mla: don't update kv cache on dummy forwards (#36282)

Signed-off-by: Itay Alroy <ialroy@nvidia.com>

vllm/model_executor/layers/attention/mla_attention.py

@@ -905,6 +905,10 @@ def unified_mla_kv_cache_update(
     the data dependency between them to ensure torch.compile preserves ordering.
     """
     forward_context = get_forward_context()
+    if forward_context.attn_metadata is None:
+        # Dummy/profile forwards should not update live KV cache pages.
+        return torch.empty(0, device=kv_c_normed.device, dtype=kv_c_normed.dtype)
+
     attn_layer = forward_context.no_compile_layers[layer_name]
     kv_cache = attn_layer.kv_cache[forward_context.virtual_engine]