gugarosa commited on
Commit
44c7c8b
1 Parent(s): 7ba0150

chore(root): Initial files upload.

Browse files
CODE_OF_CONDUCT.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # Microsoft Open Source Code of Conduct
2
+
3
+ This project has adopted the [Microsoft Open Source Code of Conduct](https://opensource.microsoft.com/codeofconduct/).
4
+
5
+ Resources:
6
+
7
+ - [Microsoft Open Source Code of Conduct](https://opensource.microsoft.com/codeofconduct/)
8
+ - [Microsoft Code of Conduct FAQ](https://opensource.microsoft.com/codeofconduct/faq/)
9
+ - Contact [[email protected]](mailto:[email protected]) with questions or concerns
LICENSE ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) Microsoft Corporation.
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE
README.md ADDED
@@ -0,0 +1,241 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ license_link: https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/resolve/main/LICENSE
4
+
5
+ language:
6
+ - multilingual
7
+ pipeline_tag: text-generation
8
+ tags:
9
+ - nlp
10
+ - code
11
+ - vision
12
+ inference:
13
+ parameters:
14
+ temperature: 0.7
15
+ widget:
16
+ - messages:
17
+ - role: user
18
+ content: <|image_1|>Can you describe what you see in the image?
19
+ ---
20
+ ## Model Summary
21
+
22
+ Phi-3 Vision is a lightweight, state-of-the-art open multimodal model built upon datasets which include - synthetic data and filtered publicly available websites - with a focus on very high-quality, reasoning dense data both on text and vision. The model belongs to the Phi-3 model family, and the multimodal version comes with 128K context length (in tokens) it can support. The model underwent a rigorous enhancement process, incorporating both supervised fine-tuning and direct preference optimization to ensure precise instruction adherence and robust safety measures.
23
+
24
+ Resources and Technical Documentation:
25
+
26
+ + [Phi-3 Microsoft Blog](https://aka.ms/phi3blog-april)
27
+ + [Phi-3 Technical Report](https://aka.ms/phi3-tech-report)
28
+ + [Phi-3 on Azure AI Studio](https://aka.ms/phi3-azure-ai)
29
+
30
+ | | Short Context | Long Context |
31
+ | ------- | ------------- | ------------ |
32
+ | Mini | 4K [[HF]](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct) ; [[ONNX]](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-onnx) ; [[GGUF]](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct-gguf) | 128K [[HF]](https://huggingface.co/microsoft/Phi-3-mini-128k-instruct) ; [[ONNX]](https://huggingface.co/microsoft/Phi-3-mini-128k-instruct-onnx)|
33
+ | Small | 8K [[HF]](https://huggingface.co/microsoft/Phi-3-small-8k-instruct) ; [[ONNX]](https://huggingface.co/microsoft/Phi-3-small-8k-instruct-onnx) | 128K [[HF]](https://huggingface.co/microsoft/Phi-3-small-128k-instruct) ; [[ONNX]](https://huggingface.co/microsoft/Phi-3-small-128k-instruct-onnx)|
34
+ | Medium | 4K [[HF]](https://huggingface.co/microsoft/Phi-3-medium-4k-instruct) ; [[ONNX]](https://huggingface.co/microsoft/Phi-3-medium-4k-instruct-onnx) | 128K [[HF]](https://huggingface.co/microsoft/Phi-3-medium-128k-instruct) ; [[ONNX]](https://huggingface.co/microsoft/Phi-3-medium-128k-instruct-onnx)|
35
+
36
+ ## Intended Uses
37
+
38
+ **Primary use cases**
39
+
40
+ The model is intended for broad commercial and research use in English. The model provides uses for general purpose AI systems and applications with visual and text input capabilities which require
41
+
42
+ 1) memory/compute constrained environments;
43
+ 2) latency bound scenarios;
44
+ 3) general image understanding;
45
+ 4) OCR;
46
+ 5) chart and table understanding.
47
+
48
+ Our model is designed to accelerate research on efficient language and multimodal models, for use as a building block for generative AI powered features.
49
+
50
+ **Use case considerations**
51
+
52
+ Our models are not specifically designed or evaluated for all downstream purposes. Developers should consider common limitations of language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fairness before using within a specific downstream use case, particularly for high-risk scenarios.
53
+ Developers should be aware of and adhere to applicable laws or regulations (including privacy, trade compliance laws, etc.) that are relevant to their use case.
54
+
55
+ Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the license the model is released under.
56
+
57
+ ## How to Use
58
+
59
+ Phi-3-vision-128K-Instruct has been integrated in the development version (4.40.2) of `transformers`. Until the official version is released through `pip`, ensure that you are doing one of the following:
60
+ * When loading the model, ensure that `trust_remote_code=True` is passed as an argument of the `from_pretrained()` function.
61
+
62
+ * Update your local `transformers` to the development version: `pip uninstall -y transformers && pip install git+https://github.com/huggingface/transformers`. The previous command is an alternative to cloning and installing from the source.
63
+
64
+ The current `transformers` version can be verified with: `pip list | grep transformers`.
65
+
66
+ Examples of required packages:
67
+ ```
68
+ flash_attn==2.5.8
69
+ numpy==1.24.4
70
+ Pillow==10.3.0
71
+ Requests==2.31.0
72
+ torch==2.3.0
73
+ torchvision==0.18.0
74
+ transformers==4.40.2
75
+ ```
76
+
77
+ Phi-3-Vision-128K-Instruct is also available in [Azure AI Studio](https://aka.ms/phi3-azure-ai).
78
+
79
+ ### Chat Format
80
+
81
+ Given the nature of the training data, the Phi-3-Vision-128K-Instruct model is best suited for prompts using the chat format as follows.
82
+ You can provide the prompt as a single image with a generic template as follow:
83
+ ```markdown
84
+ <|user|>\n<|image_1|>\n{prompt}<|end|>\n<|assistant|>\n
85
+ ```
86
+
87
+ where the model generates the text after `<|assistant|>` . In case of multi-turn conversation, the prompt can be formatted as follows:
88
+
89
+ ```markdown
90
+ <|user|>\n<|image_1|>\n{prompt_1}<|end|>\n<|assistant|>\n{response_1}<|end|>\n<|user|>\n{prompt_2}<|end|>\n<|assistant|>\n
91
+ ```
92
+
93
+ ### Sample inference code
94
+
95
+ This code snippets show how to get quickly started with running the model on a GPU:
96
+
97
+ ```python
98
+ from PIL import Image
99
+ import requests
100
+ from transformers import AutoModelForCausalLM
101
+ from transformers import AutoProcessor
102
+
103
+ model_id = "microsoft/Phi-3-vision-128k-instruct"
104
+
105
+ model = AutoModelForCausalLM.from_pretrained(model_id, device_map="cuda", trust_remote_code=True, torch_dtype="auto")
106
+
107
+ processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
108
+
109
+ messages = [
110
+ {"role": "user", "content": "<|image_1|>\nWhat is shown in this image?"},
111
+ {"role": "assistant", "content": "The chart displays the percentage of respondents who agree with various statements about their preparedness for meetings. It shows five categories: 'Having clear and pre-defined goals for meetings', 'Knowing where to find the information I need for a meeting', 'Understanding my exact role and responsibilities when I'm invited', 'Having tools to manage admin tasks like note-taking or summarization', and 'Having more focus time to sufficiently prepare for meetings'. Each category has an associated bar indicating the level of agreement, measured on a scale from 0% to 100%."},
112
+ {"role": "user", "content": "Provide insightful questions to spark discussion."}
113
+ ]
114
+
115
+ url = "https://assets-c4akfrf5b4d3f4b7.z01.azurefd.net/assets/2024/04/BMDataViz_661fb89f3845e.png"
116
+ image = Image.open(requests.get(url, stream=True).raw)
117
+
118
+ prompt = processor.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
119
+
120
+ inputs = processor(prompt, [image], return_tensors="pt").to("cuda:0")
121
+
122
+ generation_args = {
123
+ "max_new_tokens": 500,
124
+ "temperature": 0.0,
125
+ "do_sample": False,
126
+ }
127
+
128
+ generate_ids = model.generate(**inputs, eos_token_id=processor.tokenizer.eos_token_id, **generation_args)
129
+
130
+ # remove input tokens
131
+ generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
132
+ response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
133
+
134
+ print(response)
135
+ ```
136
+
137
+ Additional basic examples are provided [here](https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/resolve/main/sample_inference.py.py).
138
+
139
+
140
+ ## Responsible AI Considerations
141
+
142
+ Like other models, the Phi family of models can potentially behave in ways that are unfair, unreliable, or offensive. Some of the limiting behaviors to be aware of include:
143
+
144
+ + Quality of Service: The Phi models are trained primarily on English text. Languages other than English will experience worse performance. English language varieties with less representation in the training data might experience worse performance than standard American English.
145
+ + Representation of Harms & Perpetuation of Stereotypes: These models can over- or under-represent groups of people, erase representation of some groups, or reinforce demeaning or negative stereotypes. Despite safety post-training, these limitations may still be present due to differing levels of representation of different groups or prevalence of examples of negative stereotypes in training data that reflect real-world patterns and societal biases.
146
+ + Inappropriate or Offensive Content: These models may produce other types of inappropriate or offensive content, which may make it inappropriate to deploy for sensitive contexts without additional mitigations that are specific to the use case.
147
+ + Information Reliability: Language models can generate nonsensical content or fabricate content that might sound reasonable but is inaccurate or outdated.
148
+ + Limited Scope for Code: Majority of Phi-3 training data is based in Python and use common packages such as "typing, math, random, collections, datetime, itertools". If the model generates Python scripts that utilize other packages or scripts in other languages, we strongly recommend users manually verify all API uses.
149
+
150
+ Developers should apply responsible AI best practices and are responsible for ensuring that a specific use case complies with relevant laws and regulations (e.g. privacy, trade, etc.). Important areas for consideration include:
151
+
152
+ + Allocation: Models may not be suitable for scenarios that could have consequential impact on legal status or the allocation of resources or life opportunities (ex: housing, employment, credit, etc.) without further assessments and additional debiasing techniques.
153
+ + High-Risk Scenarios: Developers should assess suitability of using models in high-risk scenarios where unfair, unreliable or offensive outputs might be extremely costly or lead to harm. This includes providing advice in sensitive or expert domains where accuracy and reliability are critical (ex: legal or health advice). Additional safeguards should be implemented at the application level according to the deployment context.
154
+ + Misinformation: Models may produce inaccurate information. Developers should follow transparency best practices and inform end-users they are interacting with an AI system. At the application level, developers can build feedback mechanisms and pipelines to ground responses in use-case specific, contextual information, a technique known as Retrieval Augmented Generation (RAG).
155
+ + Generation of Harmful Content: Developers should assess outputs for their context and use available safety classifiers or custom solutions appropriate for their use case.
156
+ + Misuse: Other forms of misuse such as fraud, spam, or malware production may be possible, and developers should ensure that their applications do not violate applicable laws and regulations.
157
+ + Identification of individuals: models with vision capabilities may have the potential to uniquely identify individuals in images. Safety post-training steers the model to refuse such requests, but developers should consider and implement, as appropriate, additional mitigations or user consent flows as required in their respective jurisdiction, (e.g., building measures to blur faces in image inputs before processing.
158
+
159
+ ## Training
160
+
161
+ ### Model
162
+
163
+ * Architecture: Phi-3-Vision-128K-Instruct has 4.2B parameters and contains image encoder, connector, projector, and Phi-3 Mini language model.
164
+ * Inputs: Text and Image. It’s best suited for prompts using the chat format.
165
+ * Context length: 128K tokens
166
+ * GPUs: 512 H100-80G
167
+ * Training time: 1.5 days
168
+ * Training data: 500B vision and text tokens
169
+ * Outputs: Generated text in response to the input
170
+ * Dates: Our models were trained between February and April 2024
171
+ * Status: This is a static model trained on an offline text dataset with cutoff date Mar 15, 2024. Future versions of the tuned models may be released as we improve models.
172
+ * Release Type: Open weight release
173
+ * Release dates: The model weight is released on May 21, 2024.
174
+
175
+ ### Datasets
176
+
177
+ Our training data includes a wide variety of sources, and is a combination of
178
+
179
+ 1) publicly available documents filtered rigorously for quality, selected high-quality educational data and code;
180
+ 2) selected high-quality image-text interleave;
181
+ 3) newly created synthetic, “textbook-like” data for the purpose of teaching math, coding, common sense reasoning, general knowledge of the world (science, daily activities, theory of mind, etc.), newly created image data, e.g., chart/table/diagram/slides;
182
+ 4) high quality chat format supervised data covering various topics to reflect human preferences on different aspects such as instruct-following, truthfulness, honesty and helpfulness.
183
+
184
+ The data collection process involved sourcing information from publicly available documents, with a meticulous approach to filtering out undesirable documents and images. To safeguard privacy, we carefully filtered various image and text data sources to remove or scrub any potentially personal data from the training data.
185
+
186
+ More details can be found in the [Phi-3 Technical Report](https://aka.ms/phi3-tech-report).
187
+
188
+ ## Benchmarks
189
+
190
+ To understand the capabilities, we compare Phi-3 Vision-128K-Instruct with a set of models over a variety of zero-shot benchmarks using our internal benchmark platform.
191
+
192
+ |Benchmark|Phi-3 Vision-128K-In1|LlaVA-1.6 Vicuna-7B|QWEN-VL Chat|Llama3-Llava-Next-8B|Claude-3 Haiku|Gemini 1.0 Pro V|GPT-4V-Turbo|
193
+ |---------|---------------------|------------------|------------|--------------------|--------------|----------------|------------|
194
+ |MMMU|40.4|34.2|39.0|36.4|40.7|42.0|55.5| 
195
+ |MMBench|80.5|76.3|75.8|79.4|62.4|80.0|86.1|
196
+ |ScienceQA|90.8|70.6|67.2|73.7|72.0|79.7|75.7|
197
+ |MathVista|44.5|31.5|29.4|34.8|33.2|35.0|47.5|
198
+ |InterGPS|38.1|20.5|22.3|24.6|32.1|28.6|41.0|
199
+ |AI2D|76.7|63.1|59.8|66.9|60.3|62.8|74.7|
200
+ |ChartQA|81.4|55.0|50.9|65.8|59.3|58.0|62.3|
201
+ |TextVQA|70.9|64.6|59.4|55.7|62.7|64.7|68.1|
202
+ |POPE|85.8|87.2|82.6|87.0|74.4|84.2|83.7|
203
+
204
+
205
+ ## Software
206
+
207
+ * [PyTorch](https://github.com/pytorch/pytorch)
208
+ * [Transformers](https://github.com/huggingface/transformers)
209
+ * [Flash-Attention](https://github.com/HazyResearch/flash-attention)
210
+
211
+ ## Hardware
212
+ Note that by default, the Phi-3-Vision-128K model uses flash attention, which requires certain types of GPU hardware to run. We have tested on the following GPU types:
213
+ * NVIDIA A100
214
+ * NVIDIA A6000
215
+ * NVIDIA H100
216
+
217
+ ### Running on Windows or without flash attention
218
+ To enable the model on these enviroment here are steps that you may consider to follow:
219
+
220
+ Step 1: comment flash attention import code in modeling_phi3_v.py from line 52 to line 56.
221
+ ```python
222
+ # if is_flash_attn_2_available():
223
+ # from flash_attn import flash_attn_func, flash_attn_varlen_func
224
+ # from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
225
+
226
+ # _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
227
+ ```
228
+
229
+ Step 2: change _"_attn_implementation"_ from _"flash_attention_2"_ to _"eager"_ in config.json or disable flash attention when you create the model as below.
230
+
231
+ ```python
232
+ model = AutoModelForCausalLM.from_pretrained('microsoft/Phi-3-vision-128k-instruct', device_map="cuda", trust_remote_code=True, torch_dtype="auto", _attn_implementation="eager")
233
+ ```
234
+
235
+ ## License
236
+
237
+ The model is licensed under the [MIT license](https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/resolve/main/LICENSE).
238
+
239
+ ## Trademarks
240
+
241
+ This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow [Microsoft’s Trademark & Brand Guidelines](https://www.microsoft.com/en-us/legal/intellectualproperty/trademarks). Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party’s policies.
SECURITY.md ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <!-- BEGIN MICROSOFT SECURITY.MD V0.0.9 BLOCK -->
2
+
3
+ ## Security
4
+
5
+ Microsoft takes the security of our software products and services seriously, which includes all source code repositories managed through our GitHub organizations, which include [Microsoft](https://github.com/Microsoft), [Azure](https://github.com/Azure), [DotNet](https://github.com/dotnet), [AspNet](https://github.com/aspnet) and [Xamarin](https://github.com/xamarin).
6
+
7
+ If you believe you have found a security vulnerability in any Microsoft-owned repository that meets [Microsoft's definition of a security vulnerability](https://aka.ms/security.md/definition), please report it to us as described below.
8
+
9
+ ## Reporting Security Issues
10
+
11
+ **Please do not report security vulnerabilities through public GitHub issues.**
12
+
13
+ Instead, please report them to the Microsoft Security Response Center (MSRC) at [https://msrc.microsoft.com/create-report](https://aka.ms/security.md/msrc/create-report).
14
+
15
+ If you prefer to submit without logging in, send email to [[email protected]](mailto:[email protected]). If possible, encrypt your message with our PGP key; please download it from the [Microsoft Security Response Center PGP Key page](https://aka.ms/security.md/msrc/pgp).
16
+
17
+ You should receive a response within 24 hours. If for some reason you do not, please follow up via email to ensure we received your original message. Additional information can be found at [microsoft.com/msrc](https://www.microsoft.com/msrc).
18
+
19
+ Please include the requested information listed below (as much as you can provide) to help us better understand the nature and scope of the possible issue:
20
+
21
+ * Type of issue (e.g. buffer overflow, SQL injection, cross-site scripting, etc.)
22
+ * Full paths of source file(s) related to the manifestation of the issue
23
+ * The location of the affected source code (tag/branch/commit or direct URL)
24
+ * Any special configuration required to reproduce the issue
25
+ * Step-by-step instructions to reproduce the issue
26
+ * Proof-of-concept or exploit code (if possible)
27
+ * Impact of the issue, including how an attacker might exploit the issue
28
+
29
+ This information will help us triage your report more quickly.
30
+
31
+ If you are reporting for a bug bounty, more complete reports can contribute to a higher bounty award. Please visit our [Microsoft Bug Bounty Program](https://aka.ms/security.md/msrc/bounty) page for more details about our active programs.
32
+
33
+ ## Preferred Languages
34
+
35
+ We prefer all communications to be in English.
36
+
37
+ ## Policy
38
+
39
+ Microsoft follows the principle of [Coordinated Vulnerability Disclosure](https://aka.ms/security.md/cvd).
40
+
41
+ <!-- END MICROSOFT SECURITY.MD BLOCK -->
SUPPORT.md ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # TODO: The maintainer of this repo has not yet edited this file
2
+
3
+ **REPO OWNER**: Do you want Customer Service & Support (CSS) support for this product/project?
4
+
5
+ - **No CSS support:** Fill out this template with information about how to file issues and get help.
6
+ - **Yes CSS support:** Fill out an intake form at [aka.ms/onboardsupport](https://aka.ms/onboardsupport). CSS will work with/help you to determine next steps.
7
+ - **Not sure?** Fill out an intake as though the answer were "Yes". CSS will help you decide.
8
+
9
+ *Then remove this first heading from this SUPPORT.MD file before publishing your repo.*
10
+
11
+ # Support
12
+
13
+ ## How to file issues and get help
14
+
15
+ This project uses GitHub Issues to track bugs and feature requests. Please search the existing
16
+ issues before filing new issues to avoid duplicates. For new issues, file your bug or
17
+ feature request as a new Issue.
18
+
19
+ For help and questions about using this project, please **REPO MAINTAINER: INSERT INSTRUCTIONS HERE
20
+ FOR HOW TO ENGAGE REPO OWNERS OR COMMUNITY FOR HELP. COULD BE A STACK OVERFLOW TAG OR OTHER
21
+ CHANNEL. WHERE WILL YOU HELP PEOPLE?**.
22
+
23
+ ## Microsoft Support Policy
24
+
25
+ Support for this **PROJECT or PRODUCT** is limited to the resources listed above.
config.json ADDED
@@ -0,0 +1,148 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "Phi-3-vision-128k-instruct",
3
+ "architectures": [
4
+ "Phi3VForCausalLM"
5
+ ],
6
+ "attention_dropout": 0.0,
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_phi3_v.Phi3VConfig",
9
+ "AutoModelForCausalLM": "modeling_phi3_v.Phi3VForCausalLM"
10
+ },
11
+ "bos_token_id": 1,
12
+ "embd_layer": {
13
+ "embedding_cls": "image",
14
+ "hd_transform_order": "sub_glb",
15
+ "projection_cls": "mlp",
16
+ "use_hd_transform": true,
17
+ "with_learnable_separator": true
18
+ },
19
+ "eos_token_id": 2,
20
+ "hidden_act": "silu",
21
+ "hidden_size": 3072,
22
+ "img_processor": {
23
+ "image_dim_out": 1024,
24
+ "model_name": "openai/clip-vit-large-patch14-336",
25
+ "name": "clip_vision_model",
26
+ "num_img_tokens": 144
27
+ },
28
+ "initializer_range": 0.02,
29
+ "intermediate_size": 8192,
30
+ "max_position_embeddings": 131072,
31
+ "model_type": "phi3_v",
32
+ "num_attention_heads": 32,
33
+ "num_hidden_layers": 32,
34
+ "num_key_value_heads": 32,
35
+ "original_max_position_embeddings": 4096,
36
+ "rms_norm_eps": 1e-05,
37
+ "rope_scaling": {
38
+ "long_factor": [
39
+ 1.0299999713897705,
40
+ 1.0499999523162842,
41
+ 1.0499999523162842,
42
+ 1.0799999237060547,
43
+ 1.2299998998641968,
44
+ 1.2299998998641968,
45
+ 1.2999999523162842,
46
+ 1.4499999284744263,
47
+ 1.5999999046325684,
48
+ 1.6499998569488525,
49
+ 1.8999998569488525,
50
+ 2.859999895095825,
51
+ 3.68999981880188,
52
+ 5.419999599456787,
53
+ 5.489999771118164,
54
+ 5.489999771118164,
55
+ 9.09000015258789,
56
+ 11.579999923706055,
57
+ 15.65999984741211,
58
+ 15.769999504089355,
59
+ 15.789999961853027,
60
+ 18.360000610351562,
61
+ 21.989999771118164,
62
+ 23.079999923706055,
63
+ 30.009998321533203,
64
+ 32.35000228881836,
65
+ 32.590003967285156,
66
+ 35.56000518798828,
67
+ 39.95000457763672,
68
+ 53.840003967285156,
69
+ 56.20000457763672,
70
+ 57.95000457763672,
71
+ 59.29000473022461,
72
+ 59.77000427246094,
73
+ 59.920005798339844,
74
+ 61.190006256103516,
75
+ 61.96000671386719,
76
+ 62.50000762939453,
77
+ 63.3700065612793,
78
+ 63.48000717163086,
79
+ 63.48000717163086,
80
+ 63.66000747680664,
81
+ 63.850006103515625,
82
+ 64.08000946044922,
83
+ 64.760009765625,
84
+ 64.80001068115234,
85
+ 64.81001281738281,
86
+ 64.81001281738281
87
+ ],
88
+ "short_factor": [
89
+ 1.05,
90
+ 1.05,
91
+ 1.05,
92
+ 1.1,
93
+ 1.1,
94
+ 1.1,
95
+ 1.2500000000000002,
96
+ 1.2500000000000002,
97
+ 1.4000000000000004,
98
+ 1.4500000000000004,
99
+ 1.5500000000000005,
100
+ 1.8500000000000008,
101
+ 1.9000000000000008,
102
+ 2.000000000000001,
103
+ 2.000000000000001,
104
+ 2.000000000000001,
105
+ 2.000000000000001,
106
+ 2.000000000000001,
107
+ 2.000000000000001,
108
+ 2.000000000000001,
109
+ 2.000000000000001,
110
+ 2.000000000000001,
111
+ 2.000000000000001,
112
+ 2.000000000000001,
113
+ 2.000000000000001,
114
+ 2.000000000000001,
115
+ 2.000000000000001,
116
+ 2.000000000000001,
117
+ 2.000000000000001,
118
+ 2.000000000000001,
119
+ 2.000000000000001,
120
+ 2.000000000000001,
121
+ 2.1000000000000005,
122
+ 2.1000000000000005,
123
+ 2.2,
124
+ 2.3499999999999996,
125
+ 2.3499999999999996,
126
+ 2.3499999999999996,
127
+ 2.3499999999999996,
128
+ 2.3999999999999995,
129
+ 2.3999999999999995,
130
+ 2.6499999999999986,
131
+ 2.6999999999999984,
132
+ 2.8999999999999977,
133
+ 2.9499999999999975,
134
+ 3.049999999999997,
135
+ 3.049999999999997,
136
+ 3.049999999999997
137
+ ],
138
+ "type": "su"
139
+ },
140
+ "rope_theta": 10000.0,
141
+ "sliding_window": 131072,
142
+ "tie_word_embeddings": false,
143
+ "torch_dtype": "bfloat16",
144
+ "transformers_version": "4.38.1",
145
+ "use_cache": true,
146
+ "vocab_size": 32064,
147
+ "_attn_implementation": "flash_attention_2"
148
+ }
configuration_phi3_v.py ADDED
@@ -0,0 +1,217 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """ Phi-3-V model configuration"""
17
+
18
+
19
+ from transformers.configuration_utils import PretrainedConfig
20
+ from transformers.utils import logging
21
+
22
+
23
+ logger = logging.get_logger(__name__)
24
+
25
+ PHI3V_PRETRAINED_CONFIG_ARCHIVE_MAP = {
26
+ "microsoft/Phi-3-vision-128k-instruct": "https://huggingface.co/microsoft/Phi-3-vision-128k-instruct/resolve/main/config.json",
27
+ }
28
+
29
+
30
+ class Phi3VConfig(PretrainedConfig):
31
+ r"""
32
+ This is the configuration class to store the configuration of a [`Phi3VModel`]. It is used to instantiate a Phi-3
33
+ model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
34
+ defaults will yield a similar configuration to that of the
35
+ [microsoft/Phi-3-vision-128k-instruct](https://huggingface.co/microsoft/Phi-3-vision-128k-instruct).
36
+
37
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
38
+ documentation from [`PretrainedConfig`] for more information.
39
+
40
+ Args:
41
+ vocab_size (`int`, *optional*, defaults to 32064):
42
+ Vocabulary size of the Phi-3-V model. Defines the number of different tokens that can be represented by the
43
+ `inputs_ids` passed when calling [`Phi3VModel`].
44
+ hidden_size (`int`, *optional*, defaults to 3072):
45
+ Dimension of the hidden representations.
46
+ intermediate_size (`int`, *optional*, defaults to 8192):
47
+ Dimension of the MLP representations.
48
+ num_hidden_layers (`int`, *optional*, defaults to 32):
49
+ Number of hidden layers in the Transformer decoder.
50
+ num_attention_heads (`int`, *optional*, defaults to 32):
51
+ Number of attention heads for each attention layer in the Transformer decoder.
52
+ num_key_value_heads (`int`, *optional*):
53
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
54
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
55
+ `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
56
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
57
+ by meanpooling all the original heads within that group. For more details checkout [this
58
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
59
+ `num_attention_heads`.
60
+ resid_pdrop (`float`, *optional*, defaults to 0.0):
61
+ Dropout probability for mlp outputs.
62
+ embd_pdrop (`int`, *optional*, defaults to 0.0):
63
+ The dropout ratio for the embeddings.
64
+ attention_dropout (`float`, *optional*, defaults to 0.0):
65
+ The dropout ratio after computing the attention scores.
66
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
67
+ The non-linear activation function (function or string) in the decoder.
68
+ max_position_embeddings (`int`, *optional*, defaults to 4096):
69
+ The maximum sequence length that this model might ever be used with.
70
+ original_max_position_embeddings (`int`, *optional*, defaults to 4096):
71
+ The maximum sequence length that this model was trained with. This is used to determine the size of the
72
+ original RoPE embeddings when using long scaling.
73
+ initializer_range (`float`, *optional*, defaults to 0.02):
74
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
75
+ rms_norm_eps (`float`, *optional*, defaults to 1e-05):
76
+ The epsilon value used for the RMSNorm.
77
+ use_cache (`bool`, *optional*, defaults to `True`):
78
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
79
+ relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.
80
+ tie_word_embeddings (`bool`, *optional*, defaults to `False`):
81
+ Whether to tie weight embeddings
82
+ rope_theta (`float`, *optional*, defaults to 10000.0):
83
+ The base period of the RoPE embeddings.
84
+ rope_scaling (`dict`, *optional*):
85
+ The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must
86
+ contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be either `su` or `yarn` and
87
+ the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size
88
+ divided by the number of attention heads divided by 2.
89
+ bos_token_id (`int`, *optional*, defaults to 1):
90
+ The id of the "beginning-of-sequence" token.
91
+ eos_token_id (`int`, *optional*, defaults to 32000):
92
+ The id of the "end-of-sequence" token.
93
+ pad_token_id (`int`, *optional*, defaults to 32000):
94
+ The id of the padding token.
95
+ sliding_window (`int`, *optional*):
96
+ Sliding window attention window size. If `None`, no sliding window is applied.
97
+ embd_layer (`str`, *optional*, defaults to `"default"`):
98
+ The embedding layer to use. Can be either `"default"` or `"image"`. "default" uses the standard embedding for text.
99
+
100
+ Example:
101
+
102
+ ```python
103
+ >>> from transformers import Phi3VModel, Phi3VConfig
104
+
105
+ >>> # Initializing a Phi-3-V style configuration
106
+ >>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-vision-128k-instruct")
107
+
108
+ >>> # Initializing a model from the configuration
109
+ >>> model = Phi3VModel(configuration)
110
+
111
+ >>> # Accessing the model configuration
112
+ >>> configuration = model.config
113
+ ```"""
114
+
115
+ model_type = "phi3_v"
116
+ keys_to_ignore_at_inference = ["past_key_values"]
117
+
118
+ def __init__(
119
+ self,
120
+ vocab_size=32064,
121
+ hidden_size=3072,
122
+ intermediate_size=8192,
123
+ num_hidden_layers=32,
124
+ num_attention_heads=32,
125
+ num_key_value_heads=None,
126
+ resid_pdrop=0.0,
127
+ embd_pdrop=0.0,
128
+ attention_dropout=0.0,
129
+ hidden_act="silu",
130
+ max_position_embeddings=4096,
131
+ original_max_position_embeddings=4096,
132
+ initializer_range=0.02,
133
+ rms_norm_eps=1e-5,
134
+ use_cache=True,
135
+ tie_word_embeddings=False,
136
+ rope_theta=10000.0,
137
+ rope_scaling=None,
138
+ bos_token_id=1,
139
+ eos_token_id=32000,
140
+ pad_token_id=32000,
141
+ sliding_window=None,
142
+ embd_layer: str = "default",
143
+ **kwargs,
144
+ ):
145
+ self.vocab_size = vocab_size
146
+ self.hidden_size = hidden_size
147
+ self.intermediate_size = intermediate_size
148
+ self.num_hidden_layers = num_hidden_layers
149
+ self.num_attention_heads = num_attention_heads
150
+
151
+ if num_key_value_heads is None:
152
+ num_key_value_heads = num_attention_heads
153
+
154
+ self.num_key_value_heads = num_key_value_heads
155
+ self.resid_pdrop = resid_pdrop
156
+ self.embd_pdrop = embd_pdrop
157
+ self.attention_dropout = attention_dropout
158
+ self.hidden_act = hidden_act
159
+ self.max_position_embeddings = max_position_embeddings
160
+ self.original_max_position_embeddings = original_max_position_embeddings
161
+ self.initializer_range = initializer_range
162
+ self.rms_norm_eps = rms_norm_eps
163
+ self.use_cache = use_cache
164
+ self.rope_theta = rope_theta
165
+ self.rope_scaling = rope_scaling
166
+ self._rope_scaling_validation()
167
+ self.sliding_window = sliding_window
168
+ self.embd_layer = embd_layer
169
+
170
+
171
+ super().__init__(
172
+ bos_token_id=bos_token_id,
173
+ eos_token_id=eos_token_id,
174
+ pad_token_id=pad_token_id,
175
+ tie_word_embeddings=tie_word_embeddings,
176
+ **kwargs,
177
+ )
178
+
179
+ def _rope_scaling_validation(self):
180
+ """
181
+ Validate the `rope_scaling` configuration.
182
+ """
183
+ if self.rope_scaling is None:
184
+ return
185
+
186
+ if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:
187
+ raise ValueError(
188
+ "`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "
189
+ f"got {self.rope_scaling}"
190
+ )
191
+ rope_scaling_type = self.rope_scaling.get("type", None)
192
+ rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)
193
+ rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)
194
+ if rope_scaling_type is None or rope_scaling_type not in ["su", "yarn"]:
195
+ raise ValueError(f"`rope_scaling`'s type field must be one of ['su', 'yarn'], got {rope_scaling_type}")
196
+ if not (
197
+ isinstance(rope_scaling_short_factor, list)
198
+ and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)
199
+ ):
200
+ raise ValueError(
201
+ f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"
202
+ )
203
+ if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:
204
+ raise ValueError(
205
+ f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"
206
+ )
207
+ if not (
208
+ isinstance(rope_scaling_long_factor, list)
209
+ and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)
210
+ ):
211
+ raise ValueError(
212
+ f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"
213
+ )
214
+ if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:
215
+ raise ValueError(
216
+ f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"
217
+ )
image_embedding_phi3_v.py ADDED
@@ -0,0 +1,301 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ import math
17
+ import torch
18
+ import torch.nn as nn
19
+ from transformers import CLIPVisionModel, PretrainedConfig
20
+ from transformers import CLIPVisionConfig
21
+ from transformers.utils import logging
22
+ from datetime import datetime
23
+
24
+ logger = logging.get_logger(__name__)
25
+
26
+ CLIP_VIT_LARGE_PATCH14_336_CONFIG = CLIPVisionConfig(
27
+ attention_dropout=0.0,
28
+ dropout=0.0,
29
+ hidden_act="quick_gelu",
30
+ hidden_size=1024,
31
+ image_size=336,
32
+ initializer_factor=1.0,
33
+ initializer_range=0.02,
34
+ intermediate_size=4096,
35
+ layer_norm_eps=1e-05,
36
+ num_attention_heads=16,
37
+ num_channels=3,
38
+ num_hidden_layers=24,
39
+ patch_size=14,
40
+ projection_dim=768
41
+ )
42
+
43
+ class Phi3ImageEmbedding(nn.Module):
44
+ """Phi3 Image embedding."""
45
+
46
+ def __init__(self, config: PretrainedConfig, wte=None, **kwargs) -> None:
47
+ super().__init__()
48
+
49
+ # n_embed or hidden_size
50
+ hidden_size = config.n_embd if hasattr(config, 'n_embd') else config.hidden_size
51
+ if hasattr(config, 'embd_pdrop') or hasattr(config, 'embed_pdrop'):
52
+ embd_drop = config.embd_pdrop if hasattr(config, 'embd_pdrop') else config.embed_pdrop
53
+ self.drop = nn.Dropout(embd_drop)
54
+ else:
55
+ self.drop = None
56
+
57
+ self.wte = wte
58
+
59
+ if isinstance(config.img_processor, dict) and config.img_processor.get('name', None) == 'clip_vision_model':
60
+ assert 'model_name' in config.img_processor, 'model_name must be provided for CLIPVisionModel'
61
+ assert 'image_dim_out' in config.img_processor, 'image_dim_out must be provided for CLIPVisionModel'
62
+ assert 'num_img_tokens' in config.img_processor, 'num_img_tokens must be provided for CLIPVisionModel'
63
+ assert config.img_processor['model_name'] == 'openai/clip-vit-large-patch14-336'
64
+ clip_config = CLIP_VIT_LARGE_PATCH14_336_CONFIG
65
+ self.img_processor = CLIPVisionModel(clip_config)
66
+ image_dim_out = config.img_processor['image_dim_out']
67
+ self.num_img_tokens = config.img_processor['num_img_tokens']
68
+ else:
69
+ raise NotImplementedError(f'img_processor = {config.img_processor}, not implemented')
70
+
71
+ self.image_dim_out = image_dim_out
72
+ self.img_sizes = None
73
+
74
+ # global_gn and sub_gn for hd transform, serves as line separator
75
+ self.use_hd_transform = kwargs.get('use_hd_transform', False)
76
+ self.with_learnable_separator = kwargs.get('with_learnable_separator', False)
77
+ self.hd_transform_order = kwargs.get('hd_transform_order', 'glb_sub')
78
+ # with_hd_transform and with_learnable_separator should have same value
79
+ assert self.use_hd_transform == self.with_learnable_separator, 'use_hd_transform and with_learnable_separator should have same value'
80
+ if self.with_learnable_separator:
81
+ assert self.use_hd_transform, 'learnable separator is only for hd transform'
82
+ # 1024 * 4, merge spatial to channel dimension
83
+ self.glb_GN = nn.Parameter(torch.zeros([1, 1, self.image_dim_out * 4]))
84
+ self.sub_GN = nn.Parameter(torch.zeros([1, 1, 1, self.image_dim_out * 4]))
85
+ logger.info(f'learnable separator enabled for hd transform, hd_transform_order = {self.hd_transform_order}')
86
+
87
+ projection_cls = kwargs.get('projection_cls', 'linear')
88
+ if projection_cls == 'linear':
89
+ self.img_projection = nn.Linear(image_dim_out, hidden_size)
90
+ elif projection_cls == 'mlp' and self.use_hd_transform:
91
+ dim_projection = hidden_size
92
+ depth = 2
93
+ layers = [nn.Linear(image_dim_out * 4, dim_projection)]
94
+ for _ in range(1, depth):
95
+ layers.extend([nn.GELU(),
96
+ nn.Linear(dim_projection, dim_projection)])
97
+ self.img_projection = nn.Sequential(*layers)
98
+ elif projection_cls == 'mlp':
99
+ dim_projection = hidden_size
100
+ depth = 2
101
+ layers = [nn.Linear(image_dim_out, dim_projection)]
102
+ for _ in range(1, depth):
103
+ layers.extend([nn.GELU(),
104
+ nn.Linear(dim_projection, dim_projection)])
105
+ self.img_projection = nn.Sequential(*layers)
106
+ else:
107
+ raise NotImplementedError(f'projection_cls = {projection_cls}, not implemented')
108
+
109
+ self.vocab_size = config.vocab_size
110
+ self.img_features = None
111
+
112
+ if isinstance(config.img_processor, dict):
113
+ self.layer_idx = config.img_processor.get('layer_idx', -2)
114
+ self.type_feature = config.img_processor.get('type_feature', 'patch')
115
+ else:
116
+ self.layer_idx = -2
117
+ self.type_feature = 'patch'
118
+
119
+
120
+ def set_img_features(self, img_features: torch.FloatTensor) -> None:
121
+ self.img_features = img_features
122
+
123
+ def set_img_sizes(self, img_sizes: torch.LongTensor) -> None:
124
+ self.img_sizes = img_sizes
125
+
126
+ def get_img_features(self, img_embeds: torch.FloatTensor) -> torch.FloatTensor:
127
+ LAYER_IDX = self.layer_idx
128
+ TYPE_FEATURE = self.type_feature
129
+
130
+ img_processor_output = self.img_processor(img_embeds, output_hidden_states=True)
131
+ img_feature = img_processor_output.hidden_states[LAYER_IDX]
132
+
133
+ if TYPE_FEATURE == "patch":
134
+ patch_feature = img_feature[:, 1:]
135
+ return patch_feature
136
+
137
+ if TYPE_FEATURE == "cls_patch":
138
+ return img_feature
139
+
140
+ raise NotImplementedError
141
+
142
+ def forward(self, input_ids: torch.LongTensor, pixel_values: torch.FloatTensor, image_sizes=None) -> torch.FloatTensor:
143
+
144
+ MAX_INPUT_ID = int(1e9)
145
+ img_embeds = pixel_values
146
+ img_sizes = image_sizes
147
+
148
+ if self.img_features is not None:
149
+ img_embeds = self.img_features.clone()
150
+ self.img_features = None
151
+
152
+ if self.img_sizes is not None:
153
+ img_sizes = self.img_sizes
154
+
155
+ input_shape = input_ids.size()
156
+ input_ids = input_ids.view(-1, input_shape[-1])
157
+
158
+ with torch.no_grad():
159
+ positions = torch.nonzero((input_ids < 0) & (input_ids > -MAX_INPUT_ID), as_tuple=False)
160
+
161
+ select = False
162
+
163
+ if isinstance(self.img_projection, nn.Sequential):
164
+ target_device = self.img_projection[0].bias.device
165
+ target_dtype = self.img_projection[0].bias.dtype
166
+ else: # It's a single nn.Linear layer
167
+ target_device = self.img_projection.bias.device
168
+ target_dtype = self.img_projection.bias.dtype
169
+
170
+ if len(positions.tolist()) > 0:
171
+ with torch.no_grad():
172
+ g_values = abs(input_ids[positions[:, 0], positions[:, 1]])
173
+
174
+ if self.use_hd_transform and img_sizes is not None and len(img_sizes):
175
+ hd_transform = True
176
+ assert img_embeds.ndim == 5, f'img_embeds size: {img_embeds.size()}, expect 5D tensor for hd transform'
177
+ # img_embeds: (num_images, max_num_crops, 3, H, W)
178
+ # img_sizes: (num_images, 2).view(1, -1)
179
+
180
+ start_time = datetime.now()
181
+ bs = img_embeds.shape[0]
182
+ # Nx(HW)xC
183
+ img_features = self.get_img_features(img_embeds.flatten(0, 1))
184
+ base_feat_height = base_feat_width = int(img_features.shape[1] ** 0.5)
185
+
186
+ assert base_feat_height == 24 and base_feat_width == 24, f'base_feat_height: {base_feat_height}, base_feat_width: {base_feat_width}, expect 24x24 features for hd transform'
187
+
188
+ # bs x max_num_crops x (24x24) x C
189
+ img_features = img_features.view(bs, -1, base_feat_height * base_feat_width, self.image_dim_out)
190
+ C = self.image_dim_out
191
+ H = base_feat_height
192
+
193
+ output_imgs = []
194
+ output_len = []
195
+ # training is tensor, inference is list
196
+ if isinstance(img_sizes, torch.Tensor):
197
+ img_sizes = img_sizes.view(-1, 2)
198
+ for _bs in range(bs):
199
+ h, w = img_sizes[_bs]
200
+ h = h // 336
201
+ w = w // 336
202
+ B_ = h * w
203
+
204
+ # 1 x (24x24) x 1024
205
+ global_img_feature = img_features[_bs, :1]
206
+
207
+ # 1 x 12 x 12 x 4096
208
+ glb_img = global_img_feature.reshape(1,H,H,C).reshape(1,H//2,2,H//2,2,C).contiguous().permute(0,1,3,2,4,5).reshape(1,H//2,H//2,4*C).contiguous()
209
+ temp_glb_GN = self.sub_GN.repeat(1, H//2, 1, 1)
210
+
211
+ # 1 x 156 x 4096
212
+ glb_img = torch.cat([glb_img, temp_glb_GN], dim=2).reshape(1,-1,4*C)
213
+
214
+ # (max_num_crops-1) x (12x12) x C
215
+ sub_img = img_features[_bs, 1:]
216
+ # 16x574x1024
217
+ # get rid of padding sub_img
218
+ sub_img = sub_img[:B_]
219
+
220
+ # (num_crops, 12, 2, 12, 2, 1024) -> (num_crops, 12, 12, 2, 2, 1024) -> (num_crops, 12*12, 4*1024)
221
+ sub_img = sub_img.reshape(B_,H,H,C).reshape(B_,H//2,2,H//2,2,C).contiguous().permute(0,1,3,2,4,5).reshape(B_,-1,4*C).contiguous()
222
+ sub_img = sub_img.reshape(1, h, w, 12, 12, -1).permute(0,1,3,2,4,5).reshape(1,h*12,w*12,4*C)
223
+ temp_sub_GN = self.sub_GN.repeat(1, h*12, 1, 1)
224
+ sub_img = torch.cat([sub_img, temp_sub_GN], dim=2).reshape(1,-1,4*C)
225
+ # (1, num_img_tokens, 1024*4)
226
+
227
+ # glb + sub
228
+ if self.hd_transform_order == 'glb_sub':
229
+ output_imgs.append(torch.cat([glb_img, self.glb_GN, sub_img], dim=1))
230
+ elif self.hd_transform_order == 'sub_glb':
231
+ output_imgs.append(torch.cat([sub_img, self.glb_GN, glb_img], dim=1))
232
+ else:
233
+ raise NotImplementedError(f'hd_transform_order = {self.hd_transform_order}, not implemented')
234
+
235
+ temp_len = int((h*w+1)*144 + 1 + (h+1)*12)
236
+ assert temp_len == output_imgs[-1].shape[1], f'temp_len: {temp_len}, output_imgs[-1].shape[1]: {output_imgs[-1].shape[1]}'
237
+ output_len.append(temp_len)
238
+
239
+ num_img_tokens = output_len
240
+ img_set_tensor = []
241
+ for _output_img in output_imgs:
242
+ img_feature_proj = self.img_projection(_output_img.to(target_device).to(target_dtype))
243
+ img_set_tensor.append(img_feature_proj)
244
+ logger.info(f'img_embeds size: {img_embeds.size()}, image sizes: {img_sizes} loading time {datetime.now() - start_time}')
245
+ elif img_embeds.ndim == 4:
246
+ selected_g_values = g_values[::self.num_img_tokens]
247
+ assert len(img_embeds) == len(selected_g_values), f'img_embeds size: {img_embeds.size()}, selected_g_values size: {len(selected_g_values)}, selected_g_value {selected_g_values}'
248
+ start_time = datetime.now()
249
+ tt = (
250
+ self.get_img_features(img_embeds)
251
+ .to(target_device)
252
+ .to(target_dtype)
253
+ .reshape(-1, self.image_dim_out)
254
+ )
255
+ logger.info(f'img_embeds size: {img_embeds.size()}, loading time {datetime.now() - start_time}')
256
+ img_set_tensor = self.img_projection(tt) # adapted visual features.
257
+ elif img_embeds.ndim == 3:
258
+ selected_g_values = g_values[::self.num_img_tokens]
259
+ assert len(img_embeds) == len(selected_g_values), f'img_embeds size: {img_embeds.size()}, selected_g_values size: {len(selected_g_values)}, selected_g_value {selected_g_values}'
260
+ tt = (
261
+ img_embeds
262
+ .to(target_device)
263
+ .to(target_dtype)
264
+ .view(-1, self.image_dim_out)
265
+ )
266
+ img_set_tensor = self.img_projection(tt) # adapted visual features.
267
+ else:
268
+ raise NotImplementedError
269
+ select = True
270
+
271
+ with torch.no_grad():
272
+ input_ids.clamp_min_(0).clamp_max_(self.vocab_size)
273
+
274
+ hidden_states = self.wte(input_ids)
275
+
276
+ if select:
277
+ if hd_transform:
278
+ idx = 0
279
+ for i, cnt in enumerate(num_img_tokens):
280
+ hidden_states[positions[idx, 0], positions[idx, 1] : positions[idx, 1] + cnt] = (
281
+ img_set_tensor[i]
282
+ .to(hidden_states.dtype)
283
+ .to(hidden_states.device)
284
+ )
285
+ idx += cnt
286
+ else:
287
+ idx = 0
288
+ assert len(selected_g_values) * self.num_img_tokens == len(img_set_tensor), f'len(selected_g_values) * self.num_img_tokens = {len(selected_g_values) * self.num_img_tokens}, len(img_set_tensor) = {len(img_set_tensor)}'
289
+ for i, g in enumerate(selected_g_values):
290
+ cnt = self.num_img_tokens
291
+ hidden_states[positions[idx, 0], positions[idx, 1] : positions[idx, 1] + cnt] = (
292
+ img_set_tensor[i * cnt : (i + 1) * cnt]
293
+ .to(hidden_states.dtype)
294
+ .to(hidden_states.device)
295
+ )
296
+ idx += cnt
297
+
298
+ if self.drop is not None:
299
+ hidden_states = self.drop(hidden_states)
300
+
301
+ return hidden_states
image_processing_phi3_v.py ADDED
@@ -0,0 +1,274 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """Image processor class for Phi3-V."""
17
+
18
+ from typing import List, Optional, Union
19
+
20
+ import numpy as np
21
+
22
+ from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
23
+ from transformers.image_transforms import (
24
+ convert_to_rgb,
25
+ )
26
+ from transformers.image_utils import (
27
+ OPENAI_CLIP_MEAN,
28
+ OPENAI_CLIP_STD,
29
+ ImageInput,
30
+ make_list_of_images,
31
+ valid_images,
32
+ )
33
+ from transformers.utils import TensorType, is_vision_available, logging
34
+
35
+ from transformers import AutoImageProcessor
36
+
37
+ logger = logging.get_logger(__name__)
38
+
39
+
40
+ if is_vision_available():
41
+ from PIL import Image
42
+
43
+ import torch
44
+ import torchvision
45
+
46
+ def padding_336(b):
47
+ width, height = b.size
48
+ tar = int(np.ceil(height / 336) * 336)
49
+ top_padding = int((tar - height)/2)
50
+ bottom_padding = tar - height - top_padding
51
+ left_padding = 0
52
+ right_padding = 0
53
+ b = torchvision.transforms.functional.pad(b, [left_padding, top_padding, right_padding, bottom_padding], fill=[255,255,255])
54
+
55
+ return b
56
+
57
+ def calc_padded_size(width, height, padding_unit=336):
58
+ target_height = int(np.ceil(height / padding_unit) * padding_unit)
59
+ top_padding = int((target_height - height) / 2)
60
+ bottom_padding = target_height - height - top_padding
61
+ left_padding = 0
62
+ right_padding = 0
63
+ padded_width = width + left_padding + right_padding
64
+ padded_height = height + top_padding + bottom_padding
65
+ return padded_width, padded_height
66
+
67
+ def HD_transform(img, hd_num=16):
68
+ width, height = img.size
69
+ trans = False
70
+ if width < height:
71
+ img = img.transpose(Image.TRANSPOSE)
72
+ trans = True
73
+ width, height = img.size
74
+ ratio = (width/ height)
75
+ scale = 1
76
+ while scale*np.ceil(scale/ratio) <= hd_num:
77
+ scale += 1
78
+ scale -= 1
79
+ new_w = int(scale * 336)
80
+ new_h = int(new_w / ratio)
81
+
82
+ img = torchvision.transforms.functional.resize(img, [new_h, new_w],)
83
+ img = padding_336(img)
84
+ width, height = img.size
85
+ if trans:
86
+ img = img.transpose(Image.TRANSPOSE)
87
+
88
+ return img
89
+
90
+ def calc_hd_transform_size(width, height, hd_num=16):
91
+ transposed = False
92
+ if width < height:
93
+ width, height = height, width
94
+ transposed = True
95
+
96
+ ratio = width / height
97
+ scale = 1
98
+ while scale * np.ceil(scale / ratio) <= hd_num:
99
+ scale += 1
100
+ scale -= 1
101
+
102
+ new_width = int(scale * 336)
103
+ new_height = int(new_width / ratio)
104
+
105
+ padded_width, padded_height = calc_padded_size(new_width, new_height)
106
+
107
+ if transposed:
108
+ padded_width, padded_height = padded_height, padded_width
109
+
110
+ return padded_width, padded_height
111
+
112
+ def pad_to_max_num_crops_tensor(images, max_crops=5):
113
+ """
114
+ images: B x 3 x H x W, B<=max_crops
115
+ """
116
+ B, _, H, W = images.shape
117
+ if B < max_crops:
118
+ pad = torch.zeros(max_crops - B, 3, H, W, dtype=images.dtype, device=images.device)
119
+ images = torch.cat([images, pad], dim=0)
120
+ return images
121
+
122
+
123
+ class Phi3VImageProcessor(BaseImageProcessor):
124
+ r"""
125
+ Constructs a Phi3 image processor. Based on [`CLIPImageProcessor`] with incorporation of additional techniques
126
+ for processing high resolution images as explained in the [InternLM-XComposer2-4KHD](https://arxiv.org/abs/2401.16420)
127
+
128
+ Args:
129
+ image_mean (`float` or `List[float]`, *optional*, defaults to `[0.48145466, 0.4578275, 0.40821073]`):
130
+ Mean to use if normalizing the image. This is a float or list of floats the length of the number of
131
+ channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method.
132
+ image_std (`float` or `List[float]`, *optional*, defaults to `[0.26862954, 0.26130258, 0.27577711]`):
133
+ Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
134
+ number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
135
+ Can be overridden by the `image_std` parameter in the `preprocess` method.
136
+ do_convert_rgb (`bool`, *optional*, defaults to `True`):
137
+ Whether to convert the image to RGB.
138
+ """
139
+
140
+ model_input_names = ["pixel_values"]
141
+
142
+ def __init__(
143
+ self,
144
+ num_crops: int = 1,
145
+ image_mean: Optional[Union[float, List[float]]] = None,
146
+ image_std: Optional[Union[float, List[float]]] = None,
147
+ do_convert_rgb: bool = True,
148
+ **kwargs,
149
+ ) -> None:
150
+ super().__init__(**kwargs)
151
+ self.num_crops = num_crops
152
+ self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
153
+ self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
154
+ self.do_convert_rgb = do_convert_rgb
155
+
156
+ def calc_num_image_tokens(
157
+ self,
158
+ images: ImageInput
159
+ ):
160
+ """ Calculate the number of image tokens for each image.
161
+ Args:
162
+ images (`ImageInput`):
163
+ Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
164
+ passing in images with pixel values between 0 and 1, set `do_rescale=False`.
165
+ """
166
+ images = make_list_of_images(images)
167
+
168
+ if not valid_images(images):
169
+ raise ValueError(
170
+ "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
171
+ "torch.Tensor, tf.Tensor or jax.ndarray."
172
+ )
173
+
174
+ images = [image.convert('RGB') for image in images]
175
+ # (H, W, C)
176
+ elems = [HD_transform(im, hd_num = self.num_crops) for im in images]
177
+ shapes = [[im.size[1], im.size[0]] for im in elems]
178
+ num_img_tokens = [int((h//336*w//336+1)*144 + 1 + (h//336+1)*12) for h, w in shapes]
179
+ return num_img_tokens
180
+
181
+ def calc_num_image_tokens_from_image_size(self, width, height):
182
+ """
183
+ Calculate the number of image tokens for a given image size.
184
+ Args:
185
+ width (`int`): Width of the image.
186
+ height (`int`): Height of the image.
187
+ """
188
+ new_width, new_height = calc_hd_transform_size(width, height, hd_num=self.num_crops)
189
+ num_img_tokens = int((new_height // 336 * new_width // 336 + 1) * 144 + 1 + (new_height // 336 + 1) * 12)
190
+ return num_img_tokens
191
+
192
+ def preprocess(
193
+ self,
194
+ images: ImageInput,
195
+ image_mean: Optional[Union[float, List[float]]] = None,
196
+ image_std: Optional[Union[float, List[float]]] = None,
197
+ do_convert_rgb: bool = None,
198
+ return_tensors: Optional[Union[str, TensorType]] = None,
199
+ ):
200
+ """
201
+ Args:
202
+ images (`ImageInput`):
203
+ Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
204
+ passing in images with pixel values between 0 and 1, set `do_rescale=False`.
205
+ image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
206
+ Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
207
+ image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
208
+ Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
209
+ `True`.
210
+ do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
211
+ Whether to convert the image to RGB.
212
+ return_tensors (`str` or `TensorType`, *optional*):
213
+ The type of tensors to return. Can be one of:
214
+ - Unset: Return a list of `np.ndarray`.
215
+ - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
216
+ - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
217
+ - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
218
+ - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
219
+ """
220
+ image_mean = image_mean if image_mean is not None else self.image_mean
221
+ image_std = image_std if image_std is not None else self.image_std
222
+ do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
223
+
224
+ images = make_list_of_images(images)
225
+
226
+ if not valid_images(images):
227
+ raise ValueError(
228
+ "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
229
+ "torch.Tensor, tf.Tensor or jax.ndarray."
230
+ )
231
+
232
+ if do_convert_rgb:
233
+ images = [convert_to_rgb(image) for image in images]
234
+
235
+ image_sizes = []
236
+ img_processor = torchvision.transforms.Compose([
237
+ torchvision.transforms.ToTensor(),
238
+ torchvision.transforms.Normalize(image_mean, image_std)
239
+ ])
240
+
241
+ # PIL images
242
+ # HD_transform pad images to size of multiiply of 336, 336
243
+ # convert to RGB first
244
+ images = [image.convert('RGB') for image in images]
245
+ elems = [HD_transform(im, hd_num = self.num_crops) for im in images]
246
+ # tensor transform and normalize
247
+ hd_images = [img_processor(im) for im in elems]
248
+ # create global image
249
+ global_image = [torch.nn.functional.interpolate(im.unsqueeze(0).float(), size=(336, 336), mode='bicubic',).to(im.dtype) for im in hd_images]
250
+
251
+ # [(3, h, w)], where h, w is multiple of 336
252
+ shapes = [[im.size(1), im.size(2)] for im in hd_images]
253
+ num_img_tokens = [int((h//336*w//336+1)*144 + 1 + (h//336+1)*12) for h, w in shapes]
254
+ # reshape to channel dimension -> (num_images, num_crops, 3, 336, 336)
255
+ # (1, 3, h//336, 336, w//336, 336) -> (1, h//336, w//336, 3, 336, 336) -> (h//336*w//336, 3, 336, 336)
256
+ hd_images_reshape = [im.reshape(1, 3, h//336, 336, w//336, 336).permute(0,2,4,1,3,5).reshape(-1, 3, 336, 336).contiguous() for im, (h, w) in zip(hd_images, shapes)]
257
+ # concat global image and local image
258
+ hd_images_reshape = [torch.cat([_global_image] + [_im], dim=0) for _global_image, _im in zip(global_image, hd_images_reshape)]
259
+
260
+ # pad to max_num_crops
261
+ image_transformed = [pad_to_max_num_crops_tensor(im, self.num_crops+1) for im in hd_images_reshape]
262
+ image_transformed = torch.stack(image_transformed, dim=0)
263
+ image_sizes = [torch.LongTensor(_shapes) for _shapes in shapes]
264
+ padded_images = image_transformed
265
+ image_sizes = shapes
266
+
267
+ data = {"pixel_values": padded_images,
268
+ "image_sizes": image_sizes,
269
+ "num_img_tokens": num_img_tokens
270
+ }
271
+
272
+ return BatchFeature(data=data, tensor_type=return_tensors)
273
+
274
+ AutoImageProcessor.register("Phi3VImageProcessor", Phi3VImageProcessor)
model-00001-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:175b2fe918dd8bd2549e3441615ee0c6d7b1f6d638c0104a614546f55c273482
3
+ size 4944122112
model-00002-of-00002.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e61ece5a8f0c9663afa06cc22799056f5cc084fb993518bf036dc8e268fd4c94
3
+ size 3349208776
model.safetensors.index.json ADDED
@@ -0,0 +1,599 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "metadata": {
3
+ "total_size": 8293242880
4
+ },
5
+ "weight_map": {
6
+ "lm_head.weight": "model-00002-of-00002.safetensors",
7
+ "model.embed_tokens.weight": "model-00001-of-00002.safetensors",
8
+ "model.layers.0.input_layernorm.weight": "model-00001-of-00002.safetensors",
9
+ "model.layers.0.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
10
+ "model.layers.0.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
11
+ "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
12
+ "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
13
+ "model.layers.0.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
14
+ "model.layers.1.input_layernorm.weight": "model-00001-of-00002.safetensors",
15
+ "model.layers.1.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
16
+ "model.layers.1.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
17
+ "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
18
+ "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
19
+ "model.layers.1.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
20
+ "model.layers.10.input_layernorm.weight": "model-00001-of-00002.safetensors",
21
+ "model.layers.10.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
22
+ "model.layers.10.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
23
+ "model.layers.10.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
24
+ "model.layers.10.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
25
+ "model.layers.10.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
26
+ "model.layers.11.input_layernorm.weight": "model-00001-of-00002.safetensors",
27
+ "model.layers.11.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
28
+ "model.layers.11.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
29
+ "model.layers.11.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
30
+ "model.layers.11.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
31
+ "model.layers.11.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
32
+ "model.layers.12.input_layernorm.weight": "model-00001-of-00002.safetensors",
33
+ "model.layers.12.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
34
+ "model.layers.12.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
35
+ "model.layers.12.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
36
+ "model.layers.12.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
37
+ "model.layers.12.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
38
+ "model.layers.13.input_layernorm.weight": "model-00001-of-00002.safetensors",
39
+ "model.layers.13.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
40
+ "model.layers.13.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
41
+ "model.layers.13.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
42
+ "model.layers.13.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
43
+ "model.layers.13.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
44
+ "model.layers.14.input_layernorm.weight": "model-00001-of-00002.safetensors",
45
+ "model.layers.14.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
46
+ "model.layers.14.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
47
+ "model.layers.14.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
48
+ "model.layers.14.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
49
+ "model.layers.14.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
50
+ "model.layers.15.input_layernorm.weight": "model-00001-of-00002.safetensors",
51
+ "model.layers.15.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
52
+ "model.layers.15.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
53
+ "model.layers.15.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
54
+ "model.layers.15.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
55
+ "model.layers.15.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
56
+ "model.layers.16.input_layernorm.weight": "model-00001-of-00002.safetensors",
57
+ "model.layers.16.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
58
+ "model.layers.16.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
59
+ "model.layers.16.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
60
+ "model.layers.16.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
61
+ "model.layers.16.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
62
+ "model.layers.17.input_layernorm.weight": "model-00001-of-00002.safetensors",
63
+ "model.layers.17.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
64
+ "model.layers.17.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
65
+ "model.layers.17.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
66
+ "model.layers.17.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
67
+ "model.layers.17.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
68
+ "model.layers.18.input_layernorm.weight": "model-00002-of-00002.safetensors",
69
+ "model.layers.18.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
70
+ "model.layers.18.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
71
+ "model.layers.18.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
72
+ "model.layers.18.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
73
+ "model.layers.18.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
74
+ "model.layers.19.input_layernorm.weight": "model-00002-of-00002.safetensors",
75
+ "model.layers.19.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
76
+ "model.layers.19.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
77
+ "model.layers.19.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
78
+ "model.layers.19.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
79
+ "model.layers.19.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
80
+ "model.layers.2.input_layernorm.weight": "model-00001-of-00002.safetensors",
81
+ "model.layers.2.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
82
+ "model.layers.2.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
83
+ "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
84
+ "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
85
+ "model.layers.2.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
86
+ "model.layers.20.input_layernorm.weight": "model-00002-of-00002.safetensors",
87
+ "model.layers.20.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
88
+ "model.layers.20.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
89
+ "model.layers.20.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
90
+ "model.layers.20.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
91
+ "model.layers.20.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
92
+ "model.layers.21.input_layernorm.weight": "model-00002-of-00002.safetensors",
93
+ "model.layers.21.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
94
+ "model.layers.21.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
95
+ "model.layers.21.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
96
+ "model.layers.21.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
97
+ "model.layers.21.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
98
+ "model.layers.22.input_layernorm.weight": "model-00002-of-00002.safetensors",
99
+ "model.layers.22.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
100
+ "model.layers.22.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
101
+ "model.layers.22.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
102
+ "model.layers.22.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
103
+ "model.layers.22.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
104
+ "model.layers.23.input_layernorm.weight": "model-00002-of-00002.safetensors",
105
+ "model.layers.23.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
106
+ "model.layers.23.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
107
+ "model.layers.23.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
108
+ "model.layers.23.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
109
+ "model.layers.23.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
110
+ "model.layers.24.input_layernorm.weight": "model-00002-of-00002.safetensors",
111
+ "model.layers.24.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
112
+ "model.layers.24.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
113
+ "model.layers.24.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
114
+ "model.layers.24.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
115
+ "model.layers.24.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
116
+ "model.layers.25.input_layernorm.weight": "model-00002-of-00002.safetensors",
117
+ "model.layers.25.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
118
+ "model.layers.25.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
119
+ "model.layers.25.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
120
+ "model.layers.25.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
121
+ "model.layers.25.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
122
+ "model.layers.26.input_layernorm.weight": "model-00002-of-00002.safetensors",
123
+ "model.layers.26.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
124
+ "model.layers.26.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
125
+ "model.layers.26.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
126
+ "model.layers.26.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
127
+ "model.layers.26.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
128
+ "model.layers.27.input_layernorm.weight": "model-00002-of-00002.safetensors",
129
+ "model.layers.27.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
130
+ "model.layers.27.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
131
+ "model.layers.27.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
132
+ "model.layers.27.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
133
+ "model.layers.27.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
134
+ "model.layers.28.input_layernorm.weight": "model-00002-of-00002.safetensors",
135
+ "model.layers.28.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
136
+ "model.layers.28.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
137
+ "model.layers.28.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
138
+ "model.layers.28.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
139
+ "model.layers.28.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
140
+ "model.layers.29.input_layernorm.weight": "model-00002-of-00002.safetensors",
141
+ "model.layers.29.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
142
+ "model.layers.29.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
143
+ "model.layers.29.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
144
+ "model.layers.29.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
145
+ "model.layers.29.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
146
+ "model.layers.3.input_layernorm.weight": "model-00001-of-00002.safetensors",
147
+ "model.layers.3.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
148
+ "model.layers.3.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
149
+ "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
150
+ "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
151
+ "model.layers.3.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
152
+ "model.layers.30.input_layernorm.weight": "model-00002-of-00002.safetensors",
153
+ "model.layers.30.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
154
+ "model.layers.30.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
155
+ "model.layers.30.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
156
+ "model.layers.30.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
157
+ "model.layers.30.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
158
+ "model.layers.31.input_layernorm.weight": "model-00002-of-00002.safetensors",
159
+ "model.layers.31.mlp.down_proj.weight": "model-00002-of-00002.safetensors",
160
+ "model.layers.31.mlp.gate_up_proj.weight": "model-00002-of-00002.safetensors",
161
+ "model.layers.31.post_attention_layernorm.weight": "model-00002-of-00002.safetensors",
162
+ "model.layers.31.self_attn.o_proj.weight": "model-00002-of-00002.safetensors",
163
+ "model.layers.31.self_attn.qkv_proj.weight": "model-00002-of-00002.safetensors",
164
+ "model.layers.4.input_layernorm.weight": "model-00001-of-00002.safetensors",
165
+ "model.layers.4.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
166
+ "model.layers.4.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
167
+ "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
168
+ "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
169
+ "model.layers.4.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
170
+ "model.layers.5.input_layernorm.weight": "model-00001-of-00002.safetensors",
171
+ "model.layers.5.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
172
+ "model.layers.5.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
173
+ "model.layers.5.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
174
+ "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
175
+ "model.layers.5.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
176
+ "model.layers.6.input_layernorm.weight": "model-00001-of-00002.safetensors",
177
+ "model.layers.6.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
178
+ "model.layers.6.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
179
+ "model.layers.6.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
180
+ "model.layers.6.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
181
+ "model.layers.6.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
182
+ "model.layers.7.input_layernorm.weight": "model-00001-of-00002.safetensors",
183
+ "model.layers.7.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
184
+ "model.layers.7.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
185
+ "model.layers.7.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
186
+ "model.layers.7.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
187
+ "model.layers.7.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
188
+ "model.layers.8.input_layernorm.weight": "model-00001-of-00002.safetensors",
189
+ "model.layers.8.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
190
+ "model.layers.8.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
191
+ "model.layers.8.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
192
+ "model.layers.8.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
193
+ "model.layers.8.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
194
+ "model.layers.9.input_layernorm.weight": "model-00001-of-00002.safetensors",
195
+ "model.layers.9.mlp.down_proj.weight": "model-00001-of-00002.safetensors",
196
+ "model.layers.9.mlp.gate_up_proj.weight": "model-00001-of-00002.safetensors",
197
+ "model.layers.9.post_attention_layernorm.weight": "model-00001-of-00002.safetensors",
198
+ "model.layers.9.self_attn.o_proj.weight": "model-00001-of-00002.safetensors",
199
+ "model.layers.9.self_attn.qkv_proj.weight": "model-00001-of-00002.safetensors",
200
+ "model.norm.weight": "model-00002-of-00002.safetensors",
201
+ "model.vision_embed_tokens.glb_GN": "model-00001-of-00002.safetensors",
202
+ "model.vision_embed_tokens.img_processor.vision_model.embeddings.class_embedding": "model-00001-of-00002.safetensors",
203
+ "model.vision_embed_tokens.img_processor.vision_model.embeddings.patch_embedding.weight": "model-00001-of-00002.safetensors",
204
+ "model.vision_embed_tokens.img_processor.vision_model.embeddings.position_embedding.weight": "model-00001-of-00002.safetensors",
205
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.layer_norm1.bias": "model-00001-of-00002.safetensors",
206
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.layer_norm1.weight": "model-00001-of-00002.safetensors",
207
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.layer_norm2.bias": "model-00001-of-00002.safetensors",
208
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.layer_norm2.weight": "model-00001-of-00002.safetensors",
209
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.mlp.fc1.bias": "model-00001-of-00002.safetensors",
210
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.mlp.fc1.weight": "model-00001-of-00002.safetensors",
211
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.mlp.fc2.bias": "model-00001-of-00002.safetensors",
212
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.mlp.fc2.weight": "model-00001-of-00002.safetensors",
213
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
214
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
215
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
216
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
217
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
218
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
219
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
220
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.0.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
221
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.layer_norm1.bias": "model-00001-of-00002.safetensors",
222
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.layer_norm1.weight": "model-00001-of-00002.safetensors",
223
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.layer_norm2.bias": "model-00001-of-00002.safetensors",
224
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.layer_norm2.weight": "model-00001-of-00002.safetensors",
225
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.mlp.fc1.bias": "model-00001-of-00002.safetensors",
226
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.mlp.fc1.weight": "model-00001-of-00002.safetensors",
227
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.mlp.fc2.bias": "model-00001-of-00002.safetensors",
228
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.mlp.fc2.weight": "model-00001-of-00002.safetensors",
229
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
230
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
231
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
232
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
233
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
234
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
235
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
236
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.1.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
237
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.layer_norm1.bias": "model-00001-of-00002.safetensors",
238
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.layer_norm1.weight": "model-00001-of-00002.safetensors",
239
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.layer_norm2.bias": "model-00001-of-00002.safetensors",
240
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.layer_norm2.weight": "model-00001-of-00002.safetensors",
241
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.mlp.fc1.bias": "model-00001-of-00002.safetensors",
242
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.mlp.fc1.weight": "model-00001-of-00002.safetensors",
243
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.mlp.fc2.bias": "model-00001-of-00002.safetensors",
244
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.mlp.fc2.weight": "model-00001-of-00002.safetensors",
245
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
246
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
247
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
248
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
249
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
250
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
251
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
252
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.10.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
253
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.layer_norm1.bias": "model-00001-of-00002.safetensors",
254
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.layer_norm1.weight": "model-00001-of-00002.safetensors",
255
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.layer_norm2.bias": "model-00001-of-00002.safetensors",
256
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.layer_norm2.weight": "model-00001-of-00002.safetensors",
257
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.mlp.fc1.bias": "model-00001-of-00002.safetensors",
258
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.mlp.fc1.weight": "model-00001-of-00002.safetensors",
259
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.mlp.fc2.bias": "model-00001-of-00002.safetensors",
260
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.mlp.fc2.weight": "model-00001-of-00002.safetensors",
261
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
262
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
263
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
264
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
265
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
266
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
267
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
268
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.11.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
269
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.layer_norm1.bias": "model-00001-of-00002.safetensors",
270
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.layer_norm1.weight": "model-00001-of-00002.safetensors",
271
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.layer_norm2.bias": "model-00001-of-00002.safetensors",
272
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.layer_norm2.weight": "model-00001-of-00002.safetensors",
273
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.mlp.fc1.bias": "model-00001-of-00002.safetensors",
274
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.mlp.fc1.weight": "model-00001-of-00002.safetensors",
275
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.mlp.fc2.bias": "model-00001-of-00002.safetensors",
276
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.mlp.fc2.weight": "model-00001-of-00002.safetensors",
277
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
278
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
279
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
280
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
281
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
282
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
283
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
284
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.12.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
285
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.layer_norm1.bias": "model-00001-of-00002.safetensors",
286
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.layer_norm1.weight": "model-00001-of-00002.safetensors",
287
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.layer_norm2.bias": "model-00001-of-00002.safetensors",
288
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.layer_norm2.weight": "model-00001-of-00002.safetensors",
289
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.mlp.fc1.bias": "model-00001-of-00002.safetensors",
290
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.mlp.fc1.weight": "model-00001-of-00002.safetensors",
291
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.mlp.fc2.bias": "model-00001-of-00002.safetensors",
292
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.mlp.fc2.weight": "model-00001-of-00002.safetensors",
293
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
294
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
295
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
296
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
297
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
298
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
299
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
300
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.13.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
301
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.layer_norm1.bias": "model-00001-of-00002.safetensors",
302
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.layer_norm1.weight": "model-00001-of-00002.safetensors",
303
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.layer_norm2.bias": "model-00001-of-00002.safetensors",
304
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.layer_norm2.weight": "model-00001-of-00002.safetensors",
305
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.mlp.fc1.bias": "model-00001-of-00002.safetensors",
306
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.mlp.fc1.weight": "model-00001-of-00002.safetensors",
307
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.mlp.fc2.bias": "model-00001-of-00002.safetensors",
308
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.mlp.fc2.weight": "model-00001-of-00002.safetensors",
309
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
310
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
311
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
312
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
313
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
314
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
315
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
316
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.14.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
317
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.layer_norm1.bias": "model-00001-of-00002.safetensors",
318
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.layer_norm1.weight": "model-00001-of-00002.safetensors",
319
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.layer_norm2.bias": "model-00001-of-00002.safetensors",
320
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.layer_norm2.weight": "model-00001-of-00002.safetensors",
321
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.mlp.fc1.bias": "model-00001-of-00002.safetensors",
322
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.mlp.fc1.weight": "model-00001-of-00002.safetensors",
323
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.mlp.fc2.bias": "model-00001-of-00002.safetensors",
324
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.mlp.fc2.weight": "model-00001-of-00002.safetensors",
325
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
326
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
327
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
328
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
329
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
330
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
331
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
332
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.15.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
333
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.layer_norm1.bias": "model-00001-of-00002.safetensors",
334
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.layer_norm1.weight": "model-00001-of-00002.safetensors",
335
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.layer_norm2.bias": "model-00001-of-00002.safetensors",
336
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.layer_norm2.weight": "model-00001-of-00002.safetensors",
337
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.mlp.fc1.bias": "model-00001-of-00002.safetensors",
338
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.mlp.fc1.weight": "model-00001-of-00002.safetensors",
339
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.mlp.fc2.bias": "model-00001-of-00002.safetensors",
340
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.mlp.fc2.weight": "model-00001-of-00002.safetensors",
341
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
342
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
343
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
344
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
345
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
346
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
347
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
348
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.16.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
349
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.layer_norm1.bias": "model-00001-of-00002.safetensors",
350
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.layer_norm1.weight": "model-00001-of-00002.safetensors",
351
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.layer_norm2.bias": "model-00001-of-00002.safetensors",
352
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.layer_norm2.weight": "model-00001-of-00002.safetensors",
353
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.mlp.fc1.bias": "model-00001-of-00002.safetensors",
354
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.mlp.fc1.weight": "model-00001-of-00002.safetensors",
355
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.mlp.fc2.bias": "model-00001-of-00002.safetensors",
356
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.mlp.fc2.weight": "model-00001-of-00002.safetensors",
357
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
358
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
359
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
360
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
361
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
362
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
363
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
364
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.17.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
365
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.layer_norm1.bias": "model-00001-of-00002.safetensors",
366
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.layer_norm1.weight": "model-00001-of-00002.safetensors",
367
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.layer_norm2.bias": "model-00001-of-00002.safetensors",
368
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.layer_norm2.weight": "model-00001-of-00002.safetensors",
369
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.mlp.fc1.bias": "model-00001-of-00002.safetensors",
370
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.mlp.fc1.weight": "model-00001-of-00002.safetensors",
371
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.mlp.fc2.bias": "model-00001-of-00002.safetensors",
372
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.mlp.fc2.weight": "model-00001-of-00002.safetensors",
373
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
374
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
375
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
376
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
377
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
378
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
379
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
380
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.18.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
381
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.layer_norm1.bias": "model-00001-of-00002.safetensors",
382
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.layer_norm1.weight": "model-00001-of-00002.safetensors",
383
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.layer_norm2.bias": "model-00001-of-00002.safetensors",
384
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.layer_norm2.weight": "model-00001-of-00002.safetensors",
385
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.mlp.fc1.bias": "model-00001-of-00002.safetensors",
386
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.mlp.fc1.weight": "model-00001-of-00002.safetensors",
387
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.mlp.fc2.bias": "model-00001-of-00002.safetensors",
388
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.mlp.fc2.weight": "model-00001-of-00002.safetensors",
389
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
390
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
391
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
392
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
393
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
394
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
395
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
396
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.19.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
397
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.layer_norm1.bias": "model-00001-of-00002.safetensors",
398
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.layer_norm1.weight": "model-00001-of-00002.safetensors",
399
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.layer_norm2.bias": "model-00001-of-00002.safetensors",
400
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.layer_norm2.weight": "model-00001-of-00002.safetensors",
401
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.mlp.fc1.bias": "model-00001-of-00002.safetensors",
402
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.mlp.fc1.weight": "model-00001-of-00002.safetensors",
403
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.mlp.fc2.bias": "model-00001-of-00002.safetensors",
404
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.mlp.fc2.weight": "model-00001-of-00002.safetensors",
405
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
406
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
407
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
408
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
409
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
410
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
411
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
412
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.2.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
413
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.layer_norm1.bias": "model-00001-of-00002.safetensors",
414
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.layer_norm1.weight": "model-00001-of-00002.safetensors",
415
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.layer_norm2.bias": "model-00001-of-00002.safetensors",
416
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.layer_norm2.weight": "model-00001-of-00002.safetensors",
417
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.mlp.fc1.bias": "model-00001-of-00002.safetensors",
418
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.mlp.fc1.weight": "model-00001-of-00002.safetensors",
419
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.mlp.fc2.bias": "model-00001-of-00002.safetensors",
420
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.mlp.fc2.weight": "model-00001-of-00002.safetensors",
421
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
422
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
423
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
424
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
425
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
426
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
427
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
428
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.20.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
429
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.layer_norm1.bias": "model-00001-of-00002.safetensors",
430
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.layer_norm1.weight": "model-00001-of-00002.safetensors",
431
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.layer_norm2.bias": "model-00001-of-00002.safetensors",
432
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.layer_norm2.weight": "model-00001-of-00002.safetensors",
433
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.mlp.fc1.bias": "model-00001-of-00002.safetensors",
434
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.mlp.fc1.weight": "model-00001-of-00002.safetensors",
435
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.mlp.fc2.bias": "model-00001-of-00002.safetensors",
436
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.mlp.fc2.weight": "model-00001-of-00002.safetensors",
437
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
438
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
439
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
440
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
441
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
442
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
443
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
444
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.21.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
445
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.layer_norm1.bias": "model-00001-of-00002.safetensors",
446
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.layer_norm1.weight": "model-00001-of-00002.safetensors",
447
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.layer_norm2.bias": "model-00001-of-00002.safetensors",
448
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.layer_norm2.weight": "model-00001-of-00002.safetensors",
449
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.mlp.fc1.bias": "model-00001-of-00002.safetensors",
450
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.mlp.fc1.weight": "model-00001-of-00002.safetensors",
451
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.mlp.fc2.bias": "model-00001-of-00002.safetensors",
452
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.mlp.fc2.weight": "model-00001-of-00002.safetensors",
453
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
454
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
455
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
456
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
457
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
458
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
459
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
460
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.22.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
461
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.layer_norm1.bias": "model-00001-of-00002.safetensors",
462
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.layer_norm1.weight": "model-00001-of-00002.safetensors",
463
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.layer_norm2.bias": "model-00001-of-00002.safetensors",
464
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.layer_norm2.weight": "model-00001-of-00002.safetensors",
465
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.mlp.fc1.bias": "model-00001-of-00002.safetensors",
466
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.mlp.fc1.weight": "model-00001-of-00002.safetensors",
467
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.mlp.fc2.bias": "model-00001-of-00002.safetensors",
468
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.mlp.fc2.weight": "model-00001-of-00002.safetensors",
469
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
470
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
471
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
472
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
473
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
474
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
475
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
476
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.23.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
477
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.layer_norm1.bias": "model-00001-of-00002.safetensors",
478
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.layer_norm1.weight": "model-00001-of-00002.safetensors",
479
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.layer_norm2.bias": "model-00001-of-00002.safetensors",
480
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.layer_norm2.weight": "model-00001-of-00002.safetensors",
481
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.mlp.fc1.bias": "model-00001-of-00002.safetensors",
482
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.mlp.fc1.weight": "model-00001-of-00002.safetensors",
483
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.mlp.fc2.bias": "model-00001-of-00002.safetensors",
484
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.mlp.fc2.weight": "model-00001-of-00002.safetensors",
485
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
486
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
487
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
488
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
489
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
490
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
491
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
492
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.3.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
493
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.layer_norm1.bias": "model-00001-of-00002.safetensors",
494
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.layer_norm1.weight": "model-00001-of-00002.safetensors",
495
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.layer_norm2.bias": "model-00001-of-00002.safetensors",
496
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.layer_norm2.weight": "model-00001-of-00002.safetensors",
497
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.mlp.fc1.bias": "model-00001-of-00002.safetensors",
498
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.mlp.fc1.weight": "model-00001-of-00002.safetensors",
499
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.mlp.fc2.bias": "model-00001-of-00002.safetensors",
500
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.mlp.fc2.weight": "model-00001-of-00002.safetensors",
501
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
502
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
503
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
504
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
505
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
506
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
507
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
508
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.4.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
509
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.layer_norm1.bias": "model-00001-of-00002.safetensors",
510
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.layer_norm1.weight": "model-00001-of-00002.safetensors",
511
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.layer_norm2.bias": "model-00001-of-00002.safetensors",
512
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.layer_norm2.weight": "model-00001-of-00002.safetensors",
513
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.mlp.fc1.bias": "model-00001-of-00002.safetensors",
514
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.mlp.fc1.weight": "model-00001-of-00002.safetensors",
515
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.mlp.fc2.bias": "model-00001-of-00002.safetensors",
516
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.mlp.fc2.weight": "model-00001-of-00002.safetensors",
517
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
518
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
519
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
520
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
521
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
522
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
523
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
524
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.5.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
525
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.layer_norm1.bias": "model-00001-of-00002.safetensors",
526
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.layer_norm1.weight": "model-00001-of-00002.safetensors",
527
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.layer_norm2.bias": "model-00001-of-00002.safetensors",
528
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.layer_norm2.weight": "model-00001-of-00002.safetensors",
529
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.mlp.fc1.bias": "model-00001-of-00002.safetensors",
530
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.mlp.fc1.weight": "model-00001-of-00002.safetensors",
531
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.mlp.fc2.bias": "model-00001-of-00002.safetensors",
532
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.mlp.fc2.weight": "model-00001-of-00002.safetensors",
533
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
534
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
535
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
536
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
537
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
538
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
539
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
540
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.6.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
541
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.layer_norm1.bias": "model-00001-of-00002.safetensors",
542
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.layer_norm1.weight": "model-00001-of-00002.safetensors",
543
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.layer_norm2.bias": "model-00001-of-00002.safetensors",
544
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.layer_norm2.weight": "model-00001-of-00002.safetensors",
545
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.mlp.fc1.bias": "model-00001-of-00002.safetensors",
546
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.mlp.fc1.weight": "model-00001-of-00002.safetensors",
547
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.mlp.fc2.bias": "model-00001-of-00002.safetensors",
548
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.mlp.fc2.weight": "model-00001-of-00002.safetensors",
549
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
550
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
551
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
552
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
553
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
554
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
555
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
556
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.7.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
557
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.layer_norm1.bias": "model-00001-of-00002.safetensors",
558
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.layer_norm1.weight": "model-00001-of-00002.safetensors",
559
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.layer_norm2.bias": "model-00001-of-00002.safetensors",
560
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.layer_norm2.weight": "model-00001-of-00002.safetensors",
561
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.mlp.fc1.bias": "model-00001-of-00002.safetensors",
562
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.mlp.fc1.weight": "model-00001-of-00002.safetensors",
563
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.mlp.fc2.bias": "model-00001-of-00002.safetensors",
564
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.mlp.fc2.weight": "model-00001-of-00002.safetensors",
565
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
566
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
567
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
568
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
569
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
570
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
571
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
572
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.8.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
573
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.layer_norm1.bias": "model-00001-of-00002.safetensors",
574
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.layer_norm1.weight": "model-00001-of-00002.safetensors",
575
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.layer_norm2.bias": "model-00001-of-00002.safetensors",
576
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.layer_norm2.weight": "model-00001-of-00002.safetensors",
577
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.mlp.fc1.bias": "model-00001-of-00002.safetensors",
578
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.mlp.fc1.weight": "model-00001-of-00002.safetensors",
579
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.mlp.fc2.bias": "model-00001-of-00002.safetensors",
580
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.mlp.fc2.weight": "model-00001-of-00002.safetensors",
581
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.self_attn.k_proj.bias": "model-00001-of-00002.safetensors",
582
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.self_attn.k_proj.weight": "model-00001-of-00002.safetensors",
583
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.self_attn.out_proj.bias": "model-00001-of-00002.safetensors",
584
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.self_attn.out_proj.weight": "model-00001-of-00002.safetensors",
585
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.self_attn.q_proj.bias": "model-00001-of-00002.safetensors",
586
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.self_attn.q_proj.weight": "model-00001-of-00002.safetensors",
587
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.self_attn.v_proj.bias": "model-00001-of-00002.safetensors",
588
+ "model.vision_embed_tokens.img_processor.vision_model.encoder.layers.9.self_attn.v_proj.weight": "model-00001-of-00002.safetensors",
589
+ "model.vision_embed_tokens.img_processor.vision_model.post_layernorm.bias": "model-00001-of-00002.safetensors",
590
+ "model.vision_embed_tokens.img_processor.vision_model.post_layernorm.weight": "model-00001-of-00002.safetensors",
591
+ "model.vision_embed_tokens.img_processor.vision_model.pre_layrnorm.bias": "model-00001-of-00002.safetensors",
592
+ "model.vision_embed_tokens.img_processor.vision_model.pre_layrnorm.weight": "model-00001-of-00002.safetensors",
593
+ "model.vision_embed_tokens.img_projection.0.bias": "model-00001-of-00002.safetensors",
594
+ "model.vision_embed_tokens.img_projection.0.weight": "model-00001-of-00002.safetensors",
595
+ "model.vision_embed_tokens.img_projection.2.bias": "model-00001-of-00002.safetensors",
596
+ "model.vision_embed_tokens.img_projection.2.weight": "model-00001-of-00002.safetensors",
597
+ "model.vision_embed_tokens.sub_GN": "model-00001-of-00002.safetensors"
598
+ }
599
+ }
modeling_phi3_v.py ADDED
@@ -0,0 +1,1632 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """ PyTorch Phi-3-V model."""
17
+
18
+ import inspect
19
+ import math
20
+ import warnings
21
+ from typing import List, Optional, Tuple, Union
22
+
23
+ import torch
24
+ import torch.nn.functional as F
25
+ import torch.utils.checkpoint
26
+ from torch import nn
27
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
28
+
29
+ from transformers.activations import ACT2FN
30
+ from transformers.cache_utils import Cache, DynamicCache
31
+ from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
32
+ from transformers.modeling_outputs import (
33
+ BaseModelOutputWithPast,
34
+ CausalLMOutputWithPast,
35
+ SequenceClassifierOutputWithPast,
36
+ TokenClassifierOutput,
37
+ )
38
+ from transformers.modeling_utils import PreTrainedModel
39
+ from transformers.utils import (
40
+ add_code_sample_docstrings,
41
+ add_start_docstrings,
42
+ add_start_docstrings_to_model_forward,
43
+ is_flash_attn_2_available,
44
+ is_flash_attn_greater_or_equal_2_10,
45
+ logging,
46
+ replace_return_docstrings,
47
+ )
48
+ from .configuration_phi3_v import Phi3VConfig
49
+ from .image_embedding_phi3_v import Phi3ImageEmbedding
50
+
51
+
52
+ if is_flash_attn_2_available():
53
+ from flash_attn import flash_attn_func, flash_attn_varlen_func
54
+ from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
55
+
56
+ _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
57
+
58
+ logger = logging.get_logger(__name__)
59
+
60
+ _CHECKPOINT_FOR_DOC = "microsoft/Phi-3-vision-128k-instruct"
61
+ _CONFIG_FOR_DOC = "Phi3VConfig"
62
+
63
+ PHI3V_PRETRAINED_MODEL_ARCHIVE_LIST = [
64
+ "microsoft/Phi-3-vision-128k-instruct",
65
+ # See all Phi-3 models at https://huggingface.co/models?filter=Phi-3
66
+ ]
67
+
68
+
69
+ # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Phi3
70
+ class Phi3RMSNorm(nn.Module):
71
+ def __init__(self, hidden_size, eps=1e-6):
72
+ """
73
+ Phi3RMSNorm is equivalent to T5LayerNorm
74
+ """
75
+ super().__init__()
76
+ self.weight = nn.Parameter(torch.ones(hidden_size))
77
+ self.variance_epsilon = eps
78
+
79
+ def forward(self, hidden_states):
80
+ input_dtype = hidden_states.dtype
81
+ hidden_states = hidden_states.to(torch.float32)
82
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
83
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
84
+ return self.weight * hidden_states.to(input_dtype)
85
+
86
+
87
+ # Copied from transformers.models.llama.modeling_llama._get_unpad_data
88
+ def _get_unpad_data(attention_mask):
89
+ seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
90
+ indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
91
+ max_seqlen_in_batch = seqlens_in_batch.max().item()
92
+ cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
93
+ return (
94
+ indices,
95
+ cu_seqlens,
96
+ max_seqlen_in_batch,
97
+ )
98
+
99
+
100
+ # Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3
101
+ class Phi3RotaryEmbedding(nn.Module):
102
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
103
+ super().__init__()
104
+
105
+ self.dim = dim
106
+ self.max_position_embeddings = max_position_embeddings
107
+ self.base = base
108
+ self.register_buffer("inv_freq", None, persistent=False)
109
+
110
+ @torch.no_grad()
111
+ def forward(self, x, position_ids, seq_len=None):
112
+ # x: [bs, num_attention_heads, seq_len, head_size]
113
+ if self.inv_freq is None:
114
+ self.inv_freq = 1.0 / (
115
+ self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
116
+ )
117
+ inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
118
+ position_ids_expanded = position_ids[:, None, :].float()
119
+ # Force float32 since bfloat16 loses precision on long contexts
120
+ # See https://github.com/huggingface/transformers/pull/29285
121
+ device_type = x.device.type
122
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
123
+ with torch.autocast(device_type=device_type, enabled=False):
124
+ freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
125
+ emb = torch.cat((freqs, freqs), dim=-1)
126
+ cos = emb.cos()
127
+ sin = emb.sin()
128
+ return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
129
+
130
+
131
+ class Phi3SuScaledRotaryEmbedding(Phi3RotaryEmbedding):
132
+ def __init__(self, dim, config, device=None):
133
+ super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
134
+
135
+ self.short_factor = config.rope_scaling["short_factor"]
136
+ self.long_factor = config.rope_scaling["long_factor"]
137
+ self.original_max_position_embeddings = config.original_max_position_embeddings
138
+
139
+ @torch.no_grad()
140
+ def forward(self, x, position_ids, seq_len=None):
141
+ seq_len = torch.max(position_ids) + 1
142
+ if seq_len > self.original_max_position_embeddings:
143
+ ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
144
+ else:
145
+ ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
146
+
147
+ inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
148
+ self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
149
+
150
+ inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
151
+ position_ids_expanded = position_ids[:, None, :].float()
152
+
153
+ # Force float32 since bfloat16 loses precision on long contexts
154
+ # See https://github.com/huggingface/transformers/pull/29285
155
+ device_type = x.device.type
156
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
157
+ with torch.autocast(device_type=device_type, enabled=False):
158
+ freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
159
+ emb = torch.cat((freqs, freqs), dim=-1)
160
+
161
+ scale = self.max_position_embeddings / self.original_max_position_embeddings
162
+ if scale <= 1.0:
163
+ scaling_factor = 1.0
164
+ else:
165
+ scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))
166
+
167
+ cos = emb.cos() * scaling_factor
168
+ sin = emb.sin() * scaling_factor
169
+ return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
170
+
171
+
172
+ class Phi3YarnScaledRotaryEmbedding(Phi3RotaryEmbedding):
173
+ def __init__(self, dim, config, device=None):
174
+ super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
175
+
176
+ self.short_factor = config.rope_scaling["short_factor"]
177
+ self.long_factor = config.rope_scaling["long_factor"]
178
+ self.original_max_position_embeddings = config.original_max_position_embeddings
179
+
180
+ @torch.no_grad()
181
+ def forward(self, x, position_ids, seq_len=None):
182
+ seq_len = torch.max(position_ids) + 1
183
+ if seq_len > self.original_max_position_embeddings:
184
+ ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
185
+ else:
186
+ ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
187
+
188
+ inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
189
+ self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
190
+
191
+ inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
192
+ position_ids_expanded = position_ids[:, None, :].float()
193
+
194
+ # Force float32 since bfloat16 loses precision on long contexts
195
+ # See https://github.com/huggingface/transformers/pull/29285
196
+ device_type = x.device.type
197
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
198
+ with torch.autocast(device_type=device_type, enabled=False):
199
+ freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
200
+ emb = torch.cat((freqs, freqs), dim=-1)
201
+
202
+ scale = self.max_position_embeddings / self.original_max_position_embeddings
203
+ if scale <= 1.0:
204
+ scaling_factor = 1.0
205
+ else:
206
+ scaling_factor = 0.1 * math.log(scale) + 1.0
207
+
208
+ cos = emb.cos() * scaling_factor
209
+ sin = emb.sin() * scaling_factor
210
+ return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
211
+
212
+
213
+ # Copied from transformers.models.llama.modeling_llama.rotate_half
214
+ def rotate_half(x):
215
+ """Rotates half the hidden dims of the input."""
216
+ x1 = x[..., : x.shape[-1] // 2]
217
+ x2 = x[..., x.shape[-1] // 2 :]
218
+ return torch.cat((-x2, x1), dim=-1)
219
+
220
+
221
+ # Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
222
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
223
+ """Applies Rotary Position Embedding to the query and key tensors.
224
+
225
+ Args:
226
+ q (`torch.Tensor`): The query tensor.
227
+ k (`torch.Tensor`): The key tensor.
228
+ cos (`torch.Tensor`): The cosine part of the rotary embedding.
229
+ sin (`torch.Tensor`): The sine part of the rotary embedding.
230
+ position_ids (`torch.Tensor`, *optional*):
231
+ Deprecated and unused.
232
+ unsqueeze_dim (`int`, *optional*, defaults to 1):
233
+ The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
234
+ sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
235
+ that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
236
+ k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
237
+ cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
238
+ the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
239
+ Returns:
240
+ `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
241
+ """
242
+ cos = cos.unsqueeze(unsqueeze_dim)
243
+ sin = sin.unsqueeze(unsqueeze_dim)
244
+ q_embed = (q * cos) + (rotate_half(q) * sin)
245
+ k_embed = (k * cos) + (rotate_half(k) * sin)
246
+ return q_embed, k_embed
247
+
248
+
249
+ class Phi3MLP(nn.Module):
250
+ def __init__(self, config):
251
+ super().__init__()
252
+
253
+ self.config = config
254
+ self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
255
+ self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
256
+
257
+ self.activation_fn = ACT2FN[config.hidden_act]
258
+
259
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
260
+ up_states = self.gate_up_proj(hidden_states)
261
+
262
+ gate, up_states = up_states.chunk(2, dim=-1)
263
+ up_states = up_states * self.activation_fn(gate)
264
+
265
+ return self.down_proj(up_states)
266
+
267
+
268
+ # Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
269
+ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
270
+ """
271
+ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
272
+ num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
273
+ """
274
+ batch, num_key_value_heads, slen, head_dim = hidden_states.shape
275
+ if n_rep == 1:
276
+ return hidden_states
277
+ hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
278
+ return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
279
+
280
+
281
+ class Phi3Attention(nn.Module):
282
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
283
+
284
+ def __init__(self, config: Phi3VConfig, layer_idx: Optional[int] = None):
285
+ super().__init__()
286
+ self.config = config
287
+ self.layer_idx = layer_idx
288
+ if layer_idx is None:
289
+ logger.warning_once(
290
+ f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
291
+ "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
292
+ "when creating this class."
293
+ )
294
+
295
+ self.attention_dropout = config.attention_dropout
296
+ self.hidden_size = config.hidden_size
297
+ self.num_heads = config.num_attention_heads
298
+ self.head_dim = self.hidden_size // self.num_heads
299
+ self.num_key_value_heads = config.num_key_value_heads
300
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
301
+ self.max_position_embeddings = config.max_position_embeddings
302
+ self.original_max_position_embeddings = config.original_max_position_embeddings
303
+ self.rope_theta = config.rope_theta
304
+ self.rope_scaling = config.rope_scaling
305
+ self.is_causal = True
306
+
307
+ if (self.head_dim * self.num_heads) != self.hidden_size:
308
+ raise ValueError(
309
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
310
+ f" and `num_heads`: {self.num_heads})."
311
+ )
312
+
313
+ op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
314
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
315
+ self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
316
+ self._init_rope()
317
+
318
+ def _init_rope(self):
319
+ if self.rope_scaling is None:
320
+ self.rotary_emb = Phi3RotaryEmbedding(
321
+ self.head_dim,
322
+ max_position_embeddings=self.max_position_embeddings,
323
+ base=self.rope_theta,
324
+ )
325
+ else:
326
+ scaling_type = self.config.rope_scaling["type"]
327
+ if scaling_type == "su":
328
+ self.rotary_emb = Phi3SuScaledRotaryEmbedding(self.head_dim, self.config)
329
+ elif scaling_type == "yarn":
330
+ self.rotary_emb = Phi3YarnScaledRotaryEmbedding(self.head_dim, self.config)
331
+ else:
332
+ raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
333
+
334
+ def forward(
335
+ self,
336
+ hidden_states: torch.Tensor,
337
+ attention_mask: Optional[torch.Tensor] = None,
338
+ position_ids: Optional[torch.LongTensor] = None,
339
+ past_key_value: Optional[Cache] = None,
340
+ output_attentions: bool = False,
341
+ use_cache: bool = False,
342
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
343
+ logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.")
344
+
345
+ bsz, q_len, _ = hidden_states.size()
346
+
347
+ qkv = self.qkv_proj(hidden_states)
348
+ query_pos = self.num_heads * self.head_dim
349
+ query_states = qkv[..., :query_pos]
350
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
351
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
352
+
353
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
354
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
355
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
356
+
357
+ kv_seq_len = key_states.shape[-2]
358
+ if past_key_value is not None:
359
+ if self.layer_idx is None:
360
+ raise ValueError(
361
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
362
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
363
+ "with a layer index."
364
+ )
365
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
366
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
367
+
368
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
369
+
370
+ if past_key_value is not None:
371
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
372
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
373
+
374
+ # repeat k/v heads if n_kv_heads < n_heads
375
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
376
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
377
+
378
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
379
+
380
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
381
+ raise ValueError(
382
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
383
+ f" {attn_weights.size()}"
384
+ )
385
+
386
+ if attention_mask is not None:
387
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
388
+ raise ValueError(
389
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
390
+ )
391
+ attn_weights = attn_weights + attention_mask
392
+
393
+ # upcast attention to fp32
394
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
395
+ attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
396
+
397
+ attn_output = torch.matmul(attn_weights, value_states)
398
+
399
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
400
+ raise ValueError(
401
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
402
+ f" {attn_output.size()}"
403
+ )
404
+
405
+ attn_output = attn_output.transpose(1, 2).contiguous()
406
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
407
+
408
+ attn_output = self.o_proj(attn_output)
409
+
410
+ if not output_attentions:
411
+ attn_weights = None
412
+
413
+ return attn_output, attn_weights, past_key_value
414
+
415
+
416
+ class Phi3FlashAttention2(Phi3Attention):
417
+ """
418
+ Phi-3 flash attention module. This module inherits from `Phi3Attention` as the weights of the module stays
419
+ untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
420
+ flash attention and deal with padding tokens in case the input contains any of them.
421
+ """
422
+
423
+ # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
424
+ def __init__(self, *args, **kwargs):
425
+ super().__init__(*args, **kwargs)
426
+
427
+ # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
428
+ # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
429
+ # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
430
+ self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
431
+
432
+ def forward(
433
+ self,
434
+ hidden_states: torch.Tensor,
435
+ attention_mask: Optional[torch.LongTensor] = None,
436
+ position_ids: Optional[torch.LongTensor] = None,
437
+ past_key_value: Optional[Cache] = None,
438
+ output_attentions: bool = False,
439
+ use_cache: bool = False,
440
+ **kwargs,
441
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
442
+ # Phi3FlashAttention2 attention does not support output_attentions
443
+
444
+ if not _flash_supports_window_size:
445
+ logger.warning_once(
446
+ "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library."
447
+ )
448
+ raise ValueError("The current flash attention version does not support sliding window attention.")
449
+
450
+ output_attentions = False
451
+
452
+ if "padding_mask" in kwargs:
453
+ warnings.warn(
454
+ "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
455
+ )
456
+
457
+ # overwrite attention_mask with padding_mask
458
+ attention_mask = kwargs.pop("padding_mask")
459
+
460
+ bsz, q_len, _ = hidden_states.size()
461
+
462
+ qkv = self.qkv_proj(hidden_states)
463
+ query_pos = self.num_heads * self.head_dim
464
+ query_states = qkv[..., :query_pos]
465
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
466
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
467
+
468
+ # Flash attention requires the input to have the shape
469
+ # batch_size x seq_length x head_dim x hidden_dim
470
+ # therefore we just need to keep the original shape
471
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
472
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
473
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
474
+
475
+ kv_seq_len = key_states.shape[-2]
476
+ if past_key_value is not None:
477
+ if self.layer_idx is None:
478
+ raise ValueError(
479
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
480
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
481
+ "with a layer index."
482
+ )
483
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
484
+
485
+ # Because the input can be padded, the absolute sequence length depends on the max position id.
486
+ rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
487
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len)
488
+
489
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
490
+
491
+ use_sliding_windows = (
492
+ _flash_supports_window_size
493
+ and getattr(self.config, "sliding_window", None) is not None
494
+ and kv_seq_len > self.config.sliding_window
495
+ )
496
+
497
+ if past_key_value is not None:
498
+ # Activate slicing cache only if the config has a value `sliding_windows` attribute
499
+ cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
500
+ if (
501
+ getattr(self.config, "sliding_window", None) is not None
502
+ and kv_seq_len > self.config.sliding_window
503
+ and cache_has_contents
504
+ ):
505
+ slicing_tokens = 1 - self.config.sliding_window
506
+
507
+ past_key = past_key_value[self.layer_idx][0]
508
+ past_value = past_key_value[self.layer_idx][1]
509
+
510
+ past_key = past_key[:, :, slicing_tokens:, :].contiguous()
511
+ past_value = past_value[:, :, slicing_tokens:, :].contiguous()
512
+
513
+ if past_key.shape[-2] != self.config.sliding_window - 1:
514
+ raise ValueError(
515
+ f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
516
+ f" {past_key.shape}"
517
+ )
518
+
519
+ if attention_mask is not None:
520
+ attention_mask = attention_mask[:, slicing_tokens:]
521
+ attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
522
+
523
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
524
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
525
+
526
+ # repeat k/v heads if n_kv_heads < n_heads
527
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
528
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
529
+
530
+ attn_dropout = self.attention_dropout if self.training else 0.0
531
+
532
+ # In PEFT, usually we cast the layer norms in float32 for training stability reasons
533
+ # therefore the input hidden states gets silently casted in float32. Hence, we need
534
+ # cast them back in the correct dtype just to be sure everything works as expected.
535
+ # This might slowdown training & inference so it is recommended to not cast the LayerNorms
536
+ # in fp32.
537
+
538
+ if query_states.dtype == torch.float32:
539
+ if torch.is_autocast_enabled():
540
+ target_dtype = torch.get_autocast_gpu_dtype()
541
+ # Handle the case where the model is quantized
542
+ elif hasattr(self.config, "_pre_quantization_dtype"):
543
+ target_dtype = self.config._pre_quantization_dtype
544
+ else:
545
+ target_dtype = self.qkv_proj.weight.dtype
546
+
547
+ logger.warning_once(
548
+ f"The input hidden states seems to be silently casted in float32, this might be related to"
549
+ f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
550
+ f" {target_dtype}."
551
+ )
552
+
553
+ query_states = query_states.to(target_dtype)
554
+ key_states = key_states.to(target_dtype)
555
+ value_states = value_states.to(target_dtype)
556
+
557
+ # Reashape to the expected shape for Flash Attention
558
+ query_states = query_states.transpose(1, 2)
559
+ key_states = key_states.transpose(1, 2)
560
+ value_states = value_states.transpose(1, 2)
561
+
562
+ attn_output = self._flash_attention_forward(
563
+ query_states,
564
+ key_states,
565
+ value_states,
566
+ attention_mask,
567
+ q_len,
568
+ dropout=attn_dropout,
569
+ use_sliding_windows=use_sliding_windows,
570
+ )
571
+
572
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
573
+ attn_output = self.o_proj(attn_output)
574
+
575
+ if not output_attentions:
576
+ attn_weights = None
577
+
578
+ return attn_output, attn_weights, past_key_value
579
+
580
+ # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward
581
+ def _flash_attention_forward(
582
+ self,
583
+ query_states,
584
+ key_states,
585
+ value_states,
586
+ attention_mask,
587
+ query_length,
588
+ dropout=0.0,
589
+ softmax_scale=None,
590
+ use_sliding_windows=False,
591
+ ):
592
+ """
593
+ Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
594
+ first unpad the input, then computes the attention scores and pad the final attention scores.
595
+
596
+ Args:
597
+ query_states (`torch.Tensor`):
598
+ Input query states to be passed to Flash Attention API
599
+ key_states (`torch.Tensor`):
600
+ Input key states to be passed to Flash Attention API
601
+ value_states (`torch.Tensor`):
602
+ Input value states to be passed to Flash Attention API
603
+ attention_mask (`torch.Tensor`):
604
+ The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
605
+ position of padding tokens and 1 for the position of non-padding tokens.
606
+ dropout (`float`):
607
+ Attention dropout
608
+ softmax_scale (`float`, *optional*):
609
+ The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
610
+ use_sliding_windows (`bool`, *optional*):
611
+ Whether to activate sliding window attention.
612
+ """
613
+ if not self._flash_attn_uses_top_left_mask:
614
+ causal = self.is_causal
615
+ else:
616
+ # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
617
+ causal = self.is_causal and query_length != 1
618
+
619
+ # Contains at least one padding token in the sequence
620
+ if attention_mask is not None:
621
+ batch_size = query_states.shape[0]
622
+ query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
623
+ query_states, key_states, value_states, attention_mask, query_length
624
+ )
625
+
626
+ cu_seqlens_q, cu_seqlens_k = cu_seq_lens
627
+ max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
628
+
629
+ if not use_sliding_windows:
630
+ attn_output_unpad = flash_attn_varlen_func(
631
+ query_states,
632
+ key_states,
633
+ value_states,
634
+ cu_seqlens_q=cu_seqlens_q,
635
+ cu_seqlens_k=cu_seqlens_k,
636
+ max_seqlen_q=max_seqlen_in_batch_q,
637
+ max_seqlen_k=max_seqlen_in_batch_k,
638
+ dropout_p=dropout,
639
+ softmax_scale=softmax_scale,
640
+ causal=causal,
641
+ )
642
+ else:
643
+ attn_output_unpad = flash_attn_varlen_func(
644
+ query_states,
645
+ key_states,
646
+ value_states,
647
+ cu_seqlens_q=cu_seqlens_q,
648
+ cu_seqlens_k=cu_seqlens_k,
649
+ max_seqlen_q=max_seqlen_in_batch_q,
650
+ max_seqlen_k=max_seqlen_in_batch_k,
651
+ dropout_p=dropout,
652
+ softmax_scale=softmax_scale,
653
+ causal=causal,
654
+ window_size=(self.config.sliding_window, self.config.sliding_window),
655
+ )
656
+
657
+ attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
658
+ else:
659
+ if not use_sliding_windows:
660
+ attn_output = flash_attn_func(
661
+ query_states,
662
+ key_states,
663
+ value_states,
664
+ dropout,
665
+ softmax_scale=softmax_scale,
666
+ causal=causal,
667
+ )
668
+ else:
669
+ attn_output = flash_attn_func(
670
+ query_states,
671
+ key_states,
672
+ value_states,
673
+ dropout,
674
+ softmax_scale=softmax_scale,
675
+ causal=causal,
676
+ window_size=(self.config.sliding_window, self.config.sliding_window),
677
+ )
678
+
679
+ return attn_output
680
+
681
+ # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
682
+ def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
683
+ batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
684
+
685
+ # On the first iteration we need to properly re-create the padding mask
686
+ # by slicing it on the proper place
687
+ if kv_seq_len != attention_mask.shape[-1]:
688
+ attention_mask_num_tokens = attention_mask.shape[-1]
689
+ attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
690
+
691
+ indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
692
+
693
+ key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
694
+ value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
695
+
696
+ if query_length == kv_seq_len:
697
+ query_layer = index_first_axis(
698
+ query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
699
+ )
700
+ cu_seqlens_q = cu_seqlens_k
701
+ max_seqlen_in_batch_q = max_seqlen_in_batch_k
702
+ indices_q = indices_k
703
+ elif query_length == 1:
704
+ max_seqlen_in_batch_q = 1
705
+ cu_seqlens_q = torch.arange(
706
+ batch_size + 1, dtype=torch.int32, device=query_layer.device
707
+ ) # There is a memcpy here, that is very bad.
708
+ indices_q = cu_seqlens_q[:-1]
709
+ query_layer = query_layer.squeeze(1)
710
+ else:
711
+ # The -q_len: slice assumes left padding.
712
+ attention_mask = attention_mask[:, -query_length:]
713
+ query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
714
+
715
+ return (
716
+ query_layer,
717
+ key_layer,
718
+ value_layer,
719
+ indices_q,
720
+ (cu_seqlens_q, cu_seqlens_k),
721
+ (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
722
+ )
723
+
724
+
725
+ # copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
726
+ # TODO @Arthur no longer copied from LLama after static cache
727
+ class Phi3SdpaAttention(Phi3Attention):
728
+ """
729
+ Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
730
+ `Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
731
+ SDPA API.
732
+ """
733
+
734
+ # Adapted from Phi3Attention.forward
735
+ def forward(
736
+ self,
737
+ hidden_states: torch.Tensor,
738
+ attention_mask: Optional[torch.Tensor] = None,
739
+ position_ids: Optional[torch.LongTensor] = None,
740
+ past_key_value: Optional[Cache] = None,
741
+ output_attentions: bool = False,
742
+ use_cache: bool = False,
743
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
744
+ if output_attentions:
745
+ # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
746
+ logger.warning_once(
747
+ "Phi3Model is using Phi3SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
748
+ 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
749
+ )
750
+ return super().forward(
751
+ hidden_states=hidden_states,
752
+ attention_mask=attention_mask,
753
+ position_ids=position_ids,
754
+ past_key_value=past_key_value,
755
+ output_attentions=output_attentions,
756
+ use_cache=use_cache,
757
+ )
758
+
759
+ bsz, q_len, _ = hidden_states.size()
760
+
761
+ qkv = self.qkv_proj(hidden_states)
762
+ query_pos = self.num_heads * self.head_dim
763
+ query_states = qkv[..., :query_pos]
764
+ key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
765
+ value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
766
+
767
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
768
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
769
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
770
+
771
+ kv_seq_len = key_states.shape[-2]
772
+ if past_key_value is not None:
773
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
774
+ cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
775
+
776
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
777
+
778
+ if past_key_value is not None:
779
+ cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
780
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
781
+
782
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
783
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
784
+
785
+ if attention_mask is not None:
786
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
787
+ raise ValueError(
788
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
789
+ )
790
+
791
+ # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
792
+ # Reference: https://github.com/pytorch/pytorch/issues/112577.
793
+ if query_states.device.type == "cuda" and attention_mask is not None:
794
+ query_states = query_states.contiguous()
795
+ key_states = key_states.contiguous()
796
+ value_states = value_states.contiguous()
797
+
798
+ attn_output = torch.nn.functional.scaled_dot_product_attention(
799
+ query_states,
800
+ key_states,
801
+ value_states,
802
+ attn_mask=attention_mask,
803
+ dropout_p=self.attention_dropout if self.training else 0.0,
804
+ # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
805
+ is_causal=self.is_causal and attention_mask is None and q_len > 1,
806
+ )
807
+
808
+ attn_output = attn_output.transpose(1, 2).contiguous()
809
+ attn_output = attn_output.view(bsz, q_len, self.hidden_size)
810
+
811
+ attn_output = self.o_proj(attn_output)
812
+
813
+ return attn_output, None, past_key_value
814
+
815
+
816
+ PHI3_ATTENTION_CLASSES = {
817
+ "eager": Phi3Attention,
818
+ "flash_attention_2": Phi3FlashAttention2,
819
+ "sdpa": Phi3SdpaAttention,
820
+ }
821
+
822
+
823
+ class Phi3DecoderLayer(nn.Module):
824
+ def __init__(self, config: Phi3VConfig, layer_idx: int):
825
+ super().__init__()
826
+
827
+ self.config = config
828
+ self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
829
+
830
+ self.mlp = Phi3MLP(config)
831
+ self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
832
+
833
+ self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
834
+ self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
835
+ self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
836
+
837
+ def forward(
838
+ self,
839
+ hidden_states: torch.Tensor,
840
+ attention_mask: Optional[torch.Tensor] = None,
841
+ position_ids: Optional[torch.LongTensor] = None,
842
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
843
+ output_attentions: Optional[bool] = False,
844
+ use_cache: Optional[bool] = False,
845
+ **kwargs,
846
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
847
+ if "padding_mask" in kwargs:
848
+ warnings.warn(
849
+ "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
850
+ )
851
+ """
852
+ Args:
853
+ hidden_states (`torch.FloatTensor`):
854
+ input to the layer of shape `(batch, seq_len, embed_dim)`
855
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
856
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
857
+ position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
858
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
859
+ `[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
860
+ output_attentions (`bool`, *optional*):
861
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
862
+ returned tensors for more detail.
863
+ use_cache (`bool`, *optional*):
864
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
865
+ (see `past_key_values`).
866
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
867
+ """
868
+
869
+ residual = hidden_states
870
+
871
+ hidden_states = self.input_layernorm(hidden_states)
872
+
873
+ # Self Attention
874
+ attn_outputs, self_attn_weights, present_key_value = self.self_attn(
875
+ hidden_states=hidden_states,
876
+ attention_mask=attention_mask,
877
+ position_ids=position_ids,
878
+ past_key_value=past_key_value,
879
+ output_attentions=output_attentions,
880
+ use_cache=use_cache,
881
+ )
882
+
883
+ hidden_states = residual + self.resid_attn_dropout(attn_outputs)
884
+
885
+ residual = hidden_states
886
+ hidden_states = self.post_attention_layernorm(hidden_states)
887
+ hidden_states = self.mlp(hidden_states)
888
+ hidden_states = residual + self.resid_mlp_dropout(hidden_states)
889
+
890
+ outputs = (hidden_states,)
891
+
892
+ if output_attentions:
893
+ outputs += (self_attn_weights,)
894
+
895
+ if use_cache:
896
+ outputs += (present_key_value,)
897
+
898
+ return outputs
899
+
900
+
901
+ PHI3V_START_DOCSTRING = r"""
902
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
903
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
904
+ etc.)
905
+
906
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
907
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
908
+ and behavior.
909
+
910
+ Parameters:
911
+ config ([`Phi3VConfig`]):
912
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
913
+ load the weights associated with the model, only the configuration. Check out the
914
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
915
+ """
916
+
917
+
918
+ @add_start_docstrings(
919
+ "The bare Phi-3-V model outputting raw hidden-states without any specific head on top.",
920
+ PHI3V_START_DOCSTRING,
921
+ )
922
+ class Phi3VPreTrainedModel(PreTrainedModel):
923
+ config_class = Phi3VConfig
924
+ base_model_prefix = "model"
925
+ supports_gradient_checkpointing = True
926
+ _no_split_modules = ["Phi3DecoderLayer"]
927
+ _skip_keys_device_placement = "past_key_values"
928
+ _supports_flash_attn_2 = True
929
+ _supports_sdpa = False
930
+ _supports_cache_class = True
931
+
932
+ _version = "0.0.5"
933
+
934
+ def _init_weights(self, module):
935
+ std = self.config.initializer_range
936
+ if isinstance(module, nn.Linear):
937
+ module.weight.data.normal_(mean=0.0, std=std)
938
+ if module.bias is not None:
939
+ module.bias.data.zero_()
940
+ elif isinstance(module, nn.Embedding):
941
+ module.weight.data.normal_(mean=0.0, std=std)
942
+ if module.padding_idx is not None:
943
+ module.weight.data[module.padding_idx].zero_()
944
+
945
+
946
+ PHI3V_INPUTS_DOCSTRING = r"""
947
+ Args:
948
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
949
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
950
+ it.
951
+
952
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
953
+ [`PreTrainedTokenizer.__call__`] for details.
954
+
955
+ [What are input IDs?](../glossary#input-ids)
956
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
957
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
958
+
959
+ - 1 for tokens that are **not masked**,
960
+ - 0 for tokens that are **masked**.
961
+
962
+ [What are attention masks?](../glossary#attention-mask)
963
+
964
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
965
+ [`PreTrainedTokenizer.__call__`] for details.
966
+
967
+ If `past_key_values` is used, optionally only the last `input_ids` have to be input (see
968
+ `past_key_values`).
969
+
970
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
971
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
972
+ information on the default strategy.
973
+
974
+ - 1 indicates the head is **not masked**,
975
+ - 0 indicates the head is **masked**.
976
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
977
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
978
+ config.n_positions - 1]`.
979
+
980
+ [What are position IDs?](../glossary#position-ids)
981
+ past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
982
+ Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
983
+ blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
984
+ returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
985
+
986
+ Two formats are allowed:
987
+ - a [`~cache_utils.Cache`] instance;
988
+ - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
989
+ shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
990
+ cache format.
991
+
992
+ The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
993
+ legacy cache format will be returned.
994
+
995
+ If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
996
+ have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
997
+ of shape `(batch_size, sequence_length)`.
998
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
999
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
1000
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
1001
+ model's internal embedding lookup matrix.
1002
+ pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
1003
+ The tensors corresponding to the input images. Pixel values can be obtained using [`AutoImageProcessor`].
1004
+ See [`Phi3ImageProcessor.__call__`] for details.
1005
+ image_sizes (`torch.LongTensor` of shape `(batch_size, 2)`, *optional*):
1006
+ The sizes of the images in the batch, being (height, width) for each image.
1007
+ use_cache (`bool`, *optional*):
1008
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
1009
+ `past_key_values`).
1010
+ output_attentions (`bool`, *optional*):
1011
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
1012
+ tensors for more detail.
1013
+ output_hidden_states (`bool`, *optional*):
1014
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
1015
+ more detail.
1016
+ return_dict (`bool`, *optional*):
1017
+ Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
1018
+ """
1019
+
1020
+
1021
+ @add_start_docstrings(
1022
+ "The bare Phi-3-V model outputting raw hidden-states without any specific head on top.",
1023
+ PHI3V_START_DOCSTRING,
1024
+ )
1025
+ class Phi3VModel(Phi3VPreTrainedModel):
1026
+ """
1027
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Phi3DecoderLayer`]
1028
+
1029
+ Args:
1030
+ config: Phi3Config
1031
+ """
1032
+
1033
+ def __init__(self, config: Phi3VConfig):
1034
+ super().__init__(config)
1035
+ self.padding_idx = config.pad_token_id
1036
+ self.vocab_size = config.vocab_size
1037
+
1038
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
1039
+ self.embed_dropout = nn.Dropout(config.embd_pdrop)
1040
+
1041
+ self.vision_embed_tokens = None
1042
+ if isinstance(config.embd_layer, dict):
1043
+ # vision embedding layer
1044
+ embedding_config = {
1045
+ 'embedding_cls': config.embd_layer['embedding_cls'],
1046
+ **config.embd_layer
1047
+ }
1048
+ self.vision_embed_tokens = Phi3ImageEmbedding(config, wte=self.embed_tokens, **embedding_config)
1049
+ # # set wte the same for vision embedding
1050
+ # self.vision_embed_tokens.wte.weight = self.embed_tokens.weight
1051
+
1052
+ self.layers = nn.ModuleList(
1053
+ [Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
1054
+ )
1055
+ self._attn_implementation = config._attn_implementation
1056
+ self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
1057
+
1058
+ self.gradient_checkpointing = False
1059
+ # Initialize weights and apply final processing
1060
+ self.post_init()
1061
+
1062
+ def get_input_embeddings(self):
1063
+ return self.embed_tokens
1064
+
1065
+ def set_input_embeddings(self, value):
1066
+ self.embed_tokens = value
1067
+
1068
+ @add_start_docstrings_to_model_forward(PHI3V_INPUTS_DOCSTRING)
1069
+ def forward(
1070
+ self,
1071
+ input_ids: torch.LongTensor = None,
1072
+ attention_mask: Optional[torch.Tensor] = None,
1073
+ position_ids: Optional[torch.LongTensor] = None,
1074
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1075
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1076
+ pixel_values: Optional[torch.FloatTensor] = None,
1077
+ image_sizes: Optional[torch.LongTensor] = None,
1078
+ use_cache: Optional[bool] = None,
1079
+ output_attentions: Optional[bool] = None,
1080
+ output_hidden_states: Optional[bool] = None,
1081
+ return_dict: Optional[bool] = None,
1082
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
1083
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1084
+ output_hidden_states = (
1085
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1086
+ )
1087
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
1088
+
1089
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1090
+
1091
+ # retrieve input_ids and inputs_embeds
1092
+ if input_ids is not None and inputs_embeds is not None:
1093
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
1094
+ elif input_ids is not None:
1095
+ batch_size, seq_length = input_ids.shape[:2]
1096
+ elif inputs_embeds is not None:
1097
+ batch_size, seq_length = inputs_embeds.shape[:2]
1098
+ else:
1099
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
1100
+
1101
+ past_key_values_length = 0
1102
+
1103
+ if self.gradient_checkpointing and self.training:
1104
+ if use_cache:
1105
+ logger.warning_once(
1106
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
1107
+ )
1108
+ use_cache = False
1109
+
1110
+ if use_cache:
1111
+ use_legacy_cache = not isinstance(past_key_values, Cache)
1112
+ if use_legacy_cache:
1113
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
1114
+ past_key_values_length = past_key_values.get_usable_length(seq_length)
1115
+
1116
+ if position_ids is None:
1117
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
1118
+ position_ids = torch.arange(
1119
+ past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
1120
+ )
1121
+ position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
1122
+ else:
1123
+ position_ids = position_ids.view(-1, seq_length).long()
1124
+
1125
+ if inputs_embeds is None:
1126
+ if pixel_values is not None and image_sizes is not None:
1127
+ assert self.vision_embed_tokens is not None, "Vision embedding layer is not defined"
1128
+ inputs_embeds = self.vision_embed_tokens(input_ids, pixel_values=pixel_values, image_sizes=image_sizes)
1129
+ else:
1130
+ inputs_embeds = self.embed_tokens(input_ids)
1131
+
1132
+ if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
1133
+ is_padding_right = attention_mask[:, -1].sum().item() != batch_size
1134
+ if is_padding_right:
1135
+ raise ValueError(
1136
+ "You are attempting to perform batched generation with padding_side='right'"
1137
+ " this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
1138
+ " call `tokenizer.padding_side = 'left'` before tokenizing the input. "
1139
+ )
1140
+
1141
+ if self._attn_implementation == "flash_attention_2":
1142
+ # 2d mask is passed through the layers
1143
+ attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
1144
+ else:
1145
+ # 4d mask is passed through the layers
1146
+ attention_mask = _prepare_4d_causal_attention_mask(
1147
+ attention_mask,
1148
+ (batch_size, seq_length),
1149
+ inputs_embeds,
1150
+ past_key_values_length,
1151
+ sliding_window=self.config.sliding_window,
1152
+ )
1153
+
1154
+ hidden_states = inputs_embeds
1155
+
1156
+ # decoder layers
1157
+ all_hidden_states = () if output_hidden_states else None
1158
+ all_self_attns = () if output_attentions else None
1159
+ next_decoder_cache = None
1160
+
1161
+ for decoder_layer in self.layers:
1162
+ if output_hidden_states:
1163
+ all_hidden_states += (hidden_states,)
1164
+
1165
+ if self.gradient_checkpointing and self.training:
1166
+ layer_outputs = self._gradient_checkpointing_func(
1167
+ decoder_layer.__call__,
1168
+ hidden_states,
1169
+ attention_mask,
1170
+ position_ids,
1171
+ past_key_values,
1172
+ output_attentions,
1173
+ use_cache,
1174
+ )
1175
+ else:
1176
+ layer_outputs = decoder_layer(
1177
+ hidden_states,
1178
+ attention_mask=attention_mask,
1179
+ position_ids=position_ids,
1180
+ past_key_value=past_key_values,
1181
+ output_attentions=output_attentions,
1182
+ use_cache=use_cache,
1183
+ )
1184
+
1185
+ hidden_states = layer_outputs[0]
1186
+
1187
+ if use_cache:
1188
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
1189
+
1190
+ if output_attentions:
1191
+ all_self_attns += (layer_outputs[1],)
1192
+
1193
+ hidden_states = self.norm(hidden_states)
1194
+
1195
+ # add hidden states from the last decoder layer
1196
+ if output_hidden_states:
1197
+ all_hidden_states += (hidden_states,)
1198
+
1199
+ next_cache = None
1200
+ if use_cache:
1201
+ next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
1202
+ if not return_dict:
1203
+ return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
1204
+ return BaseModelOutputWithPast(
1205
+ last_hidden_state=hidden_states,
1206
+ past_key_values=next_cache,
1207
+ hidden_states=all_hidden_states,
1208
+ attentions=all_self_attns,
1209
+ )
1210
+
1211
+
1212
+ class Phi3VForCausalLM(Phi3VPreTrainedModel):
1213
+ _tied_weights_keys = ["lm_head.weight"]
1214
+
1215
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3
1216
+ def __init__(self, config):
1217
+ super().__init__(config)
1218
+ self.model = Phi3VModel(config)
1219
+ self.vocab_size = config.vocab_size
1220
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
1221
+
1222
+ # Initialize weights and apply final processing
1223
+ self.post_init()
1224
+
1225
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_input_embeddings
1226
+ def get_input_embeddings(self):
1227
+ return self.model.embed_tokens
1228
+
1229
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_input_embeddings
1230
+ def set_input_embeddings(self, value):
1231
+ self.model.embed_tokens = value
1232
+
1233
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_output_embeddings
1234
+ def get_output_embeddings(self):
1235
+ return self.lm_head
1236
+
1237
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_output_embeddings
1238
+ def set_output_embeddings(self, new_embeddings):
1239
+ self.lm_head = new_embeddings
1240
+
1241
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder
1242
+ def set_decoder(self, decoder):
1243
+ self.model = decoder
1244
+
1245
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder
1246
+ def get_decoder(self):
1247
+ return self.model
1248
+
1249
+ # Ignore copy
1250
+ @add_start_docstrings_to_model_forward(PHI3V_INPUTS_DOCSTRING)
1251
+ @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
1252
+ def forward(
1253
+ self,
1254
+ input_ids: torch.LongTensor = None,
1255
+ attention_mask: Optional[torch.Tensor] = None,
1256
+ position_ids: Optional[torch.LongTensor] = None,
1257
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1258
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1259
+ pixel_values: Optional[torch.FloatTensor] = None,
1260
+ image_sizes: Optional[torch.LongTensor] = None,
1261
+ labels: Optional[torch.LongTensor] = None,
1262
+ use_cache: Optional[bool] = None,
1263
+ output_attentions: Optional[bool] = None,
1264
+ output_hidden_states: Optional[bool] = None,
1265
+ return_dict: Optional[bool] = None,
1266
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
1267
+ r"""
1268
+ Args:
1269
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
1270
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
1271
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
1272
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
1273
+
1274
+ Returns:
1275
+
1276
+ Example:
1277
+
1278
+ ```python
1279
+ >>> from transformers import AutoTokenizer, Phi3ForCausalLM
1280
+
1281
+ >>> model = Phi3ForCausalLM.from_pretrained("microsoft/phi-3-mini-4k-instruct")
1282
+ >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-3-mini-4k-instruct")
1283
+
1284
+ >>> prompt = "This is an example script ."
1285
+ >>> inputs = tokenizer(prompt, return_tensors="pt")
1286
+
1287
+ >>> # Generate
1288
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
1289
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
1290
+ 'This is an example script .\n Certainly! Below is a sample script that demonstrates a simple task, such as calculating the sum'
1291
+ ```"""
1292
+
1293
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1294
+ output_hidden_states = (
1295
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1296
+ )
1297
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1298
+
1299
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
1300
+ outputs = self.model(
1301
+ input_ids=input_ids,
1302
+ attention_mask=attention_mask,
1303
+ position_ids=position_ids,
1304
+ past_key_values=past_key_values,
1305
+ inputs_embeds=inputs_embeds,
1306
+ pixel_values=pixel_values,
1307
+ image_sizes=image_sizes,
1308
+ use_cache=use_cache,
1309
+ output_attentions=output_attentions,
1310
+ output_hidden_states=output_hidden_states,
1311
+ return_dict=return_dict,
1312
+ )
1313
+
1314
+ hidden_states = outputs[0]
1315
+ logits = self.lm_head(hidden_states)
1316
+ logits = logits.float()
1317
+
1318
+ loss = None
1319
+ if labels is not None:
1320
+ # Shift so that tokens < n predict n
1321
+ shift_logits = logits[..., :-1, :].contiguous()
1322
+ shift_labels = labels[..., 1:].contiguous()
1323
+ # Flatten the tokens
1324
+ loss_fct = CrossEntropyLoss()
1325
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
1326
+ shift_labels = shift_labels.view(-1)
1327
+ # Enable model parallelism
1328
+ shift_labels = shift_labels.to(shift_logits.device)
1329
+ loss = loss_fct(shift_logits, shift_labels)
1330
+
1331
+ if not return_dict:
1332
+ output = (logits,) + outputs[1:]
1333
+ return (loss,) + output if loss is not None else output
1334
+
1335
+ return CausalLMOutputWithPast(
1336
+ loss=loss,
1337
+ logits=logits,
1338
+ past_key_values=outputs.past_key_values,
1339
+ hidden_states=outputs.hidden_states,
1340
+ attentions=outputs.attentions,
1341
+ )
1342
+
1343
+ # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM.prepare_inputs_for_generation
1344
+ def prepare_inputs_for_generation(
1345
+ self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, pixel_values=None, image_sizes=None, **kwargs
1346
+ ):
1347
+ if past_key_values is not None:
1348
+ if isinstance(past_key_values, Cache):
1349
+ cache_length = past_key_values.get_seq_length()
1350
+ past_length = past_key_values.seen_tokens
1351
+ max_cache_length = past_key_values.get_max_length()
1352
+ else:
1353
+ cache_length = past_length = past_key_values[0][0].shape[2]
1354
+ max_cache_length = None
1355
+
1356
+ # Keep only the unprocessed tokens:
1357
+ # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
1358
+ # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
1359
+ # input)
1360
+ if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
1361
+ input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
1362
+ # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
1363
+ # input_ids based on the past_length.
1364
+ elif past_length < input_ids.shape[1]:
1365
+ input_ids = input_ids[:, past_length:]
1366
+ # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
1367
+
1368
+ # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
1369
+ if (
1370
+ max_cache_length is not None
1371
+ and attention_mask is not None
1372
+ and cache_length + input_ids.shape[1] > max_cache_length
1373
+ ):
1374
+ attention_mask = attention_mask[:, -max_cache_length:]
1375
+
1376
+ position_ids = kwargs.get("position_ids", None)
1377
+ if attention_mask is not None and position_ids is None:
1378
+ # create position_ids on the fly for batch generation
1379
+ position_ids = attention_mask.long().cumsum(-1) - 1
1380
+ position_ids.masked_fill_(attention_mask == 0, 1)
1381
+ if past_key_values:
1382
+ position_ids = position_ids[:, -input_ids.shape[1] :]
1383
+
1384
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1385
+ if inputs_embeds is not None and past_key_values is None:
1386
+ model_inputs = {"inputs_embeds": inputs_embeds}
1387
+ else:
1388
+ model_inputs = {"input_ids": input_ids}
1389
+
1390
+ model_inputs.update(
1391
+ {
1392
+ "position_ids": position_ids,
1393
+ "past_key_values": past_key_values,
1394
+ "use_cache": kwargs.get("use_cache"),
1395
+ "attention_mask": attention_mask,
1396
+ "pixel_values": pixel_values,
1397
+ "image_sizes": image_sizes,
1398
+ }
1399
+ )
1400
+ return model_inputs
1401
+
1402
+ @staticmethod
1403
+ # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM._reorder_cache
1404
+ def _reorder_cache(past_key_values, beam_idx):
1405
+ reordered_past = ()
1406
+ for layer_past in past_key_values:
1407
+ reordered_past += (
1408
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
1409
+ )
1410
+ return reordered_past
1411
+
1412
+
1413
+ @add_start_docstrings(
1414
+ """
1415
+ The [`Phi3VModel`] with a sequence classification head on top (linear layer).
1416
+
1417
+ [`Phi3VForSequenceClassification`] uses the last token in order to do the classification, as other causal models
1418
+ (e.g. GPT-2) do.
1419
+
1420
+ Since it does classification on the last token, it requires to know the position of the last token. If a
1421
+ `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
1422
+ no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
1423
+ padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
1424
+ each row of the batch).
1425
+ """,
1426
+ PHI3V_START_DOCSTRING,
1427
+ )
1428
+ # Copied from transformers.models.llama.modeling_llama.LlamaForSequenceClassification with Llama->Phi3, LLAMA->PHI3, self.transformer->self.model, transformer_outputs->model_outputs
1429
+ class Phi3VForSequenceClassification(Phi3VPreTrainedModel):
1430
+ def __init__(self, config):
1431
+ super().__init__(config)
1432
+ self.num_labels = config.num_labels
1433
+ self.model = Phi3VModel(config)
1434
+ self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False)
1435
+
1436
+ # Initialize weights and apply final processing
1437
+ self.post_init()
1438
+
1439
+ def get_input_embeddings(self):
1440
+ return self.model.embed_tokens
1441
+
1442
+ def set_input_embeddings(self, value):
1443
+ self.model.embed_tokens = value
1444
+
1445
+ @add_start_docstrings_to_model_forward(PHI3V_INPUTS_DOCSTRING)
1446
+ def forward(
1447
+ self,
1448
+ input_ids: torch.LongTensor = None,
1449
+ attention_mask: Optional[torch.Tensor] = None,
1450
+ position_ids: Optional[torch.LongTensor] = None,
1451
+ past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None,
1452
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1453
+ pixel_values: Optional[torch.FloatTensor] = None,
1454
+ image_sizes: Optional[torch.LongTensor] = None,
1455
+ labels: Optional[torch.LongTensor] = None,
1456
+ use_cache: Optional[bool] = None,
1457
+ output_attentions: Optional[bool] = None,
1458
+ output_hidden_states: Optional[bool] = None,
1459
+ return_dict: Optional[bool] = None,
1460
+ ) -> Union[Tuple, SequenceClassifierOutputWithPast]:
1461
+ r"""
1462
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1463
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1464
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1465
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1466
+ """
1467
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1468
+
1469
+ model_outputs = self.model(
1470
+ input_ids,
1471
+ attention_mask=attention_mask,
1472
+ position_ids=position_ids,
1473
+ past_key_values=past_key_values,
1474
+ inputs_embeds=inputs_embeds,
1475
+ pixel_values=pixel_values,
1476
+ image_sizes=image_sizes,
1477
+ use_cache=use_cache,
1478
+ output_attentions=output_attentions,
1479
+ output_hidden_states=output_hidden_states,
1480
+ return_dict=return_dict,
1481
+ )
1482
+ hidden_states = model_outputs[0]
1483
+ logits = self.score(hidden_states)
1484
+
1485
+ if input_ids is not None:
1486
+ batch_size = input_ids.shape[0]
1487
+ else:
1488
+ batch_size = inputs_embeds.shape[0]
1489
+
1490
+ if self.config.pad_token_id is None and batch_size != 1:
1491
+ raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
1492
+ if self.config.pad_token_id is None:
1493
+ sequence_lengths = -1
1494
+ else:
1495
+ if input_ids is not None:
1496
+ # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
1497
+ sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
1498
+ sequence_lengths = sequence_lengths % input_ids.shape[-1]
1499
+ sequence_lengths = sequence_lengths.to(logits.device)
1500
+ else:
1501
+ sequence_lengths = -1
1502
+
1503
+ pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
1504
+
1505
+ loss = None
1506
+ if labels is not None:
1507
+ labels = labels.to(logits.device)
1508
+ if self.config.problem_type is None:
1509
+ if self.num_labels == 1:
1510
+ self.config.problem_type = "regression"
1511
+ elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
1512
+ self.config.problem_type = "single_label_classification"
1513
+ else:
1514
+ self.config.problem_type = "multi_label_classification"
1515
+
1516
+ if self.config.problem_type == "regression":
1517
+ loss_fct = MSELoss()
1518
+ if self.num_labels == 1:
1519
+ loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
1520
+ else:
1521
+ loss = loss_fct(pooled_logits, labels)
1522
+ elif self.config.problem_type == "single_label_classification":
1523
+ loss_fct = CrossEntropyLoss()
1524
+ loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
1525
+ elif self.config.problem_type == "multi_label_classification":
1526
+ loss_fct = BCEWithLogitsLoss()
1527
+ loss = loss_fct(pooled_logits, labels)
1528
+ if not return_dict:
1529
+ output = (pooled_logits,) + model_outputs[1:]
1530
+ return ((loss,) + output) if loss is not None else output
1531
+
1532
+ return SequenceClassifierOutputWithPast(
1533
+ loss=loss,
1534
+ logits=pooled_logits,
1535
+ past_key_values=model_outputs.past_key_values,
1536
+ hidden_states=model_outputs.hidden_states,
1537
+ attentions=model_outputs.attentions,
1538
+ )
1539
+
1540
+
1541
+ @add_start_docstrings(
1542
+ """
1543
+ [`Phi3VModel`] with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for
1544
+ Named-Entity-Recognition (NER) tasks.
1545
+ """,
1546
+ PHI3V_START_DOCSTRING,
1547
+ )
1548
+ # Copied from transformers.models.mpt.modeling_mpt.MptForTokenClassification with Mpt->Phi3,MPT->PHI3,self.transformer->self.model,transformer_outputs->model_outputs
1549
+ class Phi3VForTokenClassification(Phi3VPreTrainedModel):
1550
+ def __init__(self, config: Phi3VConfig):
1551
+ super().__init__(config)
1552
+ self.num_labels = config.num_labels
1553
+
1554
+ self.model = Phi3VModel(config)
1555
+ if hasattr(config, "classifier_dropout") and config.classifier_dropout is not None:
1556
+ classifier_dropout = config.classifier_dropout
1557
+ elif hasattr(config, "hidden_dropout") and config.hidden_dropout is not None:
1558
+ classifier_dropout = config.hidden_dropout
1559
+ else:
1560
+ classifier_dropout = 0.1
1561
+ self.dropout = nn.Dropout(classifier_dropout)
1562
+ self.classifier = nn.Linear(config.hidden_size, config.num_labels)
1563
+
1564
+ # Initialize weights and apply final processing
1565
+ self.post_init()
1566
+
1567
+ @add_start_docstrings_to_model_forward(PHI3V_INPUTS_DOCSTRING)
1568
+ @add_code_sample_docstrings(
1569
+ checkpoint=_CHECKPOINT_FOR_DOC,
1570
+ output_type=TokenClassifierOutput,
1571
+ config_class=_CONFIG_FOR_DOC,
1572
+ )
1573
+ def forward(
1574
+ self,
1575
+ input_ids: Optional[torch.LongTensor] = None,
1576
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
1577
+ attention_mask: Optional[torch.Tensor] = None,
1578
+ inputs_embeds: Optional[torch.Tensor] = None,
1579
+ pixel_values: Optional[torch.FloatTensor] = None,
1580
+ image_sizes: Optional[torch.LongTensor] = None,
1581
+ labels: Optional[torch.Tensor] = None,
1582
+ use_cache: Optional[bool] = None,
1583
+ output_attentions: Optional[bool] = None,
1584
+ output_hidden_states: Optional[bool] = None,
1585
+ return_dict: Optional[bool] = None,
1586
+ **deprecated_arguments,
1587
+ ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
1588
+ r"""
1589
+ labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
1590
+ Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
1591
+ config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
1592
+ `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
1593
+ """
1594
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1595
+
1596
+ model_outputs = self.model(
1597
+ input_ids,
1598
+ past_key_values=past_key_values,
1599
+ attention_mask=attention_mask,
1600
+ inputs_embeds=inputs_embeds,
1601
+ pixel_values=pixel_values,
1602
+ image_sizes=image_sizes,
1603
+ use_cache=use_cache,
1604
+ output_attentions=output_attentions,
1605
+ output_hidden_states=output_hidden_states,
1606
+ return_dict=return_dict,
1607
+ )
1608
+
1609
+ hidden_states = model_outputs[0]
1610
+ hidden_states = self.dropout(hidden_states)
1611
+ logits = self.classifier(hidden_states)
1612
+
1613
+ loss = None
1614
+ if labels is not None:
1615
+ # move labels to correct device to enable model parallelism
1616
+ labels = labels.to(logits.device)
1617
+ batch_size, seq_length = labels.shape
1618
+ loss_fct = CrossEntropyLoss()
1619
+ loss = loss_fct(
1620
+ logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length)
1621
+ )
1622
+
1623
+ if not return_dict:
1624
+ output = (logits,) + model_outputs[2:]
1625
+ return ((loss,) + output) if loss is not None else output
1626
+
1627
+ return TokenClassifierOutput(
1628
+ loss=loss,
1629
+ logits=logits,
1630
+ hidden_states=model_outputs.hidden_states,
1631
+ attentions=model_outputs.attentions,
1632
+ )
preprocessor_config.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "auto_map": {
3
+ "AutoProcessor": "processing_phi3_v.Phi3VProcessor",
4
+ "AutoImageProcessor": "image_processing_phi3_v.Phi3VImageProcessor"
5
+ },
6
+ "num_crops": 16,
7
+ "image_mean": [
8
+ 0.48145466,
9
+ 0.4578275,
10
+ 0.40821073
11
+ ],
12
+ "image_processor_type": "Phi3VImageProcessor",
13
+ "image_std": [
14
+ 0.26862954,
15
+ 0.26130258,
16
+ 0.27577711
17
+ ],
18
+ "processor_class": "Phi3VProcessor",
19
+ "num_img_tokens": 144
20
+ }
processing_phi3_v.py ADDED
@@ -0,0 +1,217 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ """
17
+ Processor class for Phi3-V.
18
+ """
19
+ import re
20
+ from typing import List, Optional, Union
21
+
22
+ import torch
23
+
24
+ import transformers
25
+ from transformers.feature_extraction_utils import BatchFeature
26
+ from transformers.image_utils import ImageInput
27
+ from transformers.processing_utils import ProcessorMixin
28
+ from transformers.tokenization_utils_base import PaddingStrategy, TextInput, TruncationStrategy
29
+ from transformers.utils import TensorType
30
+ from .image_processing_phi3_v import Phi3VImageProcessor
31
+ transformers.Phi3VImageProcessor = Phi3VImageProcessor
32
+
33
+ class Phi3VProcessor(ProcessorMixin):
34
+ r"""
35
+ Constructs a Phi3-V processor which wraps a Phi3-V image processor and a LLaMa tokenizer into a single processor.
36
+
37
+ [`Phi3VProcessor`] offers all the functionalities of [`Phi3VImageProcessor`] and [`LlamaTokenizerFast`]. See the
38
+ [`~Phi3VProcessor.__call__`] and [`~Phi3VProcessor.decode`] for more information.
39
+
40
+ Args:
41
+ image_processor ([`Phi3VImageProcessor`], *optional*):
42
+ The image processor is a required input.
43
+ tokenizer ([`LlamaTokenizerFast`], *optional*):
44
+ The tokenizer is a required input.
45
+ """
46
+
47
+ attributes = ["image_processor", "tokenizer"]
48
+ image_processor_class = "Phi3VImageProcessor"
49
+ tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
50
+ special_image_token = "<|image|>"
51
+
52
+ def __init__(self, image_processor, tokenizer):
53
+ self.image_processor = image_processor
54
+ self.tokenizer = tokenizer
55
+ self.num_img_tokens = image_processor.num_img_tokens
56
+ self.img_tokens = [f"<|image_{i+1}|>" for i in range(1000000)]
57
+
58
+ def __call__(
59
+ self,
60
+ text: Union[TextInput, List[TextInput]],
61
+ images: ImageInput = None,
62
+ padding: Union[bool, str, PaddingStrategy] = False,
63
+ truncation: Union[bool, str, TruncationStrategy] = None,
64
+ max_length=None,
65
+ return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
66
+ ) -> BatchFeature:
67
+ """
68
+ Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
69
+ and `kwargs` arguments to LlamaTokenizerFast's [`~LlamaTokenizerFast.__call__`] if `text` is not `None` to encode
70
+ the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
71
+ Phi3ImageProcessor's [`~Phi3ImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
72
+ of the above two methods for more information.
73
+
74
+ Args:
75
+ text (`str`, `List[str]`, `List[List[str]]`):
76
+ The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
77
+ (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
78
+ `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
79
+ images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
80
+ The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
81
+ tensor. Both channels-first and channels-last formats are supported.
82
+ padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
83
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding
84
+ index) among:
85
+ - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
86
+ sequence if provided).
87
+ - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
88
+ acceptable input length for the model if that argument is not provided.
89
+ - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
90
+ lengths).
91
+ max_length (`int`, *optional*):
92
+ Maximum length of the returned list and optionally padding length (see above).
93
+ truncation (`bool`, *optional*):
94
+ Activates truncation to cut input sequences longer than `max_length` to `max_length`.
95
+ return_tensors (`str` or [`~utils.TensorType`], *optional*):
96
+ If set, will return tensors of a particular framework. Acceptable values are:
97
+
98
+ - `'tf'`: Return TensorFlow `tf.constant` objects.
99
+ - `'pt'`: Return PyTorch `torch.Tensor` objects.
100
+ - `'np'`: Return NumPy `np.ndarray` objects.
101
+ - `'jax'`: Return JAX `jnp.ndarray` objects.
102
+
103
+ Returns:
104
+ [`BatchFeature`]: A [`BatchFeature`] with the following fields:
105
+
106
+ - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
107
+ - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
108
+ `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
109
+ `None`).
110
+ - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
111
+ """
112
+ if images is not None:
113
+ image_inputs = self.image_processor(images, return_tensors=return_tensors)
114
+ else:
115
+ image_inputs = {}
116
+ inputs = self._convert_images_texts_to_inputs(image_inputs, text, padding=padding, truncation=truncation, max_length=max_length, return_tensors=return_tensors)
117
+ return inputs
118
+
119
+ def calc_num_image_tokens(self, images: ImageInput):
120
+ """ Calculate the number of image tokens for each image.
121
+ Args:
122
+ images (`ImageInput`):
123
+ Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
124
+ passing in images with pixel values between 0 and 1, set `do_rescale=False`.
125
+ """
126
+ return self.image_processor.calc_num_image_tokens(images)
127
+
128
+ def calc_num_image_tokens_from_image_size(self, width, height):
129
+ """ Calculate the number of image token for an image with given width and height.
130
+ Args:
131
+ width (`int`):
132
+ Width of the image.
133
+ height (`int`):
134
+ Height of the image.
135
+ """
136
+ return self.image_processor.calc_num_image_tokens_from_image_size(width, height)
137
+
138
+
139
+ @property
140
+ def special_image_token_id(self):
141
+ return self.tokenizer.convert_tokens_to_ids(self.special_image_token)
142
+
143
+ def get_special_image_token_id(self):
144
+ return self.tokenizer.convert_tokens_to_ids(self.special_image_token)
145
+
146
+ def _convert_images_texts_to_inputs(self, images, texts, padding=False, truncation=None, max_length=None, return_tensors=None):
147
+
148
+ if not len(images):
149
+ model_inputs = self.tokenizer(texts, return_tensors=return_tensors, padding=padding, truncation=truncation, max_length=max_length)
150
+ return BatchFeature(data={**model_inputs})
151
+
152
+ pattern = r"<\|image_\d+\|>"
153
+ prompt_chunks = [self.tokenizer(chunk).input_ids for chunk in re.split(pattern, texts)]
154
+
155
+ if 'num_img_tokens' in images:
156
+ num_img_tokens = images['num_img_tokens']
157
+ else:
158
+ assert 'num_crops' in images, 'num_crops must be provided in images if num_img_tokens is not provided'
159
+ num_crops = images['num_crops']
160
+ num_img_tokens = [_num_crops * self.num_img_tokens for _num_crops in num_crops]
161
+
162
+ images, image_sizes = images['pixel_values'], images['image_sizes']
163
+
164
+ # image_tags needs to start from 1 to n
165
+ image_tags = re.findall(pattern, texts)
166
+ # image_ids = [int(s.split("|")[1].split("_")[-1]) * -1 for s in image_tags]
167
+ # image_ids_pad = [[iid]*num_img_tokens[i] for i, iid in enumerate(image_ids)]
168
+ image_ids = [int(s.split("|")[1].split("_")[-1]) for s in image_tags]
169
+ unique_image_ids = sorted(list(set(image_ids)))
170
+ # image_ids must start from 1, and must be continuous int, e.g. [1, 2, 3], cannot be [1, 4, 5]
171
+ # check the condition
172
+ assert unique_image_ids == list(range(1, len(unique_image_ids)+1)), f"image_ids must start from 1, and must be continuous int, e.g. [1, 2, 3], cannot be {unique_image_ids}"
173
+ # total images must be the same as the number of image tags
174
+ assert len(unique_image_ids) == len(images), f"total images must be the same as the number of image tags, got {len(unique_image_ids)} image tags and {len(images)} images"
175
+
176
+ image_ids_pad = [[-iid]*num_img_tokens[iid-1] for iid in image_ids]
177
+
178
+ def insert_separator(X, sep_list):
179
+ if len(X) > len(sep_list):
180
+ sep_list.append([])
181
+ return [ele for sublist in zip(X, sep_list) for ele in sublist]
182
+ input_ids = []
183
+ offset = 0
184
+ for x in insert_separator(prompt_chunks, image_ids_pad):
185
+ input_ids.extend(x[offset:])
186
+
187
+ input_ids = torch.tensor(input_ids, dtype=torch.long).unsqueeze(0)
188
+ attention_mask = (input_ids > -1000000).to(torch.long)
189
+
190
+ return BatchFeature(data={"input_ids": input_ids,
191
+ "attention_mask": attention_mask,
192
+ "pixel_values": images,
193
+ "image_sizes": image_sizes})
194
+
195
+
196
+ # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Llama
197
+ def batch_decode(self, *args, **kwargs):
198
+ """
199
+ This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
200
+ refer to the docstring of this method for more information.
201
+ """
202
+ return self.tokenizer.batch_decode(*args, **kwargs)
203
+
204
+ # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Llama
205
+ def decode(self, *args, **kwargs):
206
+ """
207
+ This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
208
+ the docstring of this method for more information.
209
+ """
210
+ return self.tokenizer.decode(*args, **kwargs)
211
+
212
+ @property
213
+ # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names
214
+ def model_input_names(self):
215
+ tokenizer_input_names = self.tokenizer.model_input_names
216
+ image_processor_input_names = self.image_processor.model_input_names
217
+ return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
sample_inference.py ADDED
@@ -0,0 +1,129 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+
3
+ from PIL import Image
4
+ import requests
5
+ import torch
6
+ from transformers import AutoModelForCausalLM
7
+ from transformers import AutoProcessor
8
+ model_path = "./"
9
+
10
+ kwargs = {}
11
+ kwargs['torch_dtype'] = torch.bfloat16
12
+
13
+ processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
14
+ model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, torch_dtype="auto").cuda()
15
+
16
+ user_prompt = '<|user|>\n'
17
+ assistant_prompt = '<|assistant|>\n'
18
+ prompt_suffix = "<|end|>\n"
19
+
20
+ #################################################### text-only ####################################################
21
+ # single-image prompt
22
+ prompt = f"{user_prompt}what is the answer for 1+1? Explain it.{prompt_suffix}{assistant_prompt}"
23
+ print(f">>> Prompt\n{prompt}")
24
+ inputs = processor(prompt, images=None, return_tensors="pt").to("cuda:0")
25
+ generate_ids = model.generate(**inputs,
26
+ max_new_tokens=1000,
27
+ eos_token_id=processor.tokenizer.eos_token_id,
28
+ )
29
+ generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
30
+ response = processor.batch_decode(generate_ids,
31
+ skip_special_tokens=True,
32
+ clean_up_tokenization_spaces=False)[0]
33
+ print(f'>>> Response\n{response}')
34
+
35
+ #################################################### text-only 2 ####################################################
36
+ # single-image prompt
37
+ prompt = f"{user_prompt}Give me the code for sloving two-sum problem.{prompt_suffix}{assistant_prompt}"
38
+ print(f">>> Prompt\n{prompt}")
39
+ inputs = processor(prompt, images=None, return_tensors="pt").to("cuda:0")
40
+ generate_ids = model.generate(**inputs,
41
+ max_new_tokens=1000,
42
+ eos_token_id=processor.tokenizer.eos_token_id,
43
+ )
44
+ generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
45
+ response = processor.batch_decode(generate_ids,
46
+ skip_special_tokens=True,
47
+ clean_up_tokenization_spaces=False)[0]
48
+ print(f'>>> Response\n{response}')
49
+
50
+
51
+ #################################################### EXAMPLE 1 ####################################################
52
+ # single-image prompt
53
+ prompt = f"{user_prompt}<|image_1|>\nWhat is shown in this image?{prompt_suffix}{assistant_prompt}"
54
+ url = "https://www.ilankelman.org/stopsigns/australia.jpg"
55
+ print(f">>> Prompt\n{prompt}")
56
+ image = Image.open(requests.get(url, stream=True).raw)
57
+ inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")
58
+ generate_ids = model.generate(**inputs,
59
+ max_new_tokens=1000,
60
+ eos_token_id=processor.tokenizer.eos_token_id,
61
+ )
62
+ generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
63
+ response = processor.batch_decode(generate_ids,
64
+ skip_special_tokens=True,
65
+ clean_up_tokenization_spaces=False)[0]
66
+ print(f'>>> Response\n{response}')
67
+
68
+ #################################################### EXAMPLE 2 ####################################################
69
+ # multiple image prompt
70
+ # Note: image tokens must start from <|image_1|>
71
+ prompt = f"{user_prompt}<|image_1|>\n<|image_2|>\n What is shown in this two images?{prompt_suffix}{assistant_prompt}"
72
+ print(f">>> Prompt\n{prompt}")
73
+ url = "https://www.ilankelman.org/stopsigns/australia.jpg"
74
+ image_1 = Image.open(requests.get(url, stream=True).raw)
75
+ url = "https://img.freepik.com/free-photo/painting-mountain-lake-with-mountain-background_188544-9126.jpg?w=2000"
76
+ image_2 = Image.open(requests.get(url, stream=True).raw)
77
+ images = [image_1, image_2]
78
+ inputs = processor(prompt, images, return_tensors="pt").to("cuda:0")
79
+ generate_ids = model.generate(**inputs,
80
+ max_new_tokens=1000,
81
+ eos_token_id=processor.tokenizer.eos_token_id,
82
+ )
83
+ generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
84
+ response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
85
+ print(f'>>> Response\n{response}')
86
+
87
+ #################################################### EXAMPLE 3 ####################################################
88
+ # chat template
89
+ chat = [
90
+ {"role": "user", "content": "<|image_1|>\nWhat is shown in this image?"},
91
+ {"role": "assistant", "content": "The image depicts a street scene with a prominent red stop sign in the foreground. The background showcases a building with traditional Chinese architecture, characterized by its red roof and ornate decorations. There are also several statues of lions, which are common in Chinese culture, positioned in front of the building. The street is lined with various shops and businesses, and there's a car passing by."},
92
+ {"role": "user", "content": "What is so special about this image"}
93
+ ]
94
+ url = "https://www.ilankelman.org/stopsigns/australia.jpg"
95
+ image = Image.open(requests.get(url, stream=True).raw)
96
+ prompt = processor.tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
97
+ # need to remove last <|endoftext|> if it is there, which is used for training, not inference. For training, make sure to add <|endoftext|> in the end.
98
+ if prompt.endswith("<|endoftext|>"):
99
+ prompt = prompt.rstrip("<|endoftext|>")
100
+
101
+ print(f">>> Prompt\n{prompt}")
102
+
103
+ inputs = processor(prompt, [image], return_tensors="pt").to("cuda:0")
104
+ generate_ids = model.generate(**inputs,
105
+ max_new_tokens=1000,
106
+ eos_token_id=processor.tokenizer.eos_token_id,
107
+ )
108
+ generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
109
+ response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
110
+ print(f'>>> Response\n{response}')
111
+
112
+
113
+ ############################# to markdown #############################
114
+ # single-image prompt
115
+ prompt = f"{user_prompt}<|image_1|>\nCan you convert the table to markdown format?{prompt_suffix}{assistant_prompt}"
116
+ url = "https://support.content.office.net/en-us/media/3dd2b79b-9160-403d-9967-af893d17b580.png"
117
+ image = Image.open(requests.get(url, stream=True).raw)
118
+ inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")
119
+
120
+ print(f">>> Prompt\n{prompt}")
121
+ generate_ids = model.generate(**inputs,
122
+ max_new_tokens=1000,
123
+ eos_token_id=processor.tokenizer.eos_token_id,
124
+ )
125
+ generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:]
126
+ response = processor.batch_decode(generate_ids,
127
+ skip_special_tokens=False,
128
+ clean_up_tokenization_spaces=False)[0]
129
+ print(f'>>> Response\n{response}')
special_tokens_map.json ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<|system|>",
4
+ "<|end|>",
5
+ "<|user|>",
6
+ "<|end|>"
7
+ ],
8
+ "bos_token": {
9
+ "content": "<s>",
10
+ "lstrip": false,
11
+ "normalized": false,
12
+ "rstrip": false,
13
+ "single_word": false
14
+ },
15
+ "eos_token": {
16
+ "content": "<|endoftext|>",
17
+ "lstrip": false,
18
+ "normalized": false,
19
+ "rstrip": false,
20
+ "single_word": false
21
+ },
22
+ "pad_token": {
23
+ "content": "<|endoftext|>",
24
+ "lstrip": false,
25
+ "normalized": false,
26
+ "rstrip": false,
27
+ "single_word": false
28
+ },
29
+ "unk_token": {
30
+ "content": "<unk>",
31
+ "lstrip": false,
32
+ "normalized": false,
33
+ "rstrip": false,
34
+ "single_word": false
35
+ }
36
+ }
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,408 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": true,
3
+ "add_eos_token": false,
4
+ "added_tokens_decoder": {
5
+ "0": {
6
+ "content": "<unk>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false,
11
+ "special": true
12
+ },
13
+ "1": {
14
+ "content": "<s>",
15
+ "lstrip": false,
16
+ "normalized": false,
17
+ "rstrip": false,
18
+ "single_word": false,
19
+ "special": true
20
+ },
21
+ "2": {
22
+ "content": "</s>",
23
+ "lstrip": false,
24
+ "normalized": false,
25
+ "rstrip": true,
26
+ "single_word": false,
27
+ "special": false
28
+ },
29
+ "32000": {
30
+ "content": "<|endoftext|>",
31
+ "lstrip": false,
32
+ "normalized": false,
33
+ "rstrip": false,
34
+ "single_word": false,
35
+ "special": true
36
+ },
37
+ "32001": {
38
+ "content": "<|assistant|>",
39
+ "lstrip": false,
40
+ "normalized": false,
41
+ "rstrip": true,
42
+ "single_word": false,
43
+ "special": true
44
+ },
45
+ "32002": {
46
+ "content": "<|placeholder1|>",
47
+ "lstrip": false,
48
+ "normalized": false,
49
+ "rstrip": true,
50
+ "single_word": false,
51
+ "special": true
52
+ },
53
+ "32003": {
54
+ "content": "<|placeholder2|>",
55
+ "lstrip": false,
56
+ "normalized": false,
57
+ "rstrip": true,
58
+ "single_word": false,
59
+ "special": true
60
+ },
61
+ "32004": {
62
+ "content": "<|placeholder3|>",
63
+ "lstrip": false,
64
+ "normalized": false,
65
+ "rstrip": true,
66
+ "single_word": false,
67
+ "special": true
68
+ },
69
+ "32005": {
70
+ "content": "<|placeholder4|>",
71
+ "lstrip": false,
72
+ "normalized": false,
73
+ "rstrip": true,
74
+ "single_word": false,
75
+ "special": true
76
+ },
77
+ "32006": {
78
+ "content": "<|system|>",
79
+ "lstrip": false,
80
+ "normalized": false,
81
+ "rstrip": false,
82
+ "single_word": false,
83
+ "special": true
84
+ },
85
+ "32007": {
86
+ "content": "<|end|>",
87
+ "lstrip": false,
88
+ "normalized": false,
89
+ "rstrip": false,
90
+ "single_word": false,
91
+ "special": true
92
+ },
93
+ "32008": {
94
+ "content": "<|placeholder5|>",
95
+ "lstrip": false,
96
+ "normalized": false,
97
+ "rstrip": true,
98
+ "single_word": false,
99
+ "special": true
100
+ },
101
+ "32009": {
102
+ "content": "<|placeholder6|>",
103
+ "lstrip": false,
104
+ "normalized": false,
105
+ "rstrip": true,
106
+ "single_word": false,
107
+ "special": true
108
+ },
109
+ "32010": {
110
+ "content": "<|user|>",
111
+ "lstrip": false,
112
+ "normalized": false,
113
+ "rstrip": false,
114
+ "single_word": false,
115
+ "special": true
116
+ },
117
+ "32011": {
118
+ "content": "<|placeholder7|>",
119
+ "lstrip": false,
120
+ "normalized": false,
121
+ "rstrip": true,
122
+ "single_word": false,
123
+ "special": true
124
+ },
125
+ "32012": {
126
+ "content": "<|placeholder8|>",
127
+ "lstrip": false,
128
+ "normalized": false,
129
+ "rstrip": true,
130
+ "single_word": false,
131
+ "special": true
132
+ },
133
+ "32013": {
134
+ "content": "<|placeholder9|>",
135
+ "lstrip": false,
136
+ "normalized": false,
137
+ "rstrip": true,
138
+ "single_word": false,
139
+ "special": true
140
+ },
141
+ "32014": {
142
+ "content": "<|placeholder10|>",
143
+ "lstrip": false,
144
+ "normalized": false,
145
+ "rstrip": true,
146
+ "single_word": false,
147
+ "special": true
148
+ },
149
+ "32015": {
150
+ "content": "<|placeholder11|>",
151
+ "lstrip": false,
152
+ "normalized": false,
153
+ "rstrip": true,
154
+ "single_word": false,
155
+ "special": true
156
+ },
157
+ "32016": {
158
+ "content": "<|placeholder12|>",
159
+ "lstrip": false,
160
+ "normalized": false,
161
+ "rstrip": true,
162
+ "single_word": false,
163
+ "special": true
164
+ },
165
+ "32017": {
166
+ "content": "<|placeholder13|>",
167
+ "lstrip": false,
168
+ "normalized": false,
169
+ "rstrip": true,
170
+ "single_word": false,
171
+ "special": true
172
+ },
173
+ "32018": {
174
+ "content": "<|placeholder14|>",
175
+ "lstrip": false,
176
+ "normalized": false,
177
+ "rstrip": true,
178
+ "single_word": false,
179
+ "special": true
180
+ },
181
+ "32019": {
182
+ "content": "<|placeholder15|>",
183
+ "lstrip": false,
184
+ "normalized": false,
185
+ "rstrip": true,
186
+ "single_word": false,
187
+ "special": true
188
+ },
189
+ "32020": {
190
+ "content": "<|placeholder16|>",
191
+ "lstrip": false,
192
+ "normalized": false,
193
+ "rstrip": true,
194
+ "single_word": false,
195
+ "special": true
196
+ },
197
+ "32021": {
198
+ "content": "<|placeholder17|>",
199
+ "lstrip": false,
200
+ "normalized": false,
201
+ "rstrip": true,
202
+ "single_word": false,
203
+ "special": true
204
+ },
205
+ "32022": {
206
+ "content": "<|placeholder18|>",
207
+ "lstrip": false,
208
+ "normalized": false,
209
+ "rstrip": true,
210
+ "single_word": false,
211
+ "special": true
212
+ },
213
+ "32023": {
214
+ "content": "<|placeholder19|>",
215
+ "lstrip": false,
216
+ "normalized": false,
217
+ "rstrip": true,
218
+ "single_word": false,
219
+ "special": true
220
+ },
221
+ "32024": {
222
+ "content": "<|placeholder20|>",
223
+ "lstrip": false,
224
+ "normalized": false,
225
+ "rstrip": true,
226
+ "single_word": false,
227
+ "special": true
228
+ },
229
+ "32025": {
230
+ "content": "<|placeholder21|>",
231
+ "lstrip": false,
232
+ "normalized": false,
233
+ "rstrip": true,
234
+ "single_word": false,
235
+ "special": true
236
+ },
237
+ "32026": {
238
+ "content": "<|placeholder22|>",
239
+ "lstrip": false,
240
+ "normalized": false,
241
+ "rstrip": true,
242
+ "single_word": false,
243
+ "special": true
244
+ },
245
+ "32027": {
246
+ "content": "<|placeholder23|>",
247
+ "lstrip": false,
248
+ "normalized": false,
249
+ "rstrip": true,
250
+ "single_word": false,
251
+ "special": true
252
+ },
253
+ "32028": {
254
+ "content": "<|placeholder24|>",
255
+ "lstrip": false,
256
+ "normalized": false,
257
+ "rstrip": true,
258
+ "single_word": false,
259
+ "special": true
260
+ },
261
+ "32029": {
262
+ "content": "<|placeholder25|>",
263
+ "lstrip": false,
264
+ "normalized": false,
265
+ "rstrip": true,
266
+ "single_word": false,
267
+ "special": true
268
+ },
269
+ "32030": {
270
+ "content": "<|placeholder26|>",
271
+ "lstrip": false,
272
+ "normalized": false,
273
+ "rstrip": true,
274
+ "single_word": false,
275
+ "special": true
276
+ },
277
+ "32031": {
278
+ "content": "<|placeholder27|>",
279
+ "lstrip": false,
280
+ "normalized": false,
281
+ "rstrip": true,
282
+ "single_word": false,
283
+ "special": true
284
+ },
285
+ "32032": {
286
+ "content": "<|placeholder28|>",
287
+ "lstrip": false,
288
+ "normalized": false,
289
+ "rstrip": true,
290
+ "single_word": false,
291
+ "special": true
292
+ },
293
+ "32033": {
294
+ "content": "<|placeholder29|>",
295
+ "lstrip": false,
296
+ "normalized": false,
297
+ "rstrip": true,
298
+ "single_word": false,
299
+ "special": true
300
+ },
301
+ "32034": {
302
+ "content": "<|placeholder30|>",
303
+ "lstrip": false,
304
+ "normalized": false,
305
+ "rstrip": true,
306
+ "single_word": false,
307
+ "special": true
308
+ },
309
+ "32035": {
310
+ "content": "<|placeholder31|>",
311
+ "lstrip": false,
312
+ "normalized": false,
313
+ "rstrip": true,
314
+ "single_word": false,
315
+ "special": true
316
+ },
317
+ "32036": {
318
+ "content": "<|placeholder32|>",
319
+ "lstrip": false,
320
+ "normalized": false,
321
+ "rstrip": true,
322
+ "single_word": false,
323
+ "special": true
324
+ },
325
+ "32037": {
326
+ "content": "<|placeholder33|>",
327
+ "lstrip": false,
328
+ "normalized": false,
329
+ "rstrip": true,
330
+ "single_word": false,
331
+ "special": true
332
+ },
333
+ "32038": {
334
+ "content": "<|placeholder34|>",
335
+ "lstrip": false,
336
+ "normalized": false,
337
+ "rstrip": true,
338
+ "single_word": false,
339
+ "special": true
340
+ },
341
+ "32039": {
342
+ "content": "<|placeholder35|>",
343
+ "lstrip": false,
344
+ "normalized": false,
345
+ "rstrip": true,
346
+ "single_word": false,
347
+ "special": true
348
+ },
349
+ "32040": {
350
+ "content": "<|placeholder36|>",
351
+ "lstrip": false,
352
+ "normalized": false,
353
+ "rstrip": true,
354
+ "single_word": false,
355
+ "special": true
356
+ },
357
+ "32041": {
358
+ "content": "<|placeholder37|>",
359
+ "lstrip": false,
360
+ "normalized": false,
361
+ "rstrip": true,
362
+ "single_word": false,
363
+ "special": true
364
+ },
365
+ "32042": {
366
+ "content": "<|placeholder38|>",
367
+ "lstrip": false,
368
+ "normalized": false,
369
+ "rstrip": true,
370
+ "single_word": false,
371
+ "special": true
372
+ },
373
+ "32043": {
374
+ "content": "<|placeholder39|>",
375
+ "lstrip": false,
376
+ "normalized": false,
377
+ "rstrip": true,
378
+ "single_word": false,
379
+ "special": true
380
+ },
381
+ "32044": {
382
+ "content": "<|image|>",
383
+ "lstrip": false,
384
+ "normalized": false,
385
+ "rstrip": true,
386
+ "single_word": false,
387
+ "special": true
388
+ }
389
+ },
390
+ "additional_special_tokens": [
391
+ "<|system|>",
392
+ "<|end|>",
393
+ "<|user|>",
394
+ "<|end|>"
395
+ ],
396
+ "bos_token": "<s>",
397
+ "chat_template": "{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{- '<|assistant|>\n' -}}{% endif %}",
398
+ "clean_up_tokenization_spaces": false,
399
+ "eos_token": "<|endoftext|>",
400
+ "legacy": false,
401
+ "model_max_length": 131072,
402
+ "pad_token": "<|endoftext|>",
403
+ "padding_side": "right",
404
+ "sp_model_kwargs": {},
405
+ "tokenizer_class": "LlamaTokenizer",
406
+ "unk_token": "<unk>",
407
+ "use_default_system_prompt": false
408
+ }