Mamba slow_forward gradient fix (#29563)

* FIX: Cached slow forward in mamba
- additionally added mamba cached test
- added unused test (mamba causal lm forward and backward)
- fixed typo: "causl" --> "causal"

* formatting

* fix: use real `slow_forward` call instead of torch module's

* add shape assertion for mixer block test

* adjust shape assertion

src/transformers/models/mamba/modeling_mamba.py

@@ -230,7 +230,7 @@ class MambaMixer(nn.Module):
 
         # 2. Convolution sequence transformation
         if cache_params is not None:
-            ssm_state = cache_params.ssm_states[self.layer_idx]
+            ssm_state = cache_params.ssm_states[self.layer_idx].clone()
             if cache_params.seqlen_offset > 0:
                 conv_state = cache_params.conv_states[self.layer_idx]                   # [batch, intermediate_size, conv_kernel_size]
                 conv_state = torch.roll(conv_state, shifts=-1, dims=-1)

tests/models/mamba/test_modeling_mamba.py

@@ -170,7 +170,7 @@ class MambaModelTester:
         self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
         self.parent.assertEqual(len(result.hidden_states), config.num_hidden_layers + 1)
 
-    def create_and_check_causl_lm(self, config, input_ids, *args):
+    def create_and_check_causal_lm(self, config, input_ids, *args):
         model = MambaForCausalLM(config)
         model.to(torch_device)
         model.eval()
@@ -197,7 +197,30 @@ class MambaModelTester:
         self.parent.assertTrue(torch.allclose(torch.cat([output_one, output_two], dim=1), output_whole, atol=1e-5))
         # TODO the orignal mamba does not support decoding more than 1 token neither do we
 
-    def create_and_check_forward_and_backwards(self, config, input_ids, *args, gradient_checkpointing=False):
+    def create_and_check_mamba_cached_slow_forward_and_backwards(
+        self, config, input_ids, *args, gradient_checkpointing=False
+    ):
+        model = MambaModel(config)
+        model.to(torch_device)
+        if gradient_checkpointing:
+            model.gradient_checkpointing_enable()
+
+        # create cache
+        cache = model(input_ids, use_cache=True).cache_params
+        cache.seqlen_offset = 0
+
+        # use cache
+        token_emb = model.embeddings(input_ids)
+        outputs = model.layers[0].mixer.slow_forward(token_emb, cache)
+
+        loss = torch.log(1 + torch.abs(outputs.sum()))
+        self.parent.assertEqual(loss.shape, ())
+        self.parent.assertEqual(outputs.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        loss.backward()
+
+    def create_and_check_mamba_lm_head_forward_and_backwards(
+        self, config, input_ids, *args, gradient_checkpointing=False
+    ):
         model = MambaForCausalLM(config)
         model.to(torch_device)
         if gradient_checkpointing:
@@ -304,12 +327,20 @@ class MambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
 
     def test_mamba_lm_head_model(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
-        self.model_tester.create_and_check_causl_lm(*config_and_inputs)
+        self.model_tester.create_and_check_causal_lm(*config_and_inputs)
 
     def test_state_equivalency(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_state_equivalency(*config_and_inputs)
 
+    def test_mamba_cached_slow_forward_and_backwards(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_mamba_cached_slow_forward_and_backwards(*config_and_inputs)
+
+    def test_mamba_lm_head_forward_and_backwards(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_mamba_lm_head_forward_and_backwards(*config_and_inputs)
+
     def test_initialization(self):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()