Add Zamba (#30950)

* Update index.md * Rebase * Rebase * Updates from make fixup * Update zamba.md * Batched inference * Update * Fix tests * Fix tests * Fix tests * Fix tests * Update docs/source/en/model_doc/zamba.md Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update docs/source/en/model_doc/zamba.md Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update configuration_zamba.py * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update modeling_zamba.py * Update modeling_zamba.py * Update modeling_zamba.py * Update configuration_zamba.py * Update modeling_zamba.py * Update modeling_zamba.py * Merge branch 'main' of https://github.com/Zyphra/transformers_zamba * Update ZambaForCausalLM * Update ZambaForCausalLM * Describe diffs with original mamba layer * Moved mamba init into `_init_weights` * Update index.md * Rebase * Rebase * Updates from make fixup * Update zamba.md * Batched inference * Update * Fix tests * Fix tests * Fix tests * Fix tests * Update docs/source/en/model_doc/zamba.md Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update docs/source/en/model_doc/zamba.md Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update configuration_zamba.py * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update modeling_zamba.py * Update modeling_zamba.py * Update modeling_zamba.py * Update configuration_zamba.py * Update modeling_zamba.py * Update modeling_zamba.py * Merge branch 'main' of https://github.com/Zyphra/transformers_zamba * Update ZambaForCausalLM * Moved mamba init into `_init_weights` * Update ZambaForCausalLM * Describe diffs with original mamba layer * make fixup fixes * quality test fixes * Fix Zamba model path * circleci fixes * circleci fixes * circleci fixes * circleci fixes * circleci fixes * circleci fixes * circleci fixes * circleci fixes * circleci fixes * Update * circleci fixes * fix zamba test from merge * fix ValueError for disabling mamba kernels * add HF copyright Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * shared_transf --> shared_transformer * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Fixes * Move attention head dim to config * Fix circle/ci tests * Update modeling_zamba.py * apply GenerationMixin inheritance change from upstream * apply import ordering * update needed transformers version for zamba Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * add contribution author * add @slow to avoid CI * Update src/transformers/models/zamba/modeling_zamba.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Define attention_hidden_size * Added doc for attention_head_size * trigger CI * Fix doc of attention_hidden_size * [run-slow] zamba * Fixed shared layer logic, swapped up<->gate in mlp * shared_transformer -> shared_transf * reformat HybridLayer __init__ * fix docstrings in zamba config * added definition of _get_input_ids_and_config * fixed formatting of _get_input_ids_and_config --------- Co-authored-by: root <root@node-4.us-southcentral1-a.compute.internal> Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> Co-authored-by: root <root@node-1.us-southcentral1-a.compute.internal> Co-authored-by: Quentin Anthony <qganthony@yahoo.com>
2024-10-04 13:28:05 -07:00
parent e3775539c8
commit f319ba16fa
18 changed files with 2939 additions and 2 deletions
--- a/tests/models/zamba/init.py
+++ b/tests/models/zamba/init.py
--- a/tests/models/zamba/test_modeling_zamba.py
+++ b/tests/models/zamba/test_modeling_zamba.py
@@ -0,0 +1,736 @@
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the PyTorch Zamba model."""
+
+import math
+import tempfile
+import unittest
+
+import pytest
+from parameterized import parameterized
+
+from transformers import AutoTokenizer, ZambaConfig, is_torch_available
+from transformers.testing_utils import (
+    require_bitsandbytes,
+    require_flash_attn,
+    require_torch,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, _config_zero_init, ids_tensor, random_attention_mask
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        ZambaForCausalLM,
+        ZambaForSequenceClassification,
+        ZambaModel,
+    )
+    from transformers.models.zamba.modeling_zamba import (
+        HybridMambaAttentionDynamicCache,
+    )
+
+
+class ZambaModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=64,
+        mamba_dt_rank=32,
+        num_hidden_layers=5,
+        attn_layer_offset=1,
+        attn_layer_period=8,
+        num_attention_heads=4,
+        num_key_value_heads=4,
+        n_mamba_heads=2,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_mamba_act="silu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.mamba_dt_rank = mamba_dt_rank
+        self.num_hidden_layers = num_hidden_layers
+        self.attn_layer_offset = attn_layer_offset
+        self.attn_layer_period = attn_layer_period
+        self.num_attention_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+        self.n_mamba_heads = n_mamba_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_mamba_act = hidden_mamba_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def get_config(self):
+        return ZambaConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            mamba_dt_rank=self.mamba_dt_rank,
+            num_hidden_layers=self.num_hidden_layers,
+            attn_layer_offset=self.attn_layer_offset,
+            attn_layer_period=self.attn_layer_period,
+            num_attention_heads=self.num_attention_heads,
+            num_key_value_heads=self.num_key_value_heads,
+            n_mamba_heads=self.n_mamba_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_mamba_act=self.hidden_mamba_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            is_decoder=True,
+            initializer_range=self.initializer_range,
+            use_mamba_kernels=False,
+        )
+
+    def prepare_config_and_inputs_for_decoder(self):
+        (
+            config,
+            input_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = self.prepare_config_and_inputs()
+
+        config.is_decoder = True
+
+        return (
+            config,
+            input_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
+        model = ZambaModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask)
+        result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_for_causal_lm(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        model = ZambaForCausalLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, labels=token_labels)
+        result = model(input_ids, attention_mask=input_mask)
+        result = model(input_ids, labels=token_labels)
+        result = model(input_ids)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        config.is_decoder = True
+        config.add_cross_attention = True
+        model = ZambaForCausalLM(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        # first forward pass
+        # Attention: Zamba needs the cache to be initialized to return a cache!
+        past_key_values = HybridMambaAttentionDynamicCache(
+            config, input_ids.shape[0], model.dtype, device=model.device
+        )
+        outputs = model(
+            input_ids,
+            attention_mask=input_mask,
+            past_key_values=past_key_values,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+            cache_position=torch.arange(
+                input_ids.shape[1], input_ids.shape[1] + next_tokens.shape[1], device=model.device
+            ),
+        )["hidden_states"][0]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
+
+        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
+    def create_and_check_for_sequence_classification(
+        self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = ZambaForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, labels=sequence_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class ZambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            ZambaModel,
+            ZambaForCausalLM,
+            ZambaForSequenceClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+    all_generative_model_classes = (ZambaForCausalLM,) if is_torch_available() else ()
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": ZambaModel,
+            "text-classification": ZambaForSequenceClassification,
+            "text-generation": ZambaForCausalLM,
+            "zero-shot": ZambaForSequenceClassification,
+        }
+        if is_torch_available()
+        else {}
+    )
+    test_headmasking = False
+    test_pruning = False
+
+    def setUp(self):
+        self.model_tester = ZambaModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ZambaConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_for_casual_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_causal_lm(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
+
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def test_initialization(self):
+        r"""
+        Overriding the test_initialization test as the A_log and D params of the Mamba block are initialized differently
+        """
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    if "A_log" in name:
+                        A = torch.arange(1, config.mamba_d_state + 1, dtype=torch.float32)[None, :]
+                        self.assertTrue(torch.allclose(param.data, torch.log(A), atol=1e-5, rtol=1e-5))
+                    elif "D" in name:
+                        # check if it's a ones like
+                        self.assertTrue(torch.allclose(param.data, torch.ones_like(param.data), atol=1e-5, rtol=1e-5))
+                    elif "x_proj" in name or "dt_proj_weight" in name:
+                        self.assertIn(
+                            ((param.data.mean() * 1e2).round() / 1e2).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized (raw value {param.data.mean()})",
+                        )
+                    elif "dt_proj_bias" in name:
+                        dt = torch.exp(
+                            torch.tensor([0, 1]) * (math.log(config.time_step_max) - math.log(config.time_step_min))
+                            + math.log(config.time_step_min)
+                        ).clamp(min=config.time_step_floor)
+                        inv_dt = dt + torch.log(-torch.expm1(-dt))
+                        if param.requires_grad:
+                            self.assertTrue(param.data.max().item() <= inv_dt[1])
+                            self.assertTrue(param.data.min().item() >= inv_dt[0])
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+
+    def test_mismatched_shapes_have_properly_initialized_weights(self):
+        r"""
+        Overriding the test_mismatched_shapes_have_properly_initialized_weights test because A_log and D params of the
+        Mamba block are initialized differently and we tested that in test_initialization
+        """
+        self.skipTest("Cumbersome and redundant for Zamba")
+
+    def test_attention_outputs(self):
+        r"""
+        Overriding the test_attention_outputs test as the Zamba model outputs attention only for its attention layers
+        """
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        seq_len = getattr(self.model_tester, "seq_length", None)
+        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
+        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
+
+        expected_num_attentions = (
+            math.ceil(
+                (self.model_tester.num_hidden_layers - self.model_tester.attn_layer_offset)
+                / self.model_tester.attn_layer_period
+            )
+            + 1
+        )
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.attentions
+            self.assertEqual(len(attentions), expected_num_attentions)
+
+            # check that output_attentions also work using config
+            del inputs_dict["output_attentions"]
+            config.output_attentions = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.attentions
+            self.assertEqual(len(attentions), expected_num_attentions)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
+            )
+            out_len = len(outputs)
+
+            # Check attention is always last and order is fine
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = True
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            added_hidden_states = 1
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs.attentions
+
+            self.assertEqual(len(self_attentions), expected_num_attentions)
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
+            )
+
+    def _get_input_ids_and_config(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        return config, input_ids, input_mask
+
+    def test_left_padding_compatibility(self):
+        r"""
+        Overriding the test_left_padding_compatibility test as the mamba layers accentuate the numerical differences
+        effect of the left padding discussed in the issue in the note. Using a more permissive tolerance value.
+        """
+        import inspect
+        # NOTE: left-padding results in small numerical differences. This is expected.
+        # See https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535
+
+        # First, filter out models that don't support left padding - generative and decoder-only.
+        # Zamba is a decoder-only architecture
+        decoder_only_classes = self.all_generative_model_classes
+
+        # Then, test left-padding
+        def _prepare_model_kwargs(input_ids, attention_mask, signature):
+            model_kwargs = {"input_ids": input_ids, "attention_mask": attention_mask}
+            if "position_ids" in signature:
+                position_ids = torch.cumsum(attention_mask, dim=-1) - 1
+                position_ids.masked_fill_(attention_mask == 0, 1)
+                model_kwargs["position_ids"] = position_ids
+            if "cache_position" in signature:
+                cache_position = torch.arange(input_ids.shape[-1], device=torch_device)
+                model_kwargs["cache_position"] = cache_position
+            return model_kwargs
+
+        for model_class in decoder_only_classes:
+            config, input_ids, attention_mask = self._get_input_ids_and_config()
+            model = model_class(config).to(torch_device).eval()
+            signature = inspect.signature(model.forward).parameters.keys()
+
+            # Without padding
+            model_kwargs = _prepare_model_kwargs(input_ids, attention_mask, signature)
+            next_logits_wo_padding = model(**model_kwargs).logits[:, -1, :]
+
+            # With left-padding (length 32)
+            pad_size = (input_ids.shape[0], 32)
+            padding = torch.ones(pad_size, dtype=input_ids.dtype, device=torch_device) * config.pad_token_id
+            padded_input_ids = torch.cat((padding, input_ids), dim=1)
+            padded_attention_mask = torch.cat((torch.zeros_like(padding), attention_mask), dim=1)
+            model_kwargs = _prepare_model_kwargs(padded_input_ids, padded_attention_mask, signature)
+            next_logits_with_padding = model(**model_kwargs).logits[:, -1, :]
+
+            # They should result in very similar logits
+            self.assertTrue(torch.allclose(next_logits_wo_padding, next_logits_with_padding, atol=3e-3))
+
+    @require_flash_attn
+    @require_torch_gpu
+    @require_bitsandbytes
+    @pytest.mark.flash_attn_test
+    @slow
+    def test_flash_attn_2_fp32_ln(self):
+        r"""
+        Overriding the test_flash_attn_2_fp32_ln test as the Zamba model, like Mixtral, doesn't support
+        right padding + use cache with FA2
+        """
+        for model_class in self.all_generative_model_classes:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            model = model_class(config)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+
+                dummy_input = inputs_dict[model.main_input_name]
+                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
+                # NOTE: Zamba does not support right padding + use_cache with FA2.
+                dummy_attention_mask[:, -1] = 1
+
+                model = model_class.from_pretrained(
+                    tmpdirname,
+                    torch_dtype=torch.float16,
+                    attn_implementation="flash_attention_2",
+                    low_cpu_mem_usage=True,
+                    load_in_4bit=True,
+                )
+
+                for _, param in model.named_parameters():
+                    # upcast only layer norms
+                    if (param.dtype == torch.float16) or (param.dtype == torch.bfloat16):
+                        param.data = param.data.to(torch.float32)
+
+                _ = model(dummy_input)
+                # with attention mask
+                _ = model(dummy_input, attention_mask=dummy_attention_mask)
+
+    @require_flash_attn
+    @require_torch_gpu
+    @pytest.mark.flash_attn_test
+    @slow
+    def test_flash_attn_2_generate_padding_right(self):
+        r"""
+        Overriding the test_flash_attn_2_generate_padding_right test as the Zamba model, like Mixtral, doesn't support
+        right padding + use cache with FA2
+        """
+        import torch
+
+        for model_class in self.all_generative_model_classes:
+            config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+            model = model_class(config)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+                model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, low_cpu_mem_usage=True).to(
+                    torch_device
+                )
+
+                dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
+                dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [1, 1, 1, 0]]).to(torch_device)
+
+                model.generate(dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False)
+
+                model = model_class.from_pretrained(
+                    tmpdirname,
+                    torch_dtype=torch.float16,
+                    attn_implementation="flash_attention_2",
+                    low_cpu_mem_usage=True,
+                ).to(torch_device)
+
+                with self.assertRaises(ValueError):
+                    _ = model.generate(
+                        dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
+                    )
+
+    @require_flash_attn
+    @require_torch_gpu
+    @pytest.mark.flash_attn_test
+    @slow
+    def test_flash_attn_2_generate_use_cache(self):
+        r"""
+        Overriding the test_flash_attn_2_generate_use_cache test as the Zamba model, like Mixtral, doesn't support
+        right padding + use cache with FA2
+        """
+        import torch
+
+        max_new_tokens = 30
+
+        for model_class in self.all_generative_model_classes:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+            dummy_input = inputs_dict[model_class.main_input_name]
+            if dummy_input.dtype in [torch.float32, torch.bfloat16]:
+                dummy_input = dummy_input.to(torch.float16)
+
+            # make sure that all models have enough positions for generation
+            if hasattr(config, "max_position_embeddings"):
+                config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
+
+            model = model_class(config)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+
+                dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
+                # NOTE: Zamba does not support right padding + use_cache with FA2.
+                dummy_attention_mask[:, -1] = 1
+
+                model = model_class.from_pretrained(
+                    tmpdirname,
+                    torch_dtype=torch.float16,
+                    attn_implementation="flash_attention_2",
+                    low_cpu_mem_usage=True,
+                ).to(torch_device)
+
+                # Just test that a large cache works as expected
+                _ = model.generate(
+                    dummy_input,
+                    attention_mask=dummy_attention_mask,
+                    max_new_tokens=max_new_tokens,
+                    do_sample=False,
+                    use_cache=True,
+                )
+
+    @require_flash_attn
+    @require_torch_gpu
+    @pytest.mark.flash_attn_test
+    @slow
+    def test_flash_attn_2_inference_equivalence_right_padding(self):
+        r"""
+        Overriding the test_flash_attn_2_inference_padding_right test as the Zamba model, like Mixtral, doesn't support
+        right padding + use cache with FA2
+        """
+        self.skipTest(reason="Zamba flash attention does not support right padding")
+
+    @unittest.skip(reason="Zamba has its own special cache type")
+    @parameterized.expand([(1, False), (1, True), (4, False)])
+    def test_new_cache_format(self, num_beams, do_sample):
+        pass
+
+
+@require_torch
+class ZambaModelIntegrationTest(unittest.TestCase):
+    model = None
+    tokenizer = None
+
+    @classmethod
+    @slow
+    def setUpClass(cls):
+        model_id = "Zyphra/Zamba-7B-v1"
+        cls.model = ZambaForCausalLM.from_pretrained(
+            model_id, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, use_mamba_kernels=False
+        )
+        cls.tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    @slow
+    def test_simple_generate(self):
+        self.model.to(torch_device)
+
+        input_ids = self.tokenizer("Hey how are you doing on this lovely evening?", return_tensors="pt")[
+            "input_ids"
+        ].to(torch_device)
+        out = self.model.generate(input_ids, do_sample=False, max_new_tokens=10)
+        output_sentence = self.tokenizer.decode(out[0, :])
+        self.assertEqual(
+            output_sentence,
+            "<s> Hey how are you doing on this lovely evening? I hope you are all doing well. I am",
+        )
+
+        with torch.no_grad():
+            logits = self.model(input_ids=input_ids).logits
+
+        EXPECTED_LOGITS_NO_GRAD = torch.tensor(
+            [
+                -7.9375,  8.1875,  1.3984, -6.0000, -7.9375, -7.9375, -7.9375, -7.9375,
+                -7.9375, -7.9375, -7.9375, -7.9375,  2.7500, 13.0625, -7.9375, -7.9375,
+                -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375,
+                -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375,
+                -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375
+            ]
+            , dtype=torch.float32)  # fmt: skip
+
+        torch.testing.assert_close(logits[0, -1, :40].cpu(), EXPECTED_LOGITS_NO_GRAD, rtol=1e-3, atol=1e-3)
+
+    @slow
+    def test_simple_batched_generate_with_padding(self):
+        self.model.to(torch_device)
+        self.tokenizer.add_special_tokens({"pad_token": "[PAD]"})
+        self.model.resize_token_embeddings(len(self.tokenizer))
+
+        inputs = self.tokenizer(
+            ["Hey how are you doing on this lovely evening?", "Tell me a story"], padding=True, return_tensors="pt"
+        ).to(torch_device)
+        out = self.model.generate(**inputs, do_sample=False, max_new_tokens=10)
+        output_sentences = self.tokenizer.batch_decode(out)
+        self.assertEqual(
+            output_sentences[0],
+            "<s> Hey how are you doing on this lovely evening? I hope you are all doing well. I am",
+        )
+        self.assertEqual(
+            output_sentences[1],
+            "[PAD][PAD][PAD][PAD][PAD][PAD]<s> Tell me a story about a time when you were in a difficult situation",
+        )
+
+        with torch.no_grad():
+            logits = self.model(input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"]).logits
+
+        EXPECTED_LOGITS_NO_GRAD_0 = torch.tensor(
+            [
+                -7.9375,  8.1250,  1.3594, -6.0000, -7.9375, -7.9375, -7.9375, -7.9375,
+                -7.9375, -7.9375, -7.9375, -7.9375,  2.7344, 13.0625, -7.9375, -7.9375,
+                -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375,
+                -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375,
+                -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375, -7.9375
+            ]
+            , dtype=torch.float32)  # fmt: skip
+
+        EXPECTED_LOGITS_NO_GRAD_1 = torch.tensor(
+            [
+               -6.3750,  3.4219,  0.6719, -5.0312, -8.5000, -8.5000, -8.5000, -8.5000,
+               -8.5000, -8.5000, -8.5000, -8.5000,  2.0625, 10.3750, -8.5000, -8.5000,
+               -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000,
+               -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000,
+               -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000, -8.5000
+            ]
+            , dtype=torch.float32)  # fmt: skip
+
+        torch.testing.assert_close(logits[0, -1, :40].cpu(), EXPECTED_LOGITS_NO_GRAD_0, rtol=1e-3, atol=1e-3)
+        torch.testing.assert_close(logits[1, -1, :40].cpu(), EXPECTED_LOGITS_NO_GRAD_1, rtol=1e-3, atol=1e-3)