3b-september-2025 upload

config.json

@@ -1,127 +1,96 @@
 {
-  "_name_or_path": "ai-sage/Giga-Embeddings-instruct",
-  "add_eos": true,
-  "add_pad_token": true,
-  "architectures": [
-    "GigarEmbedModel"
-  ],
-  "auto_map": {
-    "AutoConfig": "configuration_gigarembed.GigarEmbedConfig",
-    "AutoModel": "modeling_gigarembed.GigarEmbedModel"
-  },
-  "hidden_size": 2048,
-  "is_mask_instruction": true,
-  "latent_attention_config": {
-    "cross_dim_head": 2048,
-    "hidden_dim": 2048,
-    "latent_dim": 2048,
-    "model_type": "latent_attention"
-  },
-  "mask_type": "b",
-  "model_type": "gigarembed",
-  "padding_side": "right",
-  "text_config": {
-    "_attn_implementation_autoset": false,
     "_name_or_path": "ai-sage/Giga-Embeddings-instruct",
+    "_non_freeze_layers_idxs": null,
     "activation_checkpoint_layers_num": null,
-    "add_cross_attention": false,
+    "apply_torch_compile_to_projections": true,
     "architectures": [
-      "LlamaForCausalLM"
+        "GigarEmbedModel"
     ],
-    "attention_bias": false,
-    "attention_dropout": 0.0,
-    "attention_hidden_size": null,
-    "attention_type": "LlamaPackedAttention",
-    "bad_words_ids": null,
-    "begin_suppress_tokens": null,
-    "bos_token_id": 1,
-    "chunk_size_feed_forward": 0,
-    "cross_attention_hidden_size": null,
-    "decoder_start_token_id": null,
-    "deterministic_attention": false,
-    "diversity_penalty": 0.0,
-    "do_sample": false,
-    "early_stopping": false,
-    "encoder_no_repeat_ngram_size": 0,
-    "eos_token_id": 2,
-    "exponential_decay_length_penalty": null,
-    "finetuning_task": null,
-    "forced_bos_token_id": null,
-    "forced_eos_token_id": null,
-    "freeze_non_embed": false,
-    "fused_mlp": true,
-    "fused_mlp_checkpoint_lvl": 3,
-    "head_dim": 128,
-    "hidden_act": "silu",
-    "hidden_size": 2048,
-    "id2label": {
-      "0": "LABEL_0",
-      "1": "LABEL_1"
+    "auto_map": {
+        "AutoConfig": "configuration_gigarembed.GigarEmbedConfig",
+        "AutoModel": "modeling_gigarembed.GigarEmbedModel"
+    },
+    "latent_attention_config": {
+        "model_type": "latent_attention",
+        "num_latents_value": 512,
+        "num_cross_heads": 8,
+        "cross_dim_head": 2048,
+        "hidden_dim": 2048,
+        "latent_dim": 2048,
+        "mult": 4
     },
-    "init_device": "meta",
-    "initializer_range": 0.02,
-    "intermediate_size": 11008,
-    "is_decoder": false,
-    "is_encoder_decoder": false,
-    "label2id": {
-      "LABEL_0": 0,
-      "LABEL_1": 1
+    "hidden_size": 2048,
+    "text_config": {
+        "_name_or_path": "ai-sage/Giga-Embeddings-instruct",
+        "apply_qk_norm": true,
+        "attention_bias": false,
+        "attention_dropout": 0.0,
+        "attention_hidden_size": null,
+        "attention_type": "LlamaLatentAttention",
+        "bos_token_id": 1,
+        "delete_logits": true,
+        "deterministic_attention": false,
+        "enable_async_tp": false,
+        "eos_token_id": 2,
+        "freeze_non_embed": false,
+        "fused_mlp": true,
+        "fused_mlp_checkpoint_lvl": 3,
+        "head_dim": 64,
+        "hidden_act": "silu",
+        "hidden_size": 2048,
+        "ignore_index": -100,
+        "init_device": "meta",
+        "initializer_range": 0.02,
+        "intermediate_size": 11008,
+        "kv_lora_rank": 1024,
+        "lora_alpha": null,
+        "lora_r": null,
+        "loss_inplace_backward": false,
+        "max_position_embeddings": 4096,
+        "max_window_layers": 36,
+        "mla_config": {
+            "kv_lora_rank": 1024,
+            "q_lora_rank": 0,
+            "qk_nope_head_dim": 64,
+            "qk_rope_head_dim": 64,
+            "v_head_dim": 128
+        },
+        "mlp_bias": false,
+        "model_type": "gigar",
+        "mtp_loss_weight": 0.1,
+        "mtp_predictor_num": 1,
+        "norm_type": "LlamaRMSNorm",
+        "num_attention_heads": 16,
+        "num_hidden_layers": 36,
+        "num_key_value_heads": 16,
+        "pad_token_id": 2,
+        "parallel_embedding_type": "EmbeddingParallelEmbedding",
+        "pretraining_tp": 1,
+        "q_lora_rank": 0,
+        "qk_nope_head_dim": 64,
+        "qk_rope_head_dim": 64,
+        "rms_norm_eps": 1e-06,
+        "rope_scaling": null,
+        "rope_theta": 100000.0,
+        "skip_init_tp_modules": true,
+        "sliding_window": null,
+        "sp_split_type": "equal",
+        "tie_word_embeddings": false,
+        "tp_group": null,
+        "tp_size": 1,
+        "unk_token_id": 0,
+        "use_cache": false,
+        "use_cache_force": false,
+        "use_custom_rotary_kernel": false,
+        "use_liger": false,
+        "use_mrope": false,
+        "use_mtp": true,
+        "use_sliding_window": false,
+        "v_head_dim": 128,
+        "varlen_input": true,
+        "vocab_size": 128256,
+        "z_loss_eps": 5e-05
     },
-    "length_penalty": 1.0,
-    "loss_inplace_backward": true,
-    "max_length": 20,
-    "max_position_embeddings": 4096,
-    "max_window_layers": 36,
-    "min_length": 0,
-    "mlp_bias": false,
-    "model_type": "llama",
-    "no_repeat_ngram_size": 0,
-    "num_attention_heads": 16,
-    "num_beam_groups": 1,
-    "num_beams": 1,
-    "num_hidden_layers": 36,
-    "num_key_value_heads": 2,
-    "num_return_sequences": 1,
-    "output_attentions": false,
-    "output_hidden_states": false,
-    "output_scores": false,
-    "pad_token_id": 2,
-    "prefix": null,
-    "pretraining_tp": 1,
-    "problem_type": null,
-    "pruned_heads": {},
-    "remove_invalid_values": false,
-    "repetition_penalty": 1.0,
-    "return_dict": true,
-    "return_dict_in_generate": false,
-    "rms_norm_eps": 1e-06,
-    "rope_scaling": null,
-    "rope_theta": 1300,
-    "sep_token_id": null,
-    "sliding_window": null,
-    "sp_split_type": "equal",
-    "suppress_tokens": null,
-    "task_specific_params": null,
-    "temperature": 1.0,
-    "tf_legacy_loss": false,
-    "tie_encoder_decoder": false,
-    "tie_word_embeddings": false,
-    "tokenizer_class": null,
-    "top_k": 50,
-    "top_p": 1.0,
-    "torch_dtype": "float32",
-    "torchscript": false,
-    "tp_group": null,
-    "tp_size": 1,
-    "typical_p": 1.0,
-    "unk_token_id": 0,
-    "use_bfloat16": false,
-    "use_cache": true,
-    "use_mrope": false,
-    "use_sliding_window": false,
-    "varlen_input": false,
-    "vocab_size": 128256
-  },
-  "torch_dtype": "float32",
-  "transformers_version": "4.46.3"
+    "torch_dtype": "bfloat16",
+    "transformers_version": "4.48.0"
 }

configuration_gigarembed.py

@@ -1,12 +1,238 @@
+import warnings
+
 from typing import Literal
 from transformers import AutoConfig
-from transformers.configuration_utils import PretrainedConfig
 from transformers.models.auto import CONFIG_MAPPING
-from transformers.models.llama import LlamaConfig
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
 
 GIGAREMBED_TYPE = "gigarembed"
 LATENT_ATTENTION_TYPE = "latent_attention"
-BIDIR_LLAMA_TYPE = "bidir_llama"
+
+
+class GigarConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`GigarModel`]. It is used to instantiate an Gigar
+    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 Gigar-7B.
+
+    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 32000):
+            Vocabulary size of the Gigar model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`GigarModel`]
+        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 checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). 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. Gigar 1 supports up to 2048 tokens,
+            Gigar 2 up to 4096, CodeLlama up to 16384.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        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*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to
+            understand more about it. This value is necessary to ensure exact reproducibility of the pretraining
+            results. Please refer to [this issue](https://github.com/pytorch/pytorch/issues/76232).
+        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',
+                    'gigar3'], 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 'gigar3'. 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 'gigar3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'gigar3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, *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.
+        mlp_bias (`bool`, *optional*, defaults to `False`):
+            Whether to use a bias in up_proj, down_proj and gate_proj layers in the MLP layers.
+        head_dim (`int`, *optional*):
+            The attention head dimension. If None, it will default to hidden_size // num_attention_heads
+
+    ```python
+    >>> from transformers import GigarModel, GigarConfig
+
+    >>> # Initializing a Gigar gigar-7b style configuration
+    >>> configuration = GigarConfig()
+
+    >>> # Initializing a model from the gigar-7b style configuration
+    >>> model = GigarModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "gigar"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    # Default tensor parallel plan for base model `GigarModel`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        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,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=1,
+        eos_token_id=2,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        attention_dropout=0.0,
+        mlp_bias=False,
+        head_dim=None,
+        apply_qk_norm=False,
+        mla_config=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.rms_norm_eps = rms_norm_eps
+        self.pretraining_tp = pretraining_tp
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        self.mlp_bias = mlp_bias
+        self.head_dim = head_dim if head_dim is not None else self.hidden_size // self.num_attention_heads
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, copy it it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        self.apply_qk_norm = apply_qk_norm
+        self.mla_config = mla_config
+
+        self._validate_mla_config()
+
+    def _validate_mla_config(self):
+        if self.mla_config is None:
+            warnings.warn("MLA config is None!")
+            return
+
+        EXPECTED_KEYS = [
+            "qk_nope_head_dim",
+            "qk_rope_head_dim",
+            "v_head_dim",
+            "kv_lora_rank",
+            "q_lora_rank",
+        ]
+        if not all((key in self.mla_config for key in EXPECTED_KEYS)):
+            raise ValueError(
+                f"MLA config is expected to have the following keys {EXPECTED_KEYS} but got {self.mla_config.keys()}."
+            )
+
+        if self.mla_config["qk_nope_head_dim"] + self.mla_config["qk_rope_head_dim"] != self.mla_config["v_head_dim"]:
+            err_msg = (
+                f"QK and V head dims do not match! Got {self.mla_config['qk_nope_head_dim']} + {self.mla_config['qk_rope_head_dim']} "
+                f"= {self.mla_config['qk_rope_head_dim'] + self.mla_config['qk_nope_head_dim']} and {self.mla_config['v_head_dim']}."
+            )
+            raise ValueError(err_msg)
+
 
 class GigarEmbedConfig(PretrainedConfig):
     model_type = "gigarembed"
@@ -28,14 +254,11 @@ class GigarEmbedConfig(PretrainedConfig):
                 latent_attention_config["model_type"] if "model_type" in latent_attention_config else LATENT_ATTENTION_TYPE
             )
             latent_attention_config = CONFIG_MAPPING[latent_attention_config["model_type"]](**latent_attention_config)
-        elif latent_attention_config is None:
-            latent_attention_config = CONFIG_MAPPING[LATENT_ATTENTION_TYPE]()
 
         self.latent_attention_config = latent_attention_config
 
         if isinstance(text_config, dict):
-            text_config["model_type"] = text_config["model_type"] if "model_type" in text_config else "llama"
-            text_config = CONFIG_MAPPING[text_config["model_type"]](**text_config)
+            text_config = GigarConfig(**text_config)
         elif text_config is None:
             text_config = None
 
@@ -47,8 +270,6 @@ class GigarEmbedConfig(PretrainedConfig):
         self.mask_type = mask_type
         if "hidden_size" in kwargs:
             self.hidden_size = kwargs["hidden_size"]
-        else:
-            self.hidden_size = 2560
 
         super().__init__(**kwargs)
 
@@ -60,30 +281,26 @@ class LatentAttentionConfig(PretrainedConfig):
 
     def __init__(
         self,
-        num_latents_value: int=512,
-        num_cross_heads: int=8,
-        output_normalize: bool=True,
-        hidden_dim: int=2560,
-        latent_dim: int=2560,
-        cross_dim_head: int=2560,
+        num_latents_value: int,
+        num_cross_heads: int,
+        hidden_dim: int,
+        latent_dim: int,
+        cross_dim_head: int,
+        mult: int,
         **kwargs,
     ):
         self.num_latents_value = num_latents_value
         self.num_cross_heads = num_cross_heads
-        self.output_normalize = output_normalize
         self.hidden_dim = hidden_dim
         self.latent_dim = latent_dim
         self.cross_dim_head = cross_dim_head
-        self._attn_implementation = "eager"
+        self.mult = mult
+
+        super().__init__(**kwargs)
 
-class BidirectionalLlamaConfig(LlamaConfig):
-    model_type = BIDIR_LLAMA_TYPE
-    keys_to_ignore_at_inference = ["past_key_values"]
 
 AutoConfig.register(GIGAREMBED_TYPE, GigarEmbedConfig)
 AutoConfig.register(LATENT_ATTENTION_TYPE, LatentAttentionConfig)
-AutoConfig.register(BIDIR_LLAMA_TYPE, BidirectionalLlamaConfig)
 
 GigarEmbedConfig.register_for_auto_class()
 LatentAttentionConfig.register_for_auto_class()
-BidirectionalLlamaConfig.register_for_auto_class()

model-00001-of-00003.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f32eedfe2127f8e9507427af1d796c6547df4c8e5795e4ea8b3a22a96e782292
-size 4930720644
+oid sha256:8f58bbc242e82e65d3564fdf09a378b5fbf545e15bc8291d7a623a8873222f6f
+size 4947550528

model-00002-of-00003.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:edc8c0c52613a2712e8c65b3d8b4249b6e99622c695ee5aec698ca37a5a556d3
-size 4932780264
+oid sha256:fe27f921338efe4e1b41a27f6a1c87ca35d6ae933ebd0458ba09c329b63c9ac9
+size 4912107128

model-00003-of-00003.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8cb8d3f0fb5c526162adc18efcc9e1c13d07088602775e742037a5e53d1531b9
-size 3045246736
+oid sha256:90d75a17597fe2351022a4f8638c2509e624b25b1fc08a3d5c5204ec02d7a14b
+size 3938900016

model.safetensors.index.json

@@ -1,351 +1,520 @@
 {
   "metadata": {
-    "total_size": 12908707844
+    "total_size": 13798498304
   },
   "weight_map": {
-    "latent_attention_model.cross_attend_blocks.0.fn.to_kv.weight": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.0.fn.to_out.weight": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.0.fn.to_q.weight": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.0.norm.bias": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.0.norm.weight": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.0.norm_context.bias": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.0.norm_context.weight": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.1.fn.net.0.bias": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.1.fn.net.0.weight": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.1.fn.net.2.bias": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.1.fn.net.2.weight": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.1.norm.bias": "model-00001-of-00003.safetensors",
-    "latent_attention_model.cross_attend_blocks.1.norm.weight": "model-00001-of-00003.safetensors",
+    "latent_attention_model.cross_attend_blocks.0.to_kv.weight": "model-00001-of-00003.safetensors",
+    "latent_attention_model.cross_attend_blocks.0.to_out.weight": "model-00001-of-00003.safetensors",
+    "latent_attention_model.cross_attend_blocks.0.to_q.weight": "model-00001-of-00003.safetensors",
+    "latent_attention_model.cross_attend_blocks.1.down_proj.weight": "model-00001-of-00003.safetensors",
+    "latent_attention_model.cross_attend_blocks.1.gate_proj.weight": "model-00001-of-00003.safetensors",
+    "latent_attention_model.cross_attend_blocks.1.up_proj.weight": "model-00001-of-00003.safetensors",
     "latent_attention_model.latents": "model-00001-of-00003.safetensors",
-    "latent_attention_model.w_lexical.bias": "model-00001-of-00003.safetensors",
-    "latent_attention_model.w_lexical.weight": "model-00001-of-00003.safetensors",
-    "latent_attention_model.w_multi_vector.bias": "model-00001-of-00003.safetensors",
-    "latent_attention_model.w_multi_vector.weight": "model-00001-of-00003.safetensors",
     "model.embed_tokens.weight": "model-00001-of-00003.safetensors",
     "model.layers.0.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.0.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.0.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.1.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.1.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.1.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.10.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.10.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.10.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.10.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.10.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.10.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.10.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.10.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.11.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.11.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.11.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.11.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.11.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.11.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.12.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.12.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.12.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.12.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.13.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.13.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.13.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.13.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.14.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.14.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.14.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.15.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.15.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.15.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.16.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.16.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.16.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.16.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.16.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.16.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.16.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.16.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.16.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.17.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.17.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.17.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.17.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.17.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.17.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.17.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.17.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.17.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.18.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.18.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.18.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.18.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.18.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.18.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.18.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.18.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.18.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.19.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.19.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.19.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.19.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.19.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.19.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.19.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.19.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.19.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.2.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.2.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.2.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.20.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.20.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.20.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.20.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.20.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.20.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.20.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.20.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.20.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.21.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.21.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.21.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.21.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.21.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.21.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.21.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.21.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.21.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.22.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.22.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.22.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.22.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.22.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.22.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.22.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.22.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.22.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.23.input_layernorm.weight": "model-00002-of-00003.safetensors",
     "model.layers.23.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.23.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.23.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.23.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.23.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.23.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.23.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.23.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.24.input_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.24.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.24.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.24.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.24.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.24.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.24.self_attn.dkv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.kr_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.kv_norm.weight": "model-00002-of-00003.safetensors",
     "model.layers.24.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.24.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.24.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.input_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.mlp.gate_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.25.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.qk_k_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.qk_q_norm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.uk_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.uv_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.25.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.26.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.26.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.26.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.26.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.26.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.26.self_attn.k_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.26.self_attn.o_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.26.self_attn.q_proj.weight": "model-00002-of-00003.safetensors",
-    "model.layers.26.self_attn.v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.26.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.27.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.27.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.27.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.27.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.27.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.27.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.27.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.27.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.27.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.27.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.28.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.28.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.28.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.28.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.28.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.28.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.28.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.28.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.28.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.28.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.29.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.29.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.29.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.29.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.29.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.29.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.29.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.29.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.29.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.29.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.3.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.3.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.3.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.30.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.30.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.30.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.30.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.30.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.30.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.30.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.30.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.30.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.31.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.31.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.31.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.31.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.31.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.31.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.31.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.31.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.31.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.32.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.32.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.32.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.32.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.32.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.32.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.32.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.32.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.32.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.32.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.32.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.32.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.32.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.32.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.32.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.32.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.33.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.33.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.33.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.33.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.33.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.33.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.33.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.33.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.33.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.33.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.33.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.33.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.33.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.33.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.33.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.33.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.34.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.34.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.34.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.34.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.34.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.34.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.34.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.34.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.34.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.34.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.34.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.34.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.34.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.34.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.34.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.34.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.35.input_layernorm.weight": "model-00003-of-00003.safetensors",
     "model.layers.35.mlp.down_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.35.mlp.gate_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.35.mlp.up_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.35.post_attention_layernorm.weight": "model-00003-of-00003.safetensors",
-    "model.layers.35.self_attn.k_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.35.self_attn.dkv_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.35.self_attn.kr_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.35.self_attn.kv_norm.weight": "model-00003-of-00003.safetensors",
     "model.layers.35.self_attn.o_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.35.self_attn.q_proj.weight": "model-00003-of-00003.safetensors",
-    "model.layers.35.self_attn.v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.35.self_attn.qk_k_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.35.self_attn.qk_q_norm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.35.self_attn.uk_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.35.self_attn.uv_proj.weight": "model-00003-of-00003.safetensors",
     "model.layers.4.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.4.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.4.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.5.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.5.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.5.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.6.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.6.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.6.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.6.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.6.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.6.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.6.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.7.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.7.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.7.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.7.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.7.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.7.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.7.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.8.input_layernorm.weight": "model-00001-of-00003.safetensors",
     "model.layers.8.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.8.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.8.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.8.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.8.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.8.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.9.input_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.9.mlp.down_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00002-of-00003.safetensors",
     "model.layers.9.mlp.gate_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.9.mlp.up_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
-    "model.layers.9.self_attn.k_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.9.self_attn.dkv_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.kr_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.kv_norm.weight": "model-00001-of-00003.safetensors",
     "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00003.safetensors",
     "model.layers.9.self_attn.q_proj.weight": "model-00001-of-00003.safetensors",
-    "model.layers.9.self_attn.v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.qk_k_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.qk_q_norm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.uk_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.uv_proj.weight": "model-00001-of-00003.safetensors",
     "model.norm.weight": "model-00003-of-00003.safetensors"
   }
 }

modeling_gigarembed.py

@@ -1,139 +1,658 @@
-from typing import List, Union, Dict, Mapping, Optional, Tuple, TypedDict
+import copy
+import logging
+from typing import Callable, List, Optional, Tuple, Union, Mapping
+
 import torch
-import os
-import json
+import torch.nn as nn
 import numpy as np
 import torch.nn.functional as F
 
-from functools import partial
-from contextlib import nullcontext
-from transformers import AutoModel, PreTrainedTokenizerFast, BatchEncoding, DataCollatorWithPadding
+from einops import rearrange, repeat
+from transformers import AutoModel, AutoTokenizer
+
+from transformers.cache_utils import Cache
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import DynamicCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
 from transformers.modeling_utils import PreTrainedModel
-from transformers.models.auto import AutoTokenizer
-from transformers.models.llama.modeling_llama import LLAMA_INPUTS_DOCSTRING
-from transformers.modeling_outputs import BaseModelOutputWithPast
-from transformers.modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_attention_mask_for_sdpa
-from transformers import LlamaModel, LlamaConfig
-from transformers.cache_utils import Cache, DynamicCache
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    logging,
+from transformers.processing_utils import Unpack
+from transformers.utils import LossKwargs, add_start_docstrings, add_start_docstrings_to_model_forward, replace_return_docstrings
+
+from .configuration_gigarembed import GigarConfig, GigarEmbedConfig, LatentAttentionConfig
+
+
+logger = logging.getLogger(__name__)
+_CONFIG_FOR_DOC = "GigarEmbedConfig"
+
+
+class GigarMLP(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=config.mlp_bias)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias)
+        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 GigarRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        GigarRMSNorm 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 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)
+
+
+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.
+    """
+    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, k_embed
+
+
+from flash_attn.layers.rotary import ApplyRotaryEmb
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None):
+    q = q.transpose(1, 2).to(cos.dtype) # make it [batch_size, seqlen, nheads, headdim]
+    k = k.transpose(1, 2).to(cos.dtype) # make it [batch_size, seqlen, nheads, headdim]
+    rotary_dim = cos.shape[-1]
+    cos = cos.squeeze(1).squeeze(0)[...,: rotary_dim // 2]  # [seq_len, dim // 2]
+    sin = sin.squeeze(1).squeeze(0)[...,: rotary_dim // 2]  # [seq_len, dim // 2]
+    return ApplyRotaryEmb.apply(q, cos, sin).transpose(1, 2), ApplyRotaryEmb.apply(k, cos, sin).transpose(1, 2)
+
+
+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,
+):
+    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
+
+
+class GigarLatentAttention(nn.Module):
+    """
+    Multi-headed Latent Attention (MLA)
+
+    Check out the original paper: https://arxiv.org/pdf/2405.04434,
+    and the reference implementation: https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+    """
+
+    def __init__(self, config: GigarConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.layer_idx = layer_idx
+        self.head_dim = getattr(config, "head_dim", config.hidden_size // config.num_attention_heads)
+
+        assert config.num_attention_heads == config.num_key_value_heads, (
+            "GQA for MLA is not supported (does it even make sense?)"
+        )
+        self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.apply_qk_norm = config.apply_qk_norm
+        self.attention_dropout = config.attention_dropout
+
+        assert config.mla_config is not None
+        self.qk_nope_head_dim = config.mla_config["qk_nope_head_dim"]
+        self.qk_rope_head_dim = config.mla_config["qk_rope_head_dim"]
+        self.v_head_dim = config.mla_config["v_head_dim"]  # V has no rope part
+        self.kv_lora_rank = config.mla_config["kv_lora_rank"]
+        self.q_lora_rank = config.mla_config["q_lora_rank"]
+
+        self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
+
+        self.scaling = self.qk_head_dim**-0.5
+
+        if self.q_lora_rank == 0:
+            self.q_proj = nn.Linear(
+                self.hidden_size,
+                self.num_heads * self.qk_head_dim,
+                bias=config.attention_bias,
+            )
+        else:
+            self.dq_proj = nn.Linear(
+                self.hidden_size,
+                self.q_lora_rank,
+                bias=config.attention_bias,
+            )
+            self.q_norm = GigarRMSNorm(self.q_lora_rank)
+            self.uq_proj = nn.Linear(
+                self.q_lora_rank,
+                self.num_heads * self.qk_head_dim,
+                bias=config.attention_bias,
+            )
+
+        self.kv_norm = GigarRMSNorm(self.kv_lora_rank)
+        self.dkv_proj = nn.Linear(
+            self.hidden_size,
+            self.kv_lora_rank,
+            bias=config.attention_bias,
+        )
+        self.uk_proj = nn.Linear(
+            config.kv_lora_rank,
+            self.num_heads * self.qk_nope_head_dim,
+            bias=config.attention_bias,
+        )
+        self.uv_proj = nn.Linear(
+            config.kv_lora_rank,
+            self.num_heads * self.v_head_dim,
+            bias=config.attention_bias,
+        )
+        self.kr_proj = nn.Linear(
+            self.hidden_size,
+            self.num_heads * self.qk_rope_head_dim,
+            bias=config.attention_bias,
+        )
+
+        self.o_proj = nn.Linear(
+            self.num_heads * self.v_head_dim,
+            self.hidden_size,
+            bias=config.attention_bias,
+        )
+
+        if self.apply_qk_norm:
+            self.qk_q_norm = nn.LayerNorm(self.num_heads * self.qk_head_dim, bias=False)
+            self.qk_k_norm = nn.LayerNorm(self.num_heads * self.qk_head_dim, bias=False)
+
+        config_for_rope = copy.copy(self.config)
+        config_for_rope.head_dim = self.config.qk_rope_head_dim
+        # self.rotary_emb = GigarRotaryEmbedding(config_for_rope)
+
+        # self._compute_qkv = torch.compile(self._compute_qkv)
+        self.is_causal = False
+
+    def _compute_qkv(
+        self,
+        hidden_states: torch.Tensor,
+    ):
+        """Compute query, key, and value tensors from hidden states."""
+        bsz, seq_len, _ = hidden_states.size()
+
+        if self.q_lora_rank == 0:
+            query = self.q_proj(hidden_states)
+        else:
+            query = self.uq_proj(self.q_norm(self.dq_proj(hidden_states)))
+
+        latent = self.dkv_proj(hidden_states)
+        latent = self.kv_norm(latent)
+        k_rope = self.kr_proj(hidden_states)
+
+        k_nope = self.uk_proj(latent)
+        value = self.uv_proj(latent)
+
+        if self.apply_qk_norm:
+            query = self.qk_q_norm(query).to(query.dtype)
+            key = self.qk_k_norm(torch.cat([k_nope, k_rope], dim=-1)).to(k_nope.dtype)
+            k_nope, k_rope = torch.split(key, [k_nope.shape[-1], k_rope.shape[-1]], dim=-1)
+
+        # Reshape tensors
+        query = query.view(bsz, seq_len, self.num_heads, self.qk_head_dim).transpose(1, 2)
+        k_nope = k_nope.view(bsz, seq_len, self.num_heads, self.qk_nope_head_dim).transpose(1, 2)
+        k_rope = k_rope.view(bsz, seq_len, self.num_heads, self.qk_rope_head_dim).transpose(1, 2)
+        value = value.view(bsz, seq_len, self.num_heads, self.v_head_dim).transpose(1, 2)
+
+        q_nope, q_rope = torch.split(query, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1)
+
+        return q_nope, q_rope, k_nope, k_rope, value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor],
+        attention_mask: Optional[torch.Tensor],
+        past_key_value: Optional[Cache] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]:
+        """
+        hidden_states: [bsz, seq_len, hidden_size]
+        attention_mask: [bsz, seq_len]
+        """
+        batch_size, seq_len, _ = hidden_states.size()
+
+        q_nope, q_rope, k_nope, k_rope, value_states = self._compute_qkv(hidden_states)
+
+        # cos, sin = self.rotary_emb(q_rope, seq_len=seq_len)
+        cos, sin = position_embeddings
+        q_rope, k_rope = apply_rotary_pos_emb(q_rope, k_rope, cos, sin)
+        query_states = torch.cat([q_nope, q_rope], dim=-1)
+        key_states = torch.cat([k_nope, k_rope], dim=-1)
+
+        if past_key_value 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_value.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,
+            **kwargs,
+        )
+
+        attn_output = attn_output.reshape(batch_size, seq_len, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class GigarDecoderLayer(nn.Module):
+    def __init__(self, config: GigarConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = GigarLatentAttention(config, layer_idx)
+        self.mlp = GigarMLP(config)
+        self.input_layernorm = GigarRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = GigarRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: Optional[bool] = False,
+        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[FlashAttentionKwargs],
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        return outputs
+
+
+class GigarRotaryEmbedding(nn.Module):
+    def __init__(self, config: GigarConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        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
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, seq_len=seq_len)
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+GIGAR_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`GigarConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Gigar Model outputting raw hidden-states without any specific head on top.",
+    GIGAR_START_DOCSTRING,
 )
-from einops import rearrange, repeat
-from tqdm.auto import tqdm
-from datasets import Dataset
-from torch.utils.data import DataLoader
-from .configuration_gigarembed import GigarEmbedConfig, LatentAttentionConfig, BidirectionalLlamaConfig
+class GigarPreTrainedModel(PreTrainedModel):
+    config_class = GigarConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["GigarDecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
 
-logger = logging.get_logger(__name__)
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
 
-class GigarEmbedFeatures(TypedDict):
-    input_dict: torch.Tensor
-    attention_mask: torch.Tensor
-    pool_mask: torch.Tensor
 
-class BidirectionalLlamaModel(LlamaModel):
-    config_class = BidirectionalLlamaConfig
-    
-    def __init__(self, config: LlamaConfig):
+GIGAR_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance, see our
+            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+def gradient_checkpointing_enable(self, *args, **kwargs):
+        self.model.gradient_checkpointing_enable(*args, **kwargs)
+
+@add_start_docstrings(
+    "The bare Gigar Model outputting raw hidden-states without any specific head on top.",
+    GIGAR_START_DOCSTRING,
+)
+class GigarModel(GigarPreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GigarDecoderLayer`]
+
+    Args:
+        config: GigarConfig
+    """
+
+    def __init__(self, config: GigarConfig):
         super().__init__(config)
-        for layer in self.layers:
-            layer.self_attn.is_causal = False
-        self._attn_implementation = "eager"
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
 
-    @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [GigarDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.norm = GigarRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = GigarRotaryEmbedding(config=config)
+        self.gradient_checkpointing = False
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(GIGAR_INPUTS_DOCSTRING)
     def forward(
         self,
         input_ids: torch.LongTensor = None,
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        past_key_values: Optional[Cache] = None,
         inputs_embeds: Optional[torch.FloatTensor] = None,
         use_cache: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
     ) -> Union[Tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
         )
         use_cache = use_cache if use_cache is not None else self.config.use_cache
-
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
-        # retrieve input_ids and inputs_embeds
-        if input_ids is not None and inputs_embeds is not None:
-            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
-        elif input_ids is not None:
-            batch_size, seq_length = input_ids.shape
-        elif inputs_embeds is not None:
-            batch_size, seq_length, _ = inputs_embeds.shape
-        else:
-            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
-
-        if self.gradient_checkpointing and self.training:
-            if use_cache:
-                logger.warning_once(
-                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                )
-                use_cache = False
-
-        past_key_values_length = 0
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
 
-        if use_cache:
-            use_legacy_cache = not isinstance(past_key_values, Cache)
-            if use_legacy_cache:
-                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
-            past_key_values_length = past_key_values.get_usable_length(seq_length)
-
-        if position_ids is None:
-            device = input_ids.device if input_ids is not None else inputs_embeds.device
-            position_ids = torch.arange(
-                past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
+        if self.gradient_checkpointing and self.training and use_cache:
+            logger.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`."
             )
-            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
-        else:
-            position_ids = position_ids.view(-1, seq_length).long()
+            use_cache = False
 
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
 
-        if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
-            is_padding_right = attention_mask[:, -1].sum().item() != batch_size
-            if is_padding_right:
-                raise ValueError(
-                    "You are attempting to perform batched generation with padding_side='right'"
-                    " this may lead to unexpected behaviour for Flash Attention version of Llama. Make sure to "
-                    " call `tokenizer.padding_side = 'left'` before tokenizing the input. "
-                )
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
 
-        if self._attn_implementation == "flash_attention_2":
-            # 2d mask is passed through the layers
-            attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
-        elif self._attn_implementation == "sdpa" and not output_attentions:
-            # output_attentions=True can not be supported when using SDPA, and we fall back on
-            # the manual implementation that requires a 4D causal mask in all cases.
-            attention_mask = _prepare_4d_attention_mask_for_sdpa(
-                attention_mask, inputs_embeds.dtype
-            )
-        else:
-            # 4d mask is passed through the layers
-            attention_mask = _prepare_4d_attention_mask(
-                attention_mask, inputs_embeds.dtype,
+        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.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
             )
 
-        hidden_states = inputs_embeds
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
         
+        attention_mask = self._update_encoder_mask(attention_mask, inputs_embeds)
+
+        hidden_states = inputs_embeds
+
         # create position embeddings to be shared across the decoder layers
         position_embeddings = self.rotary_emb(hidden_states, position_ids)
 
         # decoder layers
         all_hidden_states = () if output_hidden_states else None
         all_self_attns = () if output_attentions else None
-        next_decoder_cache = None
 
-        for decoder_layer in self.layers:
+        for decoder_layer in self.layers[: self.config.num_hidden_layers]:
             if output_hidden_states:
                 all_hidden_states += (hidden_states,)
 
@@ -141,29 +660,29 @@ class BidirectionalLlamaModel(LlamaModel):
                 layer_outputs = self._gradient_checkpointing_func(
                     decoder_layer.__call__,
                     hidden_states,
-                    attention_mask,
+                    attention_mask,  # causal_mask
                     position_ids,
                     past_key_values,
                     output_attentions,
                     use_cache,
-                    position_embeddings=position_embeddings
+                    cache_position,
+                    position_embeddings,
                 )
             else:
                 layer_outputs = decoder_layer(
                     hidden_states,
-                    attention_mask=attention_mask,
+                    attention_mask=attention_mask,  # causal_mask
                     position_ids=position_ids,
                     past_key_value=past_key_values,
                     output_attentions=output_attentions,
                     use_cache=use_cache,
-                    position_embeddings=position_embeddings
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                    **flash_attn_kwargs,
                 )
 
             hidden_states = layer_outputs[0]
 
-            if use_cache:
-                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
-
             if output_attentions:
                 all_self_attns += (layer_outputs[1],)
 
@@ -173,277 +692,463 @@ class BidirectionalLlamaModel(LlamaModel):
         if output_hidden_states:
             all_hidden_states += (hidden_states,)
 
-        next_cache = None
-        if use_cache:
-            next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
-
-        if not return_dict:
-            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
-        return BaseModelOutputWithPast(
+        output = BaseModelOutputWithPast(
             last_hidden_state=hidden_states,
-            past_key_values=next_cache,
+            past_key_values=past_key_values if use_cache else None,
             hidden_states=all_hidden_states,
             attentions=all_self_attns,
         )
+        return output if return_dict else output.to_tuple()
     
-def _move_to_device(maybe_tensor, device: torch.device):
-    if torch.is_tensor(maybe_tensor):
-        return maybe_tensor.to(device, non_blocking=device.type == "cuda")
-    elif isinstance(maybe_tensor, dict):
-        return {key: _move_to_device(value, device) for key, value in maybe_tensor.items()}
-    elif isinstance(maybe_tensor, list):
-        return [_move_to_device(x, device) for x in maybe_tensor]
-    elif isinstance(maybe_tensor, tuple):
-        return tuple([_move_to_device(x, device) for x in maybe_tensor])
-    elif isinstance(maybe_tensor, Mapping):
-        return type(maybe_tensor)({k: _move_to_device(v, device) for k, v in maybe_tensor.items()})
-    else:
-        return maybe_tensor
-
-def move_to_device(sample, device: torch.device):
-    if device.type == "cpu":
-        return sample
-    
-    if len(sample) == 0:
-        return {}
-    return _move_to_device(sample, device)
-
-
-def input_transform_func(
-    tokenizer: PreTrainedTokenizerFast,
-    examples: Dict[str, List],
-    max_length: int,
-    instruction: str,
-) -> BatchEncoding:
-    examples['input_texts'] = [instruction + input_example for input_example in examples['input_texts']]
-    batch_dict = tokenizer(
-        examples['input_texts'],
-        max_length=max_length,
-        padding=True,
-        return_token_type_ids=False,
-        return_tensors="pt",
-        truncation=True)
-    return batch_dict
-
-
-class PreNorm(torch.nn.Module):
-    def __init__(self, dim, fn, context_dim = None):
-        super().__init__()
-        self.fn = fn
-        self.norm = torch.nn.LayerNorm(dim)
-        self.norm_context = torch.nn.LayerNorm(context_dim) if exists(context_dim) else None
-
-    def forward(self, x, **kwargs):
-        x = self.norm(x)
-        if exists(self.norm_context):
-            context = kwargs['context']
-            normed_context = self.norm_context(context)
-            kwargs.update(context = normed_context)
-        return self.fn(x, **kwargs)
-
-class GEGLU(torch.nn.Module):
-    def forward(self, x):
-        x, gates = x.chunk(2, dim = -1)
-        return x * F.gelu(gates)
+    def _update_encoder_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+    ):
+        # Для flash_attention_2 возвращаем исходную маску
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and (attention_mask == 0).any():
+                return attention_mask
+            return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        batch_size, sequence_length = input_tensor.shape[:2]
+
+        # 1. Создаём базовую маску без ограничений (все токены видят друг друга)
+        encoder_mask = torch.full(
+            (batch_size, 1, sequence_length, sequence_length),
+            fill_value=1.0,
+            dtype=dtype,
+            device=device
+        )
+
+        # 2. Применяем padding-маску если есть
+        if attention_mask is not None:
+            # Создаём 4D padding-маску [batch, 1, 1, seq_len]
+            padding_mask = attention_mask[:, None, None, :].to(dtype=dtype)
+
+            # Комбинируем: обнуляем позиции где padding_mask == 0
+            encoder_mask = encoder_mask * padding_mask
+
+            # Конвертируем в формат для softmax (0 = -inf)
+            min_dtype = torch.finfo(dtype).min
+            encoder_mask = encoder_mask.masked_fill(encoder_mask == 0.0, min_dtype)
 
-class FeedForward(torch.nn.Module):
+        return encoder_mask
+
+    def _update_causal_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and (attention_mask == 0.0).any():
+                return attention_mask
+            return None
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        using_static_cache = isinstance(past_key_values, StaticCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype, device = input_tensor.dtype, input_tensor.device
+        sequence_length = input_tensor.shape[1]
+        if using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            device=device,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type == "cuda"
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        **kwargs,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
+                `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            device (`torch.device`):
+                The device to plcae the 4D attention mask on.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+            )
+            if sequence_length != 1:
+                causal_mask = torch.triu(causal_mask, diagonal=1)
+            causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = causal_mask.clone()  # copy to contiguous memory for in-place edit
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+                    padding_mask, min_dtype
+                )
+
+        return causal_mask
+
+
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ...
+
+
+class GigarForCausalLM(GigarPreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = GigarModel(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_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(GIGAR_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        **kwargs: Unpack[KwargsForCausalLM],
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            num_logits_to_keep (`int`, *optional*):
+                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, GigarForCausalLM
+
+        >>> model = GigarForCausalLM.from_pretrained("meta-gigar/Gigar-2-7b-hf")
+        >>> tokenizer = AutoTokenizer.from_pretrained("meta-gigar/Gigar-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."
+        ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+            **kwargs,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+class FeedForward(nn.Module):
     def __init__(self, dim, mult = 4):
         super().__init__()
-        self.net = torch.nn.Sequential(
-            torch.nn.Linear(dim, 2 * dim * mult),
-            GEGLU(),
-            torch.nn.Linear(dim * mult, dim)
-        )
+        self.hidden_size = dim
+        self.intermediate_size = dim * mult
+        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 = nn.SiLU()
 
     def forward(self, x):
-        return self.net(x)
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
 
-def exists(val):
-    return val is not None
 
-def default(val, d):
-    return val if exists(val) else d
+class Attention(nn.Module):
+    def __init__(self, query_dimension, context_dimension=None, num_heads=8, head_dim=64):
+        super().__init__()
+        inner_dimension = head_dim * num_heads
+        context_dimension = context_dimension if context_dimension is not None else query_dimension
 
+        self.scaling_factor = head_dim ** -0.5
+        self.num_heads = num_heads
 
-class Attention(torch.nn.Module):
-    def __init__(self, query_dim, context_dim = None, heads = 8, dim_head = 64):
-        super().__init__()
-        inner_dim = dim_head * heads
-        context_dim = default(context_dim, query_dim)
-        self.scale = dim_head ** -0.5
-        self.heads = heads
+        self.to_q = nn.Linear(query_dimension, inner_dimension, bias=False)
+        self.to_kv = nn.Linear(context_dimension, inner_dimension * 2, bias=False)
+        self.to_out = nn.Linear(inner_dimension, query_dimension, bias=False)
 
-        self.to_q = torch.nn.Linear(query_dim, inner_dim, bias = False)
-        self.to_kv = torch.nn.Linear(context_dim, inner_dim * 2, bias = False)
-        self.to_out = torch.nn.Linear(inner_dim, query_dim, bias = False)
+    def forward(self, input_tensor, context=None, attention_mask=None):
+        batch_size, seq_len, _ = input_tensor.shape
+        num_heads = self.num_heads
 
-    def forward(self, x, context = None, mask = None):
-        h = self.heads
-        q = self.to_q(x)
-        context = default(context, x)
-        k, v = self.to_kv(context).chunk(2, dim = -1)
-        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h = h), (q, k, v))
+        # Project input to query
+        query = self.to_q(input_tensor)
 
-        with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_mem_efficient=True):
-            out = torch.nn.functional.scaled_dot_product_attention(q, k, v)
-        out = rearrange(out, '(b h) n d -> b n (h d)', h = h)
-        return self.to_out(out)
+        # Use input as context if not provided
+        context = input_tensor if context is None else context
+        key, value = self.to_kv(context).chunk(2, dim=-1)
+
+        # Rearrange for multi-head attention
+        query = rearrange(query, 'b n (h d) -> (b h) n d', h=num_heads)
+        key = rearrange(key, 'b n (h d) -> (b h) n d', h=num_heads)
+        value = rearrange(value, 'b n (h d) -> (b h) n d', h=num_heads)
+
+        # Compute scaled dot-product attention
+        with torch.backends.cuda.sdp_kernel(
+            enable_flash=True, 
+            enable_math=True, 
+            enable_mem_efficient=True
+        ):
+            attention_output = F.scaled_dot_product_attention(query, key, value)
+
+        # Rearrange back to original shape
+        attention_output = rearrange(attention_output, '(b h) n d -> b n (h d)', h=num_heads)
+
+        return self.to_out(attention_output)
 
 
 class LatentAttentionModel(PreTrainedModel):
     config_class = LatentAttentionConfig
 
-    def __init__(self, config: LatentAttentionConfig):
-        super().__init__(config)
-        ## cross-attention block
-        num_latents, latent_dim, cross_heads, cross_dim_head = config.num_latents_value, config.latent_dim, config.num_cross_heads, config.cross_dim_head
-        dim = config.hidden_dim
-        # init latent_attention and latents
-        self.cross_attend_blocks = torch.nn.ModuleList([
-            PreNorm(latent_dim, Attention(latent_dim, dim, heads = cross_heads, dim_head = cross_dim_head),
-                    context_dim = dim),
-            PreNorm(latent_dim, FeedForward(latent_dim)),
+    def __init__(self, configuration: LatentAttentionConfig):
+        super().__init__(configuration)
+
+        # Extract configuration parameters
+        num_latents = configuration.num_latents_value
+        latent_dimension = configuration.latent_dim
+        cross_attention_heads = configuration.num_cross_heads
+        cross_head_dimension = configuration.cross_dim_head
+        hidden_dimension = configuration.hidden_dim
+
+        # Initialize cross-attention components
+        self.cross_attend_blocks = nn.ModuleList([
+            Attention(
+                query_dimension=latent_dimension,
+                context_dimension=hidden_dimension,
+                num_heads=cross_attention_heads,
+                head_dim=cross_head_dimension
+            ),
+            FeedForward(latent_dimension)
         ])
 
-        self.w_lexical = torch.nn.Linear(latent_dim, 1)
-        self.w_multi_vector = torch.nn.Linear(latent_dim, latent_dim)
-
-        # self.output_normalize = config.output_normalize
-        self.register_parameter("latents", torch.nn.Parameter(torch.randn(num_latents, latent_dim)))
-        self._attn_implementation = "eager"
-
-    def forward(self, hiddens, attention_mask: torch.Tensor=None):
-        # cross-attention block
-        cross_attn, cross_ff = self.cross_attend_blocks
-        b, *_, device = *hiddens.shape, hiddens.device
-        x = repeat(self.latents, 'n d -> b n d', b = b)
-        output = cross_attn(hiddens, context=x, mask=attention_mask) + hiddens
-        output = cross_ff(output) + output
-        if attention_mask != None:
-            s = torch.sum(output * attention_mask.unsqueeze(-1), dim=1)
-            d = attention_mask.sum(dim=1, keepdim=True)
-            output = s / d
-            output = F.normalize(output, p=2, dim=-1)
-        return output
+        # Register learnable latents as model parameter
+        self.latents = nn.Parameter(torch.randn(num_latents, latent_dimension))
+
+    def forward(self, hidden_states, attention_mask: Optional[torch.Tensor] = None):
+        cross_attention, feed_forward = self.cross_attend_blocks
+        
+        batch_size, device = hidden_states.size(0), hidden_states.device
+        
+        # Expand latents to match batch size
+        expanded_latents = self.latents.repeat(batch_size, 1, 1)
+        
+        # Apply cross-attention with residual connection
+        attended_output = cross_attention(
+            hidden_states, context=expanded_latents, attention_mask=attention_mask) + hidden_states
+        
+        # Apply feed-forward with residual connection
+        processed_output = feed_forward(attended_output) + attended_output
+        
+        return processed_output
+
 
 class GigarEmbedModel(PreTrainedModel):
     config_class = GigarEmbedConfig
-    _no_split_modules = ["LlamaDecoderLayer", "LatentAttentionModel"]
+    _supports_flash_attn_2 = True
+    _no_split_modules = ["GigarDecoderLayer", "LatentAttentionModel"]
     
-    def __init__(self, config: GigarEmbedConfig):
-        super().__init__(config)
-        self.latent_attention_model = AutoModel.from_config(config.latent_attention_config)
-        self.model = AutoModel.from_config(
-            config.text_config,
-        ) if config.text_config is not None else None
-        self.tokenizer = AutoTokenizer.from_pretrained(config.text_config._name_or_path) if config.text_config is not None else None
-        self.padding_side = config.padding_side
-        self.is_mask_instruction = config.is_mask_instruction
-        self.add_eos = config.add_eos
-        self.mask_type = config.mask_type
-        if config.add_pad_token and self.tokenizer is not None:
+    def __init__(self, configuration: GigarEmbedConfig):
+        super().__init__(configuration)
+        
+        # Initialize latent attention model
+        self.latent_attention_model = AutoModel.from_config(
+            configuration.latent_attention_config
+        )
+
+        self.tokenizer, self.text_encoder = None, None
+        if configuration.text_config is not None:
+            # Initialize text model if provided in config
+            self.model = AutoModel.from_config(configuration.text_config)
+
+            # Initialize tokenizer if text config is available
+            self.tokenizer = AutoTokenizer.from_pretrained(
+                configuration.text_config.name_or_path
+            )
+        
+        # Set configuration parameters
+        self.padding_side = configuration.padding_side
+        self.add_eos = configuration.add_eos
+        self.mask_type = configuration.mask_type
+        
+        # Add padding token if configured
+        if configuration.add_pad_token and self.tokenizer is not None:
             self.add_pad_token()
 
     def add_pad_token(self):
         self.tokenizer.pad_token_id = 0
         self.tokenizer.padding_side = self.padding_side
-    
-    def prepare_kwargs_from_batch(self, batch_dict: dict, instruction_lens: int, device: torch.device):
-        batch_dict = move_to_device(batch_dict, device)
-        attention_mask = batch_dict['attention_mask'].clone() if 'attention_mask' in batch_dict else None
-        if (attention_mask is not None and
-            self.padding_side == "right" and
-            self.is_mask_instruction == True and
-            instruction_lens > 0):
-            # Mask out the instruction tokens for mean-pooling
-            attention_mask[:, :instruction_lens] = 0
-        features: GigarEmbedFeatures = {
-            'input_ids': torch.tensor(batch_dict.get('input_ids').to(batch_dict.get('input_ids')).long()),
-            'attention_mask': batch_dict['attention_mask'],
-            'pool_mask': attention_mask,
-        }
-        return features
 
-    @torch.no_grad()
-    def _do_encode(self,
-        prompts: List[str],
-        batch_size: int=1,
-        instruction: str="",
-        max_length: int=4096,
-        num_workers: int=32,
-        **kwargs
-    ) -> Union[np.ndarray, torch.FloatTensor]:
-        dataset: Dataset = Dataset.from_dict({'input_texts': prompts})
-        dataset.set_transform(partial(input_transform_func,
-                                      self.tokenizer,
-                                      max_length=max_length,
-                                      instruction=instruction))
-
-        data_collator = DataCollatorWithPadding(self.tokenizer)
-        data_loader = DataLoader(
-            dataset,
-            batch_size=batch_size,
-            shuffle=False,
-            drop_last=False,
-            num_workers=num_workers,
-            collate_fn=data_collator,
-            pin_memory=True)
-
-        if self.padding_side == "right" and self.is_mask_instruction == True and len(instruction) > 0:
-            instruction_lens = len(self.tokenizer.tokenize(instruction))
-        else:
-            instruction_lens = 0
-
-        encoded_embeds = []
-        device = next(self.model.parameters()).device
-        for batch_dict in tqdm(data_loader, desc='encoding', mininterval=10):
-            features = self.prepare_kwargs_from_batch(batch_dict, instruction_lens, device=device)
-            embeds=self(**features)["sentence_embeddings"].squeeze(1)
-            encoded_embeds.append(embeds)
-        encoded_embeds = torch.cat(encoded_embeds, axis=0)
-        if "return_numpy" in kwargs and kwargs.get("return_numpy"):
-            encoded_embeds = encoded_embeds.cpu().detach().numpy()
-        return encoded_embeds
-
-    def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, pool_mask: Optional[torch.Tensor]=None, 
-                return_dict: bool=True, **kwargs):
+    def gradient_checkpointing_enable(self, *args, **kwargs):
+        self.model.gradient_checkpointing_enable(*args, **kwargs)
+
+    def forward(self, input_ids: torch.Tensor, attention_mask: torch.Tensor, 
+                return_embeddings: bool = False, **kwargs):
         kwargs.pop('token_type_ids', None)
 
         with torch.autocast('cuda', dtype=torch.bfloat16):
             outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, **kwargs)
 
-            if pool_mask is None: pool_mask = attention_mask.clone()
+            last_hidden = self.latent_attention_model(outputs.last_hidden_state, attention_mask)
 
-            embeds = self.latent_attention_model(outputs.last_hidden_state, pool_mask)
+        if return_embeddings:
+            return self.mean_pool(last_hidden, attention_mask)
 
-        if not return_dict:
-            return (embeds,)
-        return {"sentence_embeddings": embeds}
-    
-    
-    @torch.no_grad()
-    def encode(self, prompts: List[str], instruction: str="", max_length: int=4096, **kwargs):
-        if self.padding_side == "right" and self.is_mask_instruction == True and len(instruction) > 0:
-            instruction_lens = len(self.tokenizer.tokenize(instruction))
-        else:
-            instruction_lens = 0
-        
-        device = next(self.model.parameters()).device
-        batch_dict = input_transform_func(self.tokenizer,
-                                          {"input_texts": [prompt for prompt in prompts]},
-                                          max_length=max_length,
-                                          instruction=instruction)
+        return last_hidden
 
-        features: GigarEmbedFeatures = self.prepare_kwargs_from_batch(batch_dict, instruction_lens, device=device)
-        return self(**features)["sentence_embeddings"].squeeze(1)
+    def mean_pool(self, last_hidden: torch.Tensor, attention_mask: torch.Tensor):
+        last_hidden = last_hidden.masked_fill(~attention_mask[..., None].bool(), 0.0)
+        embeddings = last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
+        return F.normalize(embeddings, p=2, dim=-1)
 
 
 ## AutoModel Register
+AutoModel.register(GigarConfig, GigarModel)
 AutoModel.register(GigarEmbedConfig, GigarEmbedModel)
 AutoModel.register(LatentAttentionConfig, LatentAttentionModel)
-AutoModel.register(BidirectionalLlamaConfig, BidirectionalLlamaModel)
 
 ## Register for auto class
+GigarModel.register_for_auto_class("AutoModel")
 GigarEmbedModel.register_for_auto_class("AutoModel")
 LatentAttentionModel.register_for_auto_class("AutoModel")
-BidirectionalLlamaModel.register_for_auto_class("AutoModel")

tokenizer.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bcb50618d6a02d4562ada12978a8aa9e0b6e31260f71acce28586072a9005d4a
-size 10728437
+oid sha256:0ec0a1cffcc9192f5ee3d7b273673a062918055238bda3d23cfb6d2512e947ff
+size 10728325

tokenizer_config.json

@@ -2074,8 +2074,10 @@
     }
   },
   "bos_token": "<s>",
+  "chat_template": "{%- for message in messages -%}{{ message['content'] }}{%- if not loop.last -%}{{ ' ' }}{%- endif -%}{%- endfor -%}",
   "clean_up_tokenization_spaces": true,
   "eos_token": "</s>",
+  "extra_special_tokens": {},
   "max_length": 512,
   "model_max_length": 1000000000000000019884624838656,
   "pad_to_multiple_of": null,