Ivan Nardini

AI/ML Advocate @googlecloud | Vertex AI dude | Research, Open Models, Ray & TPU | Instructor @DeepLearningAI | Startup Advisor @ycombinator x Google Cloud

[Ivan Nardini @ivnardini]

Presenting at the Next '26 Developer Keynote is one of those moments I'll remember for a long time.

Thank you to everyone who played a part.

Till the next Next!

[Ivan Nardini @ivnardini]

I will personally realize two dreams here. Going to Japan and present with さん Kaz

[Kazunori Sato @kazunori_279]

6/10開催のAnthropic主催イベントCode w/ Claudeでは、 @ivnardini と私で"Building with Claude on GoogleCloud"を担当します。すでにオンサイトは満席ですが、オンライン視聴可能なのでぜひ。
https://t.co/bIWi3Xmvee

[Ivan Nardini @ivnardini]

With v2.1.158, Anthropic shipped Auto mode in Claude Code with Google Cloud

You can now run commands in Claude Code using Claude models on Google Cloud without stopping for permission prompts every time

https://t.co/iDHnBLoqko

[Ivan Nardini @ivnardini]

Next Friday we are running a hands-on Claude Code on Google Cloud workshop together with the @AnthropicAI team in SF

Half day, Guided labs, and Live Q&A

Link
https://t.co/Lgi7B11HuB

[Ivan Nardini @ivnardini]

Github
https://t.co/xhAoEuFYpN
Docs
https://t.co/ULy5EpE0fT

[Ivan Nardini @ivnardini]

I looked into Keras Kinetic recently

Keras Kinetic is a framework that lets you run Keras and JAX workloads on Cloud TPUs by writing a training function and adding a decorator

Personally, it is one of the easiest ways I’ve seen to run a first TPU job so far

Here is a great blog post on fine tuning Gemma to speak Gen-Z slang using Kinetic

Blog
https://t.co/lqZu5MCI6U

[Ivan Nardini @ivnardini]

@graceg0ng Ofc! Great job Grace

[Ivan Nardini @ivnardini]

A good read about Cloud TPU generations

https://t.co/w16T3cO18J

[Ivan Nardini @ivnardini]

What a great series about getting started with TPUs

https://t.co/6Cd9mdhhKJ

[Ivan Nardini @ivnardini]

I spent some time testing elastic training capabilities on MaxText recently.

MaxText is Google’s open-source JAX library for the full LLM lifecycle scaling from one host to hundreds of TPU chips.

Pre-train with train method, run SFT/DPO/GRPO in the same package, and serve via vLLM.

It supports several models including Gemma, DeepSeek, Qwen, Kimi and more.

Docs
https://t.co/ppOa6xUMu9

Tutorial coming soon.

[Ivan Nardini @ivnardini]

Wrapping up the demo for Code with Claude in SF.

If you’re around, I'm happy to talk. See you tomorrow!

[Ivan Nardini @ivnardini]

Ray Serve now supports multi-host TPU slice deployments with gang scheduling.

Before, TPU slices required manual host counts and bundle replication, with no guarantee of a single co-located slice.

Now, Ray Serve uses Ray Core’s SlicePlacementGroup to pin deployments to one co-located TPU slice, matching Ray Train.

Code
https://t.co/YUSuQ27ZGe

[Ivan Nardini @ivnardini]

Anthropic released the public beta of Cowork on Third-Party Providers (3P)

Claude Desktop with Cowork and Code can now run using your own Google Cloud endpoint, billed as token consumption to your GCP project.

Docs
https://t.co/D5upoOpDzb

[Ivan Nardini @ivnardini]

Release notes
https://t.co/924W5kY7n5
Recipe
https://t.co/TH4eiuj6a4

[Ivan Nardini @ivnardini]

vLLM v0.19.1 shipped a bunch of optimizations and fixes for Gemma 4

> Gemma 4 MoE quantization support
> Eagle3 speculative decoding for faster inference
> Streaming and tool-call bug fixes for production applications

[Ivan Nardini @ivnardini]

Vertex AI Agent Engine Memory Bank just landed two features I’ve been looking for.

You can now push events yourself and decide when memories get generated.

Before, agent memory was passive. You knew conversations were flowing in, but you didn’t know when extraction happened.

Now you have
> ingest events method lets you push raw turns in per user (and force_flush if you want it now)
> generation trigger config sets idle-duration, fixed-interval, and event-count rules

Code
https://t.co/I5vRgZR8Y7

[Ivan Nardini @ivnardini]

Release notes
https://t.co/LXaVdD5132

[Ivan Nardini @ivnardini]

Claude Code adds 1-hour prompt cache support for Vertex AI.

Following interactions are now cheaper for long-running agentic coding sessions.

Under the hood, it is the ttl field on cache_control field:
{"cache_control": {"type": "ephemeral", "ttl": "1h"}}

Documentation
https://t.co/v89XgWhaLZ