Malaysia-AI

We got extra server and we decided to self host OSRM and map tiling for MY-SG-Brunei location! 1. OSRM is to calculate driving distance from points to points, and this api should compatible with https://lnkd.in/guCM8YEy, you need to replace, https://lnkd.in/gWEgT--2 -> https://lnkd.in/gNXemnBV, if you are malaysian, average requests ~500ms. (picture 1) 2. map tiling, we serve 3 styles, liberty, test and basic. i. https://lnkd.in/gzExAYfu (picture 2) ii. https://lnkd.in/gugBaYSw (picture 3) iii. https://lnkd.in/gD-BHRKs (picture 4) You can use these to add maps in your app for free! You can check simple example using leaflet, jsfiddle.net/ordxaqk0/1/ Contact us if you want contribute some styles, use editor from https://maputnik.github.io 3. Why? we used to pay hundred dollars per day just for a simple map web app, super stressful, so hopefully all of these can help someone. No limits, no registration, no cookies, no throttle. What other APIs we can serve publicly for Malaysian?

We got extra server and we decided to self host OSRM and map tiling for MY-SG-Brunei location! 1. OSRM is to calculate driving distance from points to points, and this api should compatible with https://lnkd.in/guCM8YEy, you need to replace, https://lnkd.in/gWEgT--2 -> https://lnkd.in/gNXemnBV, if you are malaysian, average requests ~500ms. (picture 1) 2. map tiling, we serve 3 styles, liberty, test and basic. i. https://lnkd.in/gzExAYfu (picture 2) ii. https://lnkd.in/gugBaYSw (picture 3) iii. https://lnkd.in/gD-BHRKs (picture 4) You can use these to add maps in your app for free! You can check simple example using leaflet, jsfiddle.net/ordxaqk0/1/ Contact us if you want contribute some styles, use editor from https://maputnik.github.io 3. Why? we used to pay hundred dollars per day just for a simple map web app, super stressful, so hopefully all of these can help someone. No limits, no registration, no cookies, no throttle. What other APIs we can serve publicly for Malaysian?

Another blog! It is about 2D Parallelism using Pytorch Distributed in Ray. https://lnkd.in/gTfAJJ6q 1. We explained and visualized the 2D Parallelism, if you have 4 layers and do 2 PP 2 TP, you required 2 TP groups and 1 PP group. So minimum 4 GPUs. 2. We do not have a node with 4 GPUs, but we got 2 nodes with each node 2 GPUs, so we use Ray to do the Parallelism. We compared with Slurm and why we decided with Ray. Super great framework from Anyscale 3. Ray natively integrated with Prometheus for real-time metrics monitoring, so when talk about Prometheus, you can do a lot of stuffs such as real-time alerts to any channels you prefer integrated with Grafana, or logging correlation! 4. And finally the 2D Parallelism, the code and we explained the data movement between the nodes and GPUs.

Great work from Aisyah Razak Ariff Nazhan Kamarul Adha Wan Adzhar Faiq Adzlan Mas Aisyah Ahmad and Ammar Azman for this preprint! We wish this work able to promote better public safety for localize LLM.

Another blog! It is about Tensor Parallelism using PyTorch. https://lnkd.in/ggtBzYzS 1. We talked about the difference between Tensor Parallelism and Pipeline Parallelism. 2. Row-Wise vs Column-Wise Parallelism. 3. Some intro about PyTorch Distributed Elastic. 4. And the data movement during the Parallelism. TP helps train or serve large AI models by spreading the work across multiple GPUs, necessary to enable experimenting with advanced models. If you are serious beyond prompting and want to understand better about Model Parallelism, TP is a good starting point.

Check out our second blog related to CUDA! Vector operation for N dimension, and N dimension can be beyond three dimension! https://lnkd.in/gQq5UCdc

Check out our first blog related to CUDA! If you understand `blockDim`, `blockIdx`, `threadIdx`, and pointers, you are good to go, CUDA already put a lot of abstractions for us to write CUDA programming. https://lnkd.in/g88_Fsfr

Thanks for using our models!

We generated beyond 100k context length synthetic Malaysian QA including CoT to verify the answers, 1. Parliament docs, pdf docs from parliament website, https://lnkd.in/eiCFV_G7 2. Ejudgement, pdf docs from Ejudgement website, https://lnkd.in/eASiV9JN 3. Hansard, pdf docs from hansard, https://lnkd.in/e_7Kf3WD 4. gov.my PDF files, pdf files under gov.my domain, https://lnkd.in/ejrG8we4 5. Maktabalbahri, https://lnkd.in/exWaadem These dataset are super useful for post-training to increase your LLM beyond 100k context length, SFT -> QLoRA Alignment -> QLoRA YaRN / ChunkLlama -> QLoRA Long Context

We generated beyond 100k context length synthetic Malaysian QA including CoT to verify the answers, 1. Parliament docs, pdf docs from parliament website, https://lnkd.in/eiCFV_G7 2. Ejudgement, pdf docs from Ejudgement website, https://lnkd.in/eASiV9JN 3. Hansard, pdf docs from hansard, https://lnkd.in/e_7Kf3WD 4. gov.my PDF files, pdf files under gov.my domain, https://lnkd.in/ejrG8we4 5. Maktabalbahri, https://lnkd.in/exWaadem These dataset are super useful for post-training to increase your LLM beyond 100k context length, SFT -> QLoRA Alignment -> QLoRA YaRN / ChunkLlama -> QLoRA Long Context