Internals of vLLM Inference In the race to turn massive LLMs into reliable, inference infra is a real competitive edge - but do you know what goes on inside a library like vLLM? For example, I didn't know it used Ray under the hood! The vLLM framework handles an inference call end-to-end within one tightly orchestrated pipeline: 1) An OpenAI-style API server offloads tokenisation to an AsyncLLM process that forwards the request via async IPC to EngineCore. 2) The Scheduler packs tokens into continuous, token-budgeted batches, assigns fixed-size KV-cache “pages,” and hands the batched tensor to Ray orchestrated ModelRunners. 3) These runners execute forward passes with FlashAttention-3 (optionally via CUDA-graph capture). 4) Afterwards freshly generated tokens stream back through IPC to be detokenised and chunked out to the client. This is quite intuitive so definitely worth deep diving further for ML Practitioners to know what is going under the hood when serving popular models (eg Llama 4). |
|
|
|---|
|
Databricks Unified ML Governance Databricks published their approach to unified ML governance at the recent SIGMOD in Berlin providing an intuitive overview of how they approach this challenge: This is quite an interesting paper which shows how Unity Catalogue provides an open control plane that places tables, files, feature sets and MLflow-registered models in the same three-level namespace (catalog.schema.asset) while enforcing uniform row-, column- and path-level security via a one-asset-per-path rule and short-lived credential vending. This makes it possible for policies to hold whether Spark references the table name or another engine reads the raw object-store path. They also provide a glimpse in adoption and industry use, with roughly 100 M tables, 550 K volumes and 400 K models for 9 K customers, sustaining ~60 K metadata calls/s with latencies comparable to an on-cluster Hive metastore. |
|
|
|---|
|
A Deep Dive Into TPUs TPUs are redefining how large-scale AI gets built and powered, and this is a fantastic deep dive into the architecture and concepts across TPUs: Unlike GPUs that juggle lots of small tasks on the fly, TPUs are streamlined “assembly lines” for matrix math. These TPUs preload data into big on-chip buffers, run it through grid-like calculator blocks, and skip the energy-hungry cache layers. A single TPU chip can link up with three others on a board, sixty-four in a rack, and thousands in a pod, which are supported by lightning-fast optical switches that let Google carve the hardware into any shape a model needs. Because the XLA compiler figures out every data move in advance, engineers who keep their models’ shapes steady get massive speed and energy savings almost automatically, however for more general use-cases workloads will likely still prefer the flexibility of GPUs. |
|
|---|
|
Google MultiModal Released On-device multimodal AI in the open source is seeing some recent leaps, and Google’s Gemma 3n is the catalyst. Gemma 3n has been able to compresses the 5-8B parameters into models that can fit in 2-3 GB VRAM. IT uses what they refer to as a "Matryoshka-style" MatFormer architecture with Per-Layer Embeddings that park token tables in CPU RAM and a KV-cache sharing that halves pre-fill latency. Google suggests a score of 1300+ on LMArena, which would make it the first sub-10 B model to rival cloud-scale performance. It is great to see the momentum and progress for open models as we pretty much see new innovation being pushed to the limits on a weekly basis. |
|
|
|---|
|
LLM Transformers as Malware Your favourite AI LLM may be hiding malware! Recent research EvilModel shows how attackers can embed full binary payloads of malware in any standard LLM for malicious attacks: This research paper shows how an attacker can inject malicious code straight into redundant weight tensors without any model performance loss. This can be done by overwriting neurons with carefully crafted 32-bit floats that scarcely dent accuracy (<1% on ImageNet models) and evade every commercial vulnerability detection engine, enabling a simple extractor script to reassemble the malware once the checkpoint is loaded. This makes any pipeline that auto-downloads or updates third-party models a potential smuggling route, so all of us as machine learning practitioners must treat external model files + code like untrusted executables, verify provenance and hashes, and apply destructive sanitisation (quantisation, pruning, or re-serialisation) before deployment! |
|
|
|---|
|
Upcoming MLOps Events The MLOps ecosystem continues to grow at break-neck speeds, making it ever harder for us as practitioners to stay up to date with relevant developments. A fantsatic way to keep on-top of relevant resources is through the great community and events that the MLOps and Production ML ecosystem offers. This is the reason why we have started curating a list of upcoming events in the space, which are outlined below. Upcoming conferences where we're speaking:
Other upcoming MLOps conferences in 2025:
In case you missed our talks:
|
|
|---|
| | |
Check out the fast-growing ecosystem of production ML tools & frameworks at the github repository which has reached over 10,000 ⭐ github stars. We are currently looking for more libraries to add - if you know of any that are not listed, please let us know or feel free to add a PR. Four featured libraries in the GPU acceleration space are outlined below. - Kompute - Blazing fast, lightweight and mobile phone-enabled GPU compute framework optimized for advanced data processing usecases.
- CuPy - An implementation of NumPy-compatible multi-dimensional array on CUDA. CuPy consists of the core multi-dimensional array class, cupy.ndarray, and many functions on it.
- Jax - Composable transformations of Python+NumPy programs: differentiate, vectorize, JIT to GPU/TPU, and more
- CuDF - Built based on the Apache Arrow columnar memory format, cuDF is a GPU DataFrame library for loading, joining, aggregating, filtering, and otherwise manipulating data.
If you know of any open source and open community events that are not listed do give us a heads up so we can add them! |
|
|---|
| | |
As AI systems become more prevalent in society, we face bigger and tougher societal challenges. We have seen a large number of resources that aim to takle these challenges in the form of AI Guidelines, Principles, Ethics Frameworks, etc, however there are so many resources it is hard to navigate. Because of this we started an Open Source initiative that aims to map the ecosystem to make it simpler to navigate. You can find multiple principles in the repo - some examples include the following: - MLSecOps Top 10 Vulnerabilities - This is an initiative that aims to further the field of machine learning security by identifying the top 10 most common vulnerabiliites in the machine learning lifecycle as well as best practices.
- AI & Machine Learning 8 principles for Responsible ML - The Institute for Ethical AI & Machine Learning has put together 8 principles for responsible machine learning that are to be adopted by individuals and delivery teams designing, building and operating machine learning systems.
- An Evaluation of Guidelines - The Ethics of Ethics; A research paper that analyses multiple Ethics principles.
- ACM's Code of Ethics and Professional Conduct - This is the code of ethics that has been put together in 1992 by the Association for Computer Machinery and updated in 2018.
If you know of any guidelines that are not in the "Awesome AI Guidelines" list, please do give us a heads up or feel free to add a pull request!
|
|
|---|
| | |
| | | | The Institute for Ethical AI & Machine Learning is a European research centre that carries out world-class research into responsible machine learning. | | | | |
|
|
|---|
|
|
This email was sent to You received this email because you are registered with The Institute for Ethical AI & Machine Learning's newsletter "The Machine Learning Engineer"
|
| | | | |
|
|
|---|
|
© 2023 The Institute for Ethical AI & Machine Learning |
|
|---|
|
|
|
The Institute for Ethical AI & Machine Learning
