I’ve been working in AI for close to a decade, starting well before today’s LLM-driven wave.

My work is grounded in understanding AI across layers — from hardware, AI models, and data to system design, reliability, and scale. I approach AI as an end-to-end engineering problem: not just modeling, but how models interact with data, infrastructure, and real-world constraints.

Alongside this, I write The AI Merge, a technical newsletter on AI foundations and systems engineering, focused on clear thinking and practical system design rather than hype.

I’m currently building an AI Systems course, centered on end-to-end system design, for engineers who want a deeper, systems-first understanding of modern AI.

🗞️ Read The AI Merge: https://read.theaimerge.com

Alex Razvant

A lot of courses out there teach AI & LLMs, but very few touch on Inference

Sure, it's easy to grasp the fundamentals of what LLMs are and how to build systems around them, but scaling out an LLM out of a lab of 1-2 H100s running it and a single concurrent user session is a big challenge.

Efficient inference is one of the largest barriers to LLMs in production.

Batching, managing memory, concurrent user sessions, GPU saturation, and sharding with/or MIG (Multi-Instance GPUs), parallelisation are key concepts for engineers to know, understand, and work with or around them.

I've had this list of resources saved for a while, which fills some of those gaps.

1. Good Courses / Videos
- CS336 Stanford (Lecture 5-11): https://buff.ly/9o4jRxl
- GPU MODE: https://buff.ly/bPEHYL0 
- LLM Inference Patterns: https://buff.ly/4ASDS5c
- LLM Inference (NVIDIA): https://buff.ly/muDMcor 

2. Repositories / Blogs
- GPU Performance Engineering (E.Andere): https://buff.ly/sfGJNMK 
- AI Performance Engineering (C.Fregly) https://buff.ly/8ey3GPa 
- Inside NVIDIA GPUs (A.Gordic): https://buff.ly/wmgNwMg 
- Understanding LLM Inference (my article): https://buff.ly/NrNU3Qz 

I don't think there's a clear, step-by-step guide on how to master AI Perf Engineering.
The best bet is to study the core pieces of the models and the hardware they run on.

Then, learn the rest on demand.

undefined undefined

Neural Bits CTA

Join 4100+ engineers and get a weekly newsletter on advanced AI/ML. 👉 https://neuralbits.substack.com/

Kubrick

Ready to start building?

--- Resources ---
📌 Kubrick Code Repository
https://github.com/multi-modal-ai/multimodal-agents-course

📌 Neural Bits Newsletter (mine)
https://multimodalai.substack.com/

📌 The Neural Maze Newsletter (Miguel's)
https://theneuralmaze.substack.com/

Senior AI/ML Engineer | Author of NeuralBits Newsletter – helping you build efficient AI Systems

Hard work always pays off, be consistent!

I'm a Senior Machine Learning Engineer with a keen interest in developing AI/ML solutions for real world problems and helping others get started on their Machine Learning Journey.

Let's connect @:
📘 https://medium.com/@alexandrurazvant
✉️ alexandrurazvant@gmail.com
🌐 https://www.neuraleaps.com

Upskill your AI Engineering!

AI Engineering Hub

Senior AI Engineer | Writing The AI Merge Newsletter | Helping Engineers Build AI Beyond Demos

The 5 structure-layers of any LLM Model.
(Understand these, and you'll understand every OSS model on HF)

With each open-source LLM released, you'll find the model weights on Hugging Face.

When you go to the Files section, you won't find a `model.xyz` file, but multiple files, usually 10-15 that compose the entire LLM.

Here's the blueprint to understand them:

1️⃣ 𝘼𝙧𝙘𝙝𝙞𝙩𝙚𝙘𝙩𝙪𝙧𝙚 𝘾𝙤𝙣𝙛𝙞𝙜𝙪𝙧𝙖𝙩𝙞𝙤𝙣 (config.json)
This one contains metadata on model architecture, layer activations, sizes, vocabulary size, number of attention heads, model precision, and more. 

The transformers library knows how to parse this config and build the model architecture. 

2️⃣ 𝙈𝙤𝙙𝙚𝙡 𝙒𝙚𝙞𝙜𝙝𝙩𝙨 (the .safetensors splits)
LLMs have B (billions) of parameters.
For safer download, model weights are split into chunks/shards.
The formats used for weights are either Binary or Safetensors.

Safetensors format is an alternative to the default Pickle serializer that PyTorch (pt) uses, as it’s vulnerable to code injection.

3️⃣ 𝙇𝙖𝙮𝙚𝙧 𝙈𝙖𝙥𝙥𝙞𝙣𝙜 (the *.index.json)
Since the models are large and weights come as part files (e.g., 0001-of-0006, 0002-of-0006, etc.), this file stores a sequential map of the model architecture.

This specifies what layers are part of what model chunk.

4️⃣ 𝙏𝙤𝙠𝙚𝙣𝙞𝙯𝙚𝙧 𝘾𝙤𝙣𝙛𝙞𝙜 (tokenizer.json)
The tokenizer config file contains metadata about the tokenizer used to train this model. 

It contains the class name used to instantiate the tokenizer, the layer names, the structure of how the inputs are processed before passing through the model, and special tokens.

The LLM generally uses 𝘴𝘱𝘦𝘤𝘪𝘢𝘭_𝘵𝘰𝘬𝘦𝘯𝘴 as markers to stop generation.

5️⃣ 𝙂𝙚𝙣𝙚𝙧𝙖𝙩𝙞𝙤𝙣 𝘾𝙤𝙣𝙛𝙞𝙜 (generation_config.json)
These configuration files contain metadata for Inference, such as Temperature and TopP/TopK thresholds or context window size, with which the model was trained. 

When you load your model using Transformers.from_pretrained(), all these files are processed to set up everything you need to actually use that model.

---
Learn AI going through Fundamentals, Engineering and Systems.
Join here (free): https://buff.ly/DnpHEXD

The 4 patterns of LLM Inference.

Deploying LLMs is not the same as deploying traditional ML models. The differences show up most clearly at inference time. 

Some principles such as optimized models, infrastructure, GPU saturation and usage or distributed systems concepts still apply.

But LLM inference adds a new layer of complexity, as Transformer models inference is way more nuanced.

With LLMs, the output is generative and not a fixed size having batch-size input = batch-size output that more traditional ML or Deep Learning models had.

LLM inference splits into two stages, Prefill and Decode, this creates 4 distinct patterns an AI Engineer/Dev should know about before optimizing or deploying.

⇢ In the prefill phase, the model processes the entire input prompt and computes the key-value (KV) cache. In benchmarks, we'll often see this metric as `pp` or `prompt processing` speed, which is fast.

⇢  In the decode phase, it autoregressively generates tokens one at a time, updating the KV cache at each step. In benchmarks, this is refered to as `tg` or text-generation speeds, and this process is memory-bound as tokens have to travel within the GPU's memory during KV cache updates.

Understanding this gives you these four distinct LLM inference patterns:

SISO -- Short Input, Short Output
→ Low latency with fast generation. If using this pattern, aim to optimize for high concurrency.

LILO --  Long Input, Long Output
→ The most compute-intensive requires distributed inference, ideally with shared KV cache across nodes or tensor parallelism.

LISO -- Long Input, Short Output
→ This one has heavy prefill and quick decode. GPUs with enough VRAM would parallelize the prefill and generate fast.

SILO -- Short Input, Long Output
→ Light prefill, which runs fast but is heavier on decode. Decode latency dominates here.

---

𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆: The longer the output, the more time is spent in the decode loop.
Before fine-tuning your own LLM and deploying it, also consider the inference pattern your system will target.

Summary:
- In LLM inference, there are 4 patterns.
- For LISO, the generation and TFTT are fast.
- On SISO, everything is fast.
- For LILO,  a lot of computing.
- For SILO, prefill is fast, generation is slow.

---

Learn AI through Fundamentals, Engineering and Systems.
Follow and join here: https://buff.ly/iqFpm9x

If you'll ever try to run LLMs locally - you'll bump into llama.cpp

Most guides tell you what to use for Edge/CPU/low-end GPU inference - but not what’s actually happening under the hood. 

No context on:
- How is the model checkpoint parsed?
- How are tensors unpacked and mapped?
- How do GGML, GGUF, and llama.cpp fit together?

✅ So I went through the codebase myself. 

I spent a few hours digging through llama.cpp, PRs, and community threads, then turned everything into a single sequence diagram that shows the flow end-to-end. 

The key pieces of the stack:
> GGML - the ML Tensor Library and execution primitives (ops, mappings, kv-cache)
> GGUF - the format for storing quantized LLM models and Quantization types.
> llama.cpp - is the C++ inference engine, building on top of GGML & GGUF, wiring everything together.

If you’re running local LLMs, this is the stack you’re relying on.

If you want to understand why things break, how quantized models are actually loaded, this breakdown will help. 

---
I share more practical advice on AI Fundamentals, Engineering and Systems.
Join here (free): https://buff.ly/iqFpm9x

You should read more code.

If you’re already a Software Engineer or AI / ML Engineer, one of the fastest ways to level up right now is to study how big AI teams write and ship - in public. 

Many top companies and major contributors to the AI space have been open-sourcing their internal frameworks, kernels, or infrastructure libraries. 

You might feel overwhelmed by how fast AI is going. But this is a great learning opportunity if you know where to look.

Learning doesn’t just happen from papers and tutorials, especially if you already build software.

You should:
- Read real production code
- Study PRs or commit history on major/popular repos.
- See how teams structure codebases, design APIs, configs, tests

Check what gets reviewed, what gets rejected, and why.
This kind of learning will upgrade your engineering intuition. 

Here are some excellent codebases to learn from.

✅ From DeepSeek:
1. FlashMLA  - A library of high-performance attention kernels.
Link: https://buff.ly/XbFnPA7 

2. DeepEP - A Communication Library for high-throughput MoE workloads.
Link: https://buff.ly/3qxMc3S

3. 3FS - A Distributed Filesystem for AI Workloads, for AI training and inference.
Link: https://buff.ly/ovfW47A

✅ From NVIDIA:
1. Dynamo - A datacenter-scale Inference Serving Framework for GenAI workloads.
Link: https://buff.ly/WyAMzBz 

2. NeMo  -  A Scalable Generative AI Framework, for training and deploying LLMs and VLMs.
Link: https://buff.ly/L9KTuJp 

✅ From vLLM:
1. The vLLM Library  - One of the most popular LLM inference engines, designed for scale.
Link: https://buff.ly/GLCfoIM

✅ From SGLang:
1. The SGLang Engine - A high-performance serving engine for LLMs & multimodal models.
Link: https://buff.ly/m4HhsWQ 

✅ From llama.cpp
1. The llama.cpp Engine - A minimal C/C++ inference engine for LLMs with broad hardware support and quantization options.
Link: https://buff.ly/u9yaXwp

📌 Don't study everything, you won't be able to.

Instead, pay attention to:
⇾ Code structure & abstractions
How are core functionalities separated (core logic vs hardware-specific optimizations)?

⇾  PRs & commit history
What context do authors provide? How do reviewers push back?

⇾  Interfaces & public APIs
How are public APIs designed?

⇾  Tests & benchmarks
How do the test suites look? What CI pipelines do they have, profiling? 

Don’t just use these libraries. 
Read them.

That will build your intuition and help you get better at your craft, either as a curious learner or an engineer as part of a larger team at your job.

---
I distill my learnings with my audience on The AI Merge.
Check it here (8700+ engineers already do that) https://buff.ly/2rwqlNL

I have a surprise for my European audience, for NVIDIA GTC 2026 (March 16–19)

🎁 I’m giving away 1 x NVIDIA DGX Spark to someone who joins me for Virtual NVIDIA GTC 2026. 

I’ve spent the last 6 months building on NVIDIA DGX Spark, and when building AI Systems locally, the Spark is an incredible machine to learn, prototype, fine-tune, build and iterate fast.

Especially if you want more control over your entire stack.

NVIDIA GTC is the premier global AI conference, where developers, and researchers come together to explore the next wave of AI across every industry and application.

✅️ Virtual GTC is free. 

𝗛𝗼𝘄 𝘁𝗼 𝗲𝗻𝘁𝗲𝗿 𝘁𝗵𝗲 𝗗𝗚𝗫 𝗦𝗽𝗮𝗿𝗸 𝗴𝗶𝘃𝗲𝗮𝘄𝗮𝘆:
𝟭. 𝗥𝗲𝗴𝗶𝘀𝘁𝗲𝗿 for GTC 2026 using my link: https://buff.ly/aKgT9BA
𝟮. 𝗔𝘁𝘁𝗲𝗻𝗱 at least 1 virtual session (keynote does not count) - virtual sessions are free
𝟯. 𝗙𝗶𝗹𝗹 𝗼𝘂𝘁 𝗺𝘆 𝗴𝗶𝘃𝗲𝗮𝘄𝗮𝘆 𝗳𝗼𝗿𝗺 𝗮𝗳𝘁𝗲𝗿 𝗮𝘁𝘁𝗲𝗻𝗱𝗶𝗻𝗴

I’ll share the giveaway form link as we get closer to GTC 2026. 𝗦𝘁𝗮𝘆 𝘁𝘂𝗻𝗲𝗱.

Here are a few sessions that caught my eye (links below):
1. AI Factories in Europe: Building the Foundations for Scalable Intelligence [S81899] 
2. Accelerate AI Through Open-Source Inference [S81902] 
3. Teach AI to Code in Every Language With NVIDIA NeMo [S82306] 
4. From Data to Meaning: Vision-Language Models Shaping the Cities of Tomorrow [S81867] 
5. Your Learning Pathway: Get Certified for Career Success [C81544] 

If you’re planning to attend, remember to register here: https://buff.ly/aKgT9BA 

Follow me here and stay tuned for the giveaway form.

Underrated advice to grow as an Engineer

1. Go Depth first (2–3 years)
Pick one stack (language + framework + system) and get good at it.

Breadth gets easy after.
I know it's tempting to know a bit out of everything, but you'll spread yourself too thin, too soon.

2. If you're a junior/mid - talk to senior/staff engineers (smartly)

Don't ask “What should I do?” 
Instead, ask: “Here’s what I tried, here’s what happened, what would you check next?”

They've seen enough systems and code to spot patterns early on.

2. Get good at Code Reviews
Express your points, think past the small code component - oversee the entire system.
Don't nitpick for code changes unless necessary. 

3. No Vibe-coding
With each feature/component you build, spend time gathering context, asking for options, plan, review, verify and write the hard paths yourself.

One practice that works for me is, after drafting a plan for my coding Agent (in Plan Mode), I ask if it's got any questions for me, before starting to work.

It usually comes up with great points.

3. Study outside work
That's an easy switch.
You do it already by consuming content, scrolling through posts and watching videos.
But make sure you focus on what actually helps you, not all the news/hype out there.

Work getting better, use AI as leverage.

A better way to manage complex configurations in your AI/ML projects.

AI/ML systems usually have more moving parts and more configuration surfaces (data, training, evaluation, inference, orchestration). Keeping track of which configuration subset was used can get difficult fast. 

Sure, as per Factor III (Config) in 12-factor apps, environment-specific config can be stored as env variables.

However, env vars alone are usually not enough. You might need hierarchical configs, runtime composition, overrides etc.

For instance, on a current project I'm working on, I have:
- 6 MLOps pipelines that build the raw_data > deployed model lineage.
- 2 edge services that ingest video, route frames to Inference Servers
- 3 MCP servers, each with their own connection details, tools and resources
- 2 Agents, wired through A2A - each with their own subset of instructions and MCP

In this setup, during testing, services often require runtime overrides (MLOps pipelines), composition or swap (training/eval), and sometimes they share the same configuration domain. 

When configuration gets more complex, I found Hydra + OmegaConf to be the way to go.

Here's what you need to know:
1. 𝗢𝗺𝗲𝗴𝗮𝗖𝗼𝗻𝗳 is a hierarchical configuration system explicitly designed for complex applications like ML pipelines. 

You can:
↳ Merge multiple configuration files directly through runtime composition.
↳ Access fields via attribute notation "config.model.optimizer" or dict notation config["model"]["optimizer"]
↳ Add Type safety against schemas such as dataclasses or Pydantic base models.

2. 𝗛𝘆𝗱𝗿𝗮 builds on top of OmegaConf, creating a full-featured framework.

You can:
↳ Group multiple YAMLs into a single configuration file and reference them directly by filename.
↳ Run a sweep, same application - multiple configurations in parallel.

Concrete examples:
1. Your hydra config could look like this (config.yaml)
``` config.yaml
defaults:
  - data: mydata
  - training: trainAB
  - inference: inferAB
  - _self_ 
```

When loading the `config.yaml` file using Hydra, we get a nested configuration object, from which we can access any field 
(cfg . inference . batch_size / cfg . training . batch_size)

2. Override any config value at runtime. 
For example, if in the config.yaml, if you have:
```
model:
 optimizer: SGD
```

At runtime, you can override the value with:

```
python train model.optimizer=Adam
```

𝗧𝗵𝗶𝘀 𝗮𝗹𝘀𝗼 𝘄𝗼𝗿𝗸𝘀 𝗳𝗼𝗿 𝗻𝗲𝘀𝘁𝗲𝗱 𝗰𝗼𝗻𝗳𝗶𝗴𝘀.
↳ Run multiple configurations simultaneously using Hydra multirun.

```
python train -m data.dataset=A1, B12, C55
```

This will launch multiple runs (one per dataset value) and create separate output directories for each run. 

---

I share more practical advice on AI Fundamentals, Engineering and Systems.
Join here (free): https://buff.ly/iqFpm9x

Qwen3.5 4B and 9B around the corner? 👀

Code is cheap nowadays. Design and planning is not. 

Some, keep measuring output by volume. 

I've been doing AI Assisted Programming for the past 2.5 years, on real codebases.

Online, I keep seeing many claim to push 10k Lines of Code a day to production and not reviewing anything.

I think that’s mostly marketing.
The more code added to a codebase, the larger the surface area.

Software isn't written once. It's designed to evolve.
What works now, should be easily changed or improved in the future.

The question you should ask: 
“Can we ship changes that are consistent and safe to release?”

A few safeguards that actually worked for me:
1. Treat AI code as proposal, not truth. Don’t accept diffs you can’t explain.
2. Put a harness on the agent. (Cursor Rules / AgentsMD / ClaudeMD / Skills / Plugins)
3. Keep Agent Rules, Skills and Plugins small, scoped and enforceable.
4. Add guardrails only after noticing Coding Agent failures.
5. Keep PRs reviewable. Minimal context + a 1-pager attached on the PR for big moves.
6. Reviews should't focus on preference. Skip nitpicks. Focus on the entire workflow.
7. Pre-commit + CI parity. Catch drift before you waste a CI cycle.
8. Write a handoff note to your tomorrow-self so you catch-up faster on the next day.

For an Engineer, AI helps, but only when you stay in control of constraints.

---
I explain complex topics in AI, AI Engineering and building AI Systems.

For free.

Join here: https://buff.ly/DnpHEXD

Six Tips on AI Assisted Programming

I make learning Data Science effortless for you | Data Science @ Bosch

Hi 👋🏻
I'm Banias, AI Engineer at Netze BW.

Applying new technologies and methods in Data Science & AI is my passion.

Through my experience and continual learning, I've gained valuable insights and tips that I love sharing with Data Scientists, Machine Learning Engineers and builders like you.

If you're looking for practical advice, helpful resources, and articles on the latest trends or best practices in the field, you're in the right place.

Let's connect and stay updated on the exciting world of Data Science together. Feel free to reach out, read my articles, and let's grow our knowledge and network. :)


Note: My views are my own and don‘t represent my employer.

Banias Baabe

Most people running local AI models have no idea which ones will actually work on their hardware.

𝗹𝗹𝗺𝗳𝗶𝘁 is a small open source CLI tool that scans your system and tells you exactly which models fit your setup.

No more downloading files that turn out to be too large to run.

Here is what it does:
✅ Evaluates 150+ models from 30 providers against your CPU, GPU, and RAM specs.
✅ Returns a ranked list of models that match your hardware.
✅ Connects with Ollama and LM Studio to detect and download models directly.
✅ Supports multi GPU setups and Mixture of Experts architectures.

It's a simple quality of life tool that saves real time for developers working with local AI.

🔗 Link to repo: github(.)com/AlexsJones/llmfit


---
♻️ Found this useful? Share it with another builder.
➕ For daily practical AI and Python posts, follow Banias Baabe.

I teach AI Engineers the concepts and tools they need to stay ahead | AI Engineer @ Netze BW

taplio

The Battle

Alex Razvant

Banias Baabe

Followers

Posts

Engagements

Likes

Performance ScoreAlex Razvant's LinkedIn Performance

Performance ScoreBanias Baabe's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Alex Razvant

Banias Baabe

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Alex Razvant's LinkedIn Performance

Performance ScoreBanias Baabe's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreAlex Razvant's LinkedIn Performance