[ROCm][Bugfix] Fix Mamba batched decode producing incorrect output (#32099)

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

vllm/model_executor/models/plamo2.py

@@ -63,6 +63,7 @@ from vllm.model_executor.models.utils import (
     maybe_prefix,
 )
 from vllm.model_executor.utils import set_weight_attrs
+from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
 from vllm.utils.torch_utils import direct_register_custom_op
 from vllm.v1.attention.backend import AttentionMetadata
@@ -414,6 +415,13 @@ class Plamo2MambaMixer(MambaBase, CustomOp):
                 conv_state_indices=state_indices_tensor_d,
             )
 
+            # ROCm: Ensure contiguous tensor for bcdt_proj linear layer.
+            # causal_conv1d_update returns a non-contiguous view (stride 8192
+            # instead of 4096 for shape [batch, 4096]), causing incorrect GEMM
+            # results when batch > 1 on ROCm.
+            if current_platform.is_rocm():
+                hidden_states_d = hidden_states_d.contiguous()
+
             B, C, dt = self._project_ssm_parameters(hidden_states_d)
 
             # 3. State Space Model sequence transformation