Upload SVDLlamaForCausalLM

config.json

@@ -0,0 +1,34 @@
+{
+  "_name_or_path": "huggingface_repos/meta-llama2-7b-svd80",
+  "architectures": [
+    "SVDLlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_svd_llama.SVDLlamaConfig",
+    "AutoModelForCausalLM": "modeling_svd_llama.SVDLlamaForCausalLM"
+  },
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 11008,
+  "max_position_embeddings": 4096,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 32,
+  "pretraining_tp": 1,
+  "ratio": 0.8,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.44.2",
+  "use_cache": true,
+  "vocab_size": 32000
+}

configuration_svd_llama.py

@@ -0,0 +1,206 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""LLaMA model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+
+
+class SVDLlamaConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`LlamaModel`]. It is used to instantiate an LLaMA
+    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 LLaMA-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 LLaMA model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`LlamaModel`]
+        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. Llama 1 supports up to 2048 tokens,
+            Llama 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',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, *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.
+
+    ```python
+    >>> from transformers import LlamaModel, LlamaConfig
+
+    >>> # Initializing a LLaMA llama-7b style configuration
+    >>> configuration = LlamaConfig()
+
+    >>> # Initializing a model from the llama-7b style configuration
+    >>> model = LlamaModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "llama"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    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,
+        ratio=1,
+        **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
+        # for svdllm
+        self.ratio = ratio
+
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move 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)
+
+        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,
+        )

generation_config.json

@@ -0,0 +1,10 @@
+{
+  "bos_token_id": 1,
+  "do_sample": true,
+  "eos_token_id": 2,
+  "max_length": 4096,
+  "pad_token_id": 0,
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "transformers_version": "4.44.2"
+}

model-00001-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:47662d973c1c827158505bc90668f476b4d04f58b97d4374d18b49e92021347b
+size 4994061160

model-00002-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9d1b9017e09f9eb927b4c49c1f6614bde6817b4f0fffe5da2bb0abace3b0cb6
+size 4994791552

model-00003-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fc41a3cd73d5bef4e9557f3922f9751bbacde5fef5cf5567f61771196ce537d
+size 896286416

model.safetensors.index.json

@@ -0,0 +1,522 @@
+{
+  "metadata": {
+    "total_size": 10885079040
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00003-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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.0.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.1.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.mlp.up_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.k_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.10.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.mlp.up_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.self_attn.k_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.11.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.mlp.up_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.self_attn.k_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.12.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.mlp.up_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.self_attn.k_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.13.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.14.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.14.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.15.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.16.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.17.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.18.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.19.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.2.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.20.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.21.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.22.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.23.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.24.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.mlp.up_v_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_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.25.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.mlp.up_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.k_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.26.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.mlp.up_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.self_attn.k_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.27.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.mlp.up_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.self_attn.k_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.28.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.input_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.mlp.down_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.mlp.down_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.mlp.gate_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.mlp.gate_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.mlp.up_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.mlp.up_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.post_attention_layernorm.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.self_attn.k_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.self_attn.k_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.self_attn.o_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.self_attn.o_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.self_attn.q_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.self_attn.v_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.29.self_attn.v_v_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.3.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.30.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.down_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.down_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.gate_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.gate_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.up_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.mlp.up_v_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_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.k_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.o_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.o_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.q_u_proj.weight": "model-00002-of-00003.safetensors",
+    "model.layers.30.self_attn.q_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.v_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.30.self_attn.v_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.input_layernorm.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.down_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.down_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.gate_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.gate_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.up_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.mlp.up_v_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_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.k_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.o_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.o_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.q_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.q_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.v_u_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.31.self_attn.v_v_proj.weight": "model-00003-of-00003.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.4.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.5.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.6.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.7.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.8.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.down_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.down_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.gate_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.gate_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.up_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.mlp.up_v_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_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.k_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.o_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.o_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.q_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.q_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.v_u_proj.weight": "model-00001-of-00003.safetensors",
+    "model.layers.9.self_attn.v_v_proj.weight": "model-00001-of-00003.safetensors",
+    "model.norm.weight": "model-00003-of-00003.safetensors"
+  }
+}

modeling_svd_llama.py

@@ -0,0 +1,356 @@
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch import nn
+from transformers.cache_utils import Cache
+from transformers.activations import ACT2FN
+from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
+from transformers.utils import logging
+from transformers import LlamaForCausalLM
+from transformers.models.llama.modeling_llama import LlamaDecoderLayer, LlamaModel, LlamaRotaryEmbedding, LlamaRMSNorm, repeat_kv, apply_rotary_pos_emb
+from component.configuration_svd_llama import SVDLlamaConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "LlamaConfig"
+
+ALL_LAYERNORM_LAYERS.append(LlamaRMSNorm)
+
+class SVDLlamaMLP(nn.Module):
+    def __init__(self, config: SVDLlamaConfig):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.ratio = config.ratio
+        self.low_rank = int(self.intermediate_size * self.hidden_size * self.ratio / (self.intermediate_size + self.hidden_size))
+        
+        self.gate_u_proj = nn.Linear(self.low_rank, self.intermediate_size, bias=config.mlp_bias)
+        self.gate_v_proj = nn.Linear(self.hidden_size, self.low_rank, bias=False)
+        
+        self.down_u_proj = nn.Linear(self.low_rank, self.hidden_size, bias=config.mlp_bias)
+        self.down_v_proj = nn.Linear(self.intermediate_size, self.low_rank, bias=False)
+        
+        self.up_u_proj = nn.Linear(self.low_rank, self.intermediate_size, bias=config.mlp_bias)
+        self.up_v_proj = nn.Linear(self.hidden_size, self.low_rank, bias=False)
+        
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        up = self.up_u_proj(self.up_v_proj(x))
+        gate = self.gate_u_proj(self.gate_v_proj(x))
+        return self.down_u_proj(self.down_v_proj(self.act_fn(gate) * up))
+
+class SVDLlamaAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config: SVDLlamaConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
+                "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.attention_dropout = config.attention_dropout
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.is_causal = True
+        self.ratio = config.ratio # 1 means no truncate, just keep normal attn
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+
+        self.q_low_rank = int(self.num_heads * self.head_dim * self.hidden_size * self.ratio / (self.num_heads * self.head_dim + self.hidden_size))
+        self.q_u_proj = nn.Linear(self.q_low_rank, self.num_heads * self.head_dim, bias=config.attention_bias)
+        self.q_v_proj = nn.Linear(self.hidden_size, self.q_low_rank, bias=False)
+
+        self.k_low_rank = int(self.num_key_value_heads * self.head_dim * self.hidden_size * self.ratio / (self.num_key_value_heads * self.head_dim + self.hidden_size))
+        self.k_u_proj = nn.Linear(self.k_low_rank, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.k_v_proj = nn.Linear(self.hidden_size, self.k_low_rank, bias=False)
+
+        self.v_low_rank = int(self.num_key_value_heads * self.head_dim * self.hidden_size * self.ratio / (self.num_key_value_heads * self.head_dim + self.hidden_size))
+        self.v_u_proj = nn.Linear(self.v_low_rank, self.num_key_value_heads * self.head_dim, bias=config.attention_bias)
+        self.v_v_proj = nn.Linear(self.hidden_size, self.v_low_rank, bias=False)
+
+        self.o_low_rank = int(self.hidden_size * self.hidden_size * self.ratio / (self.hidden_size + self.hidden_size))
+        self.o_u_proj = nn.Linear(self.o_low_rank, self.hidden_size, bias=config.attention_bias)
+        self.o_v_proj = nn.Linear(self.hidden_size, self.o_low_rank, bias=False)
+
+        # TODO (joao): remove in v4.45 (RoPE is computed in the model, not in the decoder layers)
+        self.rotary_emb = LlamaRotaryEmbedding(config=self.config)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.45
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # bsz, q_len, _ = hidden_states.size()
+
+
+        # query_states = self.q_u_proj(self.q_v_proj(hidden_states))
+        # key_states = self.k_u_proj(self.k_v_proj(hidden_states))
+        # value_states = self.v_u_proj(self.v_v_proj(hidden_states))
+
+        # query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        # key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        # value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        # if position_embeddings is None:
+        #     logger.warning_once(
+        #         "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+        #         "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+        #         "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
+        #         "removed and `position_embeddings` will be mandatory."
+        #     )
+        #     cos, sin = self.rotary_emb(value_states, position_ids)
+        # else:
+        #     cos, sin = position_embeddings
+        # query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        # 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)
+
+        # key_states = repeat_kv(key_states, self.num_key_value_groups)
+        # value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        # attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        # if attention_mask is not None:  # no matter the length, we just slice it
+        #     causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        #     attn_weights = attn_weights + causal_mask
+
+        # # upcast attention to fp32
+        # attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        # attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        # attn_output = torch.matmul(attn_weights, value_states)
+
+        # if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+        #     raise ValueError(
+        #         f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+        #         f" {attn_output.size()}"
+        #     )
+
+        # attn_output = attn_output.transpose(1, 2).contiguous()
+
+        # attn_output = attn_output.reshape(bsz, q_len, -1)
+
+        # attn_output = self.o_u_proj(self.o_v_proj(attn_output))
+
+        # if not output_attentions:
+        #     attn_weights = None
+
+        # return attn_output, attn_weights, past_key_value
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "LlamaModel is using LlamaSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+            )
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_u_proj(self.q_v_proj(hidden_states))
+        key_states = self.k_u_proj(self.k_v_proj(hidden_states))
+        value_states = self.v_u_proj(self.v_v_proj(hidden_states))
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
+                "removed and `position_embeddings` will be mandatory."
+            )
+            cos, sin = self.rotary_emb(value_states, position_ids)
+        else:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        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)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:
+            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and causal_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, -1)
+
+        attn_output = self.o_u_proj(self.o_v_proj(attn_output))
+
+        return attn_output, None, past_key_value
+
+class SVDLLaMASDPA(SVDLlamaAttention):
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.45
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if output_attentions:
+            # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
+            logger.warning_once(
+                "LlamaModel is using LlamaSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+                'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+            )
+            return super().forward(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_value,
+                output_attentions=output_attentions,
+                use_cache=use_cache,
+                cache_position=cache_position,
+                position_embeddings=position_embeddings,
+            )
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_u_proj(self.q_v_proj(hidden_states))
+        key_states = self.k_u_proj(self.k_v_proj(hidden_states))
+        value_states = self.v_u_proj(self.v_v_proj(hidden_states))
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.45 `position_ids` will be "
+                "removed and `position_embeddings` will be mandatory."
+            )
+            cos, sin = self.rotary_emb(value_states, position_ids)
+        else:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        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)
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        causal_mask = attention_mask
+        if attention_mask is not None:
+            causal_mask = causal_mask[:, :, :, : key_states.shape[-2]]
+
+        # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+        # Reference: https://github.com/pytorch/pytorch/issues/112577.
+        if query_states.device.type == "cuda" and causal_mask is not None:
+            query_states = query_states.contiguous()
+            key_states = key_states.contiguous()
+            value_states = value_states.contiguous()
+
+        # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+        # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+        is_causal = True if causal_mask is None and q_len > 1 else False
+
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=causal_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.view(bsz, q_len, -1)
+
+        attn_output = self.o_u_proj(self.o_v_proj(attn_output))
+
+        return attn_output, None, past_key_value
+
+
+class SVDLlamaDecoderLayer(LlamaDecoderLayer):
+    def __init__(self, config: SVDLlamaConfig, layer_idx: int):
+        super().__init__(config, layer_idx)
+        self.self_attn = SVDLlamaAttention(config=config, layer_idx=layer_idx)
+        self.mlp = SVDLlamaMLP(config)
+
+
+class SVDLlamaForCausalLM(LlamaForCausalLM):
+    def __init__(self, config: SVDLlamaConfig):
+        super().__init__(config)
+        self.model = LlamaModel(config)
+        self.model.layers = nn.ModuleList(
+            [SVDLlamaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.model._no_split_modules = ["SVDLlamaDecoderLayer"]
+        self._no_split_modules = ["SVDLlamaDecoderLayer"]
+        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()