Hussein Awala

[Java] developers can use any deep learning model as long as it is ONNX-compatible, without the need to rewrite the entire backend in Python. This can save valuable time for many teams, especially if your product backend is in Java. 💡 I have implemented this example, a U2Net model trained in Python can be seamlessly used in Java via ONNXRuntime Java. 🚀 - First, a mask is generated using the U2Net ONNX model. 🎭 - Then, a composite image is calculated using the inputs (image, mask). 🖼️ - Finally, the final image is generated using inputs (image, background). 🌟 #JavaDevelopment #DeepLearning #ONNX #Integration #JavaBackend

9

5 commentaires

Andrew Atkinson: PostgreSQL 17: JSON_TABLE(), MERGE with RETURNING, and Updatable Views 🚀 PostgreSQL 17 is here, and it's about as flashy as your average Nordic noir detective drama. 😉 But don't be fooled—there are some gems in this new release. Andrew Atkinson's deep dive reveals the power of JSON_TABLE() for transforming JSON to table-like query results. Need to extract the author, ISBN, and publisher from a messy JSON blob? JSON_TABLE() has your back, making data retrieval as smooth as an espresso shot on a sleepy morning. Check out Andrew's runnable examples if you don't believe me. The MERGE command also got a glow-up, embracing the RETURNING clause like a long-lost friend. Why run multiple queries when you can get everything back in one neat package? It's like the Swiss Army knife of SQL operations! And let's not forget about views! They were always updatable; now they have triggers for an encore. Update a non-admin view to promote someone to admin status—no code changes needed. Pure magic. Plus, PostgreSQL 17 saves server memory like a hoarder finding extra closet space. Your database will thank you. Dig into Andrew's article for these and more. It’s worth it. 🌟 #PostgreSQL17 #JSON #SQLMagic #AndrewAtkinson You can find the full article here: https://postgr.es/p/6F4

4

Here's a secret Snowflake tip 🤫: with the ability to host Notebooks on the new container runtime, we can now install arbitrary packages using !pip. 🔥 Yes, we can also install packages not available on the Anaconda Snowflake repo (with all the risks that come with it 😅). For instance, now you can install the Cinemagoer package and search the IMDb for all the movies containing the word "matrix" in the title 👇

89

In case you missed the release of the #MySQL Connector/Python 9.0.0, it deprecated several cursors, updated X DevAPI, and removed Debian, Solaris, and most RPM (other than Enterprise Linux). Now, you install Connector/Python with the pip utility. #python #softwareengineering #programming #devops #cloud #systemengineering #pip https://lnkd.in/gnxaGVnr

2

❓ Do you use dbt Labs in Apache Airflow? 🤗 We value your input and want to hear from YOU! 🙏🏽 Please take a few minutes to fill out this survey, and you will help improve the user experience of dbt users in Airflow: https://bit.ly/4egyJnf

80

Building a Single-node Lakehouse using 'Hudi-rs' - no JVM, Spark [New Blog] The Apache Hudi community recently released "hudi-rs" - a native Rust library for Hudi with Python bindings. This makes it possible to work with #lakehouse platform like Hudi across a range of use cases that typically don’t require distributed processing. For eg, data science & machine learning applications need frequent access to diverse datasets for exploratory data analysis/model training. Direct access to Hudi datasets for these use cases reduces the ‘wait time’ on data stakeholders, and speeds up the time-to-insight process. Traditionally, accessing these tables involves configuring Java, Spark, Hadoop, and other related dependencies. This is what "Hudi-rs" addresses & makes it easy for data consumers to get started. In this introductory blog, I go over - ✅ the internals of "Hudi-rs" with #ApacheArrow under the hood ✅ how to use it with popular single node computes such as DuckDB, Polars, Daft & Apache Datafusion Check out the link in the comments. And reach out if you want to get involved with the project! #dataengineering #softwareengineering

170

16 commentaires

MisguidedAttention : a collection of prompts to challenge the reasoning abilities of large language models. In this github repo, authors introduced  prompts with slight variations of commonly known thought experiments or paradoxes ("trick questions"), which can challenge the reasoning abilities of LLMs. 𝗘𝘅𝗽𝗲𝗿𝗶𝗺𝗲𝗻𝘁𝗮𝗹 𝗢𝗯𝘀𝗲𝗿𝘃𝗮𝘁𝗶𝗼𝗻𝘀 𝘂𝘀𝗶𝗻𝗴 𝘁𝗵𝗲𝘀𝗲 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 𝗳𝗼𝗿 𝗟𝗟𝗠 𝘁𝗲𝘀𝘁𝗶𝗻𝗴 - Expected behavior would be that the LLMs solving the problems with logical deduction, but many LLMs will mistakenly recognize the unmodified problem due to frequent occurrence in their training data. - When LLMs not working, they responded with a solution to the unmodified problem instead of going through the details step-by-step to find a solution for the modified problem. -  it's also possible to observe intertwined strings of reasoning where conflicting thoughts are alternating in the same text. -  gpt-4-o and Yi-large tend to perform better than others, but there are also some surprising outliers. 𝗣𝗿𝗼𝗯𝗹𝗲𝗺𝘀 𝘁𝗵𝗮𝘁 𝗳𝗼𝗿𝗺𝗲𝗱 𝗯𝗮𝘀𝗶𝘀 𝗼𝗳 𝗺𝗼𝗱𝗶𝗳𝗶𝗲𝗱 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 - Trolley problem: https://lnkd.in/euQHgvVg - Monty Hall problem: https://lnkd.in/e9cC9wdb - Barber paradox: https://lnkd.in/edHhQxPw - Schrödingers cat: https://lnkd.in/e_vYW9CY - Unexpected hanging paradox: https://lnkd.in/eZ642wmZ - River crossing puzzle: https://lnkd.in/eZMXJryn - Two doors problem, apparently a variant of Knights and Knaves https://lnkd.in/enQQKpcr from this movie 𝗣𝗿𝗼𝗺𝗽𝘁𝘀 𝗿𝗲𝗽𝗼: https://lnkd.in/eGP9Shes #generativeai

23

⚡️ Perf demands foresight: reach 𝟭𝟬-𝟭𝟬𝟬𝘅 perf speedup compiling Python with Codon Interpreted code offers flexibility, but it's hard to achieve high performance on-the-fly ... Codon uses LLVM to compile Python to machine code, giving you multiple 𝗮𝗱𝘃𝗮𝗻𝘁𝗮𝗴𝗲𝘀: 👉🏽 Single thread 10 to 100x or even more speedups 👉🏽 Native multithreading (unlike Python) 👉🏽 Simple GPU kernel execution 👉🏽 Packaging into a simple executable file The overhead is pretty-low so why not give a boost to your Python script with Codon?

16

Snowflake presents `Arctic` a new Beast ❄️ Key points from the announcement blog: - Arctic outperforms open-source models like DBRX, Llama3-70B, and Mixtral8x7B in enterprise tasks such as SQL generation, coding, and meticulous adherence to instructions. - What sets Arctic apart is its ability to deliver top-tier enterprise intelligence at a remarkably low training cost, coming in at under $2 million that makes it an attractive option for enterprises seeking to develop custom models. - Its exceptional training efficiency comes from a distinctive Dense-MoE Hybrid transformer architecture. - This architecture merges a 10B dense transformer with a 128×3.66B MoE MLP, resulting in optimal performance. - Additionally, Snowflake provides a comprehensive Cookbook that shares invaluable research insights and strategies for efficiently and economically constructing world-class MoE models from the training of Arctic: https://lnkd.in/gDgjg334 Btw, Snowflake claims Arctic as the premier LLM for enterprises. It becomes more important to understand the safety standards or the dark side of the model that we usually ignore. So, keep an eye out for the updates to know more about the dark-side of it.

22

🫣 I started a viral #Discord thread on production code quality in the ML community. With the rise of #AI, it's time to re-examine 🗣️: Key Practices for Production-Level Code: ✅ Basic Setup: Static code analysis, pre-commit hooks, Tox, or Nox for test automation ✅ Automated Testing: pytest, Coverage.py, Nox, pytest-mock ✅ Linting & Formatting: Flake8, Black, Bandit, Safety, pre-commit ✅ Type Annotations: mypy, pytype, Typeguard, Desert, Marshmallow ✅ Documentation: Flake8-docstrings, Darglint, Sphinx, xdoctest 🤌 Emerging tools: 🌟 prefix.dev is a game-changer for data science and AI, offering consistent and efficient solutions for managing dependencies. 💭 The Future of Coding and Engineering: From the folks building Cursor put it: ¨the hybrid engineer will have effortless control over their codebase, iterating at the speed of their judgment. By combining AI and human ingenuity, they will out-smart and out-engineer the best pure-AI systems.¨ ❓ Open Questions for the Community: - What best practices are still relevant in the age of AI-assisted development? - What other best practices should we include? - Will code quality become an AI task? More resources mentioned in the thread - #FSDL repository and https://goodresearch.dev/ - Adopt this guideline: https://12factor.net/ - Refactoring: https://refactoring.guru/ - Code system analysis: https://lnkd.in/ggQ5Q5Sh - Code smell cheat: https://lnkd.in/gUzbghKy More: https://lnkd.in/gnnSkUtv

9

2 commentaires

Snowflake released "Arctic", a fully open-source LLM which is on-par or better than both Llama 3 8B and Llama 2 70B on all metrics, while using less than ½ of the training compute budget. It uses a unique Dense-MoE Hybrid transformer architecture with 128 experts and 480B parameters. However, it adopts a top-2 gating to choose 17B active parameters. Enterprise tasks such as SQL Generation, Coding, and Instruction Following are performed with an accuracy level close to Llama3 70B, despite the significantly lower training cost (about one-seventeenth) Despite using 7x less computing than Databricks DBRX, it remains competitive in Language Understanding and Reasoning while being better in Math. Weights and code (as well as data recipes and research insights) are open-source under Apache 2.0 licence. The model is downloadable from HuggingFace: https://lnkd.in/da9VfVja GitHub repo (inference and fine-tuning recipes): https://lnkd.in/dYC2qG9G Cookbook: https://lnkd.in/denfNJX5 Live demo: https://lnkd.in/dkttBF2b https://lnkd.in/db_DtKrG Blog Post: https://lnkd.in/d946D2X9

43

2 commentaires

Why is PostgreSQL popular? Check out this article: #softwareengineering #postgres #devops #itmanagement #datascience #dataanalytics https://lnkd.in/gCq2zzZW

8

1 commentaire

Want to test Meta, Google, and Mistral models all on one platform ? **Groq.com** Why Groq? Because speed matters. GroqCloud supports a diverse array of models, enabling you to select the best for your needs: ➡ llama3-8b-8192 by Meta (Context Window: 8,192 tokens) ➡ llama3-70b-8192 by Meta (Context Window: 8,192 tokens) ➡ mixtral-8x7b-32768 by Mistral (Context Window: 32,768 tokens) ➡ gemma-7b-it by Google (Context Window: 8,192 tokens) Groq's commitment to GenAI inference speed is transforming real-time AI applications with its LPU. What is an LPU? ➡ Language Processing Unit, is a revolutionary processing system designed specifically for LANGUAGE applications. Unlike traditional CPUs (Central Processing Units) and GPUs (Graphics Processing Units), which handle a broad range of tasks. LPUs are optimized for the rapid processing demands of LLMs This specialized architecture allows for more efficient handling of sequential data, significantly reducing the time per word processed. ➡ That means you get faster outputs You can run Llama-2 70B at over 300 tokens per second per user!

24

One particularly powerful thing about open-source LLMs with Ollama and Timescale's pgai for #Postgres is that you can now host your own models without needing to pay or share private data with closed-sourced providers like OpenAI.

24

FlexAttention: Simplifying Attention in PyTorch _________________________________________________________________ FlexAttention is a new PyTorch API that combines the flexibility of PyTorch with the performance of optimized attention methods. It allows researchers to implement various attention mechanisms easily without needing to write custom kernels. Key Points: - Flexibility with Performance: -- Traditional attention methods offer high performance but limited flexibility. -- FlexAttention provides a flexible API that supports many attention variants with minimal effort. - How It Works: -- Standard Attention: Calculates attention scores using queries, keys, and values, normalizes them, and computes the weighted sum of values. -- FlexAttention: Allows for modifications to the attention scores through a flexible scoring function, enabling customization for different types of attention mechanisms. - Examples of Attention Variants: -- Relative Position Encoding: Adds positional information to attention scores based on relative positions. -- ALiBi Bias: Applies a bias to attention scores to account for position differences, enhancing model performance. -- Causal Mask: Masks future positions in the sequence to ensure the model only attends to previous positions, useful for autoregressive models. Handling Sparsity: -- BlockMask: Efficiently manages sparse attention masks, such as those used in causal attention, by creating masks that prevent unnecessary computations. Learn More: https://lnkd.in/g58pxB4A

8

Breaking down Apache XTable's Architecture. Apache XTable (Incubating) is an omni-directional translation layer on top of open table formats such as Apache Hudi, Apache Iceberg & Delta Lake. It is NOT ❌ a new table format! Essentially what we are doing is this: SOURCE ---> (read metadata) ---> XTable's Model ---> write into TARGET We read the metadata from the SOURCE table format, put it as a unified representation & write the metadata in the TARGET format. * Note that we are only touching metadata, not the actual data files (such as #Parquet) with XTable. Let's breakdown its architecture. XTable’s architecture consists of three key components: 1. Conversion Source: ✅ These are table format specific modules responsible for reading metadata from the source ✅ They extract information like schema, transactions, partitions & translate it into XTable’s unified internal representation 2. Conversion Logic: ✅ This is the central processing unit of XTable ✅ It orchestrates the entire translation process, including initializing of all components, managing sources and targets, among other critical things 3. Conversion Target: ✅ These mirror the source readers ✅ They take the internal representation of the metadata & maps it to the target format’s metadata structure Blog in comments for a detailed read. #dataengineering #softwareengineering

73

3 commentaires

Dynamic Watermarking With IMGProxy and Apache APISIX https://lnkd.in/duYw4NNS Last week, I described how to add a dynamic watermark to your images on the JVM. I didn't find any library, so I had to develop the feature, or, more precisely, an embryo of a feature, by myself. Depending on your tech stack, you must search for an existing library or roll up your sleeves. For example, Rust offers such an out-of-the-box library. Worse, this approach might be impossible to implement if you don't have access to the source image. Another alternative is to use ready-made components, namely imgproxy and Apache APISIX. I already combined them to resize images on-the-fly.

1

Check out my latest talk on "Peak Performance in Python" at this year's PyConDE Open to feedback!

20