Upload BolmoForCausalLM

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 +1,74 @@
+{
+  "add_expanded_embeddings": true,
+  "architectures": [
+    "BolmoForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_bolmo.BolmoConfig",
+    "AutoModelForCausalLM": "modeling_bolmo.BolmoForCausalLM"
+  },
+  "bos_token_id": 1,
+  "boundary_predictor_lookahead": 1,
+  "boundary_threshold": "sample:0",
+  "dtype": "float32",
+  "eos_token_id": 1,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "local_intermediate_size": 2816,
+  "local_rms_norm_eps": 1e-05,
+  "max_position_embeddings": 65536,
+  "model_type": "bolmo",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 16,
+  "num_local_decoder_layers": 4,
+  "num_local_encoder_layers": 1,
+  "num_local_heads": 16,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "sliding_window": 4096,
+  "subword_vocab_size": 100278,
+  "tie_word_embeddings": false,
+  "tokenizer_config": {
+    "bos_token_id": 1,
+    "bpe_token_end_id": 3,
+    "eos_token_id": 1,
+    "original_identifier": "allenai/dolma2-tokenizer",
+    "pad_token_id": 0,
+    "special_tokens": [
+      "<pad>",
+      "<bos>",
+      "<eos>",
+      "<bpe_token_end>"
+    ],
+    "special_tokens_first": true,
+    "vocab_size": 520
+  },
+  "transformers_version": "4.57.3",
+  "use_cache": true,
+  "vocab_size": 520
+}

configuration_bolmo.py

@@ -0,0 +1,235 @@
+from dataclasses import asdict
+from typing import Any
+
+from transformers.configuration_utils import PretrainedConfig, layer_type_validation
+from transformers.modeling_rope_utils import rope_config_validation
+from olmo_core.nn.blt.hf.tokenization_bolmo import ByteTokenizerConfig
+
+class BolmoConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Olmo3Model`]. It is used to instantiate an OLMo3
+    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 [allenai/OLMo-3-0725-1B](https://huggingface.co/allenai/OLMo-3-0725-1B).
+
+    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 50304):
+            Vocabulary size of the Olmo3 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Olmo3Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 11008):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details, check out [this
+            paper](https://huggingface.co/papers/2305.13245). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with.
+        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). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*, defaults to 1):
+            Padding token id.
+        bos_token_id (`int`, *optional*):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 50279):
+            End of stream token id.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`list[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+            The epsilon used by the rms normalization layers.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Size of the sliding window for sliding window attention.
+        layer_types (`list`, *optional*):
+            Attention pattern for each layer. Defaults to sliding window attention
+            for 3 out of 4 layers, and full attention for every 4th layer.
+
+    ```python
+    >>> from transformers import Olmo3Model, Olmo3Config
+
+    >>> # Initializing a Olmo3 7B style configuration
+    >>> configuration = Olmo3Config()
+
+    >>> # Initializing a model from the Olmo3 7B style configuration
+    >>> model = Olmo3Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "bolmo"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise_rep",  # we need to replicate here due to the added norm on q and k
+        "layers.*.self_attn.k_proj": "colwise_rep",  # we need to replicate here due to the added norm on q and k
+        "layers.*.self_attn.v_proj": "colwise_rep",  # we need to replicate here due to the added norm on q and k
+        "layers.*.self_attn.o_proj": "rowwise_rep",  # we need to replicate here due to the added norm on q and k
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=520,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        use_cache=True,
+        pad_token_id=1,
+        bos_token_id=None,
+        eos_token_id=50279,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        rms_norm_eps=1e-5,
+        sliding_window=4096,
+        layer_types=None,
+        # bolmo config
+        add_expanded_embeddings: bool = True,
+        boundary_predictor_lookahead: int = 1,
+        boundary_threshold: str = "sample:0",
+        num_local_encoder_layers: int = 1,
+        num_local_decoder_layers: int = 4,
+        num_local_heads: int = 16,
+        local_intermediate_size: int = 5504,
+        local_rms_norm_eps=1e-5,
+        subword_vocab_size: int = 100278, # dolma2_tokenizer subword vocab size
+        tokenizer_config: ByteTokenizerConfig | dict[str, Any] | None = None,
+        **kwargs,
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self._rope_scaling_validation()
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+
+        self.rms_norm_eps = rms_norm_eps
+
+        self.sliding_window = sliding_window
+        self.layer_types = layer_types
+        if self.layer_types is None:
+            self.layer_types = [
+                "sliding_attention" if (i + 1) % 4 != 0 else "full_attention" for i in range(self.num_hidden_layers)
+            ]
+        layer_type_validation(self.layer_types)
+
+        # bolmo configuration
+        self.add_expanded_embeddings = add_expanded_embeddings
+        self.boundary_predictor_lookahead = boundary_predictor_lookahead
+        self.boundary_threshold = boundary_threshold
+        self.num_local_encoder_layers = num_local_encoder_layers
+        self.num_local_decoder_layers = num_local_decoder_layers
+        self.num_local_heads = num_local_heads
+        self.local_intermediate_size = local_intermediate_size
+        self.local_rms_norm_eps = local_rms_norm_eps
+        self.subword_vocab_size = subword_vocab_size
+
+        if tokenizer_config is None:
+            self.tokenizer_config = asdict(ByteTokenizerConfig.bolmo())
+        elif isinstance(tokenizer_config, ByteTokenizerConfig):
+            self.tokenizer_config = asdict(tokenizer_config)
+        else:
+            self.tokenizer_config = tokenizer_config
+
+    def _rope_scaling_validation(self):
+        """
+        Validate the `rope_scaling` configuration.
+        """
+        rope_config_validation(self)
+
+__all__ = ["BolmoConfig"]

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 50279,
+  "pad_token_id": 1,
+  "transformers_version": "4.57.3"
+}

model-00001-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5facf4f9f174a3420bac74dfb6c0b626b0a0f7d61478bea2fd6c731f8bb0e34b
+size 4998875928

model-00002-of-00002.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:307b7023df9cf08d41c5c02b599f805b2774e30b7a612cfacbf1a776dd3135be
+size 876802264

model.safetensors.index.json

@@ -0,0 +1,271 @@
+{
+  "metadata": {
+    "total_parameters": 1468911776,
+    "total_size": 5875647104
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00002-of-00002.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.12.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.12.post_feedforward_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.12.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.post_feedforward_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.post_feedforward_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.mlp.up_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.post_feedforward_layernorm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.self_attn.k_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.self_attn.q_norm.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00002.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.post_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.k_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.q_norm.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.in_projection.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.in_projection.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.initial_norm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.pre_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.pre_xlstm_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.fgate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.fgate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.igate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.igate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.k.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.multihead_norm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.ogate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.out_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.q.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.0.xlstm.v.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.pre_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.pre_xlstm_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.fgate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.fgate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.igate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.igate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.k.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.multihead_norm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.ogate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.out_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.q.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.1.xlstm.v.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.pre_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.pre_xlstm_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.fgate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.fgate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.igate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.igate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.k.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.multihead_norm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.ogate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.out_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.q.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.2.xlstm.v.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.pre_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.pre_xlstm_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.fgate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.fgate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.igate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.igate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.k.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.multihead_norm.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.ogate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.out_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.q.weight": "model-00001-of-00002.safetensors",
+    "model.local_decoder.layers.3.xlstm.v.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.boundary_predictor_module.k_proj_layer.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.boundary_predictor_module.q_proj_layer.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.byte_embedding.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.mlp.gate_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.mlp.up_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.pre_feedforward_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.pre_xlstm_layernorm.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.fgate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.fgate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.igate_preact.bias": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.igate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.k.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.multihead_norm.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.ogate_preact.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.out_proj.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.q.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.layers.0.xlstm.v.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.out_projection.bias": "model-00001-of-00002.safetensors",
+    "model.local_encoder.out_projection.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.post_last_block_norm.weight": "model-00001-of-00002.safetensors",
+    "model.local_encoder.subword_embedding.weight": "model-00001-of-00002.safetensors",
+    "model.norm.weight": "model-00002-of-00002.safetensors"
+  }
+}

modeling_bolmo.py

@@ -0,0 +1,1295 @@
+import copy
+from typing import Callable, Optional, Union, cast
+import math
+
+import torch
+import torch.nn as nn
+from torch.nn import functional as F
+
+from transformers.utils.generic import TransformersKwargs
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.integrations import use_kernel_forward_from_hub
+from transformers.masking_utils import create_causal_mask, create_sliding_window_causal_mask
+from transformers.modeling_layers import GradientCheckpointingLayer
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.processing_utils import Unpack
+from transformers.utils import auto_docstring, can_return_tuple
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import check_model_inputs
+
+from olmo_core.nn.blt.hf.configuration_bolmo import BolmoConfig
+from olmo_core.nn.blt.hf.tokenization_bolmo import ByteTokenizerConfig
+from olmo_core.nn.blt.hf.utils_bolmo import compute_boundary_mask, pad_right, pad_left, MaskState
+
+from xlstm.xlstm_large.model import mLSTMLayer, mLSTMLayerConfig, mLSTMLayerStateType, soft_cap, mLSTMBackendConfig
+
+
+@use_kernel_forward_from_hub("RMSNorm")
+class BolmoRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        BolmoRMSNorm is equivalent to T5LayerNorm
+        """
+        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)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs: Unpack[TransformersKwargs],
+):
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query.dtype)
+    attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """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`, *optional*):
+            Deprecated and unused.
+        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.
+    """
+    q_type, k_type = q.dtype, k.dtype
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed.to(q_type), k_embed.to(k_type)
+
+
+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)
+
+
+class BolmoAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: BolmoConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = config.attention_dropout
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(
+            config.hidden_size, config.num_attention_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.k_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.v_proj = nn.Linear(
+            config.hidden_size, config.num_key_value_heads * self.head_dim, bias=config.attention_bias
+        )
+        self.o_proj = nn.Linear(
+            config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
+        )
+        self.q_norm = BolmoRMSNorm(config.num_attention_heads * self.head_dim, config.rms_norm_eps)
+        self.k_norm = BolmoRMSNorm(config.num_key_value_heads * self.head_dim, config.rms_norm_eps)
+        assert config.layer_types is not None
+        self.attention_type = config.layer_types[layer_idx]
+        self.sliding_window = config.sliding_window if self.attention_type == "sliding_attention" else None
+
+    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_values: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor]]:
+        input_shape = hidden_states.shape[:-1]
+        hidden_shape = (*input_shape, -1, self.head_dim)
+
+        query_states = self.q_norm(self.q_proj(hidden_states))
+        key_states = self.k_norm(self.k_proj(hidden_states))
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(hidden_shape).transpose(1, 2)
+        key_states = key_states.view(hidden_shape).transpose(1, 2)
+        value_states = value_states.view(hidden_shape).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_values is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            sliding_window=self.sliding_window,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+        return attn_output, attn_weights
+
+
+class BolmoMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class BolmoDecoderLayer(GradientCheckpointingLayer):
+    def __init__(self, config: BolmoConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = BolmoAttention(config=config, layer_idx=layer_idx)
+
+        self.mlp = BolmoMLP(config)
+        self.post_attention_layernorm = BolmoRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_feedforward_layernorm = BolmoRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    @deprecate_kwarg("past_key_value", new_name="past_key_values", version="4.58")
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> torch.Tensor:
+        residual = hidden_states
+        attn_out, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = self.post_attention_layernorm(attn_out)
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        mlp_out = self.mlp(hidden_states)
+        hidden_states = self.post_feedforward_layernorm(mlp_out)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class BolmoBoundaryPredictor(nn.Module):
+    def __init__(self, config: BolmoConfig):
+        super().__init__()
+
+        self.d_model = config.hidden_size
+        self.boundary_threshold = config.boundary_threshold
+        self.boundary_predictor_lookahead = config.boundary_predictor_lookahead
+        self.q_proj_layer = nn.Linear(self.d_model, self.d_model, bias=False)
+        self.k_proj_layer = nn.Linear(self.d_model, self.d_model, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+        epsilon: float = 1e-3,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        if self.boundary_predictor_lookahead == 0:
+            # do not use the same rep for k and v, use current and one before as in H-Net + pad with negative to the left
+            cos_sim = torch.cat([
+                torch.ones((hidden_states.shape[0], 1), device=hidden_states.device, dtype=hidden_states.dtype) * -1,
+                torch.einsum(
+                    "b l d, b l d -> b l",
+                    F.normalize(self.q_proj_layer(hidden_states[:, :-1]), dim=-1),
+                    F.normalize(self.k_proj_layer(hidden_states[:, 1:]), dim=-1),
+                )
+            ], dim=1)
+        else:
+            cos_sim = torch.einsum(
+                "b l d, b l d -> b l",
+                F.normalize(self.q_proj_layer(hidden_states[:, :-self.boundary_predictor_lookahead]), dim=-1),
+                F.normalize(self.k_proj_layer(hidden_states[:, self.boundary_predictor_lookahead:]), dim=-1),
+            )
+        boundary_logprobs = torch.log1p(-cos_sim.float().clip(max=1.0 - epsilon)) - math.log(2)
+        POSITIVE_LOGPROB = 0.0
+        NEGATIVE_LOGPROB = -100_000
+        if sequence_start_indices is None:
+            boundary_logprobs[:, 0] = POSITIVE_LOGPROB
+        else:
+            pad_mask = torch.arange(boundary_logprobs.shape[1], device=boundary_logprobs.device)[None, :] < sequence_start_indices[:, None]
+            boundary_logprobs = boundary_logprobs.masked_fill(pad_mask, NEGATIVE_LOGPROB)
+            boundary_logprobs[torch.arange(len(boundary_logprobs), device=boundary_logprobs.device), sequence_start_indices] = POSITIVE_LOGPROB
+
+        boundary_logprobs = F.pad(boundary_logprobs, (0, self.boundary_predictor_lookahead), "constant", NEGATIVE_LOGPROB)
+        boundary_mask = compute_boundary_mask(boundary_logprobs, self.boundary_threshold)
+
+        return boundary_logprobs, boundary_mask
+
+
+class BolmoXLSTMLayer(mLSTMLayer):
+    def __init__(self, config: BolmoConfig):
+        super().__init__(mLSTMLayerConfig(
+            embedding_dim=config.hidden_size,
+            num_heads=config.num_local_heads,
+            mlstm_backend=mLSTMBackendConfig(
+                chunkwise_kernel="chunkwise--triton_limit_chunk",
+                sequence_kernel="native_sequence__triton",
+                step_kernel="triton",
+                mode="train",
+                return_last_states=True,
+                autocast_kernel_dtype="float32",
+            )
+        ))
+
+    # original forward adapted to support sequence_start_indices
+    # i.e. set the forget gate to zero at the start of sequence
+    def _original_forward(
+        self, x: torch.Tensor,
+        state: mLSTMLayerStateType | None = None,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+    ) -> tuple[torch.Tensor, mLSTMLayerStateType | None]:
+        assert x.ndim == 3, f"Input must have shape [B, S, D], got {x.shape}"
+        B, S, _ = x.shape
+        if self.config.weight_mode == "single":
+            q = self.q(x)
+            k = self.k(x)
+            v = self.v(x)
+            o_preact = self.ogate_preact(x)
+            i_preact = soft_cap(
+                self.igate_preact(x), cap_value=self.config.gate_soft_cap
+            )
+            f_preact = soft_cap(
+                self.fgate_preact(x), cap_value=self.config.gate_soft_cap
+            )
+        elif self.config.weight_mode == "fused":
+            qkv_opreact = self.qkv_opreact(x)
+            q, k, v, o_preact = torch.tensor_split(
+                qkv_opreact,
+                (
+                    self.qk_dim,
+                    2 * self.qk_dim,
+                    2 * self.qk_dim + self.v_dim,
+                ),
+                dim=-1,
+            )
+
+            if_preact = soft_cap(
+                self.ifgate_preact(x), cap_value=self.config.gate_soft_cap
+            )
+            i_preact, f_preact = torch.tensor_split(
+                if_preact, (self.config.num_heads,), dim=-1
+            )
+        else:
+            raise ValueError(f"Unknown weight_mode: {self.config.weight_mode}")
+
+        q = q.reshape(B, S, self.config.num_heads, -1).transpose(1, 2)
+        k = k.reshape(B, S, self.config.num_heads, -1).transpose(1, 2)
+        v = v.reshape(B, S, self.config.num_heads, -1).transpose(1, 2)
+
+        if sequence_start_indices is not None:
+            f_preact[torch.arange(B, device=f_preact.device), sequence_start_indices] = -100_000
+
+        i_preact = i_preact.transpose(1, 2)
+        f_preact = f_preact.transpose(1, 2)
+        if state is None:
+            c_initial, n_initial, m_initial = None, None, None
+        else:
+            c_initial, n_initial, m_initial = state
+
+        h, state = self.mlstm_backend(
+            q=q,
+            k=k,
+            v=v,
+            i=i_preact,
+            f=f_preact,
+            c_initial=c_initial,
+            n_initial=n_initial,
+            m_initial=m_initial,
+        )
+        expected_h_shape = (
+            B,
+            self.config.num_heads,
+            S,
+            self.v_dim // self.config.num_heads,
+        )
+        assert (
+            h.shape == expected_h_shape
+        ), f"Got {h.shape}, expected {expected_h_shape}"
+
+        h = h.transpose(1, 2)
+        h_norm = self.multihead_norm(h)
+        h_norm = h_norm.reshape(B, S, -1)
+
+        h_out = self.ogate_act_fn(o_preact) * h_norm
+
+        y = self.out_proj(h_out)
+        return y, state
+
+    def forward(  # type: ignore
+        self,
+        x: torch.Tensor,
+        past_key_values: Optional[dict] = None,
+        use_cache: bool = False,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+        cache_mask: Optional[MaskState] = None
+    ):
+        if self.training:
+            self.mlstm_backend.config.mode = "train"
+        else:
+            self.mlstm_backend.config.mode = "inference"
+
+        if use_cache:
+            assert past_key_values is not None
+
+            prev_mode = self.mlstm_backend.config.mode
+            state = past_key_values.get("state", None)
+
+            if cache_mask is not None:
+                state_for_model = cast(mLSTMLayerStateType, tuple(cache_mask.selective_get(x, inv=True) for x in state) if state is not None else None)
+            else:
+                state_for_model = state
+
+            h, new_state = self._original_forward(
+                x,
+                state=state_for_model,
+                sequence_start_indices=sequence_start_indices
+            )
+            assert new_state is not None
+
+            if state is None or cache_mask is None:
+                state = new_state
+            else:
+                if cache_mask is not None:
+                    for i in range(len(state)):
+                        cache_mask.selective_put(new_state[i], state[i], inv=True)
+
+            past_key_values["state"] = state
+            self.mlstm_backend.config.mode = prev_mode
+
+            return h
+        else:
+            h, _ = super().forward(x)
+            return h
+
+class BolmoLocalLayer(nn.Module):
+    def __init__(self, config: BolmoConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+
+        self.act_fn = ACT2FN[config.hidden_act]
+
+        self.xlstm = BolmoXLSTMLayer(config)
+
+        local_mlp_config = copy.deepcopy(config)
+        local_mlp_config.intermediate_size = config.local_intermediate_size
+        self.mlp = BolmoMLP(local_mlp_config)
+
+        self.pre_xlstm_layernorm = BolmoRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.pre_feedforward_layernorm = BolmoRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+        past_key_values: Optional[dict] = None,
+        use_cache: Optional[bool] = False,
+        cache_mask: Optional[MaskState] = None,
+    ) -> torch.Tensor:
+        residual = hidden_states
+        xlstm_out = self.xlstm(self.pre_xlstm_layernorm(hidden_states), sequence_start_indices=sequence_start_indices, past_key_values=past_key_values["xlstm"] if past_key_values is not None else None, use_cache=use_cache, cache_mask=cache_mask)
+        hidden_states = residual + xlstm_out
+
+        # Fully Connected
+        residual = hidden_states
+        ffn_out = self.mlp(self.pre_feedforward_layernorm(hidden_states))
+        hidden_states = residual + ffn_out
+
+        return hidden_states
+
+
+class BolmoLocalEncoder(nn.Module):
+    def __init__(self, config: BolmoConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.add_expanded_embeddings = config.add_expanded_embeddings
+
+        self.byte_embedding = nn.Embedding(
+            config.vocab_size,
+            self.hidden_size,
+        )
+        if self.add_expanded_embeddings:
+            self.subword_embedding = nn.Embedding(
+                config.subword_vocab_size,
+                self.hidden_size,
+            )
+        else:
+            self.subword_embedding = None
+
+        self.layers = nn.ModuleList(
+            [BolmoLocalLayer(config) for _ in range(config.num_local_encoder_layers)]
+        )
+
+        self.post_last_block_norm = BolmoRMSNorm(
+            self.hidden_size,
+            config.local_rms_norm_eps,
+        )
+        self.out_projection = nn.Linear(
+            self.hidden_size,
+            self.hidden_size,
+            bias=True,
+        )
+
+        self.boundary_predictor_module = BolmoBoundaryPredictor(config)
+
+        self.has_cache = False
+
+    def prepare_inference_cache(self, batch_size: int):
+        device = next(self.parameters()).device
+        self.has_cache = True
+
+        self.cache_seqlens = 0
+        self.last_h = torch.zeros((batch_size, self.hidden_size), dtype=self.out_projection.weight.dtype, device=device)
+        self.layer_states = [{"xlstm": {}} for _ in range(len(self.layers))]
+
+    def free_inference_cache(self):
+        self.has_cache = False
+        if hasattr(self, "cache_seqlens"):
+            del self.cache_seqlens
+        if hasattr(self, "last_h"):
+            del self.last_h
+        if hasattr(self, "layer_states"):
+            del self.layer_states
+
+    def _embed(self, tokens, expanded_input_ids: Optional[torch.Tensor] = None):
+        embeddings = self.byte_embedding(tokens)
+        if self.add_expanded_embeddings:
+            assert expanded_input_ids is not None and self.subword_embedding is not None
+            embeddings = embeddings + self.subword_embedding(expanded_input_ids)
+
+        return embeddings
+
+    def _pool(
+        self,
+        h: torch.Tensor,
+        boundary_mask: torch.Tensor | None,
+        n_patches: int,
+        boundary_state: Optional[MaskState] = None,
+    ):
+        if self.has_cache and self.cache_seqlens > 0:
+            assert boundary_state is not None
+            if boundary_state.all():
+                assert h.shape[1] == 1
+                reduced_h = h
+            else:
+                reduced_h = h[[], :, :]
+        else:
+            assert boundary_mask is not None
+
+            L = h.shape[1]
+            token_idx = (
+                torch.arange(L, device=h.device)[None, :] + (~boundary_mask).long() * L  # type: ignore
+            )
+            seq_sorted_indices = torch.argsort(token_idx, dim=1)
+            index = seq_sorted_indices[:, :n_patches, None].expand(
+                -1, -1, h.shape[-1]
+            )
+
+            reduced_h = torch.gather(
+                h,
+                dim=1,
+                index=index,
+            )
+
+        return reduced_h
+
+    def forward(
+        self,
+        input_ids,
+        true_boundary_mask: Optional[torch.Tensor] = None,
+        boundary_state: Optional[MaskState] = None,
+        pad_state: Optional[MaskState] = None,
+        expanded_input_ids: Optional[torch.Tensor] = None,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+    ):
+        embeddings = self._embed(input_ids, expanded_input_ids)
+
+        # pass through encoder layers
+        if self.has_cache and self.cache_seqlens > 0:
+            assert pad_state is not None
+
+            # step those batch positions which are not currently idle (i.e. at a boundary position)
+            # if all batch positions are idle, skip the step entirely
+            # all positions being idle only happens if fuse_boundaries=False. In this case, the step where we
+            # obtain a new representation from the global model will have all positions for the local encoder being idle.
+            if not pad_state.all():
+                h = pad_state.selective_get(embeddings, inv=True)
+
+                for i, block in enumerate(self.layers):
+                    h = block(h, past_key_values=self.layer_states[i], use_cache=True, cache_mask=pad_state)
+
+                if self.post_last_block_norm is not None:
+                    h = self.post_last_block_norm(h)
+
+                pad_state.selective_put(h[:, -1, :], self.last_h, inv=True)
+
+            h = self.last_h.unsqueeze(1)
+        else:
+            h = embeddings
+            for i, block in enumerate(self.layers):
+                if self.has_cache:
+                    use_cache = True
+                    past_key_values = self.layer_states[i]
+                else:
+                    use_cache = False
+                    past_key_values = None
+
+                h = block(h, past_key_values=past_key_values, use_cache=use_cache, sequence_start_indices=sequence_start_indices)
+
+            if self.post_last_block_norm is not None:
+                h = self.post_last_block_norm(h)
+
+            if self.has_cache:
+                self.last_h.copy_(h[:, -1, :])
+
+        if not self.has_cache or self.cache_seqlens == 0: # only used for prefill
+            boundary_logprobs, boundary_mask = self.boundary_predictor_module(
+                h,
+                sequence_start_indices=sequence_start_indices,
+            )
+            if boundary_state is not None:
+                # can't predict through encoder - must be through prev local decoder step
+                boundary_mask[:, -1] = boundary_state.mask
+        else:
+            boundary_logprobs = boundary_mask = None
+
+        # overwrite with true boundaries
+        if true_boundary_mask is not None:
+            boundary_mask = true_boundary_mask
+
+        patch_embeddings = self._pool(
+            h=h,
+            boundary_mask=boundary_mask,
+            n_patches=int(cast(torch.Tensor, boundary_mask).sum(-1).max().item()) if boundary_mask is not None else 1,
+            boundary_state=boundary_state,
+        )
+        patch_embeddings = self.out_projection(patch_embeddings)
+
+        if self.has_cache:
+            self.cache_seqlens += input_ids.shape[1]
+
+        return h, patch_embeddings, boundary_logprobs, boundary_mask
+
+
+class BolmoLocalDecoder(nn.Module):
+    def __init__(self, config: BolmoConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+
+        self.initial_norm = BolmoRMSNorm(
+            self.hidden_size,
+            eps=config.local_rms_norm_eps,
+        )
+
+        self.in_projection = nn.Linear(
+            self.hidden_size,
+            self.hidden_size,
+            bias=True,
+        )
+
+        self.layers = nn.ModuleList(
+            [BolmoLocalLayer(config) for _ in range(config.num_local_decoder_layers)]
+        )
+
+        self.has_cache = False
+
+    def prepare_inference_cache(self, batch_size: int):
+        device = next(self.parameters()).device
+        self.has_cache = True
+
+        self.cache_seqlens = 0
+        self.last_value = torch.zeros((batch_size, self.hidden_size), dtype=self.in_projection.weight.dtype, device=device)
+        self.layer_states = [{"xlstm": {}} for _ in range(len(self.layers))]
+
+    def free_inference_cache(self):
+        self.has_cache = False
+        if hasattr(self, "cache_seqlens"):
+            del self.cache_seqlens
+        if hasattr(self, "last_value"):
+            del self.last_value
+        if hasattr(self, "layer_states"):
+            del self.layer_states
+
+    def _depool(
+        self,
+        embeds: torch.Tensor,
+        patch_embeds: torch.Tensor,
+        boundary_mask: Optional[torch.Tensor],
+        boundary_state: Optional[MaskState] = None,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if self.has_cache and self.cache_seqlens > 0:
+            assert boundary_state is not None
+
+            if patch_embeds.numel() > 0:
+                # we got a new value from the global model, so must be at boundary position
+                h_patch = patch_embeds[:, -1:, :]
+                h = embeds + h_patch
+
+                self.last_value.copy_(h_patch[:, -1])
+            else:
+                h = embeds + self.last_value.unsqueeze(1)
+
+            # skip pad positions until we get a new value from the global model
+            if patch_embeds.numel() == 0:
+                h = boundary_state.selective_get(h, inv=True)
+            else:
+                boundary_state = None
+
+            if h.shape[0] > 0:
+                for i, layer in enumerate(self.layers):
+                    h = layer(h, past_key_values=self.layer_states[i], use_cache=True, cache_mask=boundary_state)
+
+            self.cache_seqlens += h.shape[1]
+
+            return h
+        else:
+            assert boundary_mask is not None
+
+            h_patch = patch_embeds 
+            prepool_out = h_patch
+
+            # TODO(benjaminm): clipping is problematic if it happens too much; track clip %.
+            plug_back_idx = (torch.cumsum(boundary_mask, dim=1) - 1).clip(min=0, max=prepool_out.shape[1] - 1)
+            depool_out = torch.gather(
+                prepool_out,
+                dim=1,
+                index=plug_back_idx.unsqueeze(-1).expand(-1, -1, self.hidden_size),
+            )
+                
+            depool_out_modulated = depool_out
+            h = depool_out_modulated + embeds
+
+            for i, layer in enumerate(self.layers):
+                if self.has_cache:
+                    use_cache = True
+                    past_key_values = self.layer_states[i]
+                else:
+                    use_cache = False
+                    past_key_values = None
+
+                h = layer(h, past_key_values=past_key_values, use_cache=use_cache, sequence_start_indices=sequence_start_indices)
+
+            if self.has_cache:
+                self.last_value.copy_(prepool_out[:, -1])
+                self.cache_seqlens += h.shape[1]
+
+            return h
+
+    def forward(
+        self,
+        embeds: torch.Tensor,
+        patch_embeds: torch.Tensor,
+        boundary_state: Optional[MaskState],
+        boundary_mask: torch.Tensor | None,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        h = self.in_projection(embeds)
+        h_patch = self.initial_norm(patch_embeds)
+
+        return self._depool(
+            embeds=h,
+            patch_embeds=h_patch,
+            boundary_mask=boundary_mask,
+            boundary_state=boundary_state,
+            sequence_start_indices=sequence_start_indices,
+        )
+
+
+class BolmoRotaryEmbedding(nn.Module):
+    inv_freq: torch.Tensor  # fix linting for `register_buffer`
+
+    def __init__(self, config: BolmoConfig, device=None, rope_type: Optional[str] = None):
+        super().__init__()
+        if rope_type is not None:
+            self.rope_type = rope_type
+        elif hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict):
+            # BC: "rope_type" was originally "type"
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        assert self.rope_type is not None
+
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    @dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        position_ids_expanded = position_ids[:, None, :].float()
+
+        device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos() * self.attention_scaling
+            sin = emb.sin() * self.attention_scaling
+            return cos, sin
+
+
+@auto_docstring
+class BolmoPreTrainedModel(PreTrainedModel):
+    config: BolmoConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["BolmoDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+
+    _can_compile_fullgraph = True
+    _supports_attention_backend = True
+    _can_record_outputs = {
+        "hidden_states": BolmoDecoderLayer,
+        "attentions": BolmoAttention,
+    }
+
+
+@auto_docstring
+class BolmoModel(BolmoPreTrainedModel):
+    def __init__(self, config: BolmoConfig):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.local_encoder = BolmoLocalEncoder(config)
+        self.local_decoder = BolmoLocalDecoder(config)
+
+        self.layers = nn.ModuleList(
+            [BolmoDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = BolmoRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
+        self.rotary_embs = nn.ModuleDict(
+            {
+                "sliding_attention": BolmoRotaryEmbedding(config=config, rope_type="default"),
+                "full_attention": BolmoRotaryEmbedding(config=config),
+            }
+        )
+
+        self.tokenizer_config = ByteTokenizerConfig(**config.tokenizer_config)
+        self._tokenizer = None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.local_encoder.byte_embedding
+
+    def set_input_embeddings(self, value: nn.Embedding):  # type: ignore
+        self.local_encoder.byte_embedding = value
+
+    @property
+    def tokenizer(self):
+        if self._tokenizer is None:
+            self._tokenizer = self.tokenizer_config.build()
+
+        return self._tokenizer
+
+    def prefill_boundary_prediction_forward(
+        self,
+        input_ids: torch.Tensor,
+        expanded_input_ids: Optional[torch.LongTensor] = None,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+        last_token_is_boundary: bool = False,
+        **kwargs,
+    ) -> torch.Tensor:
+        _, _, _, boundary_mask = self.local_encoder.forward(  # type: ignore
+            input_ids,
+            expanded_input_ids=expanded_input_ids,
+            boundary_state=MaskState(torch.full((input_ids.shape[0],), fill_value=last_token_is_boundary, device=input_ids.device, dtype=torch.bool)),
+            pad_state=MaskState(torch.zeros((input_ids.shape[0],), device=input_ids.device, dtype=torch.bool)),
+            sequence_start_indices=sequence_start_indices,
+        )
+
+        return cast(torch.Tensor, boundary_mask)
+
+    @check_model_inputs()
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        expanded_input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        boundary_mask: Optional[torch.Tensor] = None,
+        boundary_state: Optional[MaskState] = None,
+        pad_state: Optional[MaskState] = None,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> BaseModelOutputWithPast:
+        batch_size = input_ids.shape[0]
+        device = input_ids.device
+
+        if self.local_encoder.add_expanded_embeddings and expanded_input_ids is None and input_ids is not None:
+            # not optimized
+            expanded_input_ids_list: list[torch.Tensor] = []
+            for example_idx in range(batch_size):
+                expanded_input_ids_list.append(torch.tensor(self.tokenizer.expand_byte_ids(input_ids[example_idx].tolist()), dtype=torch.long, device=device))
+            expanded_input_ids = pad_right(expanded_input_ids_list, value=self.tokenizer.pad_token_id, multiple_of=1)  # type: ignore
+
+        h_byte, h_patch, _, boundary_mask = self.local_encoder(
+            input_ids=input_ids,
+            expanded_input_ids=expanded_input_ids,
+            true_boundary_mask=boundary_mask,
+            boundary_state=boundary_state,
+            pad_state=pad_state,
+        )
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache(config=self.config)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position: torch.Tensor = torch.arange(
+                past_seen_tokens, past_seen_tokens + h_patch.shape[1], device=device
+            )
+
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)  # type: ignore
+
+        # It may already have been prepared by e.g. `generate`
+        if not isinstance(causal_mask_mapping := attention_mask, dict):
+            # Prepare mask arguments
+            mask_kwargs = {
+                "config": self.config,
+                "input_embeds": h_patch,
+                "attention_mask": attention_mask,
+                "cache_position": cache_position,
+                "past_key_values": past_key_values,
+                "position_ids": position_ids,
+            }
+            # Create the masks
+            causal_mask_mapping = {
+                "full_attention": create_causal_mask(**mask_kwargs),
+                "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs),
+            }
+
+        position_embeddings_mapping = {
+            "sliding_attention": self.rotary_embs["sliding_attention"](h_byte, position_ids),
+            "full_attention": self.rotary_embs["full_attention"](h_byte, position_ids),
+        }
+
+        if h_patch.numel() > 0:
+            # we need to convert from right-pad to left-pad and back for prefill
+            # since flash attention expects left-pad and local/enc dec expect right-pad global tokens
+            # should add better left-pad support but this only affects prefill so OK for now
+            # although super inefficient!
+            if boundary_mask is not None: # prefill
+                n_boundaries = boundary_mask.sum(-1)
+
+                for i, current_n_boundaries in enumerate(n_boundaries):
+                    h_patch[i, -current_n_boundaries:] = h_patch[i, :current_n_boundaries].clone()
+
+            h_patch_after_global = h_patch
+
+            for decoder_layer in self.layers[: self.config.num_hidden_layers]:
+                h_patch_after_global = decoder_layer(
+                    h_patch_after_global,
+                    attention_mask=causal_mask_mapping[decoder_layer.self_attn.attention_type],
+                    position_ids=position_ids,
+                    past_key_values=past_key_values,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings_mapping[decoder_layer.self_attn.attention_type],
+                    **kwargs,
+                )
+
+            if boundary_mask is not None: # prefill
+                n_boundaries = boundary_mask.sum(-1)
+
+                for i, current_n_boundaries in enumerate(n_boundaries):
+                    h_patch_after_global[i, :current_n_boundaries] = h_patch_after_global[i, -current_n_boundaries:].clone()
+        else:
+            h_patch_after_global = h_patch
+
+        h_out = self.local_decoder.forward(  # type: ignore
+            embeds=h_byte,
+            patch_embeds=h_patch_after_global,
+            boundary_mask=boundary_mask,
+            boundary_state=boundary_state,
+            sequence_start_indices=sequence_start_indices,
+        )
+        h_out = self.norm(h_out)
+
+        return BaseModelOutputWithPast(
+            last_hidden_state=h_out,
+            past_key_values=past_key_values,
+        )
+
+
+@auto_docstring
+class BolmoForCausalLM(BolmoPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = BolmoModel(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+    
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings: nn.Linear):
+        self.lm_head = new_embeddings
+
+    @can_return_tuple
+    @auto_docstring
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        expanded_input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        boundary_mask: Optional[torch.Tensor] = None,
+        boundary_state: Optional[MaskState] = None,
+        pad_state: Optional[MaskState] = None,
+        sequence_start_indices: Optional[torch.Tensor] = None,
+        logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> CausalLMOutputWithPast:
+        r"""
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, BolmoForCausalLM
+
+        >>> model = BolmoForCausalLM.from_pretrained("meta-olmo3/Bolmo-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-olmo3/Bolmo-2-7b-hf")
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+        outputs: BaseModelOutputWithPast = self.model(
+            input_ids=input_ids,
+            expanded_input_ids=expanded_input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            cache_position=cache_position,
+            use_cache=use_cache,
+            boundary_mask=boundary_mask,
+            boundary_state=boundary_state,
+            pad_state=pad_state,
+            sequence_start_indices=sequence_start_indices,
+            **kwargs,
+        )
+
+        hidden_states = cast(torch.Tensor, outputs.last_hidden_state)
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        return CausalLMOutputWithPast(
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+    def generate(self, input_ids: list[list[int]], max_new_tokens: int = 20):
+        expand_input_ids = self.model.local_encoder.add_expanded_embeddings
+        batch_size = len(input_ids)
+
+        if expand_input_ids:
+            expanded_input_ids = []
+
+            for i in range(len(input_ids)):
+                expanded_input_ids.append(torch.tensor(self.model.tokenizer.expand_byte_ids(input_ids[i]), device=self.device, dtype=torch.long))
+
+            expanded_input_ids = pad_left(expanded_input_ids, value=self.model.tokenizer.pad_token_id, multiple_of=1)  # type: ignore
+        else:
+            expanded_input_ids = None
+
+        byte_input_ids: torch.Tensor = pad_left([torch.tensor(x, device=self.device, dtype=torch.long) for x in input_ids], value=self.model.tokenizer.pad_token_id, multiple_of=1)
+
+        sequence_start_indices = (byte_input_ids == self.model.tokenizer.pad_token_id).sum(-1)
+        batch_size, prompt_len = byte_input_ids.shape
+        finished = torch.zeros(batch_size, dtype=torch.bool, device=self.device)
+
+        boundary_offset = self.model.tokenizer.offset + 256
+        eos = self.model.tokenizer.eos_token_id
+
+        boundary_mask = self.model.prefill_boundary_prediction_forward(  # type: ignore
+            byte_input_ids,
+            expanded_input_ids=expanded_input_ids,
+            sequence_start_indices=sequence_start_indices,
+        )
+
+        self.model.local_encoder.prepare_inference_cache(batch_size)
+        self.model.local_decoder.prepare_inference_cache(batch_size)
+
+        # roll back by one and force decoding to account for lookahead
+        boundary_mask = boundary_mask[:, :-1]
+        # need to roll one byte back and force decoding to detect whether the last byte is a boundary
+        forced_decoding_ids = byte_input_ids[:, -1].cpu().tolist()
+        byte_input_ids = byte_input_ids[:, :-1]
+        expanded_input_ids = expanded_input_ids[:, :-1] if expanded_input_ids is not None else None
+        # stays the same unless last token is pad.
+        sequence_start_indices = (byte_input_ids == self.model.tokenizer.pad_token_id).sum(-1)
+
+        # output container
+        generated = byte_input_ids
+
+        max_n_prefill_patches = boundary_mask.sum(-1).max().item()
+        tokens_generated_plus_prefilled = max_n_prefill_patches
+        bytes_generated = 0
+
+        max_length = max_n_prefill_patches + max_new_tokens
+
+        # generation state
+        boundary_state = MaskState(boundary_mask[:, -1].clone())
+        pad_state = MaskState(torch.zeros(batch_size, dtype=torch.bool, device=self.device))
+        next_tokens = torch.full((batch_size,), self.model.tokenizer.bpe_token_end_id, device=self.device, dtype=torch.long)  # type: ignore
+        non_boundary_generated_tokens = [[byte_input_ids[example_idx, -1].item()] for example_idx in range(batch_size)]
+        bytes_since_boundary = (boundary_mask.flip(1).cumsum(-1) == 0).sum(-1)
+        is_first_forward = True
+        global_past_key_values = None
+
+        # TODO: impl
+        stop_token_sequences = []
+
+        while not ((max_length is not None and tokens_generated_plus_prefilled >= max_length) or finished.all()):
+            input_ids_for_model = (
+                generated
+                if is_first_forward
+                else torch.tensor([x[-1] for x in non_boundary_generated_tokens], device=generated.device, dtype=generated.dtype).unsqueeze(1)
+            )
+            assert not (
+                (input_ids_for_model == self.model.tokenizer.bpe_token_end_id) |
+                (input_ids_for_model >= boundary_offset)
+            ).any().item()  # type: ignore
+            if expand_input_ids:
+                expanded_input_ids_for_model = torch.zeros_like(input_ids_for_model)
+                for i in range(input_ids_for_model.shape[0]):
+                    expanded_input_ids_for_model[i, :] = torch.tensor(self.model.tokenizer.expand_byte_ids(
+                        generated[i, :].tolist(),
+                        n_last=input_ids_for_model.shape[1],
+                    ), device=expanded_input_ids_for_model.device, dtype=expanded_input_ids_for_model.dtype)
+            else:
+                expanded_input_ids_for_model = None
+
+            out = self.forward(  # type: ignore
+                input_ids_for_model,
+                expanded_input_ids=expanded_input_ids_for_model,
+                boundary_mask=boundary_mask if is_first_forward else None,
+                boundary_state=boundary_state,
+                pad_state=pad_state,
+                sequence_start_indices=sequence_start_indices,
+                logits_to_keep=1,
+                use_cache=True,
+                past_key_values=global_past_key_values,
+            )
+            next_token_logits = cast(torch.Tensor, out.logits)
+            global_past_key_values = out.past_key_values
+
+            if boundary_state.all():
+                # new token, must not be boundary
+                bytes_since_boundary[:] = 0
+            else:
+                boundary_state.selective_add(1, bytes_since_boundary, inv=True)
+
+            if any(x is not None for x in forced_decoding_ids):
+                # only supported for the first token atm, so len(next_token_logits) == batch_size
+                assert len(next_token_logits) == batch_size and is_first_forward
+                for example_idx in range(batch_size):
+                    forced_decoding_id = forced_decoding_ids[example_idx]
+
+                    if forced_decoding_id is not None:
+                        no_boundary_logit = next_token_logits[example_idx, 0, forced_decoding_id].item()
+                        boundary_logit = next_token_logits[example_idx, 0, forced_decoding_id + boundary_offset].item()
+
+                        next_token_logits[example_idx, 0, :] = -100_000
+                        next_token_logits[example_idx, 0, forced_decoding_id] = no_boundary_logit
+                        next_token_logits[example_idx, 0, forced_decoding_id + boundary_offset] = boundary_logit
+
+                        forced_decoding_ids[example_idx] = None  # only force once
+
+            # TODO: impl non-greedy
+            new_next_tokens = next_token_logits.squeeze(1).argmax(dim=-1)
+
+            if boundary_state.all():
+                tokens_generated_plus_prefilled += 1
+
+                next_tokens = new_next_tokens
+                next_tokens_cpu = next_tokens.cpu()
+                for example_idx in range(batch_size):
+                    next_token_cpu = next_tokens_cpu[example_idx].item()
+
+                    if next_token_cpu >= boundary_offset:
+                        next_token_cpu -= boundary_offset
+
+                    non_boundary_generated_tokens[example_idx].append(next_token_cpu)
+            else:
+                next_tokens[:] = self.model.tokenizer.bpe_token_end_id # type: ignore
+                boundary_state.selective_put(new_next_tokens, next_tokens, inv=True)
+                next_tokens_cpu = next_tokens.cpu()
+
+                for example_idx in range(batch_size):
+                    next_token_cpu = next_tokens_cpu[example_idx].item()
+
+                    if not boundary_state.cpu_mask[example_idx].item():
+                        if next_token_cpu >= boundary_offset:
+                            next_token_cpu -= boundary_offset
+
+                        non_boundary_generated_tokens[example_idx].append(next_token_cpu)
+
+            is_first_forward = False
+
+            boundary_state = MaskState(
+                (next_tokens == self.model.tokenizer.bpe_token_end_id) |
+                (next_tokens >= boundary_offset) |
+                finished
+            )  # type: ignore
+            pad_state = MaskState(
+                (next_tokens == self.model.tokenizer.bpe_token_end_id) |
+                finished
+            )
+
+            # Force EOS for (previously) finished sequences
+            next_tokens = torch.where(finished, torch.full_like(next_tokens, eos), next_tokens)
+
+            # Append next tokens
+            generated = torch.cat([generated, next_tokens.unsqueeze(-1)], dim=1)
+
+            # Handle finished sequences
+            stop_hit = next_tokens.eq(eos) | next_tokens.eq(eos + boundary_offset)
+
+            # Also check for stop tokens if provided
+            # TODO(benjaminm): this is very annoying due to the boundaries
+            # make better
+            if len(stop_token_sequences) > 0:
+                # TODO: implement
+                raise NotImplementedError("stop_token_sequences not implemented yet for Bolmo generation.")
+
+            finished |= stop_hit
+            bytes_generated += 1
+
+__all__ = ["BolmoForCausalLM", "BolmoModel", "BolmoPreTrainedModel"]