Fix inter-chunk recurrence in the Zamba2 / Nemotron-H Mamba2 slow path

src/transformers/models/nemotron_h/modeling_nemotron_h.py

@@ -551,9 +551,8 @@ def torch_forward(self, input_states, cache_params: Cache | None=None, attention
             states = torch.cat([previous_states, states], dim=1)
             decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
 
-            states_permuted = states.permute(0, 2, 1, 3, 4)
-            result = (decay_chunk[..., None, None] * states_permuted[:, :, None, ...]).sum(dim=2)
-            new_states = result.permute(0, 2, 1, 3, 4)
+            decay_chunk = decay_chunk.transpose(1, 3)
+            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
             states, ssm_state = new_states[:, :-1], new_states[:, -1]
 
             # Compute state -> output conversion per chunk

src/transformers/models/zamba2/modeling_zamba2.py

@@ -839,9 +839,8 @@ def torch_forward(self, input_states, cache_params: Cache | None=None, attention
             states = torch.cat([previous_states, states], dim=1)
             decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
 
-            states_permuted = states.permute(0, 2, 1, 3, 4)
-            result = (decay_chunk[..., None, None] * states_permuted[:, :, None, ...]).sum(dim=2)
-            new_states = result.permute(0, 2, 1, 3, 4)
+            decay_chunk = decay_chunk.transpose(1, 3)
+            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
             states, ssm_state = new_states[:, :-1], new_states[:, -1]
 
             # Compute state -> output conversion per chunk

src/transformers/models/zamba2/modular_zamba2.py

@@ -627,9 +627,8 @@ def torch_forward(self, input_states, cache_params: Cache | None=None, attention
             states = torch.cat([previous_states, states], dim=1)
             decay_chunk = torch.exp(segment_sum(nn.functional.pad(A_cumsum[:, :, :, -1], (1, 0))))
 
-            states_permuted = states.permute(0, 2, 1, 3, 4)
-            result = (decay_chunk[..., None, None] * states_permuted[:, :, None, ...]).sum(dim=2)
-            new_states = result.permute(0, 2, 1, 3, 4)
+            decay_chunk = decay_chunk.transpose(1, 3)
+            new_states = (decay_chunk[..., None, None] * states[:, :, None, ...]).sum(dim=1)
             states, ssm_state = new_states[:, :-1], new_states[:, -1]
 
             # Compute state -> output conversion per chunk

tests/models/nemotron_h/test_modeling_nemotron_h.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 """Testing suite for the PyTorch NemotronH model."""
 
+import copy
 import tempfile
 import unittest
 
@@ -356,6 +357,41 @@ def create_and_check_nemotron_h_chunked_prefill(self, config, input_ids, *args,
             msg=f"Max diff: {(ref_first - under_test_first).abs().max().item():.6f}",
         )
 
+    def create_and_check_nemotron_h_slow_path_multi_chunk(self, config, input_ids, *args):
+        """
+        Regression test for the inter-chunk recurrence of the Mamba2 slow (torch) path: a
+        single chunked forward over a multi-chunk sequence must reproduce a token-by-token
+        recurrent decode.
+
+        The mixer is fed a large-magnitude input so the SSM state is O(1). With a normal
+        activation scale the inter-chunk contribution is ~1e-7, i.e. below any tolerance,
+        which is why neither `create_and_check_mamba2_slow_vs_fast_forward` (single chunk,
+        needs the kernels) nor a plain `prepare_config_and_inputs` input surfaces the bug.
+        A small `chunk_size` lets a short sequence span several chunks. Runs on CPU without
+        the fast-path kernels.
+        """
+        config = copy.deepcopy(config)
+        config.chunk_size = 8
+        model = NemotronHModel(config).eval().to(torch_device)
+        mixer = next(layer.mixer for layer in model.layers if getattr(layer, "block_type", None) == "linear_attention")
+
+        seq_len = 4 * config.chunk_size + 1
+        torch.manual_seed(0)
+        hidden_states = 100.0 * torch.randn(1, seq_len, config.hidden_size, device=torch_device)
+
+        with torch.no_grad():
+            chunked = mixer.torch_forward(hidden_states)
+            cache = DynamicCache(config=config)
+            recurrent = torch.cat(
+                [mixer.torch_forward(hidden_states[:, t : t + 1], cache_params=cache) for t in range(seq_len)],
+                dim=1,
+            )
+
+        max_diff = (chunked - recurrent).abs().max().item()
+        self.parent.assertLess(
+            max_diff, 1e-3, f"slow-path chunked forward disagrees with recurrent decode: {max_diff}"
+        )
+
     def prepare_config_and_inputs_for_common(self):
         config_and_inputs = self.prepare_config_and_inputs()
         (
@@ -520,6 +556,11 @@ def test_mamba2_slow_vs_fast_forward(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_mamba2_slow_vs_fast_forward(*config_and_inputs)
 
+    def test_mamba2_slow_path_multi_chunk(self):
+        """The Mamba2 slow path must reproduce the token-by-token recurrence across chunks."""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_nemotron_h_slow_path_multi_chunk(*config_and_inputs)
+
     def test_attention_outputs(self):
         r"""
         Overriding the test_attention_outputs test as the NemotronH model outputs attention only for its attention layers

tests/models/zamba2/test_modeling_zamba2.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 """Testing suite for the PyTorch Zamba model."""
 
+import copy
 import tempfile
 import unittest
 
@@ -298,6 +299,53 @@ def create_and_check_zamba2_chunked_prefill(self, config, input_ids, *args, devi
             msg=f"Max diff: {(ref_first - under_test_first).abs().max().item():.6f}",
         )
 
+    def create_and_check_zamba2_slow_vs_fast_forward(self, config, input_ids, *args):
+        """Slow vs fast path check guarded by require kernels to enable fast path"""
+        model = Zamba2Model(config)
+        model.eval()
+        model.to(torch_device)
+
+        mamba_mixer = next(layer.mamba for layer in model.layers if hasattr(layer, "mamba"))
+        hidden_states = model.embed_tokens(input_ids)
+        outputs_fast = mamba_mixer.cuda_kernels_forward(hidden_states)
+        outputs_slow = mamba_mixer.torch_forward(hidden_states)
+        self.parent.assertTrue(torch.allclose(outputs_fast, outputs_slow, atol=1e-3, rtol=1e-3))
+
+    def create_and_check_zamba2_slow_path_multi_chunk(self, config, input_ids, *args):
+        """
+        Regression test for the inter-chunk recurrence of the Mamba2 slow (torch) path: a
+        single chunked forward over a multi-chunk sequence must reproduce a token-by-token
+        recurrent decode.
+
+        The mixer is fed a large-magnitude input so the SSM state is O(1). With a normal
+        activation scale the inter-chunk contribution is ~1e-7, i.e. below any tolerance,
+        which is why neither `create_and_check_zamba2_slow_vs_fast_forward` (single chunk,
+        needs the kernels) nor a plain `prepare_config_and_inputs` input surfaces the bug.
+        A small `chunk_size` lets a short sequence span several chunks. Runs on CPU without
+        the fast-path kernels.
+        """
+        config = copy.deepcopy(config)
+        config.chunk_size = 8
+        model = Zamba2Model(config).eval().to(torch_device)
+        mixer = next(layer.mamba for layer in model.layers if hasattr(layer, "mamba"))
+
+        seq_len = 4 * config.chunk_size + 1
+        torch.manual_seed(0)
+        hidden_states = 100.0 * torch.randn(1, seq_len, config.hidden_size, device=torch_device)
+
+        with torch.no_grad():
+            chunked = mixer.torch_forward(hidden_states)
+            cache = DynamicCache(config=config)
+            recurrent = torch.cat(
+                [mixer.torch_forward(hidden_states[:, t : t + 1], cache_params=cache) for t in range(seq_len)],
+                dim=1,
+            )
+
+        max_diff = (chunked - recurrent).abs().max().item()
+        self.parent.assertLess(
+            max_diff, 1e-3, f"slow-path chunked forward disagrees with recurrent decode: {max_diff}"
+        )
+
     def prepare_config_and_inputs_for_common(self):
         config_and_inputs = self.prepare_config_and_inputs()
         (
@@ -351,6 +399,17 @@ def setUp(self):
         self.model_tester = Zamba2ModelTester(self)
         self.config_tester = ConfigTester(self, config_class=Zamba2Config, hidden_size=32)
 
+    @require_torch_accelerator
+    @require_kernels
+    def test_mamba2_slow_vs_fast_forward(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_zamba2_slow_vs_fast_forward(*config_and_inputs)
+
+    def test_mamba2_slow_path_multi_chunk(self):
+        """The Mamba2 slow path must reproduce the token-by-token recurrence across chunks."""
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_zamba2_slow_path_multi_chunk(*config_and_inputs)
+
     @unittest.skip("We need at leat 3 layers to test weight tying!")
     def test_num_layers_is_small(self):
         pass