Upload MiniCPMV

README.md

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

config.json

@@ -0,0 +1,54 @@
+{
+  "_name_or_path": "full_model",
+  "architectures": [
+    "MiniCPMV"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_minicpm.MiniCPMVConfig",
+    "AutoModel": "openbmb/MiniCPM-Llama3-V-2_5--modeling_minicpmv.MiniCPMV",
+    "AutoModelForCausalLM": "modeling_minicpmv.MiniCPMV"
+  },
+  "batch_vision_input": true,
+  "bos_token_id": 128000,
+  "drop_vision_last_layer": false,
+  "eos_token_id": 128001,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "image_size": 448,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "max_position_embeddings": 8192,
+  "mlp_bias": false,
+  "mm_use_im_start_end": true,
+  "model_type": "minicpmv",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 8,
+  "patch_size": 14,
+  "pretraining_tp": 1,
+  "query_num": 96,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 500000.0,
+  "slice_config": {
+    "max_slice_nums": 9,
+    "model_type": "minicpmv"
+  },
+  "slice_mode": true,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.41.2",
+  "use_cache": true,
+  "vision_config": {
+    "hidden_size": 1152,
+    "image_size": 980,
+    "intermediate_size": 4304,
+    "model_type": "idefics2",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "patch_size": 14
+  },
+  "vocab_size": 128256
+}

configuration_minicpm.py

@@ -0,0 +1,113 @@
+# coding=utf-8
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" MiniCPM model configuration"""
+import os
+from typing import Union
+
+from transformers.utils import logging
+from transformers import LlamaConfig, PretrainedConfig
+from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionConfig
+
+logger = logging.get_logger(__name__)
+
+
+class MiniCPMVSliceConfig(PretrainedConfig):
+    model_type = "minicpmv"
+
+    def __init__(
+        self,
+        patch_size=14,
+        max_slice_nums=9,
+        scale_resolution=448,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.patch_size = patch_size
+        self.max_slice_nums = max_slice_nums
+        self.scale_resolution = scale_resolution
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        if config_dict.get("model_type") == "minicpmv":
+            config_dict = config_dict["slice_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+
+class MiniCPMVConfig(LlamaConfig):
+    model_type = "minicpmv"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    default_vision_config = {
+        "hidden_size": 1152,
+        "image_size": 980,
+        "intermediate_size": 4304,
+        "model_type": "idefics2",
+        "num_attention_heads": 16,
+        "num_hidden_layers": 27,
+        "patch_size": 14,
+    }
+
+    def __init__(
+        self,
+        use_cache=True,
+        query_num=64,
+        image_size=448,
+        drop_vision_last_layer=True,
+        batch_vision_input=True,
+        slice_config=None,
+        vision_config=None,
+        **kwargs,
+    ):
+        self.use_cache = use_cache
+        self.query_num = query_num
+        self.image_size = image_size
+        self.drop_vision_last_layer = drop_vision_last_layer
+        self.batch_vision_input = batch_vision_input
+
+        if slice_config is None:
+            self.slice_config = MiniCPMVSliceConfig(max_slice_nums=1)
+        else:
+            self.slice_config = MiniCPMVSliceConfig(**slice_config)
+        self.slice_mode = True
+
+        # same as HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
+        if vision_config is None:
+            self.vision_config = Idefics2VisionConfig(**self.default_vision_config)
+            logger.info("vision_config is None, using default vision config")
+        elif isinstance(vision_config, dict):
+            self.vision_config = Idefics2VisionConfig(**vision_config)
+        elif isinstance(vision_config, Idefics2VisionConfig):
+            self.vision_config = vision_config
+
+        self.patch_size = self.vision_config.patch_size
+
+        super().__init__(**kwargs)

generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 128000,
+  "eos_token_id": 128001,
+  "transformers_version": "4.41.2"
+}

model-00001-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2226d735f021174620a0119802e7b8fa5a9398db98d1ae2439cdb39365d8491a
+size 4886466296

model-00002-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:002d366e0660bc0b27ce5154c85a8d10a50772886dfb9e0a378584df5aa896ab
+size 4832007640

model-00003-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30b338ce4bfbd431ba5a1c2d9227e412cbcb921fed063a37d6495bb665b9898e
+size 4999813320

model-00004-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb8b635cfa530bf039e6256665550fedb777d14ed4d47e6fb895453bcd485efb
+size 4999813336

model-00005-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e82f796fa848b135accfb227dc7074aa75608dad25b6d0bfc51191d9ff26b61
+size 4832007688

model-00006-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2f6eb553265900f7c28eeca523e8dbc3440d8823024967a15c1f0926aaa5d07
+size 4999813336

model-00007-of-00007.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0564d9fd568e043bbeb9f3f1b1951b194e5a54d5aa5209d8791922f89ba2589f
+size 4598534152

model.safetensors.index.json

@@ -0,0 +1,748 @@
+{
+  "metadata": {
+    "total_size": 34148369344
+  },
+  "weight_map": {
+    "llm.lm_head.weight": "model-00007-of-00007.safetensors",
+    "llm.model.embed_tokens.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.input_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.down_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.gate_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.mlp.up_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.post_attention_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.k_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.o_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.q_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.0.self_attn.v_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.input_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.down_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.gate_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.mlp.up_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.post_attention_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.k_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.o_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.q_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.1.self_attn.v_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.10.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.10.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.11.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.12.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.13.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.14.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.14.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.14.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.15.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.15.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.16.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.17.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.18.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.input_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.mlp.down_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.mlp.gate_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.mlp.up_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.post_attention_layernorm.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.19.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.2.input_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.down_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.gate_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.mlp.up_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.post_attention_layernorm.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.k_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.o_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.q_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.2.self_attn.v_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.20.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.k_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.o_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.q_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.20.self_attn.v_proj.weight": "model-00004-of-00007.safetensors",
+    "llm.model.layers.21.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.21.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.22.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.23.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.input_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.mlp.down_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.post_attention_layernorm.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.24.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.25.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.mlp.gate_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.25.mlp.up_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.25.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.k_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.o_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.q_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.25.self_attn.v_proj.weight": "model-00005-of-00007.safetensors",
+    "llm.model.layers.26.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.26.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.27.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.28.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.29.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.3.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.k_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.o_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.q_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.3.self_attn.v_proj.weight": "model-00001-of-00007.safetensors",
+    "llm.model.layers.30.input_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.30.mlp.down_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.30.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.30.mlp.up_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.30.post_attention_layernorm.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.30.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.30.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.30.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.30.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.31.input_layernorm.weight": "model-00007-of-00007.safetensors",
+    "llm.model.layers.31.mlp.down_proj.weight": "model-00007-of-00007.safetensors",
+    "llm.model.layers.31.mlp.gate_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.31.mlp.up_proj.weight": "model-00007-of-00007.safetensors",
+    "llm.model.layers.31.post_attention_layernorm.weight": "model-00007-of-00007.safetensors",
+    "llm.model.layers.31.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.31.self_attn.o_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.31.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.31.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
+    "llm.model.layers.4.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.4.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.5.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.6.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.input_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.down_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.post_attention_layernorm.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.7.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.mlp.gate_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.mlp.up_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.k_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.o_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.q_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.8.self_attn.v_proj.weight": "model-00002-of-00007.safetensors",
+    "llm.model.layers.9.input_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.down_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.gate_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.mlp.up_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.post_attention_layernorm.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.k_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.o_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.q_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.layers.9.self_attn.v_proj.weight": "model-00003-of-00007.safetensors",
+    "llm.model.norm.weight": "model-00007-of-00007.safetensors",
+    "resampler.attn.in_proj_bias": "model-00007-of-00007.safetensors",
+    "resampler.attn.in_proj_weight": "model-00007-of-00007.safetensors",
+    "resampler.attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "resampler.attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "resampler.kv_proj.weight": "model-00007-of-00007.safetensors",
+    "resampler.ln_kv.bias": "model-00007-of-00007.safetensors",
+    "resampler.ln_kv.weight": "model-00007-of-00007.safetensors",
+    "resampler.ln_post.bias": "model-00007-of-00007.safetensors",
+    "resampler.ln_post.weight": "model-00007-of-00007.safetensors",
+    "resampler.ln_q.bias": "model-00007-of-00007.safetensors",
+    "resampler.ln_q.weight": "model-00007-of-00007.safetensors",
+    "resampler.proj": "model-00007-of-00007.safetensors",
+    "resampler.query": "model-00007-of-00007.safetensors",
+    "vpm.embeddings.patch_embedding.bias": "model-00007-of-00007.safetensors",
+    "vpm.embeddings.patch_embedding.weight": "model-00007-of-00007.safetensors",
+    "vpm.embeddings.position_embedding.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.0.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.1.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.10.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.11.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.12.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.13.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.14.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.15.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.16.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.17.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.18.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.19.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.2.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.20.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.21.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.22.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.23.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.24.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.25.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.26.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.3.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.4.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.5.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.6.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.7.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.8.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.layer_norm1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.layer_norm1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.layer_norm2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.layer_norm2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.mlp.fc1.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.mlp.fc1.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.mlp.fc2.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.mlp.fc2.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
+    "vpm.encoder.layers.9.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
+    "vpm.post_layernorm.bias": "model-00007-of-00007.safetensors",
+    "vpm.post_layernorm.weight": "model-00007-of-00007.safetensors"
+  }
+}

modeling_minicpmv.py

@@ -0,0 +1,702 @@
+import math
+from typing import List, Optional
+import json
+import torch
+import torchvision
+from threading import Thread
+from copy import deepcopy
+from PIL import Image
+from torchvision import transforms
+from transformers import LlamaTokenizer, LlamaPreTrainedModel, LlamaForCausalLM, AutoModel, PreTrainedTokenizerFast, TextIteratorStreamer
+from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer
+
+from .configuration_minicpm import MiniCPMVConfig
+from .resampler import Resampler
+
+IMAGENET_INCEPTION_MEAN = (0.5, 0.5, 0.5) # timm.data.IMAGENET_INCEPTION_MEAN
+IMAGENET_INCEPTION_STD = (0.5, 0.5, 0.5)  # timm.data.IMAGENET_INCEPTION_STD
+
+class MiniCPMVPreTrainedModel(LlamaPreTrainedModel):
+    config_class = MiniCPMVConfig
+
+
+class MiniCPMV(MiniCPMVPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.llm = LlamaForCausalLM(config)
+        self.vpm = self.init_vision_module()
+        self.vision_dim = self.vpm.embed_dim
+        self.embed_dim = self.llm.config.hidden_size
+        self.resampler = self.init_resampler(self.embed_dim, self.vision_dim)
+        self.transform = self.init_transform()
+
+    def init_vision_module(self):
+        # same as HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
+        model = Idefics2VisionTransformer(self.config.vision_config)
+        if self.config.drop_vision_last_layer:
+            model.encoder.layers = model.encoder.layers[:-1]
+
+        setattr(model, 'embed_dim', model.embeddings.embed_dim)
+        setattr(model, 'patch_size', model.embeddings.patch_size)
+
+        return model
+
+    def init_resampler(self, embed_dim, vision_dim):
+        return Resampler(
+            num_queries=self.config.query_num,
+            embed_dim=embed_dim,
+            num_heads=embed_dim // 128,
+            kv_dim=vision_dim,
+            adaptive=True
+        )
+
+    def init_transform(self):
+        return transforms.Compose(
+            [
+                transforms.ToTensor(),
+                transforms.Normalize(
+                    mean=IMAGENET_INCEPTION_MEAN, std=IMAGENET_INCEPTION_STD
+                ),
+            ]
+        )
+
+    def get_input_embeddings(self):
+        return self.llm.get_input_embeddings()
+
+    def set_input_embeddings(self, value):
+        self.llm.embed_tokens = value
+
+    def get_vllm_embedding(self, data):
+        if 'vision_hidden_states' not in data:
+            dtype = self.vpm.embeddings.position_embedding.weight.dtype
+            device = self.vpm.embeddings.position_embedding.weight.device
+            tgt_sizes = data['tgt_sizes']
+            pixel_values_list = data['pixel_values']
+            vision_hidden_states = []
+            all_pixel_values = []
+            img_cnt = []
+            for pixel_values in pixel_values_list:
+                img_cnt.append(len(pixel_values))
+                all_pixel_values.extend([i.flatten(end_dim=1).permute(1, 0) for i in pixel_values])
+
+            # exist image
+            if all_pixel_values:
+                tgt_sizes = torch.vstack(tgt_sizes).type(torch.int32)
+
+                if self.config.batch_vision_input:
+                    max_patches = torch.max(tgt_sizes[:, 0] * tgt_sizes[:, 1])
+
+                    all_pixel_values = torch.nn.utils.rnn.pad_sequence(all_pixel_values, batch_first=True,
+                                                                       padding_value=0.0)
+                    B, L, _ = all_pixel_values.shape
+                    all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
+
+                    patch_attn_mask = torch.zeros((B, 1, max_patches), dtype=torch.bool, device=device)
+                    for i in range(B):
+                        patch_attn_mask[i, :tgt_sizes[i][0] * tgt_sizes[i][1]] = True
+
+                    vision_embedding = self.vpm(all_pixel_values.type(dtype), patch_attention_mask=patch_attn_mask).last_hidden_state
+                    vision_embedding = self.resampler(vision_embedding, tgt_sizes)
+                else:
+                    # get vision_embedding foreach
+                    vision_embedding = []
+                    for single_tgt_size, single_pixel_values in zip(tgt_sizes, all_pixel_values):
+                        single_pixel_values = single_pixel_values.unsqueeze(0)
+                        B, L, _ = single_pixel_values.shape
+                        single_pixel_values = single_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
+                        single_vision_embedding = self.vpm(single_pixel_values.type(dtype)).last_hidden_state
+                        single_vision_embedding = self.resampler(single_vision_embedding, single_tgt_size.unsqueeze(0))
+                        vision_embedding.append(single_vision_embedding)
+                    vision_embedding = torch.vstack(vision_embedding)
+
+                start = 0
+                for pixel_values in pixel_values_list:
+                    img_cnt = len(pixel_values)
+                    if img_cnt > 0:
+                        vision_hidden_states.append(vision_embedding[start: start + img_cnt])
+                        start += img_cnt
+                    else:
+                        vision_hidden_states.append([])
+            else: # no image
+                if self.training:
+                    dummy_image = torch.zeros(
+                        (1, 3, 224, 224),
+                        device=device, dtype=dtype
+                    )
+                    tgt_sizes = torch.Tensor([[(224 // self.config.patch_size), math.ceil(224 / self.config.patch_size)]]).type(torch.int32)
+                    dummy_feature = self.resampler(self.vpm(dummy_image).last_hidden_state, tgt_sizes)
+                else:
+                    dummy_feature = []
+                for _ in range(len(pixel_values_list)):
+                    vision_hidden_states.append(dummy_feature)
+
+        else:
+            vision_hidden_states = data['vision_hidden_states']
+
+        if hasattr(self.llm.config, 'scale_emb'):
+            vllm_embedding = self.llm.model.embed_tokens(data['input_ids']) * self.llm.config.scale_emb
+        else:
+            vllm_embedding = self.llm.model.embed_tokens(data['input_ids'])
+
+        vision_hidden_states = [i.type(vllm_embedding.dtype) if isinstance(
+            i, torch.Tensor) else i for i in vision_hidden_states]
+
+        bs = len(data['input_ids'])
+        for i in range(bs):
+            cur_vs_hs = vision_hidden_states[i]
+            if len(cur_vs_hs) > 0:
+                cur_vllm_emb = vllm_embedding[i]
+                cur_image_bound = data['image_bound'][i]
+                if len(cur_image_bound) > 0:
+                    image_indices = torch.stack(
+                        [torch.arange(r[0], r[1], dtype=torch.long) for r in cur_image_bound]
+                    ).to(vllm_embedding.device)
+
+                    cur_vllm_emb.scatter_(0, image_indices.view(-1, 1).repeat(1, cur_vllm_emb.shape[-1]),
+                                          cur_vs_hs.view(-1, cur_vs_hs.shape[-1]))
+                elif self.training:
+                    cur_vllm_emb += cur_vs_hs[0].mean() * 0
+
+        return vllm_embedding, vision_hidden_states
+
+    def forward(self, data, **kwargs):
+        vllm_embedding, vision_hidden_states = self.get_vllm_embedding(data)
+        position_ids = data["position_ids"]
+        if position_ids.dtype != torch.int64:
+            position_ids = position_ids.long()
+
+        return self.llm(
+            input_ids=None,
+            position_ids=position_ids,
+            inputs_embeds=vllm_embedding,
+            **kwargs
+        )
+
+    def _convert_to_tensors(
+        self, tokenizer, input_ids, max_inp_length: Optional[int] = None
+    ):
+        if max_inp_length is not None:
+            input_ids = input_ids[:max_inp_length]
+        input_ids = torch.tensor(input_ids, dtype=torch.int32)
+
+        image_start_tokens = torch.where(input_ids == tokenizer.im_start_id)[0]
+        # 跳过 im_start
+        image_start_tokens += 1
+        image_end_tokens = torch.where(input_ids == tokenizer.im_end_id)[0]
+        valid_image_nums = max(len(image_start_tokens), len(image_end_tokens))
+        image_bound = torch.hstack(
+            [
+                image_start_tokens[:valid_image_nums].unsqueeze(-1),
+                image_end_tokens[:valid_image_nums].unsqueeze(-1),
+            ]
+        )
+
+        model_input = {}
+        model_input["input_ids"] = input_ids.unsqueeze(0).to(self.device)
+        model_input["image_bound"] = image_bound
+
+        return model_input
+
+    def _process_list(
+        self, tokenizer, input_id_list, max_inp_length: Optional[int] = None
+    ):
+        pad_keys = ["input_ids"]
+        input_tensors = []
+        for input_ids in input_id_list:
+            input_tensors.append(
+                self._convert_to_tensors(tokenizer, input_ids, max_inp_length)
+            )
+        padded = {}
+        for key in pad_keys:
+            padded[key] = pad(input_tensors, key, padding_side="left").to(self.device)
+        padded["image_bound"] = [i["image_bound"] for i in input_tensors]
+        return padded
+
+    def _decode(self, inputs_embeds, tokenizer, **kwargs):
+        terminators = [
+            tokenizer.eos_token_id,
+            tokenizer.convert_tokens_to_ids("<|eot_id|>")
+        ]
+        output = self.llm.generate(
+            inputs_embeds=inputs_embeds,
+            pad_token_id=0,
+            eos_token_id=terminators,
+            **kwargs
+        )
+        return self._decode_text(output, tokenizer)
+    
+    def _decode_stream(self, inputs_embeds, tokenizer, **kwargs):
+        terminators = [
+            tokenizer.eos_token_id,
+            tokenizer.convert_tokens_to_ids("<|eot_id|>")
+        ]
+        streamer = TextIteratorStreamer(tokenizer=tokenizer)
+        generation_kwargs = {
+            'inputs_embeds': inputs_embeds,
+            'pad_token_id': 0,
+            'eos_token_id': terminators,
+            'streamer': streamer
+        }
+        generation_kwargs.update(kwargs)
+
+        thread = Thread(target=self.llm.generate, kwargs=generation_kwargs)
+        thread.start()
+    
+        return streamer
+
+    def _decode_text(self, result_ids, tokenizer):
+        result_text = []
+        for result in result_ids:
+            result = result[result != 0]
+            if result[0] == tokenizer.bos_id:
+                result = result[1:]
+            if result[-1] == tokenizer.eos_id or result[-1] == tokenizer.eot_id:
+                result = result[:-1]
+            result_text.append(tokenizer.decode(result).strip())
+        return result_text
+
+    def slice_image(self, image):
+        return slice_image(
+            image,
+            self.config.slice_config.max_slice_nums,
+            self.config.slice_config.scale_resolution,
+            self.config.slice_config.patch_size,
+        )
+
+    def get_slice_image_placeholder(self, image, tokenizer):
+        image_placeholder = (
+            tokenizer.im_start
+            + tokenizer.unk_token * self.config.query_num
+            + tokenizer.im_end
+        )
+
+        slice_images = []
+
+        source_image, patches, best_grid = slice_image(
+            image,
+            self.config.slice_config.max_slice_nums,
+            self.config.slice_config.scale_resolution,
+            self.config.slice_config.patch_size,
+        )
+
+        slice_images.append(source_image)
+        final_placeholder = image_placeholder
+
+        if len(patches) > 0:
+            for i in range(len(patches)):
+                for j in range(len(patches[0])):
+                    slice_images.append(patches[i][j])
+
+            final_placeholder += get_grid_placeholder(
+                tokenizer, best_grid, self.config.query_num
+            )
+
+        return slice_images, final_placeholder
+
+    def reshape_by_patch(self, image_tensor):
+        """
+        :param image_tensor: shape [3, H, W]
+        :param patch_size:
+        :return: [3, patch_size, HW/patch_size]
+        """
+        patch_size = self.config.patch_size
+        patches = torch.nn.functional.unfold(
+            image_tensor,
+            (patch_size, patch_size),
+            stride=(patch_size, patch_size)
+        )
+
+        patches = patches.reshape(image_tensor.size(0), patch_size, patch_size, -1)
+        patches = patches.permute(0, 1, 3, 2).reshape(image_tensor.size(0), patch_size, -1)
+        return patches
+
+    def generate(
+        self,
+        input_id_list=None,
+        img_list=None,
+        tgt_sizes=None,
+        tokenizer=None,
+        max_inp_length: Optional[int] = None,
+        vision_hidden_states=None,
+        return_vision_hidden_states=False,
+        stream=False,
+        **kwargs
+    ):
+
+        assert input_id_list is not None
+        bs = len(input_id_list)
+        if img_list == None:
+            img_list = [[] for i in range(bs)]
+        assert bs == len(img_list)
+
+        model_inputs = self._process_list(tokenizer, input_id_list, max_inp_length)
+
+        if vision_hidden_states is None:
+            pixel_values = []
+            for i in range(bs):
+                img_inps = []
+                for img in img_list[i]:
+                    img_inps.append(img.to(self.device))
+                if img_inps:
+                    pixel_values.append(img_inps)
+                else:
+                    pixel_values.append([])
+            model_inputs["pixel_values"] = pixel_values
+            model_inputs['tgt_sizes'] = tgt_sizes
+        else:
+            model_inputs["vision_hidden_states"] = vision_hidden_states
+
+        with torch.inference_mode():
+            (
+                model_inputs["inputs_embeds"],
+                vision_hidden_states,
+            ) = self.get_vllm_embedding(model_inputs)
+
+            if stream:
+                result = self._decode_stream(model_inputs["inputs_embeds"], tokenizer, **kwargs)
+            else:
+                result = self._decode(model_inputs["inputs_embeds"], tokenizer, **kwargs)
+
+        if return_vision_hidden_states:
+            return result, vision_hidden_states
+
+        return result
+
+    def chat(
+        self,
+        image,
+        msgs,
+        tokenizer,
+        vision_hidden_states=None,
+        max_new_tokens=1024,
+        sampling=True,
+        max_inp_length=2048,
+        system_prompt='',
+        stream=False,
+        **kwargs
+    ):
+        if isinstance(msgs, str):
+            msgs = json.loads(msgs)
+
+        copy_msgs = deepcopy(msgs)
+        assert len(copy_msgs) > 0, 'msgs is empty'
+        assert sampling or not stream, 'if use stream mode, make sure sampling=True'
+
+        if image is not None and isinstance(copy_msgs[0]['content'], str):
+            copy_msgs[0]['content'] = [image, copy_msgs[0]['content']]
+
+        images = []
+        tgt_sizes = []
+        for i, msg in enumerate(copy_msgs):
+            role = msg["role"]
+            content = msg["content"]
+            assert role in ["user", "assistant"]
+            if i == 0:
+                assert role == "user", "The role of first msg should be user"
+            if isinstance(content, str):
+                content = [content]
+
+            cur_msgs = []
+            for c in content:
+                if isinstance(c, Image.Image):
+                    image = c
+                    if self.config.slice_mode:
+                        slice_images, image_placeholder = self.get_slice_image_placeholder(
+                            image, tokenizer
+                        )
+                        cur_msgs.append(image_placeholder)
+                        for slice_image in slice_images:
+                            slice_image = self.transform(slice_image)
+                            H, W = slice_image.shape[1:]
+                            images.append(self.reshape_by_patch(slice_image))
+                            tgt_sizes.append(torch.Tensor([H // self.config.patch_size, W // self.config.patch_size]).type(torch.int32))
+                    else:
+                        images.append(self.transform(image))
+                        cur_msgs.append(
+                            tokenizer.im_start
+                            + tokenizer.unk_token * self.config.query_num
+                            + tokenizer.im_end
+                        )
+                elif isinstance(c, str):
+                    cur_msgs.append(c)
+
+
+            msg['content'] = '\n'.join(cur_msgs)
+        if tgt_sizes:
+            tgt_sizes = torch.vstack(tgt_sizes)
+
+        if system_prompt:
+            sys_msg = {'role': 'system', 'content': system_prompt}
+            copy_msgs = [sys_msg] + copy_msgs
+
+        input_ids = tokenizer.apply_chat_template(copy_msgs, tokenize=True, add_generation_prompt=False)
+
+        if sampling:
+            generation_config = {
+                "top_p": 0.8,
+                "top_k": 100,
+                "temperature": 0.7,
+                "do_sample": True,
+                "repetition_penalty": 1.05
+            }
+        else:
+            generation_config = {
+                "num_beams": 3,
+                "repetition_penalty": 1.2,
+            }
+
+        generation_config.update(
+            (k, kwargs[k]) for k in generation_config.keys() & kwargs.keys()
+        )
+
+        with torch.inference_mode():
+            res, vision_hidden_states = self.generate(
+                input_id_list=[input_ids],
+                max_inp_length=max_inp_length,
+                img_list=[images],
+                tgt_sizes=[tgt_sizes],
+                tokenizer=tokenizer,
+                max_new_tokens=max_new_tokens,
+                vision_hidden_states=vision_hidden_states,
+                return_vision_hidden_states=True,
+                stream=stream,
+                **generation_config
+            )
+
+        if stream:
+            def stream_gen():
+                for text in res:
+                    text = text.replace(tokenizer.eot_token, '').replace(tokenizer.eos_token, '')
+                    yield text
+            return stream_gen()
+
+        else:
+            answer = res[0]
+            return answer
+
+
+class PreTrainedTokenizerFastWrapper(PreTrainedTokenizerFast):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.eot_token = "<|eot_id|>"
+        self.im_start = "<image>"
+        self.im_end = "</image>"
+        self.ref_start = "<ref>"
+        self.ref_end = "</ref>"
+        self.box_start = "<box>"
+        self.box_end = "</box>"
+        self.quad_start = "<quad>"
+        self.quad_end = "</quad>"
+        self.slice_start = "<slice>"
+        self.slice_end = "</slice>"
+
+    @property
+    def eos_id(self):
+        return self.eos_token_id
+
+    @property
+    def bos_id(self):
+        return self.bos_token_id
+
+    @property
+    def unk_id(self):
+        return self.unk_token_id
+
+    @property
+    def eot_id(self):
+        return self.convert_tokens_to_ids(self.eot_token)
+
+    @property
+    def im_start_id(self):
+        return self.convert_tokens_to_ids(self.im_start)
+
+    @property
+    def im_end_id(self):
+        return self.convert_tokens_to_ids(self.im_end)
+
+    @staticmethod
+    def escape(text: str) -> str:
+        return text
+
+    @staticmethod
+    def unescape(text: str) -> str:
+        return text
+
+
+def pad(orig_items, key, max_length=None, padding_value=0, padding_side="left"):
+    items = []
+    if isinstance(orig_items[0][key], list):
+        assert isinstance(orig_items[0][key][0], torch.Tensor)
+        for it in orig_items:
+            for tr in it[key]:
+                items.append({key: tr})
+    else:
+        assert isinstance(orig_items[0][key], torch.Tensor)
+        items = orig_items
+
+    batch_size = len(items)
+    shape = items[0][key].shape
+    dim = len(shape)
+    assert dim <= 3
+    if max_length is None:
+        max_length = 0
+    max_length = max(max_length, max(item[key].shape[-1] for item in items))
+    min_length = min(item[key].shape[-1] for item in items)
+    dtype = items[0][key].dtype
+
+    if dim == 1:
+        return torch.cat([item[key] for item in items], dim=0)
+    elif dim == 2:
+        if max_length == min_length:
+            return torch.cat([item[key] for item in items], dim=0)
+        tensor = torch.zeros((batch_size, max_length), dtype=dtype) + padding_value
+    else:
+        tensor = (
+            torch.zeros((batch_size, max_length, shape[-1]), dtype=dtype)
+            + padding_value
+        )
+
+    for i, item in enumerate(items):
+        if dim == 2:
+            if padding_side == "left":
+                tensor[i, -len(item[key][0]) :] = item[key][0].clone()
+            else:
+                tensor[i, : len(item[key][0])] = item[key][0].clone()
+        elif dim == 3:
+            if padding_side == "left":
+                tensor[i, -len(item[key][0]) :, :] = item[key][0].clone()
+            else:
+                tensor[i, : len(item[key][0]), :] = item[key][0].clone()
+
+    return tensor
+
+
+def slice_image(
+    image, max_slice_nums=9, scale_resolution=448, patch_size=14, never_split=False
+):
+    original_size = image.size
+    original_width, original_height = original_size
+    log_ratio = math.log(original_width / original_height)
+    ratio = original_width * original_height / (scale_resolution * scale_resolution)
+    multiple = min(math.ceil(ratio), max_slice_nums)
+
+    source_image = None
+    best_grid = None
+    patches = []
+
+    if multiple <= 1 or never_split:
+        # dont need to slice, upsample
+        best_size = find_best_resize(
+            original_size, scale_resolution, patch_size, allow_upscale=True
+        )
+        source_image = image.resize(best_size, Image.Resampling.BICUBIC)
+    else:
+        candidate_split_grids_nums = []
+        for i in [multiple - 1, multiple, multiple + 1]:
+            if i == 1 or i > max_slice_nums:
+                continue
+            candidate_split_grids_nums.append(i)
+
+        # source image, down-sampling and ensure divided by patch_size
+        best_resize = find_best_resize(original_size, scale_resolution, patch_size)
+        source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
+        candidate_grids = []
+
+        # find best grid
+        for split_grids_nums in candidate_split_grids_nums:
+            m = 1
+            while m <= split_grids_nums:
+                if split_grids_nums % m == 0:
+                    candidate_grids.append([m, split_grids_nums // m])
+                m += 1
+
+        best_grid = [1, 1]
+        min_error = float("inf")
+        for grid in candidate_grids:
+            error = abs(log_ratio - math.log(grid[0] / grid[1]))
+            if error < min_error:
+                best_grid = grid
+                min_error = error
+
+        refine_size = get_refine_size(
+            original_size, best_grid, scale_resolution, patch_size, allow_upscale=True
+        )
+
+        refine_image = image.resize(refine_size, Image.Resampling.BICUBIC)
+        patches = split_to_patches(refine_image, best_grid)
+
+    return source_image, patches, best_grid
+
+
+def ensure_divide(length, patch_size):
+    return max(round(length / patch_size) * patch_size, patch_size)
+
+
+def find_best_resize(original_size, scale_resolution, patch_size, allow_upscale=False):
+    width, height = original_size
+    if (width * height > scale_resolution * scale_resolution) or allow_upscale:
+        r = width / height
+        height = int(scale_resolution / math.sqrt(r))
+        width = int(height * r)
+    best_width = ensure_divide(width, patch_size)
+    best_height = ensure_divide(height, patch_size)
+    return (best_width, best_height)
+
+
+def get_refine_size(
+    original_size, grid, scale_resolution, patch_size, allow_upscale=False
+):
+    width, height = original_size
+    grid_x, grid_y = grid
+
+    refine_width = ensure_divide(width, grid_x)
+    refine_height = ensure_divide(height, grid_y)
+
+    grid_width = refine_width / grid_x
+    grid_height = refine_height / grid_y
+
+    best_grid_size = find_best_resize(
+        (grid_width, grid_height),
+        scale_resolution,
+        patch_size,
+        allow_upscale=allow_upscale,
+    )
+
+    refine_size = (best_grid_size[0] * grid_x, best_grid_size[1] * grid_y)
+
+    return refine_size
+
+
+def split_to_patches(image, grid):
+    patches = []
+    width, height = image.size
+    grid_x = int(width / grid[0])
+    grid_y = int(height / grid[1])
+
+    for i in range(0, height, grid_y):
+        images = []
+        for j in range(0, width, grid_x):
+            box = (j, i, j + grid_x, i + grid_y)
+            patch = image.crop(box)
+            images.append(patch)
+        patches.append(images)
+
+    return patches
+
+
+def get_grid_placeholder(tokenizer, grid, query_num):
+    image_placeholder = (
+        tokenizer.im_start + tokenizer.unk_token * query_num + tokenizer.im_end
+    )
+
+    cols = grid[0]
+    rows = grid[1]
+    slices = []
+    for i in range(rows):
+        lines = []
+        for j in range(cols):
+            lines.append(image_placeholder)
+        slices.append("".join(lines))
+    slice_placeholder = tokenizer.slice_start + "\n".join(slices) + tokenizer.slice_end
+    return slice_placeholder

resampler.py

@@ -0,0 +1,163 @@
+from functools import partial
+import numpy as np
+
+import torch
+from torch import nn
+from torch.nn.init import trunc_normal_
+
+def get_2d_sincos_pos_embed(embed_dim, image_size):
+    """
+    image_size: image_size or (image_height, image_width)
+    return:
+    pos_embed: [image_height, image_width, embed_dim]
+    """
+    if isinstance(image_size, int):
+        grid_h_size, grid_w_size = image_size, image_size
+    else:
+        grid_h_size, grid_w_size = image_size[0], image_size[1]
+
+    grid_h = np.arange(grid_h_size, dtype=np.float32)
+    grid_w = np.arange(grid_w_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[0])  # (H, W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[1])  # (H, W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=-1)  # (H, W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid_new(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (H, W)
+    out: (H, W, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    out = np.einsum('hw,d->hwd', pos, omega)  # (H, W, D/2), outer product
+
+    emb_sin = np.sin(out)  # (H, W, D/2)
+    emb_cos = np.cos(out)  # (H, W, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=-1)  # (H, W, D)
+    return emb
+
+
+class Resampler(nn.Module):
+    """
+    A 2D perceiver-resampler network with one cross attention layers by
+       given learnable queries and 2d sincos pos_emb
+    Outputs:
+        A tensor with the shape of (batch_size, num_queries, embed_dim)
+    """
+
+    def __init__(
+            self,
+            num_queries,
+            embed_dim,
+            num_heads,
+            kv_dim=None,
+            norm_layer=partial(nn.LayerNorm, eps=1e-6),
+            adaptive=False,
+            max_size=(70, 70),
+    ):
+        super().__init__()
+        self.num_queries = num_queries
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.adaptive = adaptive
+        self.max_size = max_size
+
+        self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
+        trunc_normal_(self.query, std=.02)
+
+        if kv_dim is not None and kv_dim != embed_dim:
+            self.kv_proj = nn.Linear(kv_dim, embed_dim, bias=False)
+        else:
+            self.kv_proj = nn.Identity()
+
+        self.attn = nn.MultiheadAttention(embed_dim, num_heads)
+        self.ln_q = norm_layer(embed_dim)
+        self.ln_kv = norm_layer(embed_dim)
+
+        self.ln_post = norm_layer(embed_dim)
+        self.proj = nn.Parameter((embed_dim ** -0.5) * torch.randn(embed_dim, embed_dim))
+
+        self._set_2d_pos_cache(self.max_size)
+        self.apply(self._init_weights)
+
+    def _set_2d_pos_cache(self, max_size, device='cpu'):
+        pos_embed = torch.from_numpy(get_2d_sincos_pos_embed(self.embed_dim, max_size)).float().to(device)
+        self.register_buffer("pos_embed", pos_embed, persistent=False)
+
+    def _adjust_pos_cache(self, tgt_sizes, device):
+        max_h = torch.max(tgt_sizes[:, 0])
+        max_w = torch.max(tgt_sizes[:, 1])
+        if max_h > self.max_size[0] or max_w > self.max_size[1]:
+            self.max_size = [max(max_h, self.max_size[0]), max(max_w, self.max_size[1])]
+            self._set_2d_pos_cache(self.max_size, device)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def forward(self, x, tgt_sizes=None):
+        assert x.shape[0] == tgt_sizes.shape[0]
+        bs = x.shape[0]
+
+        device = x.device
+        dtype = x.dtype
+
+        patch_len = tgt_sizes[:, 0] * tgt_sizes[:, 1]
+
+        self._adjust_pos_cache(tgt_sizes, device=device)
+
+        max_patch_len = torch.max(patch_len)
+        key_padding_mask = torch.zeros((bs, max_patch_len), dtype=torch.bool, device=device)
+
+        pos_embed = []
+        for i in range(bs):
+            tgt_h, tgt_w = tgt_sizes[i]
+            pos_embed.append(self.pos_embed[:tgt_h, :tgt_w, :].reshape((tgt_h * tgt_w, -1)).to(dtype))  # patches * D
+            key_padding_mask[i, patch_len[i]:] = True
+
+        pos_embed = torch.nn.utils.rnn.pad_sequence(
+            pos_embed, batch_first=True, padding_value=0.0).permute(1, 0, 2)  # BLD => L * B * D
+
+        x = self.kv_proj(x)  # B * L * D
+        x = self.ln_kv(x).permute(1, 0, 2)  # L * B * D
+
+        q = self.ln_q(self.query)  # Q * D
+
+        out = self.attn(
+            self._repeat(q, bs),  # Q * B * D
+            x + pos_embed,  # L * B * D +  L * B * D
+            x,
+            key_padding_mask=key_padding_mask)[0]
+        #  out: Q * B * D
+        x = out.permute(1, 0, 2)  # B * Q * D
+
+        x = self.ln_post(x)
+        x = x @ self.proj
+        return x
+
+    def _repeat(self, query, N: int):
+        return query.unsqueeze(1).repeat(1, N, 1)