mitzen

 

  The emerging 2026 picture is: Optimizer : Muon (or Muon + Polar Express) is replacing Adam as the go-to for training efficiency — ~2x cheaper for same quality Fine-tuning : QLoRA is being pushed further with papers like LowRA, squeezing more out of your 30–48GB machine Reasoning : GRPO variants (S-GRPO, Training-Free GRPO) are making DeepSeek-R1-style reasoning training accessible on consumer hardware Unsloth (already implements many of these optimizations) + QLoRA + Qwen2.5-Coder is still your fastest path, but keep an eye on Muon support landing in Unsloth/TRL, which would meaningfully speed up your training run The most relevant 2026 papers for cheap training 1. "The Polar Express" — ICLR 2026 Honorable Mention (Amsel, Persson, Musco, Gower) This is the most formally recognized paper of 2026 directly about training efficiency. It introduces a new method for computing the polar decomposition used in the Muon optimizer for training deep neural networks, using only m...

 AzApi helps solving the issue for you when you have to make a manual REST API call to azure. It is better than null script apprach whereby your:- State fully managed No external script dependency or PowerShell requirement Handles the API version explicitly Native depends_on support Idempotent by design Besides it is often recommended to use " a zapi provider" > ARM template deployment > null resource " This is an actual scenario when configuring Azure Foundry capability host - we have to initiate REST API call as there's no other sdk available.  To be able to do that we will use azapi providers and the sample code are shown here:- terraform {   required_providers {     azapi = {       source   = " azure/azapi "       version = " ~> 2.0 "     }     azurerm = {       source   = " hashicorp/azurerm "       version = " ~> 3.0 "     }   }...

 Here a some Bicep sample for setting up Azure Foundry https://github.com/microsoft-foundry/foundry-samples

Ran into this error while trying to setup my LLM agent to do tool calling. In this case, it is " google_search ".  in raise_error_async raise ClientError(status_code, response_json, response) google.genai.errors.ClientError: 400 Bad Request. {'message': '{\n  "error": {\n    "code": 400,\n    "message": "Built-in tools ({google_search}) and Function Calling cannot be combined in the same request. Please choose one to continue.",\n    "status": "INVALID_ARGUMENT"\n  }\n}\n', 'status': 'Bad Request'} The gist of the problem - ADK framework doesn't like me putting Agent model together with tool.  To resolve this issue, I have change my agent from Agent ( from google . adk import Agent )  to LlmAgent ( from google . adk . agents import LlmAgent ) 💥Error code      researcher = Agent (     name = " researcher " ,     model = model_name ,     description = " Condu...

Graphql fragment are used to simplified and provide shortcuts definitions to specified the same fields in a graphql query.  In this example, we provide defined ReportFields to consist of propertyType, id and status. Then we hook it up in the actual gteReportsByUserId query as shown here:-  query GetReportsByUserId ( $ includeStatus : Boolean ! ) {     getReportsByUserId ( userId : " 123 " , limit : 10 ) {     status     reports {       ... ReportFields     }   } } fragment ReportFields on RetrievedReport {   propertyType   id   status @ include ( if : $ includeStatus ) } Save us some typings in our query.

This is an example of graphql @include directive. The purpose of the @include directive is to allow a field to be retrieve or not based on a condition. In this case, it is determined by $inculdeStatus. Given that we are passing in $includeStatus = false, the field status will be excluded. query GetReportsByUserId ( $ includeStatus : Boolean ! ) {   getReportsByUserId ( userId : " 123 " , limit : 10 ) {     status     reports {       propertyType       id       status @ include ( if : $ includeStatus )     }   } } -------------------------------------------------- variables {   " userId " : " 123 " ,   " limit " : 10 ,   " includeStatus " : false } We can also use the @skip directive to skip field from being returned  query GetReportsByUserId ( $ includeStatus : Boolean ! ) {   getReportsByUserId ( userId : " 123 " , limit : 10 ) {     status   ...

We can make use of google adk to call ollama local llm agent or other inference runtime of your choice. This integration requires LiteLLM connector. Key libraries to install first and this basically will install LiteLLM  uv add google-adk[extensions] Here is an example of the code that we can use to integrate a local LLM spinned up using the code here:- from google . adk import Agent from google . adk . agents import LlmAgent from google . adk . models . lite_llm import LiteLlm # https://adk.dev/agents/models/litellm/ # https://adk.dev/agents/models/ollama/ root_agent = Agent (     model = LiteLlm ( model = " ollama_chat/gemma3:latest " ),     name = " dice_agent " ,     description =(         " hello world agent that can roll a dice of 8 sides and check prime "         " numbers. "     ),     instruction = """       You roll dice and answer questions about the outco...

When rebasing commit history via git rebase -i we can use the following command. Example use-case # In the todo list: edit abc1234 Fix typo Here are the most useful commands for removing or modifying commits : When you run git rebase -i , Git opens a todo list where each commit is prefixed with a command. These are the commands you can use (aliases in parentheses): Command Alias What it does pick p Apply the commit as-is (default) reword r Apply commit, but pause to edit only the commit message edit e Apply commit, then pause so you can change files and run git commit --amend squash s Combine this commit into the one above it & merge both commit messages fixup f Combine this commit into the one above it & discard this commit's message drop d Delete the commit entirely break b Pause the rebase here (useful for manual testing before continuing) exec x Run a shell command at this point in the history label / reset / merge l / t / m Advanced commands for complex rebases ...