Fix Mamba slow path bug with dtype mismatch. (huggingface#32691)

* Fix Mamba slow path bug with dtype mismatch.

* Update test_modeling_mamba.py

* Improve style.

* Fix issue with cache position of dtype mismatch test.

* Change test for slow path.

* Revert changes.

* Switch to buggy code and add test to catch it.

* Fix the dtype mismatch bug and add test code to verify it.

* Fix minor bug with test.

* Fix incorrect dtype of model output.

* Fix incorrect dtype of cache.

* Fix incorrect dtype of ssm cache.

* Fix incorrect dtype of conv state.

* Remove assertion for ssm state.

* Add assertion for conv state dtype.

* Fix all issues with dtype mismatch test.

src/transformers/cache_utils.py

@@ -1797,7 +1797,7 @@ def update_conv_state(
         cache_position = cache_position.clamp(0, self.conv_kernel_size - 1)
 
         conv_state = conv_state.roll(shifts=-1, dims=-1)
-        conv_state[:, :, cache_position] = new_conv_state.to(conv_state.device)
+        conv_state[:, :, cache_position] = new_conv_state.to(device=conv_state.device, dtype=conv_state.dtype)
         self.conv_states[layer_idx].zero_()
         self.conv_states[layer_idx] += conv_state
         return self.conv_states[layer_idx]

tests/models/mamba/test_modeling_mamba.py

@@ -421,6 +421,30 @@ def recursive_check(tuple_object, dict_object):
     def test_beam_sample_generate(self):
         pass
 
+    def test_dtype_mismatch_handled_in_cache(self):
+        config, input_ids, *args = self.model_tester.prepare_config_and_inputs()
+        model = MambaModel(config)
+        model.to(torch_device).to(torch.float16)
+        model.eval()
+
+        # Create cache with float32 dtype
+        cache_params = MambaCache(config, batch_size=input_ids.size(0), dtype=torch.float32, device=torch_device)
+
+        # If code is correct, no error occurs and test passes
+        outputs = model(
+            input_ids,
+            cache_params=cache_params,
+            use_cache=True,
+            cache_position=torch.arange(0, config.conv_kernel, device=input_ids.device),
+        )
+
+        self.assertIsNotNone(outputs)
+        self.assertIsNotNone(outputs.last_hidden_state)
+        self.assertEqual(
+            outputs.last_hidden_state.shape,
+            (self.model_tester.batch_size, self.model_tester.seq_length, self.model_tester.hidden_size),
+        )
+
 
 @require_torch
 class MambaIntegrationTests(unittest.TestCase):