[Bugfix][ROCm][CI] Give the AITER MLA decode metadata stub its MLA dims (#52139)

Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com>

tests/kernels/attention/test_rocm_aiter_mla_decode_metadata.py

@@ -82,6 +82,7 @@ def _build_decode_metadata():
         create_vllm_config,
     )
     from vllm.config.vllm import set_current_vllm_config
+    from vllm.model_executor.layers.attention.mla_attention import get_mla_dims
     from vllm.v1.attention.backends.registry import AttentionBackendEnum
     from vllm.v1.kv_cache_interface import MLAAttentionSpec
     from vllm.v1.worker.workspace import init_workspace_manager
@@ -110,11 +111,20 @@ def _build_decode_metadata():
 
     builder_cls = AttentionBackendEnum.ROCM_AITER_MLA.get_class().get_builder_cls()
 
-    # The builder reads layer.prefill_backend from static_forward_context; a
-    # stub with the attribute is enough for metadata construction.
+    # The builder reads prefill_backend and the MLA latent dims off the layer,
+    # so the stub carries both; the dims come from the model config to stay in
+    # step with the checkpoint.
+    mla_dims = get_mla_dims(vllm_config.model_config)
     layer_name = "placeholder"
     vllm_config.compilation_config.static_forward_context[layer_name] = (
-        types.SimpleNamespace(prefill_backend=torch.empty((1,)))
+        types.SimpleNamespace(
+            prefill_backend=torch.empty((1,)),
+            q_lora_rank=mla_dims.q_lora_rank,
+            kv_lora_rank=mla_dims.kv_lora_rank,
+            qk_nope_head_dim=mla_dims.qk_nope_head_dim,
+            qk_rope_head_dim=mla_dims.qk_rope_head_dim,
+            v_head_dim=mla_dims.v_head_dim,
+        )
     )
 
     init_workspace_manager(device)