[i - 1]

vllm/model_executor/models/qwen2_vl.py

@@ -453,6 +453,11 @@ def forward(
             ).contiguous()
         elif self.attn_backend == _Backend.TORCH_SDPA:
             # Execute attention entry by entry for speed & less VRAM.
+            from vllm.platforms import current_platform
+            if current_platform.is_rocm():
+                q = q.contiguous()
+                k = k.contiguous()
+                v = v.contiguous()
             outputs = []
             for i in range(1, len(cu_seqlens)):
                 start_idx = cu_seqlens[i - 1]