Note: set _attn_implementation='eager' if you don't have flash_attn installed

model = AutoModelForCausalLM.from_pretrained( model_id, device_map="cuda", trust_remote_code=True, torch_dtype="auto", _attn_implementation='flash_attention_2'
)

for best performance, use num_crops=4 for multi-frame, num_crops=16 for single-frame.

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True, num_crops=4 )

images = [] placeholder = ""

Note: if OOM, you might consider reduce number of frames in this example.

for i in range(1,20): url = f"https://image.slidesharecdn.com/azureintroduction-191206101932/75/Introduction-to-Microsoft-Azure-Cloud-{i}-2048.jpg" images.append(Image.open(requests.get(url, stream=True).raw)) placeholder += f"<|image_{i}|>\n"

messages = [ {"role": "user", "content": placeholder+"Summarize the deck of slides."}, ]

prompt = processor.tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True )

inputs = processor(prompt, images, return_tensors="pt").to("cuda:0")

generation_args = { "max_new_tokens": 1000, "temperature": 0.0, "do_sample": False, }

generate_ids = model.generate(**inputs, eos_token_id=processor.tokenizer.eos_token_id, **generation_args )

remove input tokens

generate_ids = generate_ids[:, inputs['input_ids'].shape[1]:] response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]

print(response)

Notes:
+ to achieve best performances we suggest to set _num_crops=4_ for multi-frame and _num_crops=16_ for single-frame.
+ to turn off flash_attention users can set __attn_implementation='eager'_

## Responsible AI Considerations

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:  
* 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.   
* 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. 
* 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. 
* Information Reliability: Language models can generate nonsensical content or fabricate content that might sound reasonable but is inaccurate or outdated.  
* 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.   

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: 

* 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.
* 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. 
* 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).   
* Generation of Harmful Content: Developers should assess outputs for their context and use available safety classifiers or custom solutions appropriate for their use case. 
* 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.
* 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).   

## Training

### Models

**Architecture:** Phi-3.5-vision has 4.2B parameters and contains image encoder, connector, projector, and Phi-3 Mini language model.<br>
**Inputs:** Text and Image. It’s best suited for prompts using the chat format.<br>
**Context length:** 128K tokens<br>
**GPUs:** 256 A100-80G<br>
**Training time:** 6 days<br>
**Training data:** 500B tokens (vision tokens + text tokens)<br>
**Outputs:** Generated text in response to the input<br>
**Dates:** Trained between July and August 2024<br>
**Status:** This is a static model trained on an offline text dataset with cutoff date March 15, 2024. Future versions of the tuned models may be released as we improve models.<br>
**Release date:** August 2024<br>

### Data Overview

Our training data includes a wide variety of sources, and is a combination of 
1) publicly available documents filtered rigorously for quality, selected high-quality educational data and code;
2) selected high-quality image-text interleave data;
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, newly created multi-image and video data, e.g., short video clips/pair of two similar images;
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.

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. More details about data can be found in the [Phi-3 Technical Report](https://arxiv.org/pdf/2404.14219).

### How to finetune?
We recommend user to take a look at the [Phi-3 CookBook finetuning recipe for Vision](https://github.com/microsoft/Phi-3CookBook/blob/main/md/04.Fine-tuning/FineTuning_Vision.md)

## Benchmarks

To understand the capabilities, we compare Phi-3.5-vision with a set of models over a variety of zero-shot benchmarks using our internal benchmark platform. At the high-level overview of the model quality on representative benchmarks:

| Category | Benchmark | Phi-3.5-vision-instruct | Intern-VL-2-4B | Intern-VL-2-8B | Gemini-1.5-Flash | GPT-4o-mini 2024-7-18 | Claude-3.5-Sonnet | Gemini-1.5-Pro  | GPT-4o 2024-5-13 |
|--|--|--|--|--|--|--|--|--|--|
| Popular aggregated benchmark | MMMU (val) | 43.0 | 44.22 | 46.33 | 49.33 | 52.1 | 52.67 | 54.11 | 61.78 |
| | MMBench (dev-en) | 81.9 | 83.4 | 87.0 | 85.7 | 83.8 | 82.3 | 87.9 | 88.4 |
| Visual scientific knowledge reasoning | ScienceQA (img-test) | 91.3 | 94.9 | 95.9 | 84.5 | 84.0 | 73.8 | 86.0 | 88.5 |
| Visual math reasoning | MathVista (testmini) | 43.9 | 53.7 | 51.1 | 55.3 | 38.8 | 54.0 | 57.4 | 54.4 |
| | InterGPS (test) | 36.3 | 45.6 | 53.2 | 39.4 | 39.9 | 45.6 | 58.2 | 46.9 |
| Chart reasoning | AI2D (test) | 78.1 | 77.3 | 81.4 | 78.4 | 75.2 | 68.9 | 75.6 | 82.8 |
| | ChartQA (test) | 81.8 | 78.8 | 80.4 | 57.6 | 54.5 | 73.2 | 68.2 | 64.0 |
| Document Intelligence | TextVQA (val) | 72.0 | 66.2 | 68.8 | 67.4 | 70.9 | 70.5 | 64.5 | 75.6 |
| Object visual presence verification | POPE (test) | 86.1 | 83.3 | 84.2 | 86.1 | 83.6 | 76.6 | 89.3 | 87.0 |

## Safety Evaluation and Red-Teaming

**Approach** 
The Phi-3 family of models has adopted a robust safety post-training approach. This approach leverages a variety of both open-source and in-house generated datasets. 
The overall technique employed to do the safety alignment is a combination of SFT (Supervised Fine-Tuning) and RLHF (Reinforcement Learning from Human Feedback) approaches
by utilizing human-labeled and synthetic English-language datasets, including publicly available datasets focusing on helpfulness and harmlessness as well as various 
questions and answers targeted to multiple safety categories.

**Safety Evaluation**
We leveraged various evaluation techniques including red teaming, adversarial conversation simulations, and safety evaluation benchmark datasets to evaluate Phi-3.5 
models' propensity to produce undesirable outputs across multiple risk categories. Several approaches were used to compensate for the limitations of one approach alone. 
Please refer to the [technical report](https://arxiv.org/pdf/2404.14219) for more details of our safety alignment.


## Software
* [PyTorch](https://github.com/pytorch/pytorch)
* [Transformers](https://github.com/huggingface/transformers)
* [Flash-Attention](https://github.com/HazyResearch/flash-attention)

## Hardware
Note that by default, the Phi-3.5-Mini-Instruct model uses flash attention, which requires certain types of GPU hardware to run. We have tested on the following GPU types:
* NVIDIA A100
* NVIDIA A6000
* NVIDIA H100
  
## License
The model is licensed under the [MIT license](./LICENSE).

## Trademarks
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.

## Data Summary
https://huggingface.co/microsoft/Phi-3.5-vision-instruct/blob/main/data_summary_card.md