Upload VoRAForCausalLM

README.md

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

attention_mask.py

@@ -0,0 +1,81 @@
+from typing import Optional
+
+import torch
+
+
+def _make_causal_mask(
+    attention_mask: torch.Tensor, dtype: torch.dtype, device: torch.device
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = attention_mask.shape
+    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min, device=device), device=device)
+    mask_cond = torch.arange(mask.size(-1), device=device)
+    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
+    mask = mask.to(dtype)
+
+    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len)
+
+
+def _make_2dvison_mask(column_mask, dtype: torch.dtype, device: torch.device):
+    """
+    """
+    bsz, seq_length = column_mask.shape
+    cross_mask = torch.zeros((bsz, 1, seq_length, seq_length), dtype=dtype, device=device)
+
+    # 找到连续的 1 的区间
+    start = None
+    for bsz_idx in range(bsz):
+        for i in range(seq_length):
+            if column_mask[bsz_idx, i] == 1:
+                if start is None:
+                    start = i
+            else:
+                if start is not None:
+                    # 填充区间
+                    cross_mask[bsz_idx, 0, start:i, start:i] = 1
+                    start = None
+
+        # 处理最后一个区间
+        if start is not None:
+            cross_mask[bsz_idx, 0, start:seq_length, start:seq_length] = 1
+
+    return cross_mask
+
+
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+    inverted_mask = 1.0 - expanded_mask
+
+    return inverted_mask.masked_fill_(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+def make_mask(attention_mask: torch.Tensor, dtype: torch.dtype=None, device: torch.device=None, mode: str="default", vision_mask: torch.Tensor=None, ):
+    if dtype is None:
+        dtype = attention_mask.dtype
+    if device is None:
+        device = attention_mask.device      
+    expanded_attn_mask = _expand_mask(attention_mask, dtype).to(device)
+    causal_mask = _make_causal_mask(attention_mask, dtype, device).to(device)
+    if mode == "default":
+        return attention_mask
+    else:
+        assert vision_mask is not None, "vision_mask is None"
+        vision_mask = vision_mask.to(device)
+        bsz, seq_length = attention_mask.shape
+        vision_mask_bg = vision_mask[:, None, :, None]
+        vision_mask_2d = _make_2dvison_mask(vision_mask, dtype, device)
+        if mode == "bidirectional":
+            mask = expanded_attn_mask + causal_mask
+            mask = mask.clone().masked_fill_(vision_mask_2d.to(torch.bool), 0)
+            return mask 
+        else:
+            raise NotImplementedError(f"mode {mode} is not implemented")

aux_vision.py

@@ -0,0 +1,102 @@
+import torch
+import torch.nn as nn
+from transformers import CLIPVisionModel, AutoModel
+
+from .configuration_vora import VoRAConfig
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-5):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+
+class CosineLoss(nn.Module):
+    def __init__(self, reduction='mean'):
+        super(CosineLoss, self).__init__()
+        self.reduction = reduction
+
+    @staticmethod
+    def interpolate_tokens_2d(self, teacher_tokens, target_size):
+        """
+        Interpolate teacher tokens to the target size using bilinear interpolation.
+        """
+        # teacher_tokens shape is (batch_size, height, width, feature_dim)
+        teacher_tokens = teacher_tokens.permute(0, 3, 1, 2)  # Convert to (batch_size, feature_dim, height, width)
+        interpolated = torch.nn.functional.interpolate(teacher_tokens, size=target_size, mode='bilinear', align_corners=True).flatten(2)  # Flatten height and width dimensions
+        return interpolated.permute(0, 2, 1)  # Convert back to (batch_size, new_height * new_width, feature_dim)
+
+    def forward(self, input: torch.Tensor, target: torch.Tensor, input_shape=None, target_shape=None) -> torch.Tensor:
+        if input_shape is not None and target_shape is not None:
+            input = input.reshape((input.shape[0], ) + input_shape + (-1, ))
+            input = self.interpolate_tokens_2d(input, target_shape)
+
+        cos_sim = nn.functional.cosine_similarity(input, target, dim=1)
+        loss = 1 - cos_sim
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        else:
+            return loss
+
+
+class AuxVision(nn.Module):
+    def __init__(self,
+        config: VoRAConfig = None,
+    ):
+        super().__init__()
+        self.skip_aux_cls = config.skip_aux_cls  # whether to skip the cls token in ViT
+        # ---------------- Setup Aux Model ----------------
+        if 'clip' in config.aux_vision.lower():
+            self.aux_model = CLIPVisionModel.from_pretrained(config.aux_vision)
+            vision_hidden_size = self.aux_model.vision_model.config.hidden_size
+            num_hidden_layers = self.aux_model.vision_model.config.num_hidden_layers
+        else:
+            self.aux_model = AutoModel.from_pretrained(config.aux_vision, trust_remote_code=True)
+            vision_hidden_size = self.aux_model.config.hidden_size
+            num_hidden_layers = self.aux_model.config.num_hidden_layers
+        for name, param in self.aux_model.named_parameters():
+            param.requires_grad = False
+        # -------------------------------------------------
+
+        # ---------------- Setup Aux Heads ----------------
+        self.aux_layers = list(range(num_hidden_layers))
+        for layer_id in self.aux_layers:
+            self.add_module(f"aux_layer_{layer_id}", self.build_aux_layer(config.hidden_size, vision_hidden_size))
+        # -------------------------------------------------
+
+        self.loss_function = CosineLoss()
+        self.loss_keys = [f"loss_aux_layer_{layer_id}" for layer_id in self.aux_layers]
+
+    def build_aux_layer(self, llm_hidden_size, vit_hidden_size):
+        return nn.Sequential(
+                    RMSNorm(llm_hidden_size),
+                    nn.Linear(
+                        llm_hidden_size,
+                        vit_hidden_size,
+                        bias=False,
+                    )
+                )
+
+    def forward(self, frames, llm_hidden_states, vision_mask):
+        vision_hidden_states = self.aux_model(frames, output_hidden_states=True).hidden_states
+        losses = {}
+        for layer_idx in self.aux_layers:
+            aux_hidden_states = getattr(self, f"aux_layer_{layer_idx}")(llm_hidden_states[layer_idx][vision_mask == 1])
+            start_id = 1 if self.skip_aux_cls else 0
+            aux_loss = self.loss_function(vision_hidden_states[layer_idx][:, start_id:].reshape(aux_hidden_states.shape), aux_hidden_states)
+            losses[f"loss_aux_layer_{layer_idx}"] = aux_loss
+        return losses

config.json

@@ -0,0 +1,55 @@
+{
+  "architectures": [
+    "VoRAForCausalLM"
+  ],
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_vora.VoRAConfig",
+    "AutoModelForCausalLM": "modeling_vora.VoRAForCausalLM"
+  },
+  "aux_vision": "",
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 3584,
+  "image_size": 448,
+  "initializer_range": 0.02,
+  "intermediate_size": 18944,
+  "llm": "Qwen/Qwen2.5-7B-Instruct",
+  "lora": {
+    "layers": 24,
+    "r": 1024,
+    "target_modules": [
+      "self_attn.q_proj",
+      "self_attn.k_proj",
+      "self_attn.v_proj",
+      "self_attn.o_proj",
+      "mlp.up_proj",
+      "mlp.gate_proj",
+      "mlp.down_proj"
+    ]
+  },
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "vora",
+  "num_attention_heads": 28,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 4,
+  "patch_size": 14,
+  "pretrained": "/root/.cache/huggingface/hub/models--Hon-Wong--VoRA-7B-Instruct/snapshots/6598d808aebd0bac624665f92b3433f2d311afd0/",
+  "reuse_aux_vision_embedding_layers": "",
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000.0,
+  "skip_aux_cls": false,
+  "sliding_window": 131072,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.50.3",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vision_attention_mask": "bidirectional",
+  "vision_embedding": "AIMv2Embedding",
+  "vision_embedding_intermediate_size": 1536,
+  "vocab_size": 152064
+}

configuration_vora.py

@@ -0,0 +1,38 @@
+from typing import Any
+
+from transformers.configuration_utils import PretrainedConfig
+
+__all__ = ["VoRAConfig"]
+
+
+class VoRAConfig(PretrainedConfig):
+    model_type = "vora"
+    _auto_class = "AutoConfig"
+
+    def __init__(
+        self,
+        llm: str = "",
+        aux_vision: str = "",
+        skip_aux_cls: bool = False,
+        reuse_aux_vision_embedding_layers: str = "",
+        lora: dict = {},
+        image_size: int = 448,
+        vision_embedding: str = "AIMv2",
+        vision_embedding_intermediate_size: int = 1536,
+        patch_size: int = 14,
+        vision_attention_mask: str = "bidirectional",
+        rms_norm_eps: float = 1e-5,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.llm = llm
+        self.aux_vision = aux_vision
+        self.skip_aux_cls = skip_aux_cls
+        self.reuse_aux_vision_embedding_layers = reuse_aux_vision_embedding_layers
+        self.lora = lora
+        self.image_size = image_size
+        self.vision_embedding = vision_embedding
+        self.vision_embedding_intermediate_size = vision_embedding_intermediate_size
+        self.patch_size = patch_size
+        self.vision_attention_mask = vision_attention_mask
+        self.rms_norm_eps = rms_norm_eps

generation_config.json

@@ -0,0 +1,14 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.05,
+  "temperature": 0.7,
+  "top_k": 20,
+  "top_p": 0.8,
+  "transformers_version": "4.50.3"
+}

lora.py

@@ -0,0 +1,104 @@
+import torch
+import types
+import math
+from torch import nn
+import torch.nn.functional as F
+
+
+QWEN2_TARGET_MODULES = [
+    "self_attn.q_proj",
+    "self_attn.k_proj",
+    "self_attn.v_proj",
+    "self_attn.o_proj",
+    "mlp.up_proj",
+    "mlp.gate_proj",
+    "mlp.down_proj",
+]
+
+
+class LoRALayer(nn.Linear):
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int,
+        r: int = 1024,
+        **kwargs
+    ):
+        nn.Linear.__init__(self, in_features, out_features)
+        # we elimate lora_alpha here bc we find it unnecessary in VoRA
+        if r < 0:
+            self.forward = self.naive_forward
+        else:
+            self.lora_A = nn.Linear(in_features, r, bias=False)
+            self.lora_B = nn.Linear(r, out_features, bias=False)
+            nn.init.kaiming_uniform_(self.lora_A.weight, a=math.sqrt(5))
+            nn.init.zeros_(self.lora_B.weight)
+
+    def forward(self, x: torch.Tensor):
+        intermediate = F.linear(x, self.weight, bias=self.bias)
+        result = intermediate + self.lora_B(self.lora_A(x))
+        return result
+
+    def naive_forward(self, x: torch.Tensor):
+        return F.linear(x, self.weight, bias=self.bias)
+
+
+def _get_submodules(self, key):
+    parent = self.get_submodule(".".join(key.split(".")[:-1]))
+    target_name = key.split(".")[-1]
+    target = self.get_submodule(key)
+    return parent, target, target_name
+
+
+def _find_and_replace(self, lora_params):
+    target_modules = lora_params["target_modules"]
+
+    for llm_module_name in target_modules:
+        parent, target, target_name = self._get_submodules(llm_module_name)
+        bias = target.bias is not None
+        vora_layer = LoRALayer(
+            target.in_features,
+            target.out_features,
+            bias=bias,
+            **lora_params
+        )
+        self._replace_module(parent, target_name, vora_layer, target)
+
+
+def _replace_module(self, parent_module, child_name, new_module, old_module):
+    setattr(parent_module, child_name, new_module)
+    new_module.weight = old_module.weight
+    if old_module.bias is not None:
+        new_module.bias = old_module.bias
+    if getattr(old_module, "state", None) is not None:
+        new_module.state = old_module.state
+        new_module.to(old_module.weight.device)
+
+
+def apply_lora(llm, lora_params={"layers": "all", "r": 1024, "target_modules": QWEN2_TARGET_MODULES}):
+    llm_num_layers = llm.config.num_hidden_layers
+    total_layers = lora_params.get("layers", "all")
+
+    # -------------------- validation check ---------------------
+    if isinstance(total_layers, str):
+        if total_layers.lower() == "all":
+            total_layers = list(range(llm_num_layers))
+    else:
+        assert isinstance(total_layers, int), "total_layers must be an integer or 'all'"
+        total_layers = list(range(total_layers))
+    # -------------------- validation check ---------------------
+
+    # -------------------- replace llm layers ---------------------
+    for i in total_layers:
+        llm_layer = llm.model.layers[i]
+        llm_layer._get_submodules = types.MethodType(_get_submodules, llm_layer)
+        llm_layer._find_and_replace = types.MethodType(_find_and_replace, llm_layer)
+        llm_layer._replace_module = types.MethodType(_replace_module, llm_layer)
+        llm_layer._find_and_replace(lora_params)
+
+
+if __name__ == "__main__":
+    from transformers import LlamaForCausalLM, CLIPVisionModel, AutoModel
+    llama = LlamaForCausalLM.from_pretrained("/mnt/bn/wh-data/data/models/llama2_7b_hf_chat")
+    apply_lora(llama)
+    print(llama)

model-00001-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fd23c14b1d5b7707a7c23e91f62bd2f8f6889e1886d8d2ad4777be1ba72cd05b
+size 4992767056

model-00002-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:601430b3763dfacfbcce9291ad7672caf4d780231f1d52a631314332762f27dd
+size 4996305072

model-00003-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f431425f27f62b81874df6c0447ebb053ee36099700dd9c9d1390640a770c6cc
+size 4842155392

model-00004-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0c72caac26378342d5f6544ec444b55ff2bc5f971e1ec1df98306b57dc18dfc
+size 4842122976

model-00005-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f227aa97c65250b736bfd9ea5cc3bcf1db287fa2bcbd79b7d110f6368bb56df
+size 4842122968

model-00006-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bd0654e7c43092c37905ffecbb1d70b9e69b8d6c5e305784e4a817120912036
+size 4996305176

model-00007-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ecdf1c467318ad0f64632a60543ab6f6c3201acaa04199f305c8a5fba99e67f
+size 4954233672

model-00008-of-00008.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9d44433adb6162ce553ba38efa185a0a952a947b5ec74ac1c4dfe3cb124e2c8
+size 4891179120

model.safetensors.index.json

@@ -0,0 +1,783 @@
+{
+  "metadata": {
+    "total_size": 39357097984
+  },
+  "weight_map": {
+    "llm.lm_head.weight": "model-00008-of-00008.safetensors",
+    "llm.model.embed_tokens.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.input_layernorm.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.down_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.down_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.down_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.down_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.gate_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.gate_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.gate_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.gate_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.up_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.up_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.up_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.mlp.up_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.post_attention_layernorm.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.k_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.k_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.k_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.k_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.o_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.o_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.o_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.o_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.q_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.q_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.q_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.q_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.v_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.v_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.v_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.0.self_attn.v_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.input_layernorm.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.down_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.down_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.down_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.down_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.gate_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.gate_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.gate_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.gate_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.up_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.up_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.up_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.mlp.up_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.post_attention_layernorm.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.k_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.k_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.k_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.k_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.o_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.o_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.o_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.o_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.q_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.q_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.q_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.q_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.v_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.v_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.v_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.1.self_attn.v_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.10.input_layernorm.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.down_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.down_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.down_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.down_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.gate_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.gate_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.gate_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.gate_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.up_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.up_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.up_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.mlp.up_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.post_attention_layernorm.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.k_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.k_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.k_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.k_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.o_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.o_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.o_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.o_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.q_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.q_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.q_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.q_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.v_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.v_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.v_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.10.self_attn.v_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.input_layernorm.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.down_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.down_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.down_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.down_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.gate_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.gate_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.gate_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.gate_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.up_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.up_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.up_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.mlp.up_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.post_attention_layernorm.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.k_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.k_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.k_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.k_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.o_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.o_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.o_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.o_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.q_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.q_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.q_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.q_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.v_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.v_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.v_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.11.self_attn.v_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.input_layernorm.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.down_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.down_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.down_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.down_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.gate_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.gate_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.gate_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.gate_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.up_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.up_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.up_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.mlp.up_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.post_attention_layernorm.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.k_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.k_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.k_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.k_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.o_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.o_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.o_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.o_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.q_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.q_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.q_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.q_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.v_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.v_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.v_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.12.self_attn.v_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.input_layernorm.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.mlp.down_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.mlp.down_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.mlp.down_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.mlp.down_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.mlp.gate_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.mlp.gate_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.mlp.gate_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.mlp.gate_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.mlp.up_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.mlp.up_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.mlp.up_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.mlp.up_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.post_attention_layernorm.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.k_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.k_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.k_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.k_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.o_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.o_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.o_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.o_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.q_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.q_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.q_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.q_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.v_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.v_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.v_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.13.self_attn.v_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.14.input_layernorm.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.down_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.down_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.down_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.down_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.gate_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.gate_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.gate_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.gate_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.up_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.up_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.up_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.mlp.up_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.post_attention_layernorm.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.k_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.k_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.k_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.k_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.o_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.o_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.o_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.o_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.q_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.q_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.q_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.q_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.v_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.v_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.v_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.14.self_attn.v_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.input_layernorm.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.down_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.down_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.down_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.down_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.gate_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.gate_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.gate_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.gate_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.up_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.up_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.up_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.mlp.up_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.post_attention_layernorm.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.k_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.k_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.k_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.k_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.o_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.o_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.o_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.o_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.q_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.q_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.q_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.q_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.v_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.v_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.v_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.15.self_attn.v_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.input_layernorm.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.down_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.down_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.down_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.down_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.gate_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.gate_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.gate_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.gate_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.up_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.up_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.up_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.mlp.up_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.post_attention_layernorm.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.k_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.k_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.k_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.k_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.o_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.o_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.o_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.o_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.q_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.q_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.q_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.q_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.v_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.v_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.v_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.16.self_attn.v_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.input_layernorm.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.down_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.down_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.down_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.down_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.gate_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.gate_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.gate_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.gate_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.up_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.up_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.up_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.mlp.up_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.post_attention_layernorm.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.k_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.k_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.k_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.k_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.o_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.o_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.o_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.o_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.q_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.q_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.q_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.q_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.v_proj.bias": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.v_proj.lora_A.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.v_proj.lora_B.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.17.self_attn.v_proj.weight": "model-00005-of-00008.safetensors",
+    "llm.model.layers.18.input_layernorm.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.down_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.down_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.down_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.down_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.gate_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.gate_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.gate_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.gate_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.up_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.up_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.up_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.mlp.up_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.post_attention_layernorm.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.k_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.k_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.k_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.k_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.o_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.o_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.o_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.o_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.q_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.q_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.q_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.q_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.v_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.v_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.v_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.18.self_attn.v_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.input_layernorm.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.down_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.down_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.down_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.down_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.gate_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.gate_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.gate_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.gate_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.up_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.up_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.up_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.mlp.up_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.post_attention_layernorm.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.k_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.k_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.k_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.k_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.o_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.o_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.o_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.o_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.q_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.q_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.q_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.q_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.v_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.v_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.v_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.19.self_attn.v_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.2.input_layernorm.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.down_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.down_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.down_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.down_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.gate_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.gate_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.gate_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.gate_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.up_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.up_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.up_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.mlp.up_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.post_attention_layernorm.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.k_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.k_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.k_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.k_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.o_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.o_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.o_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.o_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.q_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.q_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.q_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.q_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.v_proj.bias": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.v_proj.lora_A.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.v_proj.lora_B.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.2.self_attn.v_proj.weight": "model-00001-of-00008.safetensors",
+    "llm.model.layers.20.input_layernorm.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.down_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.down_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.down_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.down_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.gate_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.gate_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.gate_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.gate_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.up_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.up_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.up_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.mlp.up_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.post_attention_layernorm.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.k_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.k_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.k_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.k_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.o_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.o_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.o_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.o_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.q_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.q_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.q_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.q_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.v_proj.bias": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.v_proj.lora_A.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.v_proj.lora_B.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.20.self_attn.v_proj.weight": "model-00006-of-00008.safetensors",
+    "llm.model.layers.21.input_layernorm.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.down_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.down_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.down_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.down_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.gate_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.gate_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.gate_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.gate_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.up_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.up_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.up_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.mlp.up_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.post_attention_layernorm.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.k_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.k_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.k_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.o_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.o_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.o_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.q_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.q_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.q_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.v_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.v_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.v_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.21.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.input_layernorm.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.down_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.down_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.down_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.down_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.gate_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.gate_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.gate_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.gate_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.up_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.up_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.up_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.mlp.up_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.post_attention_layernorm.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.k_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.k_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.k_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.o_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.o_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.o_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.q_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.q_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.q_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.v_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.v_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.v_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.22.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.input_layernorm.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.down_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.down_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.down_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.down_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.gate_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.gate_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.gate_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.gate_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.up_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.up_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.up_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.mlp.up_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.post_attention_layernorm.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.k_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.k_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.k_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.o_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.o_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.o_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.q_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.q_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.q_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.v_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.v_proj.lora_A.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.v_proj.lora_B.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.23.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.input_layernorm.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.mlp.down_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.mlp.gate_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.mlp.up_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.post_attention_layernorm.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.self_attn.k_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.self_attn.q_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.self_attn.v_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.24.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.25.input_layernorm.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.25.mlp.down_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.25.mlp.gate_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.25.mlp.up_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.25.post_attention_layernorm.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.25.self_attn.k_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.25.self_attn.k_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.25.self_attn.o_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.25.self_attn.q_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.25.self_attn.q_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.25.self_attn.v_proj.bias": "model-00007-of-00008.safetensors",
+    "llm.model.layers.25.self_attn.v_proj.weight": "model-00007-of-00008.safetensors",
+    "llm.model.layers.26.input_layernorm.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.mlp.down_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.mlp.gate_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.mlp.up_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.post_attention_layernorm.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.self_attn.k_proj.bias": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.self_attn.k_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.self_attn.o_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.self_attn.q_proj.bias": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.self_attn.q_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.self_attn.v_proj.bias": "model-00008-of-00008.safetensors",
+    "llm.model.layers.26.self_attn.v_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.input_layernorm.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.mlp.down_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.mlp.gate_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.mlp.up_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.post_attention_layernorm.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.self_attn.k_proj.bias": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.self_attn.k_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.self_attn.o_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.self_attn.q_proj.bias": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.self_attn.q_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.self_attn.v_proj.bias": "model-00008-of-00008.safetensors",
+    "llm.model.layers.27.self_attn.v_proj.weight": "model-00008-of-00008.safetensors",
+    "llm.model.layers.3.input_layernorm.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.down_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.down_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.down_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.down_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.gate_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.gate_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.gate_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.gate_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.up_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.up_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.up_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.mlp.up_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.post_attention_layernorm.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.k_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.k_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.k_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.k_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.o_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.o_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.o_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.o_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.q_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.q_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.q_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.q_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.v_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.v_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.v_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.3.self_attn.v_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.input_layernorm.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.down_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.down_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.down_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.down_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.gate_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.gate_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.gate_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.gate_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.up_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.up_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.up_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.mlp.up_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.post_attention_layernorm.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.k_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.k_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.k_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.k_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.o_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.o_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.o_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.o_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.q_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.q_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.q_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.q_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.v_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.v_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.v_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.4.self_attn.v_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.input_layernorm.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.down_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.down_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.down_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.down_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.gate_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.gate_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.gate_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.gate_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.up_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.up_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.up_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.mlp.up_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.post_attention_layernorm.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.k_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.k_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.k_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.k_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.o_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.o_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.o_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.o_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.q_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.q_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.q_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.q_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.v_proj.bias": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.v_proj.lora_A.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.v_proj.lora_B.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.5.self_attn.v_proj.weight": "model-00002-of-00008.safetensors",
+    "llm.model.layers.6.input_layernorm.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.down_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.down_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.down_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.down_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.gate_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.gate_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.gate_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.gate_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.up_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.up_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.up_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.mlp.up_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.post_attention_layernorm.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.k_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.k_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.k_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.k_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.o_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.o_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.o_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.o_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.q_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.q_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.q_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.q_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.v_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.v_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.v_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.6.self_attn.v_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.input_layernorm.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.down_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.down_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.down_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.down_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.gate_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.gate_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.gate_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.gate_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.up_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.up_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.up_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.mlp.up_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.post_attention_layernorm.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.k_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.k_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.k_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.k_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.o_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.o_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.o_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.o_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.q_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.q_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.q_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.q_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.v_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.v_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.v_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.7.self_attn.v_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.input_layernorm.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.down_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.down_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.down_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.down_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.gate_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.gate_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.gate_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.gate_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.up_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.up_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.up_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.mlp.up_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.post_attention_layernorm.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.k_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.k_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.k_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.k_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.o_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.o_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.o_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.o_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.q_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.q_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.q_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.q_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.v_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.v_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.v_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.8.self_attn.v_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.input_layernorm.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.9.mlp.down_proj.bias": "model-00004-of-00008.safetensors",
+    "llm.model.layers.9.mlp.down_proj.lora_A.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.9.mlp.down_proj.lora_B.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.9.mlp.down_proj.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.9.mlp.gate_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.mlp.gate_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.mlp.gate_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.mlp.gate_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.mlp.up_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.mlp.up_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.mlp.up_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.mlp.up_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.post_attention_layernorm.weight": "model-00004-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.k_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.k_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.k_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.k_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.o_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.o_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.o_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.o_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.q_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.q_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.q_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.q_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.v_proj.bias": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.v_proj.lora_A.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.v_proj.lora_B.weight": "model-00003-of-00008.safetensors",
+    "llm.model.layers.9.self_attn.v_proj.weight": "model-00003-of-00008.safetensors",
+    "llm.model.norm.weight": "model-00008-of-00008.safetensors",
+    "vision_embedding.out_proj.weight": "model-00008-of-00008.safetensors",
+    "vision_embedding.patchifier.norm.weight": "model-00008-of-00008.safetensors",
+    "vision_embedding.patchifier.proj.bias": "model-00008-of-00008.safetensors",
+    "vision_embedding.patchifier.proj.weight": "model-00008-of-00008.safetensors",
+    "vision_embedding.pos_embed": "model-00008-of-00008.safetensors"
+  }
+}

modeling_vora.py

@@ -0,0 +1,290 @@
+import torch
+import torch.distributed as dist
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    PreTrainedModel,
+    PretrainedConfig,
+)
+
+from .attention_mask import make_mask
+from .configuration_vora import VoRAConfig
+from .vision_embedding import *  # hacking, let transformers find vision_embedding
+from . import vision_embedding as VB
+from .lora import apply_lora
+from .vora_generation_utils import (
+    VoraGenerationMixin,
+    custom_prepare_4d_causal_attention_mask_with_cache_position,
+)
+
+try:
+    from utils import logging
+except:
+    from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class VoRAForCausalLM(PreTrainedModel):
+    config_class = VoRAConfig
+    _auto_class = 'AutoModelForCausalLM'
+    supports_gradient_checkpointing = True
+    supports_report_metrics: bool = True
+
+    def __init__(self, config: PretrainedConfig = VoRAConfig()):
+        super().__init__(config)
+        self.config = config
+        # -------------- Setup LLM ---------------------
+        self.llm = AutoModelForCausalLM.from_pretrained(config.llm)
+        self.tokenizer = AutoTokenizer.from_pretrained(config.llm)
+        self.llm.__class__ = type(self.llm.__class__.__name__, (self.llm.__class__, VoraGenerationMixin), {})
+        self.llm.model._prepare_4d_causal_attention_mask_with_cache_position = staticmethod(custom_prepare_4d_causal_attention_mask_with_cache_position)
+
+        self.config.update(self.llm.config.to_dict())
+
+        # -------------- Setup LoRA -------------------
+        if config.lora:
+            for _, param in self.llm.named_parameters():
+                param.requires_grad = False
+            apply_lora(self.llm, config.lora)
+        # ----------------------------------------------
+
+        # ------------ Setup Vision Embedding ----------
+        self.vision_embedding = getattr(VB, config.vision_embedding)(self.config)  # setup after llm so that we know the hiddensize
+        # ----------------------------------------------
+
+        # ------------- Setup Aux Vision ---------------
+        self.enable_aux_vision = False
+        if config.aux_vision:
+            from .aux_vision import AuxVision
+            self.enable_aux_vision = True
+            self.aux_vision = AuxVision(self.config)
+            if config.reuse_aux_vision_embedding_layers:
+                weights = getattr(self.aux_vision.aux_model, config.reuse_aux_vision_embedding_layers).state_dict()
+                msg = self.vision_embedding.load_state_dict(weights, strict=False)
+                logger.info(f"Loaded aux vision weights: {msg}")
+        # ----------------------------------------------
+        # print trainable prameters and total parameters so that we can check if we are loading the correct model
+        #logger.info("Trainable parameters:")
+        print("Trainable parameters:")
+        for name, param in self.named_parameters():
+            if param.requires_grad:
+                print(f"{name}: {param.numel()}")
+        print(f"Total parameters: {sum(p.numel() for p in self.parameters())}")
+
+    def detach_and_gather_loss(self, loss, dtype, device):
+        if not dist.is_initialized():
+            return loss.item()
+        gathered_loss = [torch.tensor(0.0, dtype=loss.dtype).to(device) for _ in range(dist.get_world_size())]
+        dist.all_gather(gathered_loss, loss.detach().clone())
+        avg_gathered_loss = torch.mean(torch.stack(gathered_loss))
+        return avg_gathered_loss.item()
+
+    def _encode_vision(self, images, n_frames):
+        # TODO: we need a more elegant way here to deal with mixed image and pure text training
+        if images.size(0) > 0:
+            vision_embeds = self.vision_embedding(images)
+        else:
+            # FIXME: hacking for deepspeed training
+            # we feed a dummy image tensor (1, 3, H, W) into vision_encoder when training a pure-text batch
+            images = images.new_zeros((1, *images.shape[1:]))
+            vision_embeds = self.vision_embedding(images)[0:0]
+        vision_embeds = vision_embeds.split(n_frames, dim=0)
+        attention_mask = [torch.ones(feature.size()[:-1], dtype=torch.long).to(feature.device) for feature in vision_embeds]
+        vision_targets = [torch.ones(feature.size(), dtype=torch.long).to(feature.device).fill_(-100) for feature in attention_mask]
+
+        image_shapes = images.shape[-2:]
+
+        return vision_embeds, attention_mask, vision_targets, image_shapes
+
+    def _concat_embedding(self, vision_encode_out, batch, vision_placeholder_index, left_padding=False):
+        """ concat vision and text
+        """
+
+        vision_embeds, vision_atts, vision_targets, _ = vision_encode_out
+
+        input_embeds = []
+        attention_mask = []
+        targets = []
+        vision_mask = []  # set vision token as 1, text token as 0
+
+        for cur_batch_idx, cur_input_ids in enumerate(batch["input_ids"]):
+            cur_vision_embeds = vision_embeds[cur_batch_idx]
+            cur_vision_attn = vision_atts[cur_batch_idx]
+            cur_vision_targets = vision_targets[cur_batch_idx]
+            cur_attn_masks = batch["attention_mask"][cur_batch_idx]
+
+            image_token_indices = torch.where(cur_input_ids == vision_placeholder_index)[0]
+            cur_image_num = len(image_token_indices)
+            image_token_indices = list(image_token_indices) + [cur_input_ids.shape[0]]
+
+            cur_input_embeds = []
+            cur_attention_mask = []
+            cur_target = []
+            cur_vision_mask = []
+
+            # convert text before 1st <image> to embedding
+            image_token_index = image_token_indices[0]
+
+            cur_input_embeds.append(
+                self.llm.get_input_embeddings()(cur_input_ids[:image_token_index]),
+            )
+            cur_attention_mask.append(
+                cur_attn_masks[:image_token_index],
+            )
+            cur_vision_mask.append(
+                torch.zeros_like(cur_attn_masks[:image_token_index]).to(cur_attn_masks.device),
+            )
+            if "labels" in batch:
+                cur_target.append(
+                    batch["labels"][cur_batch_idx, :image_token_index],
+                )
+
+            if batch.get("vison_placeholder_mode", 0) == 1:
+                assert cur_image_num <= 1, "multiple video input is not supported"
+                cur_vision_embeds = cur_vision_embeds.unsqueeze(0)
+                cur_vision_attn = cur_vision_attn.unsqueeze(0)
+                cur_vision_targets = cur_vision_targets.unsqueeze(0)
+            assert cur_image_num == len(cur_vision_embeds), \
+                f"Size mismatch! cur_image_num: {cur_image_num}, len(cur_vision_embeds): {len(cur_vision_embeds)} {len(cur_vision_embeds)} \
+                    in {batch['prompt'][cur_batch_idx]} & {batch['gt'][cur_batch_idx]} & {batch['input_ids'][cur_batch_idx]}"
+            # convert each <image> xxx group into embedding
+            text_embedding = self.llm.get_input_embeddings()(cur_input_ids.relu())
+            for i in range(0, cur_image_num):
+                image_token_index = image_token_indices[i]
+                cur_input_embeds.extend([
+                    cur_vision_embeds[i],
+                    text_embedding[image_token_index+1:image_token_indices[i+1]]
+                ])
+                cur_attention_mask.extend([
+                    cur_vision_attn[i],
+                    cur_attn_masks[image_token_index+1:image_token_indices[i+1]]
+                ])
+                cur_vision_mask.extend([
+                    torch.ones_like(cur_vision_attn[i]).to(cur_vision_attn[i].device),
+                    torch.zeros_like(cur_attn_masks[image_token_index+1:image_token_indices[i+1]]).to(cur_vision_attn[i].device),
+                ])
+                if "labels" in batch:
+                    cur_target.extend([
+                        cur_vision_targets[i],
+                        batch["labels"][cur_batch_idx, image_token_index+1:image_token_indices[i+1]],
+                    ])
+
+            input_embeds.append(torch.cat(cur_input_embeds))
+            attention_mask.append(torch.cat(cur_attention_mask))
+            vision_mask.append(torch.cat(cur_vision_mask))
+            if "labels" in batch:
+                targets.append(torch.cat(cur_target))
+
+        # padding
+        n_tokens = [embed.shape[0] for embed in input_embeds]
+
+        max_token = max(n_tokens)
+
+        for i in range(len(input_embeds)):
+            if max_token > n_tokens[i]:
+                self.pad_id = self.tokenizer.pad_token_id or self.tokenizer.eos_token_id
+                pad_token = torch.tensor([self.pad_id] * (max_token - n_tokens[i]))
+                pad_embedding = self.llm.get_input_embeddings()(pad_token.to(batch["attention_mask"][i].device))
+                pad_attention = torch.zeros(pad_embedding.shape[0], dtype=torch.long).to(batch["attention_mask"][i].device)
+                pad_targets = torch.ones(pad_attention.size(), dtype=torch.long).to(batch["attention_mask"][i].device).fill_(-100)
+
+                if left_padding:
+                    input_embeds[i] = torch.cat([pad_embedding, input_embeds[i]])
+                    attention_mask[i] = torch.cat([pad_attention, attention_mask[i]])
+                    vision_mask[i] = torch.cat([pad_attention, vision_mask[i]])
+                    if "labels" in batch:
+                        targets[i] = torch.cat([pad_targets, targets[i]])
+                else:
+                    input_embeds[i] = torch.cat([input_embeds[i], pad_embedding])
+                    attention_mask[i] = torch.cat([attention_mask[i], pad_attention])
+                    vision_mask[i] = torch.cat([vision_mask[i], pad_attention])
+                    if "labels" in batch:
+                        targets[i] = torch.cat([targets[i], pad_targets])
+
+        inputs_embeds = torch.stack(input_embeds, dim=0).type(self.llm.dtype)
+        attention_mask = torch.stack(attention_mask, dim=0)
+        vision_mask = torch.stack(vision_mask, dim=0).to(attention_mask.device)
+
+        if len(targets) > 0:
+            targets = torch.stack(targets, dim=0)
+
+        attention_mask = make_mask(
+            attention_mask,
+            mode=self.config.vision_attention_mask,
+            vision_mask=vision_mask,
+            dtype=inputs_embeds.dtype
+        )
+
+        return inputs_embeds, attention_mask, targets, vision_mask
+
+    def forward(self, **batch):
+        # -------------- Vision/Text Embedding ----------
+        vision_placeholder_index = batch.pop("vision_placeholder_index")
+        images, n_frames = batch["frames"], batch["n_frames"]
+        vision_encode_out = self._encode_vision(images, n_frames)
+        inputs_embeds, attention_mask, targets, vision_mask = self._concat_embedding(
+            vision_encode_out, batch, vision_placeholder_index)
+        # -----------------------------------------------
+
+        outputs = self.llm(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            labels=targets,
+            return_dict=True,
+            output_hidden_states=True,
+        )
+
+        llm_loss = outputs.loss
+        device = llm_loss.device
+        dtype = llm_loss.dtype
+
+        metrics = {}
+
+        metrics["llm_loss"] = self.detach_and_gather_loss(llm_loss, dtype, device)
+        if self.enable_aux_vision:
+            if images.size(0) > 0:
+                aux_losses = self.aux_vision(images, outputs.hidden_states, vision_mask)
+            else:
+                # FIXME: hacking for deepspeed training
+                aux_losses = {key: torch.tensor(0., dtype=dtype).to(device) for key in self.aux_vision.loss_keys}
+
+            aux_loss = torch.tensor(0., dtype=dtype).to(device)
+            n_aux = 0
+            for _aux_key, _aux_loss in aux_losses.items():
+                aux_loss += _aux_loss
+                n_aux += 1
+                metrics[_aux_key] = self.detach_and_gather_loss(_aux_loss, dtype, device)
+            aux_loss /= n_aux
+
+            outputs.loss = aux_loss + llm_loss
+        metrics["total_loss"] = self.detach_and_gather_loss(outputs.loss, dtype, device)
+        self.report_metrics(**metrics)
+
+        return outputs
+
+    def generate(self, batch, **generate_params):
+
+        with torch.amp.autocast(
+            enabled=(self.device != torch.device("cpu")),
+            device_type=self.device.type,
+        ):
+            # get vision token
+            vision_placeholder_index = batch.pop("vision_placeholder_index")
+
+            # get vision features
+            images, n_frames = batch["frames"], batch["n_frames"]
+            vision_encode_out = self._encode_vision(images, n_frames)
+
+            inputs_embeds, attention_mask, _, _ = self._concat_embedding(
+                vision_encode_out, batch, vision_placeholder_index, left_padding=False)
+
+        outputs = self.llm.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            output_attentions=True,
+            **generate_params
+        )
+
+        return outputs

vision_embedding.py

@@ -0,0 +1,70 @@
+import torch
+import torch.nn as nn
+
+from .configuration_vora import VoRAConfig
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+
+class AIMv2PatchEmbed(nn.Module):
+    def __init__(self, config: VoRAConfig):
+        super().__init__()
+        self.proj = nn.Conv2d(
+            3,
+            config.vision_embedding_intermediate_size,
+            kernel_size=(config.patch_size, config.patch_size),
+            stride=(config.patch_size, config.patch_size),
+        )
+        self.norm = RMSNorm(config.vision_embedding_intermediate_size, eps=config.rms_norm_eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x).flatten(2).transpose(1, 2)
+        x = self.norm(x)
+        return x
+
+
+class AIMv2Embedding(nn.Module):
+    def __init__(self,
+        config: VoRAConfig = None,
+    ):
+        super().__init__()
+        hidden_size = config.hidden_size
+        num_patches = (config.image_size // config.patch_size) ** 2
+        self.config = config
+
+        self.patchifier = AIMv2PatchEmbed(config)
+        self.pos_embed = nn.Parameter(torch.zeros((1, num_patches, config.vision_embedding_intermediate_size)))
+        self.out_proj = nn.Linear(config.vision_embedding_intermediate_size, hidden_size, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, C, H, W = x.shape
+        h_token = H // self.config.patch_size
+        w_token = W // self.config.patch_size
+        tokens = self.patchifier(x)
+        _, N, _ = tokens.shape
+        pos_embed = self.pos_embed.to(tokens.device)
+
+        if N <= pos_embed.size(1):
+            tokens = tokens + pos_embed[:, :N]
+        else:
+            pos_embed = pos_embed.view(1, int(pos_embed.size(1)**0.5), int(pos_embed.size(1)**0.5), -1).permute(0, 3, 1, 2)
+            pos_embed = nn.functional.interpolate(pos_embed, size=(h_token, w_token), mode='bilinear', align_corners=False).permute(0, 2, 3, 1)
+            pos_embed = pos_embed.view(1, N, pos_embed.size(-1))
+            tokens = tokens + pos_embed
+
+        return self.out_proj(tokens)

vora_generation_utils.py

@@ -0,0 +1,101 @@
+from typing import Any, Dict, Optional
+
+import torch
+from transformers import GenerationMixin
+from transformers.cache_utils import Cache
+from transformers.utils import ModelOutput
+
+
+class VoraGenerationMixin(GenerationMixin):
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        past_key_values: Optional[Cache] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
+    ):
+        if attention_mask is not None and attention_mask.ndim == 4:
+            attention_mask_2d = (attention_mask[:, 0, :, :] == 0).any(dim=1).long().to(attention_mask.device)
+            model_input = super().prepare_inputs_for_generation(
+                input_ids,
+                past_key_values=past_key_values,
+                attention_mask=attention_mask_2d,
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                **kwargs,
+            )
+            model_input['attention_mask'] = attention_mask
+            return model_input
+        else:
+            return super().prepare_inputs_for_generation(
+                input_ids,
+                past_key_values=past_key_values,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                **kwargs,
+            )
+
+    def _update_model_kwargs_for_generation(
+        self,
+        outputs: ModelOutput,
+        model_kwargs: Dict[str, Any],
+        is_encoder_decoder: bool = False,
+        num_new_tokens: int = 1,
+    ) -> Dict[str, Any]:
+        if "attention_mask" in model_kwargs and model_kwargs["attention_mask"].ndim == 4:
+            attention_mask = model_kwargs.pop("attention_mask")
+            model_kwargs = super()._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=is_encoder_decoder, num_new_tokens=num_new_tokens
+            )
+            bs, _, seq_len, tgt_len = attention_mask.shape
+            dtype = attention_mask.dtype
+            min_dtype = torch.finfo(dtype).min
+            new_col = attention_mask.new_zeros((bs, 1, seq_len, 1)).fill_(min_dtype)
+            new_row = attention_mask.new_zeros((bs, 1, 1, tgt_len + 1))
+            model_kwargs["attention_mask"] = torch.cat([
+                torch.cat([attention_mask, new_col], dim=-1),
+                new_row
+            ], dim=2)
+            return model_kwargs
+        else:
+            return super()._update_model_kwargs_for_generation(
+                outputs, model_kwargs, is_encoder_decoder=is_encoder_decoder, num_new_tokens=num_new_tokens
+            )
+
+
+def custom_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,
+):
+    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[:, :, -sequence_length:, -target_length:]
+    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