Mixture of Experts and Switch Transformer

Mixture of Experts (MoE) is an innovative horizontal scaling technique employed to the basic Transformer architecture. The Feed Forward (FFN) Layer of the Transformer is replaced with a MoE layer which is a collection of N-Experts (each one a seperate FFN) in parallel. The MoE also includes a Router layer with a gating logic (learnt) to decide the expert(s) to route the token to.

One of the early MoE based Transformers was the Switch Transformer (https://arxiv.org/abs/2101.03961) with a MoE routing layer. The Switch Transformer specifically includes logic to enable balancing of token loads across the different Experts in order to prevent hot-spots where only a few experts end up handling a majority of the tokens. This also leads to a second issue with the other experts remain untrained through training therby rendering them useless for inference.

There are several sota MoE implementations available on the different ML platforms. The Keras-io examples has one Switch Transformer. The code text_classification_switch_transformer_pytorch.py is a PyTorch port of the same code with couple of changes done to make the code modular and resolve issues with super.init call and position_in_expert.

Further, a much simpler SwitchRouter combined implementation is done in SwitchTransformerUtil.SimpleSwitchRoute(). The code flow is:

• Compute gateLogits, with option to add Noise to load balance during training
• Compute weights & selectedExperts indexes of the topK experts 
• Compute auxLoss to be minimized for balancing load across experts
• Finally, for every expert, fetch weights, invoke expert to get the outputs
• Also drop tokens beyond expert capacity threshold

Fairly straightforward!

References

• https://newsletter.theaiedge.io/p/mixture-of-experts-early-sparse-moe?utm_source=publication-search
• https://medium.com/@pilliudayaditya1207/understanding-mixture-of-experts-switch-transformers-load-balancing-vs-mixtral-s-natural-balance-25ed528cadfe
• https://huggingface.co/blog/NormalUhr/moe-balance
• https://newsletter.maartengrootendorst.com/p/a-visual-guide-to-mixture-of-experts

Fine Tuning Text Classification Model

Fine tuning is a technique employed to a base fully trained model (foundation) and to retraining/ repurposing it to meet some different objective(s). The key aspect of fine tuning being that it is not a complete/ full retraining of the base model. It's done on a much smaller training budget keeping the weights of the original model intact, and bringing in a much smaller additional set of trainable weights known as adapters. 

These adapter weights are typically of Low Rank matrices thus the name LoRA (Low Rank Adapters). With every round of training only these LoRA weights get updated which the weights from the base model stay frozen. Since final weights are additive so the corresponding fine trained LoRA model equation:

    output = f(W_base*x + b_base + B*A*x), where for any given input x

    W_base, b_base: Base model weights & bias which remain fixed
    B, A: Low Rank Adapter weights of a small rank (r), which are trained during fine tuning 
     f: Activation Function 

In the example TextClassificationFineTuningLora.py the working of the LoRA adapter for fine tuning a Text Classification model is demonstrated.

Fine Tuning Objective

• The base model is a Sentiment Classifier on the Imdb database which labels data as either 1 (Positive) or 0 (Negative).
• The base model is repurposed as a Exaggeration detector that labels the data as either 1 (Has-Exaggerations) or 0 (No Exaggerations).
• Exaggerations are any reviews containing the exaggerations defined in: TextClassificationTorchUtilities.exaggerations. Also see: TextClassificationTorchUtilities.buildExaggerationDataset().

Fine Tuning Details

• The base model had 2.67 Mn total parameters of which 8.86 Lakh paramters were trainable. For fine tuning these 8.86 Lakh parameters are all frozen.
• The LoRA model is employed to every trainable layer of the base model. Each trainable layer of the base model is set to enable_lora(rank=4). This results in total trainable parameters of just ~30.6K.
• After fine tuning the model is able to identify Exaggerations with an accuracy in the high 90's.

Knowledge Distillation

Knowledge distillation from a trained large Teacher model to a smaller Student model is a very popular technique in the ML. Distillation helps to train a Student model which despite being much smaller and compressed shows performance comparable to the larger Teacher model.

The other advantage of Distillation is that the Student model requires a much smaller set of labelled training data (<10%) since it's essentially trying to match the output of the Teacher during training. The Distillation loss is a function of the difference between the prediction of the Student (y_pred) & the Teacher models (teacher_pred) for every training input (x). Kullback-Leibler divergence (KLDivergence) loss between student_pred (y_pred) & teacher_pred is a common pick for the Distillation loss.

For a working example of Distillation refer to TextClassificationDistillation.py which is distilled from a Keras Text Classification model in Torch. The original Text Classification Teacher model had several Convolution layers which have been replaced by a Dense layer. Also the Input Embedding layer's ouput dimension has been reduced from 128 to 32. 

The original Text Classification model (Teacher) had ~2.67 Mn parameters (8.9 Lakh trainable) and was trained with 25K data samples. The distilled Student model has only ~1.6 Lakh parameters (~18%) and was trained using 2.5K samples (~10%). In terms of the size of the saved models the Teacher model is 10.2MB vs 0.6 MB of the student. There was only a marginal 4% drop in accuracy seen with the Student model on the held-out test data.

Fig 1: Text Classification - Teacher Model	Fig 2: Keras Text Classification - Student Model

Text Classification from Scratch using PyTorch

The AI/ ML development framework Keras 3x supports in recent times has got support for Torch & Jax backends, in addition to Tensorflow. However, given Keras's Tensorflow legacy large sections of the code are deeply integerated with Tensorflow. 

One such piece of code is text_classification_from_scratch.py from the keras-io/ examples project. Without tensorflow this piece of code simply won't run!

Here's text_classification_torch.py a pure Torch/ PyTorch port of the same code. The bits that needed modification:

• Removing all tensorflow related imports
• Loading the Imdb text files in "grain" format in place of "tf" format, by passing the appropriate param: 

    keras.utils->text_dataset_from_directory(format="grain") 

Also grain needs to be installed:

    pip3 install grain 

• For building Vocab, Tokenizer, Vectorizing use torchtext:

    pip3 install torchtext

• Few other changes such as ensure max_features constraint's honoured, text is standardized, padded, and so on   

Guardrails & Guard-Llm's

With wide scale adoption of Llm's & Agentic models in production, there's also a pressing need to verify both the inputs & output for GenAI use cases. This should ideally be done in real-time just before serving the response to the end user. This would ensure that no invalid, harmful, hateful, confidential, etc content goes through in either direction. Guardrails are the answer to that very problem.

The simple idea with Guardrails is to apply intelligent input/ output filters that can sanitize and filter out both bad requests/ responses from getting through. There are many ways of implementing Guardrails as pattern based, rule engines, etc. Though these have worked so far, in an ever changing Agentic world it's now up to the self learning guard Llm's to judge & flag! 

Guard llm's are specifically trained to flag out harmful content. One such implementation is llama-guard which flags out violations of any of the ML Commons AI Safety Taxonomies/ Categories.

An implementation of the guard-llm can be found in the ApiCaller project. More specifically the ApiCaller->invokeWithGuardrails():

•  First calls a local Ollama model with sanitized input to get a response
•  Then calls the isSafe() method with the received response
•  isSafe() internally makes a call to a different Ollama model llama-guard which flags out the content as safe/ unsafe

Check the TestApiCaller.py test case for better clarity.

References

• https://mlcommons.org/2024/04/mlc-aisafety-v0-5-poc/
• https://www.ibm.com/think/tutorials/llm-guardrails
• https://ollama.com/library/llama-guard3

Agent2Agent (A2A) with a2a-sdk and Http2

Continuing with A2A evaluation next up is a2a-sdk (unrelated to previously evaluated a2a-server). This evaluation is largely based on getting the hello world from the a2a-samples project working as per the instruction of a2a-protocol. With additional, integration with other Http2 based non Python clients.

(I) Installation

pip install a2a-sdk 

# uvicorn python-dotenv (packages existing) 

# For Http2 support 

pip install hypercorn 

pip install h2==4.2.0 (See Issue 1 at the end & the bug details) 

git clone https://github.com/a2aproject/a2a-samples.git -b main --depth 1

(II) Replace uvicorn server with hypercorn (support for Http2) 

The a2a-samples make use of the uvicorn python server. However, uvicorn is a Http1.x compliant server and doesn't support Http2. Keep seeing the following messages if client requests from Http2: 

"WARNING:  Unsupported upgrade request. "

In order to support a wider & more updated category of clients, uvicorn is replaced with a hypercorn which is Http2 compliant.

In order to switch to hypercorn, the following changes are done to _main_.py of helloworld python project: 

#import uvicorn
 

# Use Hypercorn for Http2
import asyncio
from hypercorn.config import Config
from hypercorn.asyncio import serve

 ....

    config = Config()
    config.bind="127.0.0.1:8080"  # Binds to all interfaces on port 8080

    asyncio.run(serve(server.build(), config))
   # uvicorn.run(server.build(), host='127.0.0.1', port=8080, log-level='debug') 

(III) Run helloworld

python a2a-samples/samples/python/agents/helloworld/__main__.py 

(IV) View AgentCard

Open in the browser or via curl:

curl http:///127.0.0.1:8080/.well-known/agent-card.json

Response: 

{"capabilities":{"streaming":true},"defaultInputModes":["text"],"defaultOutputModes":["text"],"description":"Just a hello world agent","name":"Hello World Agent","preferredTransport":"JSONRPC","protocolVersion":"0.3.0","skills":[{"description":"just returns hello world","examples":["hi","hello world"],"id":"hello_world","name":"Returns hello world","tags":["hello world"]}],"supportsAuthenticatedExtendedCard":true,"url":"http://127.0.0.1:8080/","version":"1.0.0"} 

For the Authorized Extended Agent Card:

curl -H "Authorization: Bearer dummy-token-for-extended-card" --http2 http://127.0.0.1:8080/agent/authenticatedExtendedCard 

Response: 

{"capabilities":{"streaming":true},"defaultInputModes":["text"],"defaultOutputModes":["text"],"description":"The full-featured hello world agent for authenticated users.","name":"Hello World Agent - Extended Edition","preferredTransport":"JSONRPC","protocolVersion":"0.3.0","skills":[{"description":"just returns hello world","examples":["hi","hello world"],"id":"hello_world","name":"Returns hello world","tags":["hello world"]},{"description":"A more enthusiastic greeting, only for authenticated users.","examples":["super hi","give me a super hello"],"id":"super_hello_world","name":"Returns a SUPER Hello World","tags":["hello world","super","extended"]}],"supportsAuthenticatedExtendedCard":true,"url":"http://127.0.0.1:8080/","version":"1.0.1"} 

(V) Send/ Receive message to Agent

curl -H "Content-Type: application/json"  http:///127.0.0.1:8080 -d '{"jsonrpc":"2.0","id":"ee22f765-0253-40a0-a29f-c786b090889d","method":"message/send","params":{"message":{"role":"user","parts":[{"text":"hello there!","kind":"text"}],"messageId":"ccaf4715-712e-40c6-82bc-634a7a7136f2","kind":"message"},"configuration":{"blocking":false}}}' 

Response: 

 {"id":"ee22f765-0253-40a0-a29f-c786b090889d","jsonrpc":"2.0","result":{"kind":"message","messageId":"d813fed8-58cd-4337-8295-6282930d4d4e","parts":[{"kind":"text","text":"Hello World"}],"role":"agent"}}

(VI) Send/ Receive via Http2

curl -iv --http2 http://127.0.0.1:8080/.well-known/agent-card.json

& 

curl -iv --http2  -H "Content-Type: application/json"  http://127.0.0.1:8080 -d '{"jsonrpc":"2.0","id":"ee22f765-0253-40a0-a29f-c786b090889d","method":"message/send","params":{"message":{"role":"user","parts":[{"text":"dragons and wizards","kind":"text"}],"messageId":"ccaf4715-712e-40c6-82bc-634a7a7136f2","kind":"message"},"configuration":{"blocking":false}}}'

(The responses are the same as shown above)

(VII) Send/ Receive from Java client

TBD

(VIII) Issues 

Issue 1: Compatibility issue with hypercorn (ver=0.17.3) & latest h2 (ver=4.3.0)

Ran in to the issue in the mentioned here:

    |   File "/home/algo/Tools/venv/langvang/lib/python3.13/site-packages/hypercorn/protocol/h2.py", line 138, in initiate
    |     event = h2.events.RequestReceived()
    | TypeError: RequestReceived.__init__() missing 1 required keyword-only argument: 'stream_id' 

Issue was resolved by downgrading to h2 (ver=4.2.0).

 

Codegen LLMs

One GenAI feature making headlines is coding. GenAI apps are getting better with reading and writing code (codegen) in various programming languages such as Python, Java, C++ and so on.

On the evaluation side there are all kinds of benchmarks and leaderboards that track progress on codegen. Additionally, aspects of usability, platform support, IDE integration, etc are all key factors for using codegen.

In terms of local evaluations, Ollama provides handy options. With Ollama it's easy to download and run LLMs from various providers (Llama, Gemma, etc). Most now support codegen and readily follow instructions in a chat to churn out basic level Python code:

•  Llama3.2
• Gemma3
• Codegemma 
• Falcon3
• Starcoder 

 

Langchain4j

LangChain is one of the leading python based AI/ ML, agentic modelling and integration frameworks. Langchain (and allied frameworks like LangGraph) allow integration with almost all LLMs, python libraries and tools out there. 

Langchain4j is its Java couterpart. Langchain4j allows LLM integrations and workflows to be built using pure Java constructs. It primarily operates as a Java client to the various Api's exposed by the different LLM provides such as OpenAi, Azure, Bedrock, Gemini and so on. 

Langchain4j has covered a lot of ground in terms of the supported modules from both the Python and the Java ecosystems. It's actively supported and should be one for the long run.. 

To get a feel for Langchain4j on a local LLM try out langchain4j-ollama. 

This will get: 

    Java langchain4j-ollama to talk to  

        -> Ollama (deployed locally) 

                -> Hosting the llama3.2:1b  model  

(I) Get a local Ollama up & running

Refer to the previous post regarding installing getting Ollama running locally. Once done, you should have a llama3.2:1b model running & ready to chat locally on:

    http://127.0.0.1:11434 

(II) Download & build langchain4j-ollama project

Clone langchain4j-ollama project & build:

    cd </download/folder/langchain4j-ollama> 

    mvn install 

(III) Run langchain4j-ollama tests

Run a couple of the langchain4j-ollama integration tests. Start with OllamaChatModelIT.java. Make sure to update the Model_Name value to llama3.2:1b downloaded in step (I) above:

         static final String MODEL_NAME = "llama3.2:1b";

That's about it for getting the three pieces integrated & chatting! 

Ollama for Local Llm

Continuing with the theme of running Llm's locally, it was time ideal to go with Ollama which has gained significant ground over the one odd year. While tools like Gpt4All, Llamafile, etc are all based on llama.cpp, Ollama has its own model format. 

More importantly Ollama supports older CPU's (non-AVX), something stopped by llama.cpp and the others. 

(I) Installation

Ollama installation is on a box with Ubuntu 25.10 with the latest libraries installed.

    sudo snap install ollama

(II) Pull Ollama model

        ollama pull llama3.2:1b

• To view the downloaded models using 'ollama list' command:

        ollama list  

            |__ 

NAME                          ID              SIZE      MODIFIED   
llama3.2:1b                   baf6a787fdff    1.3 GB    ... 

• Models by default get downloaded to:/var/snap/ollama/common/models, verified via:

        sudo snap get ollama 

        |__

Key                   Value
context-length        
cuda-visible-devices  
debug                 0
flash-attention       0
host                  127.0.0.1:11434
models                /var/snap/ollama/common/models
origins               *://localhost 

To change the default location the config models needs to be updated: 

        sudo snap set ollama models=/UPDATED/MODELS/LOCATION/PATH

 (III) Run/ chat with downloaded model:

    ollama run llama3.2:1b

        |__ 

 >>> bye
Bye.

>>> what is 80 + 10?
80 + 10 = 90.

(IV) Install/ Run any locally downloaded GGUF model

Ollama also provides the option to run any downloaded model GGUF locally. These are models not downloaded via ollama pull (ref step (II)) but models downloaded from Hugging face, etc.  

A simple modelfile needs to be prepared with one line instruction:

        FROM </GGUF/FILE/DOWNLOAD/LOCATION> 

Next the Ollama create command is to be run using the modelfile:

        ollama create <CUSTOM_MODEL_NAME> -f  </modefile/LOCATION>

 With that a your downloaded model (GGUF) file would be available for run from Ollama and show up in: 

        ollama list    

(Note: There's a known issue with the template of downloaded model.)

(V) Ollama API

Ollama server by default listens on the end point: http://127.0.0.1:11434.  

Through the endpoint various Ollama APIs are available for chatting, generating completions, list models, show details, running models, version, push, pull, etc with the installed models.

(VI) Remove Models

To remove any downloaded models run the 'ollama rm' command:

        ollama rm llama3.2:1b

(VII) Stop Ollama

• Stopping/ unloading of just the running model can be effected via an Ollama API call with keep_alive=0, along with an empty message: 

        curl http://127.0.0.1:11434/api/chat -d '{"model": "llama3.2:1b","messages": [],"keep_alive": 0}'

• On the other hand, stopping of Ollama service is unimplemented. So a hard kill is the only option that works (will also unload all running models):

        sudo pkill -9 ollama

     Snap will however restart Ollama snap the moment it is killed (recovery/ restart).

• To completely shutdown/ disable Ollama:

        sudo snap disable ollama 

 

Gpt4All on Ubuntu-20

Notes from a rather tough, yet futile, attempt at getting Gpt4All to run locally on an old Ubuntu20.04 box, with Python-3.8.

* Pip Install: First up, the gpt4All installed via pip (ver 2..8.2) has changes incompatible with current/ recent model files & gguf (llama3.2, etc). Causing type, value, keyword, attribute errors etc at different stages of installation & execution.

* Custom Build: Alt. is to download the latest Gpt4All & build it.

This leads to issues with Ubuntu 20.04 library's being outdated/ missing & the hardware being outdated :

• GLIBC_2.32 not found 
• GLIBCXX_3.4.29 not found
• CMake 3.23 or higher is required.  You are running version 3.16.3
• Vulkan not found, version incompatible with Gpu, etc
• CUDA Toolkit not found. 
• CPU does not support AVX 

Anyway, after a lot of false steps the build did succeed with the following flags set did succeed:

    cmake -B build -DCMAKE_BUILD_TYPE=Rel -DLLMODEL_CUDA=OFF -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON 

     Build files have been written to: .../gpt4all/gpt4all-backend/build 

 

Even after all that there were issues popping up with getting Llms to run from libraries like langchain, pygpt4all and so on. Clearly indicating that it was time to bid adieu to Ubuntu 20.04 & upgrade to more recent and better supported versions. 

References

• https://python.langchain.com/docs/how_to/local_llms/
• https://askubuntu.com/questions/1393285/how-to-install-glibcxx-3-4-29-on-ubuntu-20-04
• https://stackoverflow.com/questions/71940179/error-lib-x86-64-linux-gnu-libc-so-6-version-glibc-2-34-not-found 

Mlflow Java client

Mlflow is a leading open source framework for managing AI/ ML workflows. Mlflow allows tracking, monitoring and generally visualizing end-to-end ML project lifecycles. A handy ops side tool that improves over interpretability of AI/ ML projects.

Key Mlflow concepts include ML Projects, Models on which several Runs of Experiments conducted to name a few. Experiments can also be Tagged with meaningful humanly relevant labels.

While Mlflow is a Python native library with integrations with all the leading Python AI/ ML frameworks such as OpenAI, Langchain, Llamaindex, etc there are also Mlflow API endpoints for wider portability. 

There is also a specific Mlflow Java Api for use from the Java ecosystem. The corresponding Mlflow Java client (maven plugin, etc) works well with the API. To get started with the mlflow using Java:

(I) Install mlflow (Getting started guide)

        $ pip install mlflow 

 This installs mlflow to the users .local folder:

        ~/.local/bin/mlflow 

(II) Start Local mlflow server (simple without authentication)

        $ mlflow server --host 127.0.0.1 --port 8080

mlflow server should be running on 

        http://127.0.0.1:8080

(III) Download mlflower repo (sample Java client code)

Next clone the mlflower repo which has some sample code showing working of the mlflow Java client. 

• The class Mlfclient shows a simple use case of Creating an Experiment:

            client.createExperiment(experimentName);

Followed by a few runs of logging some Parameters, Metrics, Artifacts:

     run.logParam();

      run.logMetric();

       run.logArtifact()

 

• Run Hierarchy: Class NestedMlfClient shows nesting hierarchy of Mlflow runs
  

        Parent Run -> Child Run -> Grand Child Run ->.... & so on

(IV) Start Local mlflow server (with eBasic Authentication)

While authentication is crucial for managing workflows, Mlflow only provided Basic Auth till very recently. Version 3.5 onwards has better support for various auth provides, SSO, etc. For now only mlflow Basic Auth integration is shown.

           # Start server with Basic Auth
            mlflow server --host 127.0.0.1 --port 8080 --app-name basic-auth

Like previously, mlflow server should start running on

            http://127.0.0.1:8080

Only requiring a login credential this time to access the page. The default admin credentials are mentioned on mlflow basic-auth-http.

• The class BasicAuthMlfclient shows the Java client using BasicMlflowHostCreds to connect to Mlflow with basic auth. 

            new MlflowClient(new BasicMlflowHostCreds(TRACKING_URI, USERNAME, PASSWORD));

(V) Deletes Soft/ Hard

• Experiments, Runs, etc created within mlflow can be deleted from the ui (& client). The deletes are however only Soft, and get stored somewhere in a Recycle Bin, not visible on the UI.
•  Hard/ permanent deletes can be effected from the mlflow cli

    # Set mlflow server tracking uri 

    export MLFLOW_TRACKING_URI=http://127.0.0.1:8080

    # Clear garbage

    mlflow gc

  (VI) Issues

• MlflowContext.withActiveRun() absorbs exception without any logs, simply sets the run status to RunStatus.FAILED. 
• So incase runs show failure on the mlflow UI, its best to put explicit try-catch on the client to find the cause.

• Unable to upload artifacts since cli looks for python (& not python3) on path to run. 
• Error message: Failed to exec 'python -m mlflow.store.artifact.cli', needed to access artifacts within the non-Java-native artifact store at 'mlflow-artifacts:
• The dev box (Ubuntu ver 20.04) has python3 (& not python) installed.
• Without changing the dev box a simple fix is to set/ export the environment variable MLFLOW_PYTHON_EXECUTABLE (within the IDE, shell, etc) to whichever python lib is installed on the box:

               MLFLOW_PYTHON_EXECUTABLE=/usr/bin/python3 

 

So with that keep the AI/ Ml projects flowing!

AI/ML '25

• GenAI
    - Text: Chat, Q&A, Compose, Summarize, Think, Search, Insights, Research
     - Image: Gen, Identify, Search (Image-Image, Text-Image, etc), Label, Multimodal
    - Code gen
    - Research: Projects, Science, Breakthroughs
    - MoE

• Agentic
    - Workflows: GenAI, DNN, Scripts, Tools, etc combined to fulfil Objectives
        -- Auto-Generated Plans & Objectives  
    - Standardization: MCP (API), Interoperability, Protocols
    - RAG
    - Tools: Websearch, DB, Invoke API/ Tools/ LLM, etc

• Context
    - Fresh/ Updated
    - Length: Cost vs Speed trade-off
    - RAG
    - VectorDB (Similarity/ Relevance)
    - Memory enhanced

• Fine Tune
    - Foundation models (generalists) -> Specialists
    - LoRA
    - Inference time scaling (compute, tuning, etc)
    - Prompts

• Multimodal: Text, Audio, Video, Image, Graph, Sensors

• Safety/ Security
    - Output Quality: Relevance, Accuracy, Correctness, Evaluation (Automated Rating, Ranking, JudgeLLM, etc)
        -- Hallucination
    - Privacy, Data Leak, Backdoor, Jailbreak
    - Guard Rails

AI Agentic Frameworks

With prolification of AI Agents, it's only logical that there will be attempts at standardization and building protocols & frameworks:

• MCP covered previously
• Any-Agent from Mozilla.ai to switch between agents, vendors, clouds, etc
• Agent2Agent interoperability protocol

On Quantization

• Speed vs Accuracy trade off.
• Reduce costs on storage, compute, operations .
• Speed up output generation, inference, etc.
• Work with lower precision data.
• Cast/ map data from Int32, Float32, etc 32-bit or higher precision to lower precision data types such as 16-bit Brain Float (BFloat16) or 4-bit (NFloat)/ int4 or int8, etc.
• East mapping Float32 (1-bit Sign, 7-bit Exponent, 23-bit Mantissa) => BFloat16 (1-bit Sign, 7-bit Exponent, 7-bit Mantissa). Just discard the higher 16-bits of mantissa. No overflow!
• Straightforward mapping work out max, min, data distribution, mean, variance, etc & then sub-divide into equally sized buckets based on bit size of the lower precision data type. E.g int4 (4-bit) => 2^4 = 16 buckets. 
  
• Handle outliers, data skew which can mess up the mapping, yet lead to loss of useful info if discarded randomly.
• Work out Bounds wrt Loss of Accuracy.

LLMs, AI/ ML side:

• https://newsletter.theaiedge.io/p/reduce-ai-model-operational-costs

Lucene, Search side:

• https://www.elastic.co/search-labs/blog/scalar-quantization-101
• https://www.elastic.co/search-labs/blog/scalar-quantization-in-lucene

Speculative Decoding

• Ensemble of Weak + Strong model
  
• Weak model has a quick first go at generating tokens/ inference (potentials)
• Followed by the Strong, but slow model which catches up & uses the outputs of the weak model, samples them, grades them, accepting/ rejecting them to generate the final output
• Overall making inferences via LLMs quicker and cheaper

More to follow..

• https://pytorch.org/blog/hitchhikers-guide-speculative-decoding/ 
• https://www.baseten.co/blog/a-quick-introduction-to-speculative-decoding/
• https://research.google/blog/looking-back-at-speculative-decoding/
• https://medium.com/ai-science/speculative-decoding-make-llm-inference-faster-c004501af120

Model Context Protocol (MCP)

Standardization Protocol for AI agents. Enables them to act, inter-connect, process, parse, invoke functions. In other words to Crawl, Browse, Search, click, etc. 

MCP re-uses well known client-server architecture using JSON-RPC. 

Apps use MCP Clients -> MCP Servers (abstracts the service)

Kind of API++ for an AI world!

The Big Book of LLM

A book by Damien Benveniste of AIEdge. Though a work in progress, chapters 2 - 4 available for preview are fantastic. 

Look forward to a paperback edition, which I certainly hope to own...

Mozilla.ai

Mozilla pedigree, AI focus, Open-source, Dev oriented.

Blueprint Hub: Mozilla.ai's Hub of open-source templtaized customizable AI solutions for developers.

Lumigator: Platform for model evaluation and selection. Consists a Python FastAPI backend for AI lifecycle management & capturing workflow data useful for evaluation.

Scaling Laws

Quick notes around Chinchilla Scaling Law/ Limits & beyond for DeepLearning and LLMs.

Factors

• Model size (N)
• Dataset size (D)
• Training Cost (aka Compute) (C)
• Test Cross-entropy loss (L)

The intuitive way,

• Larger data will need a larger model, and have higher training cost. In other words, N, D, C all increase together, not necessarily linearly, could be exponential, log-linear, etc.
• Likewise Loss is likely to increase for larger datasets. So an inverse relationship between L & D (& the rest).
• Tying them into equations would be some constants (scaling, exponential, alpha, beta, etc), unknown for now (identified later).
  

Beyond common sense, the theoretical foundations linking the factors aren't available right now. Perhaps the nature of the problem is it's hard (NP).

The next best thing then, is to somehow work out the relationships/ bounds empirically. To work with existing Deep Learning models, LLMs, etc using large data sets spanning TB/ PB of data, Trillions of parameters, etc using large compute budget cumulatively spanning years.

Papers by Hestness & Narang, Kaplan, Chinchilla are all attempts along the empirical route. So are more recent papers like Mosaic, DeepSeek, MoE, Llam3, Microsoft among many others. 

Key take away being,

• The scale & bounds are getting larger over time. 
• Models from a couple of years back, are found to be grossly under-trained in terms of volumes of training data used. They should have been trained on an order of magnitude larger training data for an optimal training, without risk of overfitting.
• Conversely, the previously used data volumes are suited to much smaller models (SLMs), with inference capabilities similar to those older LLMs.

References

• https://en.wikipedia.org/wiki/Neural_scaling_law
• https://lifearchitect.ai/chinchilla/
• https://medium.com/@raniahossam/chinchilla-scaling-laws-for-large-language-models-llms-40c434e4e1c1
• https://bigscience.huggingface.co/blog/what-language-model-to-train-if-you-have-two-million-gpu-hours
• https://medium.com/nlplanet/two-minutes-nlp-scaling-laws-for-neural-language-models-add6061aece7
• https://lifearchitect.ai/the-sky-is-bigger/

Diffusion Models

Diffusion

•     Forward, Backward (Learning), Sampling (Random)    
•     Continous Diffusion
•     VAE, Denoising Autoencoder
•     Markov Chains
•     U-Net
•     DALL-E (OpenAI), Stable Diffusion,
•     Imagen, Muse, VEO (Google)
•     LLaDa, Mercury Coder (Inception)

Non-equilibrium Thermodynamics

•     Langevin dynamics
•     Thermodynamic Equilibrium - Boltzmann Distribution
•     Wiener Process - Multidimensional Brownian Motion
•     Energy Based Models

Gaussian Noise

•     Denoising
•     Noise/ Variance Schedule
•     Derivation by Reparameterization

Variational Inference    

•     Denoising Diffusion Probabilistic Model (DDPM)
•     Noise Prediction Networks    
•     Denoising Diffusion Implicit Model (DDIM)

Loss Functions

•     Variational Lower Bound (VLB)
•     Evidence Lower Bound (ELBO)
•     Kullback-Leibler divergence (KL divergence)
•     Mean Squared Error (MSE)

Score Based Generative Model

•     Annealing
•     Noise conditional score network (NCSN)
•     Equivalence: DDPM and Score BBased Generative Models

Conditional (Guided) Generation

•     Classifier Guidance    
•     Classifier Free Guidance (CFG)

Latent Varible Generative Model

•     Latent Diffusion Model (LDM)
•     Lower Dimension (Latent) Space

References:

• https://en.wikipedia.org/wiki/Diffusion_model
• https://www.assemblyai.com/blog/diffusion-models-for-machine-learning-introduction
• https://www.ibm.com/think/topics/diffusion-models
• https://hackernoon.com/what-is-a-diffusion-llm-and-why-does-it-matter
• Large Language Diffusion Models (LLaDA): https://arxiv.org/abs/2502.09992