handle weight sharing with init_on_device

src/accelerate/big_modeling.py

@@ -132,7 +132,14 @@ def register_empty_parameter(module, name, param):
             param_cls = type(module._parameters[name])
             kwargs = module._parameters[name].__dict__
             kwargs["requires_grad"] = param.requires_grad
-            module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
+            # When we have a case of tensor2 = tensor1, it would call the set_attr
+            # of param, which in turn would call the register_parameter API.
+            # In this case, the new param is already on meta-device, since it was moved
+            # previously when it was initialized. Hence, when resetting, you can
+            # directly assign that tensor instead of re-init. If you re-init you would
+            # lose the relationship.
+            module._parameters[name] = param if param.device == device else \
+                            param_cls(module._parameters[name].to(device), **kwargs)
 
     def register_empty_buffer(module, name, buffer, persistent=True):
         old_register_buffer(module, name, buffer, persistent=persistent)

tests/test_big_modeling.py

@@ -188,6 +188,18 @@ def test_init_empty_weights(self):
         assert module.weight.device == torch.device("cpu")
         assert module.running_mean.device == torch.device("cpu")
 
+    def test_init_empty_weights_with_tie_embedding(self):
+        with init_empty_weights():
+            module = torch.nn.ModuleList([torch.nn.Embedding(12, 12), torch.nn.Linear(12, 12)])
+            # tie embedding
+            module[0].weight = module[1].weight
+
+            from transformers.models import Qwen2Config, Qwen2ForCausalLM
+
+            qwen2 = Qwen2ForCausalLM(Qwen2Config(tie_word_embeddings=True))
+        assert module[0].weight is module[1].weight
+        assert qwen2.lm_head.weight is qwen2.model.embed_tokens.weight
+
     def test_init_empty_weights_very_large_model(self):
         # This is a 100 billion parameters model.
         with init_empty_weights():