Upload JetMoEForCausalLM

README.md

@@ -0,0 +1,199 @@
+---
+library_name: transformers
+tags: []
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]

config.json

@@ -0,0 +1,92 @@
+{
+  "_attn_implementation_internal": "flash_attention_2",
+  "_commit_hash": "e4d9ef770a272f072cac70d16dd1ae113f5989ae",
+  "_name_or_path": "jetmoe/jetmoe-8b-chat",
+  "activation_function": "silu",
+  "add_cross_attention": false,
+  "architectures": [
+    "JetMoEForCausalLM"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_jetmoe.JetMoEConfig",
+    "AutoModelForCausalLM": "modeling_jetmoe.JetMoEForCausalLM"
+  },
+  "aux_loss_coef": 0.01,
+  "bad_words_ids": null,
+  "begin_suppress_tokens": null,
+  "bias": true,
+  "bos_token_id": 1,
+  "chunk_size_feed_forward": 0,
+  "cross_attention_hidden_size": null,
+  "decoder_start_token_id": null,
+  "diversity_penalty": 0.0,
+  "do_sample": false,
+  "early_stopping": false,
+  "encoder_no_repeat_ngram_size": 0,
+  "eos_token_id": 2,
+  "exponential_decay_length_penalty": null,
+  "ffn_hidden_size": 5632,
+  "finetuning_task": null,
+  "forced_bos_token_id": null,
+  "forced_eos_token_id": null,
+  "glu": true,
+  "id2label": {
+    "0": "LABEL_0",
+    "1": "LABEL_1"
+  },
+  "initializer_range": 0.01,
+  "is_decoder": false,
+  "is_encoder_decoder": false,
+  "kv_channels": 128,
+  "label2id": {
+    "LABEL_0": 0,
+    "LABEL_1": 1
+  },
+  "layer_norm_epsilon": 1e-05,
+  "length_penalty": 1.0,
+  "max_length": 20,
+  "min_length": 0,
+  "model_type": "jetmoe",
+  "moe_num_experts": 8,
+  "moe_top_k": 2,
+  "n_embd": 2048,
+  "n_head": 16,
+  "n_layer": 24,
+  "n_positions": 4096,
+  "no_repeat_ngram_size": 0,
+  "num_beam_groups": 1,
+  "num_beams": 1,
+  "num_key_value_heads": 8,
+  "num_layers": 24,
+  "num_return_sequences": 1,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "output_scores": false,
+  "pad_token_id": null,
+  "prefix": null,
+  "problem_type": null,
+  "pruned_heads": {},
+  "remove_invalid_values": false,
+  "repetition_penalty": 1.0,
+  "return_dict": true,
+  "return_dict_in_generate": false,
+  "rms_norm_eps": 1e-05,
+  "rope_theta": 10000.0,
+  "rotary_percent": 1.0,
+  "sep_token_id": null,
+  "suppress_tokens": null,
+  "task_specific_params": null,
+  "temperature": 1.0,
+  "tf_legacy_loss": false,
+  "tie_encoder_decoder": false,
+  "tie_word_embeddings": true,
+  "tokenizer_class": null,
+  "top_k": 50,
+  "top_p": 1.0,
+  "torchscript": false,
+  "transformers_version": null,
+  "typical_p": 1.0,
+  "use_bfloat16": false,
+  "use_cache": true,
+  "vocab_size": 32000
+}

configuration_jetmoe.py

@@ -0,0 +1,269 @@
+""" JetMoE model configuration"""
+from collections import OrderedDict
+from typing import Any, List, Mapping, Optional
+
+from transformers import PreTrainedTokenizer, TensorType, is_torch_available
+from transformers.configuration_utils import PretrainedConfig
+from transformers.onnx import OnnxConfigWithPast, PatchingSpec
+from transformers.utils import logging
+import torch.nn.init as init
+import json
+
+logger = logging.get_logger(__name__)
+
+
+class JetMoEConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`JetMoEModel`]. It is used to instantiate a
+    JetMoE model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the JetMoE
+    [jetmoe-small](https://huggingface.co/jetmoe-small) architecture. Configuration objects
+    inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from
+    [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 50400):
+            Vocabulary size of the JetMoE model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`JetMoEModel`].
+        n_positions (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        n_embd (`int`, *optional*, defaults to 4096):
+            Dimensionality of the embeddings and hidden states.
+        n_layer (`int`, *optional*, defaults to 28):
+            Number of hidden layers in the Transformer encoder.
+        n_head (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        rotary_dim (`int`, *optional*, defaults to 64):
+            Number of dimensions in the embedding that Rotary Position Embedding is applied to.
+        n_inner (`int`, *optional*, defaults to None):
+            Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
+        activation_function (`str`, *optional*, defaults to `"gelu_new"`):
+            Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`.
+        resid_pdrop (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        embd_pdrop (`int`, *optional*, defaults to 0.1):
+            The dropout ratio for the embeddings.
+        attn_pdrop (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention.
+        layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
+            The epsilon to use in the layer normalization layers.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+
+    Example:
+
+    ```python
+    >>> from transformers import JetMoEConfig, JetMoEModel
+
+    >>> # Initializing a JetMoE 6B configuration
+    >>> configuration = JetMoEConfig()
+
+    >>> # Initializing a model (with random weights) from the configuration
+    >>> model = JetMoEModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "jetmoe"
+    attribute_map = {
+        "max_position_embeddings": "n_positions",
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "num_layers",
+    }
+
+    def __init__(
+        self,
+        vocab_size=50295,
+        hidden_size=1024,
+        num_layers=24,
+        num_attention_heads=16,
+        kv_channels = 128,
+        ffn_hidden_size=2048,
+        max_position_embeddings=4096,
+        rotary_percent=1.0,
+        activation_function="silu",
+        glu=True,
+        moe_num_experts=8,
+        moe_top_k=2,
+        use_cache=True,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=True,
+        bias=True,
+        rope_theta=10000.0,
+        rms_norm_eps=1e-6,
+        initializer_range=0.01,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_layers = num_layers
+        self.num_attention_heads = num_attention_heads
+        self.kv_channels = kv_channels
+        self.ffn_hidden_size = ffn_hidden_size
+        self.max_position_embeddings = max_position_embeddings
+        self.rotary_percent = rotary_percent
+        self.activation_function = activation_function
+        self.glu = glu
+        self.moe_num_experts = moe_num_experts
+        self.moe_top_k = moe_top_k
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+
+        self.init_method = init.xavier_uniform_
+        self.output_layer_init_method = init.xavier_uniform_
+        self.bias = bias
+        self.rope_theta = rope_theta
+        self.rms_norm_eps = rms_norm_eps
+
+        super().__init__(
+            bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
+        )
+
+    def to_dict(self):
+        """Returns a dictionary representation of the config, excluding non-serializable attributes."""
+        return {k: v for k, v in self.__dict__.items() if k not in ['init_method', 'output_layer_init_method', 'torch_dtype', '_pre_quantization_dtype', 'quantization_config']}
+
+    def to_json_string(self, use_diff=False):
+        """Serializes this instance to a JSON string, excluding non-serializable attributes.
+        
+        Args:
+            use_diff (bool): Whether to use differences with the default config. This argument is
+                             accepted for compatibility with the transformers library but is not
+                             used in this custom implementation.
+        """
+        config_dict = self.to_dict()  # Assuming you have a to_dict method as shown earlier
+        return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
+
+class JetMoEOnnxConfig(OnnxConfigWithPast):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        task: str = "default",
+        patching_specs: List[PatchingSpec] = None,
+        use_past: bool = False,
+    ):
+        """
+        Initialize the JetMoEOnnxConfig.
+
+        Args:
+            config (PretrainedConfig): Pretrained model configuration.
+            task (str): Task description.
+            patching_specs (List[PatchingSpec]): List of patching specifications.
+            use_past (bool): Whether to use past tokens in the configuration.
+        """
+        super().__init__(config, task=task, patching_specs=patching_specs, use_past=use_past)
+        if not getattr(self._config, "pad_token_id", None):
+            # TODO: how to do that better?
+            self._config.pad_token_id = 0
+
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        """
+        Define the input mappings.
+
+        Returns:
+            Mapping[str, Mapping[int, str]]: Input mappings.
+        """
+        common_inputs = OrderedDict({"input_ids": {0: "batch", 1: "sequence"}})
+        if self.use_past:
+            self.fill_with_past_key_values_(common_inputs, direction="inputs")
+            common_inputs["attention_mask"] = {0: "batch", 1: "past_sequence + sequence"}
+        else:
+            common_inputs["attention_mask"] = {0: "batch", 1: "sequence"}
+
+        return common_inputs
+
+    @property
+    def num_layers(self) -> int:
+        """
+        Get the number of layers.
+
+        Returns:
+            int: Number of layers.
+        """
+        return self._config.n_layer
+
+    @property
+    def num_attention_heads(self) -> int:
+        """
+        Get the number of attention heads.
+
+        Returns:
+            int: Number of attention heads.
+        """
+        return self._config.n_head
+
+    def generate_dummy_inputs(
+        self,
+        tokenizer: PreTrainedTokenizer,
+        batch_size: int = -1,
+        seq_length: int = -1,
+        is_pair: bool = False,
+        framework: Optional[TensorType] = None,
+    ) -> Mapping[str, Any]:
+        """
+        Generate dummy inputs for testing.
+
+        Args:
+            tokenizer (PreTrainedTokenizer): Pretrained tokenizer.
+            batch_size (int): Batch size.
+            seq_length (int): Sequence length.
+            is_pair (bool): Whether the input is a pair.
+            framework (Optional[TensorType]): Tensor framework.
+
+        Returns:
+            Mapping[str, Any]: Dummy inputs.
+        """
+        common_inputs = super(OnnxConfigWithPast, self).generate_dummy_inputs(
+            tokenizer, batch_size=batch_size, seq_length=seq_length, is_pair=is_pair, framework=framework
+        )
+
+        # We need to order the input in the way they appears in the forward()
+        ordered_inputs = OrderedDict({"input_ids": common_inputs["input_ids"]})
+
+        # Need to add the past_keys
+        if self.use_past:
+            if not is_torch_available():
+                raise ValueError("Cannot generate dummy past_keys inputs without PyTorch installed.")
+            else:
+                import torch
+
+                batch, seqlen = common_inputs["input_ids"].shape
+                # Not using the same length for past_key_values
+                past_key_values_length = seqlen + 2
+                past_shape = (
+                    batch,
+                    self.num_attention_heads,
+                    past_key_values_length,
+                    self._config.hidden_size // self.num_attention_heads,
+                )
+                ordered_inputs["past_key_values"] = [
+                    (torch.zeros(past_shape), torch.zeros(past_shape)) for _ in range(self.num_layers)
+                ]
+
+        ordered_inputs["attention_mask"] = common_inputs["attention_mask"]
+        if self.use_past:
+            mask_dtype = ordered_inputs["attention_mask"].dtype
+            ordered_inputs["attention_mask"] = torch.cat(
+                [ordered_inputs["attention_mask"], torch.ones(batch, past_key_values_length, dtype=mask_dtype)], dim=1
+            )
+
+        return ordered_inputs
+
+    @property
+    def default_onnx_opset(self) -> int:
+        """
+        Get the default ONNX opset version.
+
+        Returns:
+            int: Default ONNX opset version.
+        """
+        return 13

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 2,
+  "transformers_version": "4.39.0"
+}

model-00001-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abcc8c363bff8e2c81a5aa32825bbdef3f02d306bedc6f0be09ed2724f77104a
+size 4879573992

model-00002-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:704135057509ea3d46459fbecdc4fb09d978dcca63ad5305af565f2620c1af1b
+size 4933084288

model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4208b59359347e1976e859c803c5bf88f18fafeee9ec02f97277d2028e01b86c
+size 4933084344

model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc65610116e723056a06f19f4f7b94379c0c701500f0e86eb51ed97c33cf8f14
+size 2298765576

model.safetensors.index.json

@@ -0,0 +1,273 @@
+{
+  "metadata": {
+    "total_size": 17044475904
+  },
+  "weight_map": {
+    "model.embed_tokens.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.bias": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attention.experts.bias": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.bias": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attention.experts.bias": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.bias": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attention.experts.bias": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.bias": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attention.experts.bias": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.bias": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attention.experts.bias": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.bias": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.13.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attention.experts.bias": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.14.mlp.bias": "model-00003-of-00004.safetensors",
+    "model.layers.14.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.14.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.14.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.14.self_attention.experts.bias": "model-00003-of-00004.safetensors",
+    "model.layers.14.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.14.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.14.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.14.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.mlp.bias": "model-00003-of-00004.safetensors",
+    "model.layers.15.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.self_attention.experts.bias": "model-00003-of-00004.safetensors",
+    "model.layers.15.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.15.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.bias": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.self_attention.experts.bias": "model-00003-of-00004.safetensors",
+    "model.layers.16.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.bias": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attention.experts.bias": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.bias": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attention.experts.bias": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.bias": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attention.experts.bias": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.bias": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attention.experts.bias": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.bias": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.output_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.20.mlp.router.layer.weight": "model-00004-of-00004.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attention.experts.bias": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attention.experts.input_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attention.experts.output_linear.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attention.experts.router.layer.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attention.kv_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.21.mlp.bias": "model-00004-of-00004.safetensors",
+    "model.layers.21.mlp.input_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.21.mlp.output_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.21.mlp.router.layer.weight": "model-00004-of-00004.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.21.self_attention.experts.bias": "model-00004-of-00004.safetensors",
+    "model.layers.21.self_attention.experts.input_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.21.self_attention.experts.output_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.21.self_attention.experts.router.layer.weight": "model-00004-of-00004.safetensors",
+    "model.layers.21.self_attention.kv_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.mlp.bias": "model-00004-of-00004.safetensors",
+    "model.layers.22.mlp.input_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.mlp.output_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.mlp.router.layer.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.self_attention.experts.bias": "model-00004-of-00004.safetensors",
+    "model.layers.22.self_attention.experts.input_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.self_attention.experts.output_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.self_attention.experts.router.layer.weight": "model-00004-of-00004.safetensors",
+    "model.layers.22.self_attention.kv_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.mlp.bias": "model-00004-of-00004.safetensors",
+    "model.layers.23.mlp.input_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.mlp.output_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.mlp.router.layer.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.self_attention.experts.bias": "model-00004-of-00004.safetensors",
+    "model.layers.23.self_attention.experts.input_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.self_attention.experts.output_linear.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.self_attention.experts.router.layer.weight": "model-00004-of-00004.safetensors",
+    "model.layers.23.self_attention.kv_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.bias": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attention.experts.bias": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.bias": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attention.experts.bias": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.bias": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attention.experts.bias": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.mlp.bias": "model-00001-of-00004.safetensors",
+    "model.layers.6.mlp.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attention.experts.bias": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attention.experts.input_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attention.experts.output_linear.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attention.experts.router.layer.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.self_attention.kv_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.bias": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attention.experts.bias": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.bias": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attention.experts.bias": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.bias": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attention.experts.bias": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attention.experts.input_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attention.experts.output_linear.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attention.experts.router.layer.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attention.kv_proj.weight": "model-00002-of-00004.safetensors",
+    "model.norm.weight": "model-00004-of-00004.safetensors"
+  }
+}

modeling_jetmoe.py

@@ -0,0 +1,1399 @@
+""" PyTorch JetMoE model."""
+
+from typing import List, Optional, Tuple, Union
+import warnings, math
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss
+from torch.nn import functional as F
+
+import megablocks
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast, 
+    CausalLMOutputWithPast,
+    SequenceClassifierOutputWithPast,
+    dataclass
+)
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    add_start_docstrings, 
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    replace_return_docstrings, 
+    logging
+)
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa
+from transformers.cache_utils import Cache, DynamicCache
+from .configuration_jetmoe import JetMoEConfig
+from jetmoe_model.utils import moe
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_func, flash_attn_varlen_func
+    from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "jetmoe"
+_CONFIG_FOR_DOC = "JetMoEConfig"
+
+
+@dataclass
+class JetMoEBaseModelOutputWithPast(BaseModelOutputWithPast):
+    """
+    Base class for model's outputs that may also contain a past key/values (to speed up sequential decoding).
+
+    Args:
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+
+            If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1,
+            hidden_size)` is output.
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            `config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
+            input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    last_hidden_state: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    aux_loss: Optional[torch.FloatTensor] = None
+
+
+@dataclass
+class JetMoECausalLMOutputWithPast(CausalLMOutputWithPast):
+    """
+    Base class for causal language model (or autoregressive) outputs.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Language modeling loss (for next-token prediction).
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            `past_key_values` input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    aux_loss: Optional[torch.FloatTensor] = None
+
+
+@dataclass
+class JetMoESequenceClassifierOutputWithPast(SequenceClassifierOutputWithPast):
+    """
+    Base class for outputs of sentence classification models.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (`torch.FloatTensor` of shape `(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+            `past_key_values` input) to speed up sequential decoding.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    aux_loss: Optional[torch.FloatTensor] = None
+
+
+# Copied from transformers.models.llama.modeling_llama._get_unpad_data
+def _get_unpad_data(attention_mask):
+    seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
+    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
+    max_seqlen_in_batch = seqlens_in_batch.max().item()
+    cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
+    return (
+        indices,
+        cu_seqlens,
+        max_seqlen_in_batch,
+    )
+
+class JetMoERMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        JetMoERMSNorm module
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+# copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding
+class JetMoERotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Build here to make `torch.jit.trace` work.
+        self._set_cos_sin_cache(
+            seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
+        )
+
+    def _set_cos_sin_cache(self, seq_len, device, dtype):
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
+
+        freqs = torch.outer(t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
+
+        return (
+            self.cos_cached[:seq_len].to(dtype=x.dtype),
+            self.sin_cached[:seq_len].to(dtype=x.dtype),
+        )
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=2):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`):
+            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+            used to pass offsetted position ids when working with a KV-cache.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos[position_ids].unsqueeze(unsqueeze_dim)
+    sin = sin[position_ids].unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class JetMoEAttention(nn.Module):
+    """
+    Multi-headed attention from 'Attention Is All You Need' paper.
+    """
+
+    def __init__(self, config: JetMoEConfig, layer_idx: Optional[int] = None):
+        """
+        Initialize the JetMoEAttention module.
+
+        Args:
+            config: Configuration object with model hyperparameters.
+        """
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.is_causal = True
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.top_k = config.moe_top_k
+
+        self.kv_projection_size = config.kv_channels * config.num_attention_heads
+        self.num_key_value_heads = config.num_attention_heads
+        self.num_heads = self.num_key_value_heads * self.top_k
+        self.hidden_size_per_attention_head = config.kv_channels
+
+        self.experts = moe.MoE(
+            input_size=config.hidden_size,
+            hidden_size=self.kv_projection_size,
+            num_experts=config.moe_num_experts,
+            top_k=config.moe_top_k,
+            glu=False
+        )
+
+        self.kv_proj = torch.nn.Linear(
+            config.hidden_size, self.kv_projection_size * 2, bias=False
+            )
+        
+        self.rotary_emb = JetMoERotaryEmbedding(
+            config.kv_channels,
+            max_position_embeddings=config.max_position_embeddings,
+            base=config.rope_theta,
+        )
+
+    # def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+    #     return tensor.view(bsz, seq_len, self.num_attention_heads, self.hidden_size_per_attention_head).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states, aux_loss = self.experts.map(hidden_states)
+        key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.hidden_size_per_attention_head).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.hidden_size_per_attention_head).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.hidden_size_per_attention_head).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids, unsqueeze_dim=1)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = key_states.repeat(1, self.top_k, 1, 1)
+        value_states = value_states.repeat(1, self.top_k, 1, 1)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.hidden_size_per_attention_head)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        # attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.hidden_size_per_attention_head):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.hidden_size_per_attention_head)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.top_k, self.kv_projection_size)
+
+        attn_output = self.experts.reduce(attn_output)
+        attn_output = attn_output.view(bsz, q_len, -1)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value, aux_loss
+    
+
+# copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->JetMoE
+class JetMoESdpaAttention(JetMoEAttention):
+    """
+    JetMoE attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+    `JetMoEAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+    SDPA API.
+    """
+
+    # Adapted from JetMoEAttention.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "JetMoEModel is using JetMoESdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+            )
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states, aux_loss = self.experts.map(hidden_states)
+        key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.hidden_size_per_attention_head).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.hidden_size_per_attention_head).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.hidden_size_per_attention_head).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids, unsqueeze_dim=1)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        key_states = key_states.repeat(1, self.top_k, 1, 1)
+        value_states = value_states.repeat(1, self.top_k, 1, 1)
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and attention_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=0.0,
+            # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+            is_causal=self.is_causal and attention_mask is None and q_len > 1,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.top_k, self.kv_projection_size)
+
+        attn_output = self.experts.reduce(attn_output)
+        attn_output = attn_output.view(bsz, q_len, -1)
+
+        return attn_output, None, past_key_value, aux_loss
+
+
+class JetMoEFlashAttention2(JetMoEAttention):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: Optional[torch.FloatTensor],
+        attention_mask: Optional[torch.FloatTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        output_attentions: Optional[bool] = False,
+        **kwargs,
+    ) -> Union[
+        Tuple[torch.Tensor, Tuple[torch.Tensor]],
+        Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],
+    ]:
+        """
+        Forward pass of the JetMoEAttention module.
+
+        Args:
+            hidden_states (Optional[torch.FloatTensor]): Input hidden states.
+            attention_mask (Optional[torch.FloatTensor]): Attention mask.
+            layer_past (Optional[Tuple[torch.Tensor]]): Past layer state.
+            use_cache (Optional[bool]): Whether to use cached states.
+            output_attentions (Optional[bool]): Whether to output attention weights.
+
+        Returns:
+            Union[Tuple[torch.Tensor, Tuple[torch.Tensor]], Optional[Tuple[...]]]: Tuple containing outputs.
+        """
+        #assert attention_mask is None, "attention_mask is not supported"
+        assert output_attentions is False, "output_attentions is not supported"
+
+        B, T, C = hidden_states.size() # batch size, sequence length, embedding dimensionality (hidden_size)
+
+        # calculate query, key, values 
+        query_layer, aux_loss = self.experts.map(hidden_states)
+        key_layer, value_layer = self.kv_proj(hidden_states).chunk(2, dim=-1)
+
+        query_layer = query_layer.view(B, T, self.num_heads, self.hidden_size_per_attention_head) # (B, T, k * nh, hs)
+        key_layer = key_layer.view(B, T, self.num_key_value_heads, self.hidden_size_per_attention_head) # (B, T, nh, hs)
+        value_layer = value_layer.view(B, T, self.num_key_value_heads, self.hidden_size_per_attention_head) # (B, T, nh, hs)
+
+        kv_seq_len = key_layer.shape[1]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_layer, seq_len=kv_seq_len)
+        query_layer, key_layer = apply_rotary_pos_emb(query_layer, key_layer, cos, sin, position_ids)
+
+        # query_layer = query_layer.contiguous()
+        # expand the key_layer and value_layer [sk, b, ng, hn] -> [sk, b, np, hn]
+        key_layer = key_layer.repeat(1, 1, self.top_k, 1)
+        value_layer = value_layer.repeat(1, 1, self.top_k, 1)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            # print(self.layer_idx, key_layer.size())
+            key_layer = key_layer.transpose(1, 2)
+            value_layer = value_layer.transpose(1, 2)
+            key_layer, value_layer = past_key_value.update(key_layer, value_layer, self.layer_idx, cache_kwargs)
+            key_layer = key_layer.transpose(1, 2)
+            value_layer = value_layer.transpose(1, 2)
+
+        context_layer = self._flash_attention_forward(
+            query_layer,
+            key_layer,
+            value_layer,
+            attention_mask,
+            T,
+        ) 
+
+        # output projection
+        y = self.experts.reduce(context_layer.reshape(T, B, self.top_k, self.kv_projection_size))
+        y = y.view(B, T, C) # re-assemble all head outputs side by side
+
+        if not output_attentions:
+            attn_weights = None
+
+        return y, attn_weights, past_key_value, aux_loss
+
+    def _flash_attention_forward(
+        self,
+        query_states,
+        key_states,
+        value_states,
+        attention_mask,
+        query_length,
+        dropout=0.0,
+        softmax_scale=None,
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`float`):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states,
+                key_states,
+                value_states,
+                dropout,
+                softmax_scale=softmax_scale,
+                causal=causal
+            )
+
+        return attn_output
+
+
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
+JETMOE_ATTENTION_CLASSES = {
+    "eager": JetMoEAttention,
+    "flash_attention_2": JetMoEFlashAttention2,
+    "sdpa": JetMoESdpaAttention,
+}
+
+
+class JetMoEBlock(nn.Module):
+    def __init__(self, config: JetMoEConfig, layer_idx: Optional[int] = None):
+        """
+        Initialize the JetMoEBlock module.
+
+        Args:
+            config: Configuration object with model hyperparameters.
+        """
+        super().__init__()
+        self.input_layernorm = JetMoERMSNorm(config.hidden_size)
+        #self.self_attention = JetMoEAttention(config, layer_idx)
+        self.self_attention = JETMOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
+        self.post_attention_layernorm = JetMoERMSNorm(config.hidden_size)
+
+        moe_args = megablocks.layers.arguments.from_megatron(config)
+        moe_args.activation_fn = F.silu
+        moe_args.return_bias = False
+        # self.mlp = megablocks.layers.dmoe.dMoE(moe_args)
+        self.mlp = moe.MoE(
+            input_size=config.hidden_size,
+            hidden_size=config.ffn_hidden_size,
+            num_experts=config.moe_num_experts,
+            activation=F.silu,
+            top_k=config.moe_top_k,
+            glu=config.glu
+        )
+
+    def forward(
+        self,
+        hidden_states: Optional[torch.FloatTensor],
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        **kwargs,
+    ) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
+        """
+        Forward pass of the JetMoEBlock module.
+
+        Args:
+            hidden_states (Optional[torch.FloatTensor]): Input hidden states.
+            layer_past (Optional[Tuple[torch.Tensor]]): Past layer state.
+            attention_mask (Optional[torch.FloatTensor]): Attention mask.
+            head_mask (Optional[torch.FloatTensor]): Head mask.
+            use_cache (Optional[bool]): Whether to use cached states.
+            output_attentions (Optional[bool]): Whether to output attention weights.
+
+        Returns:
+            Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
+            Tuple containing outputs or optional attention weights.
+        """
+        # Self Attention
+        attn_output, self_attn_weights, present_key_value, att_aux_loss = self.self_attention(
+            hidden_states=self.input_layernorm(hidden_states),
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+        )
+
+        hidden_states = hidden_states + attn_output
+        x_mlp, mlp_aux_loss = self.mlp(self.post_attention_layernorm(hidden_states))
+        hidden_states = hidden_states + x_mlp
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        outputs += (att_aux_loss + mlp_aux_loss,)
+
+        return outputs
+
+
+
+class JetMoEPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = JetMoEConfig
+    base_model_prefix = "transformer"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["JetMoEBlock"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_cache_class = True
+
+    def __init__(self, *inputs, **kwargs):
+        """
+        Initialize the JetMoEPreTrainedModel.
+
+        Args:
+            *inputs: Variable length input arguments.
+            **kwargs: Keyword arguments.
+        """
+        super().__init__(*inputs, **kwargs)
+
+        self.gradient_checkpointing = False
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, (nn.Linear,)):
+            # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    # def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs={}):
+    #     for module in self.modules():
+    #         if hasattr(module, "gradient_checkpointing"):
+    #             self._set_gradient_checkpointing(
+    #                 module, True, gradient_checkpointing_kwargs
+    #             )
+
+    # def gradient_checkpointing_disable(self):
+    #     for module in self.modules():
+    #         if hasattr(module, "gradient_checkpointing"):
+    #             self._set_gradient_checkpointing(
+    #                 module, False
+    #             )
+
+    # def _set_gradient_checkpointing(
+    #     self,
+    #     module,
+    #     value=False,
+    #     gradient_checkpointing_kwargs={"use_reentrant": False},
+    # ):
+    #     """
+    #     Set gradient checkpointing for the JetMoEModel.
+
+    #     Args:
+    #         module: The module for which gradient checkpointing is set.
+    #         value (bool): Whether to enable gradient checkpointing.
+    #     """
+    #     self._gradient_checkpointing_func = checkpoint
+    #     self.gradient_checkpointing = True
+    #     if isinstance(module, JetMoEModel):
+    #         module.gradient_checkpointing = value
+    #         module.gradient_checkpointing_kwargs = gradient_checkpointing_kwargs
+    #         module._gradient_checkpointing_func = checkpoint
+
+MODULEFORMER_START_DOCSTRING = r"""
+    This model is a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) sub-class. Use
+    it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`JetMoEConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+MODULEFORMER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoProcenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        head_mask (`torch.FloatTensor` of shape `(num_attention_heads,)` or `(n_layer, num_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `({0}, hidden_dim)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare JetMoE Model outputting raw hidden-states without any specific head on top.",
+    MODULEFORMER_START_DOCSTRING,
+)
+class JetMoEModel(JetMoEPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`JetMoEBlock`]
+
+    Args:
+        config: JetMoEConfig
+    """
+
+    def __init__(self, config: JetMoEConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [JetMoEBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self._attn_implementation = config._attn_implementation
+        self.norm = JetMoERMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(MODULEFORMER_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+        elif inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        past_key_values_length = 0
+
+        if use_cache:
+            use_legacy_cache = not isinstance(past_key_values, Cache)
+            if use_legacy_cache:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
+
+        if position_ids is None:
+            device = input_ids.device if input_ids is not None else inputs_embeds.device
+            position_ids = torch.arange(
+                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+            )
+            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+        else:
+            position_ids = position_ids.view(-1, seq_length).long()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
+            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+            if is_padding_right:
+                raise ValueError(
+                    "You are attempting to perform batched generation with padding_side='right'"
+                    " this may lead to unexpected behaviour for Flash Attention version of JetMoE. Make sure to "
+                    " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                )
+
+        if self._attn_implementation == "flash_attention_2":
+            # 2d mask is passed through the layers
+            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
+        elif self._attn_implementation == "sdpa" and not output_attentions:
+            # output_attentions=True can not be supported when using SDPA, and we fall back on
+            # the manual implementation that requires a 4D causal mask in all cases.
+            attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+            )
+        else:
+            # 4d mask is passed through the layers
+            attention_mask = _prepare_4d_causal_attention_mask(
+                attention_mask,
+                (batch_size, seq_length),
+                inputs_embeds,
+                past_key_values_length,
+            )
+
+        hidden_states = inputs_embeds
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        aux_loss = 0
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            # hidden_states: Optional[torch.FloatTensor],
+            # position_ids: Optional[torch.LongTensor] = None,
+            # past_key_value: Optional[Tuple[torch.Tensor]] = None,
+            # attention_mask: Optional[torch.FloatTensor] = None,
+            # output_attentions: Optional[bool] = False,
+            # use_cache: Optional[bool] = False,
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    #decoder_layer.__call__,
+                    decoder_layer,
+                    hidden_states,
+                    position_ids,
+                    past_key_values,
+                    attention_mask,
+                    output_attentions,
+                    use_cache,
+                    use_reentrant=False,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            aux_loss += layer_outputs[-1]
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = None
+        if use_cache:
+            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return JetMoEBaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            aux_loss=aux_loss,
+        )
+
+
+class JetMoEForCausalLM(JetMoEPreTrainedModel):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = JetMoEModel(config)
+        self.vocab_size = config.vocab_size
+        self.aux_loss_coef = getattr(config, 'aux_loss_coef', 0.01)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(MODULEFORMER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+        """
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        logits = self.lm_head(hidden_states)
+        logits = logits.float()
+
+        loss = None
+        if labels is not None:
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Ensure tensors are on the same device
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(shift_logits, shift_labels)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        if labels is not None and self.model.training:
+            loss += self.aux_loss_coef * outputs.aux_loss.to(loss.device)
+
+        return JetMoECausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            aux_loss=outputs.aux_loss,
+        )
+
+    def prepare_inputs_for_generation(
+        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
+    ):
+        # Omit tokens covered by past_key_values
+        if past_key_values is not None:
+            if isinstance(past_key_values, Cache):
+                cache_length = past_key_values.get_seq_length()
+                past_length = past_key_values.seen_tokens
+                max_cache_length = past_key_values.get_max_length()
+            else:
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
+
+            # Keep only the unprocessed tokens:
+            # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
+            # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
+            # input)
+            if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
+                input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
+            # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
+            # input_ids based on the past_length.
+            elif past_length < input_ids.shape[1]:
+                input_ids = input_ids[:, past_length:]
+            # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
+
+            # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
+            if (
+                max_cache_length is not None
+                and attention_mask is not None
+                and cache_length + input_ids.shape[1] > max_cache_length
+            ):
+                attention_mask = attention_mask[:, -max_cache_length:]
+
+        position_ids = kwargs.get("position_ids", None)
+        if attention_mask is not None and position_ids is None:
+            # create position_ids on the fly for batch generation
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -input_ids.shape[1] :]
+
+        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+        if inputs_embeds is not None and past_key_values is None:
+            model_inputs = {"inputs_embeds": inputs_embeds}
+        else:
+            model_inputs = {"input_ids": input_ids}
+
+        model_inputs.update(
+            {
+                "position_ids": position_ids,
+                "past_key_values": past_key_values,
+                "use_cache": kwargs.get("use_cache"),
+                "attention_mask": attention_mask,
+            }
+        )
+        return model_inputs
+
+    @staticmethod
+    def _reorder_cache(past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (
+                tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
+            )
+        return reordered_past
+
+
+@add_start_docstrings(
+    """
+    The JetMoE Model transformer with a sequence classification head on top (linear layer).
+
+    [`JetMoEForSequenceClassification`] uses the last token in order to do the classification, as other causal models
+    (e.g. GPT-2) do.
+
+    Since it does classification on the last token, it requires to know the position of the last token. If a
+    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
+    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
+    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
+    each row of the batch).
+    """,
+    MODULEFORMER_START_DOCSTRING,
+)
+# Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->JetMoE, LLAMA->MODULEFORMER
+class JetMoEForSequenceClassification(JetMoEPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.model = JetMoEModel(config)
+        self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(MODULEFORMER_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
+            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.model(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = transformer_outputs[0]
+        logits = self.score(hidden_states)
+
+        if input_ids is not None:
+            batch_size = input_ids.shape[0]
+        else:
+            batch_size = inputs_embeds.shape[0]
+
+        if self.config.pad_token_id is None and batch_size != 1:
+            raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
+        if self.config.pad_token_id is None:
+            sequence_lengths = -1
+        else:
+            if input_ids is not None:
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
+            else:
+                sequence_lengths = -1
+
+        pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
+
+        loss = None
+        if labels is not None:
+            labels = labels.to(logits.device)
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
+                    self.config.problem_type = "single_label_classification"
+                else:
+                    self.config.problem_type = "multi_label_classification"
+
+            if self.config.problem_type == "regression":
+                loss_fct = MSELoss()
+                if self.num_labels == 1:
+                    loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
+                else:
+                    loss = loss_fct(pooled_logits, labels)
+            elif self.config.problem_type == "single_label_classification":
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
+            elif self.config.problem_type == "multi_label_classification":
+                loss_fct = BCEWithLogitsLoss()
+                loss = loss_fct(pooled_logits, labels)
+        if not return_dict:
+            output = (pooled_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return JetMoESequenceClassifierOutputWithPast(
+            loss=loss,
+            logits=pooled_logits,
+            past_key_values=transformer_outputs.past_key_values,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+            aux_loss=transformer_outputs.aux_loss,
+        )