Reaching out to mentors for project 12 "Enable popular Keras hub GenAI/LLM pipelines for the OpenVINO backend in Keras 3 workflow and optimize" GSOC'2025 #29228
Replies: 5 comments 10 replies
-
Beta Was this translation helpful? Give feedback.
-
|
In your proposal/application, you should respond to these questions you raised:) Try to justify the selected pipelines for enabling. Study optimization techniques for LLMs inference provided by OpenVINO GenAI, Keras and try to design approach to combine these solutions to reach the best performance. Best regards, |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for the guidance! I’ve started working on my proposal and am diving deeper into optimization techniques for LLM inference in OpenVINO and Keras. I’d really appreciate your feedback as I refine my approach, and I’ll be reaching out for guidance and reviews from time to time. |
Beta Was this translation helpful? Give feedback.
-
|
I’ve sent an email requesting a review of my initial proposal to ensure I’m on the right track. I also asked about the possibility of improving our communication, perhaps via Discord or any platform you prefer. Looking forward to your thoughts! |
Beta Was this translation helpful? Give feedback.
-
|
I’ve sent an email requesting a review of my initial proposal to ensure I’m on the right track. I also asked about the possibility of improving our communication, perhaps via Discord or any platform you prefer. Looking forward to your thoughts! |
Beta Was this translation helpful? Give feedback.
-
|
@adrianboguszewski , @mlukasze |
Beta Was this translation helpful? Give feedback.
-
|
Some information important to candidates for this project. In order to stand out of crowd: Try to create prototype to show that GPT-2 or other GenAI pipeline is working. It will a good addition to your proposal. Cnadidates having such prototype will be connected with me via Discord and will show case it privately. Do not disclose it prototype publicly. Use less ChatGPT (that is obvious when people use it), more do real experiments and describe results in proposal. Try to answer on questions: what kind of quantization we will use: one that is provided by Keras; or NNCF, or something is already available on Keras Hub. Describe all of these scenaries on your proposal. |
Beta Was this translation helpful? Give feedback.
-
|
Hi @rkazants, |
Beta Was this translation helpful? Give feedback.
-
|
@adrianboguszewski , |
Beta Was this translation helpful? Give feedback.
-
|
@Mohamed-Ashraf273 Roman is out of office, so please talk to other mentors (@mvafin @popovaan) |
Beta Was this translation helpful? Give feedback.
-
|
I hope you're doing well. I’m interested in applying for Project 12: "Enable Popular Keras Hub GenAI/LLM Pipelines for OpenVINO Backend" in GSoC 2025. I have already drafted a proposal. I would greatly appreciate your guidance and feedback on both. I’ve tried reaching out to mentors but haven’t connected yet. If possible, I’d appreciate it if you could follow up with me and provide insights. I’d be happy to discuss this further via Discord or LinkedIn, whichever is more convenient for you. Looking forward to your response! Best regards, |
Beta Was this translation helpful? Give feedback.
-
|
Hi @Mohamed-Ashraf273 you can contact me with your questions. If you prefer discord: my alias is |
Beta Was this translation helpful? Give feedback.
-
|
@mvafin |
Beta Was this translation helpful? Give feedback.
-
|
Beta Was this translation helpful? Give feedback.
-
|
@mvafin |
Beta Was this translation helpful? Give feedback.
-
Dear Roman Kazantsev @rkazants , Maxim Vafin @mvafin , Anastasia Popova, and Andrei Kochin,
I'm Mohamed, a third-year Computer Engineering student, and I am highly interested in working on and contributing to the "Enable Popular Keras Hub GenAI/LLM Pipelines for OpenVINO Backend" project.
Here is what I have learned so far:
I have experience working with Python, TensorFlow, Keras, and I've fine-tuned a lot of models and achieved acceptable performance, which I have used in my previous machine-learning projects.
I contributed to OpenVINO's Keras 3 backend #20945, #20934 and OpenVINO's PyTorch Frontend #29142. Through this, I have gained insight into OpenVINO’s execution flow and backend integration.
I have a solid understanding of model inference optimization and how backend implementations affect performance, and I have already started learning KV-cache and LoRA adapters and I am eager to explore them further.
I have a few queries:
Are there specific GenAI/LLM models from Keras Hub that you recommend focusing on first?
Apart from KV-cache and LoRA, would other optimization techniques like quantization or model distillation be relevant for this project?
What OpenVINO benchmarks or performance targets should we aim for, especially in comparison to TensorFlow, PyTorch, and JAX?
I would love to discuss these ideas further and understand how I can contribute effectively to the project.
@adrianboguszewski , @mlukasze
Could you please help connect me with the mentors for further discussions?
Best regards,
Mohamed
Beta Was this translation helpful? Give feedback.
All reactions