TechPrufer
Get started
Google
Google logoGoogle · Feature trail

TensorFlow Machine Learning Framework

21 posts · Jun 2016 – Aug 2026

What this trail covers

The TensorFlow ecosystem continues to evolve with advancements in model building and training, specifically focusing on efficient checkpointing strategies for large-scale ML workloads. Continuous checkpointing in Orbax and MaxText optimizes reliability by intelligently initiating asynchronous saves upon completion of the preceding operation, minimizing performance degradation and resource wastage from hardware failures or preemptions. This approach is particularly beneficial for large-scale trai. This post introduces essential debugging and profiling tools for JAX on Cloud TPUs, including libtpu, JAX/jaxlib, verbose logging flags, the TPU Monitoring Library for programmatic insights into hardware performance, and the tpu-info CLI for real-time utilization metrics. It details how to access logs and configure debugging flags across TPU VMs, and provides examples of using the monitoring library and tpu-info.

All Google trails
  1. Decoding cosmic signals with deep learning and Keras- Google Developers BlogThis post details the application of deep learning, specifically using Keras, to analyze complex data from astroparticle physics experiments like the Pierre Auger Observatory, Cherenkov Telescope Array, and IceCube Observatory. It highlights how deep neural networks can replace traditional event reconstruction methods by learning complex, non-linear correlations from simulated detector signals (waveforms) to reconstruct primary particle properties (energy, mass) with unprecedented precision, effectively utilizing the full spatio-temporal structure of the data.
  2. Enterprise-Grade Precision for Long-Context Multimodal Embedding Inference on Cloud TPU- Google Developers BlogThis post details the integration of native TPU support into vLLM for optimized LLM serving, addressing production bottlenecks in scaling embedding pipelines. It introduces engineering optimizations for long-context multimodal embedding inference on TPUs, including hardware-safe tensor alignment, hardened materialization with lazy-loading and compilation pre-warming, and a hybrid StepPool architecture. The post also presents a minimal Python example for initializing Qwen3-Embedding-8B on TPU and provides rigorous mathematical parity evaluations comparing TPU outputs against other XPU golden references, demonstrating high cosine similarity scores.
  3. HeyGen x Google Cloud: Bringing Avatar IV to TPUs- Google Developers BlogThis post details the optimization of the Avatar IV AI video generation model for Google Cloud's Trillium (v6e) TPUs. Key contributions include: porting the PyTorch model to JAX via torchax, developing custom Pallas kernels for attention variants, implementing FSDP weight sharding and Ulysses sequence parallelism across an eight-chip mesh, and optimizing the pipeline by hiding collectives through head grouping, deleting sparse attention masks by aligning blocks to frame boundaries, and unchaining the softmax inner loop using precomputed bounds derived from the Cauchy-Schwarz inequality. These optimizations resulted in a 1.86x speedup and comparable performance to an 8xH100 setup.
  4. Run Ray on TPU, Part 2: Ray AI libraries- Google Developers BlogThis post details the integration of Ray libraries (Ray Serve, Ray Data, and JaxTrainer) with Google Cloud TPUs. It explains how to configure Ray Serve for TPU inference using vLLM and the `topology` field for large, sharded models. It introduces `iter_jax_batches()` in Ray Data for efficient JAX array feeding to TPUs and describes JaxTrainer for distributed JAX/Flax training on TPUs, emphasizing the `topology` configuration and in-worker JAX imports. It also highlights official TPU Docker images and enhanced Ray Dashboard metrics for TPUs.
  5. Run Ray on TPU, Part 1: The foundations- Google Developers BlogThis post introduces the integration of Ray with Google Cloud TPUs, making TPUs a first-class accelerator in Ray. It details how Ray's task-and-actor model, along with GKE orchestration, handles the complexities of TPU slices and topologies. Key components discussed include the Ray Operator add-on for GKE, the Ray TPU webhook for labeling TPU hosts, and the `ray.util.tpu.slice_placement_group()` function for reserving entire TPU slices. The post emphasizes that existing Ray libraries and development processes can be used with TPUs, with future parts of the series covering Ray AI libraries like Ray Data, JaxTrainer, and Ray Serve on TPUs.
  7. Systems Engineering Playbook: Optimizing Qwen 3.5-397B MoE on Ironwood (TPU7x)- Google Developers BlogThis post details the systems engineering playbook for optimizing the Qwen 3.5-397B MoE model on Ironwood (TPU v7x). It introduces a modular, model-agnostic optimization strategy using reusable JAX/Pallas kernels and hardware-aware cost models to achieve scalability. Key contributions include: 1. Detailed analysis of Qwen 3.5's hybrid architecture (Gated DeltaNet, Grouped Query Attention, Sparse MoE) and its mathematical formulations. 2. Establishment of a multi-dimensional evaluation matrix for prefill-heavy and decode-heavy workloads under varying concurrency tiers. 3. Co-design of a hybrid sharding scheme (8-way Attention Batch Sharding + 8-way Expert Parallelism) to overcome hardware constraints imposed by GQA and MoE layers, replacing traditional Tensor Parallelism. 4. Deep dive into distributed collective sequences for cross-device token routing. 5. Reporting significant performance uplifts (3.1x for decode-heavy, 4.7x for prefill-heavy workloads) and integration with open-source serving frameworks like vLLM and SGLang.
  8. Unlocking the Power of the TPU Stack: Introducing our new Developer Hub- Google Developers BlogIntroduces the TPU Developer Hub, a new educational resource for model builders, optimizers, and developers using Google Cloud TPUs. The hub provides guidance on TPU hardware architecture, software stack capabilities (including PyTorch migration and XLA), tracing/debugging/observability with XProf, parallelism/optimization strategies (Pallas kernels, KV cache offloading), and networking/security for distributed training and inference. It emphasizes code-first learning with open-source recipes and documentation, designed to be agent-ingestion friendly.
  9. Supercharging LLM inference on Google TPUs: Achieving 3X speedups with diffusion-style speculative decoding- Google Developers BlogThis post details the successful implementation of DFlash, a diffusion-style speculative decoding method, on Google TPUs within the vLLM TPU inference framework. The UCSD research team achieved an average 3.13x increase in tokens per second on TPU v5p, with peak speedups reaching nearly 6x for complex math tasks. Key technical contributions include a dual-cache solution for attention to bridge DFlash's non-causal block diffusion with paged attention, intelligent context management with power-of-2 padding for efficient host-TPU communication, and re-engineering the proposer to synchronize with the true accepted token count to bridge metadata gaps in TPU inference. Benchmarking demonstrated DFlash's superiority over autoregressive speculative decoding methods like EAGLE-3 on TPU v5p, achieving a 2.29x end-to-end serving speedup. The discovery of 'K-Flat' verification highlights that on high-end hardware like TPU v5p, the cost of verifying longer token blocks is nearly free, shifting the bottleneck to draft quality.
  10. TorchTPU: Running PyTorch Natively on TPUs at Google Scale- Google Developers BlogThis post introduces TorchTPU, a new stack enabling PyTorch to run natively on Google's TPUs. Key contributions include an "Eager First" philosophy with three distinct eager modes (Debug, Strict, Fused), integration with torch.compile using XLA and StableHLO as the backend, support for custom kernels via Pallas and JAX, and improved handling of Multi-Process Multi-Device (MPMD) execution for distributed training. It also highlights hardware-specific optimizations for TPUs.
  11. Boost Training Goodput: How Continuous Checkpointing Optimizes Reliability in Orbax and MaxText- Google Developers BlogThis post introduces continuous checkpointing as a feature within Orbax and MaxText, designed to improve the reliability and efficiency of large-scale ML training. It details the trade-offs between infrequent and frequent checkpointing, highlighting how continuous checkpointing mitigates risks from hardware failures and preemptions with minimal performance degradation. The post provides configuration examples for MaxText and benchmarks demonstrating reduced checkpoint intervals and resource preservation. It also outlines more advanced Orbax features like custom save decision policies and preservation policies with minimum intervals, and emphasizes the importance of co-locating storage with training clusters for optimal performance.
  12. A Developer's Guide to Debugging JAX on Cloud TPUs: Essential Tools and Techniques- Google Developers BlogIntroduces essential debugging and profiling tools for JAX on Cloud TPUs: libtpu (TPU Runtime), JAX/jaxlib (framework), verbose logging flags (TPU_VMODULE, TPU_MIN_LOG_LEVEL, TF_CPP_MIN_LOG_LEVEL, TPU_STDERR_LOG_LEVEL), TPU Monitoring Library (programmatic metrics like duty_cycle_pct), and tpu-info CLI (real-time memory and utilization). Details log file locations and retrieval methods across TPU VMs. Provides example commands and code snippets for enabling logging, using the monitoring library, and running tpu-info.
  13. Gemma explained: EmbeddingGemma Architecture and Recipe- Google Developers BlogThis post details the architecture and training recipe of EmbeddingGemma, a new text embedding model derived from Gemma 3. It explains how Gemma 3 was adapted into an encoder-decoder architecture and initialized from its encoder. The post elaborates on the formation of embeddings through a transformer encoder, mean pooling, and two dense layers followed by normalization. It describes the three weighted loss functions used for training: Noise-Contrastive Estimation (NCE) loss for similarity and contrast, Global Orthogonal Regularizer (GOR) for spreading embeddings, and Geometric Embedding Distillation from a Gemini model. Matryoshka Representation Learning (MRL) is introduced, allowing for variable embedding dimensionality. The development recipe includes encoder-decoder training, pre fine-tuning, fine-tuning with hard negatives, model souping, and quantization-aware training (QAT).
  14. Beyond backpropagation: JAX's symbolic power unlocks new frontiers in scientific computing- Google Developers BlogThis post details how researchers are leveraging JAX's composable transformations and Taylor mode automatic differentiation to solve high-order Partial Differential Equations (PDEs) with neural networks. It highlights the limitations of traditional frameworks optimized for backpropagation in computing high-order derivatives and explains how JAX's Taylor mode AD enables efficient computation in a single forward pass. The post introduces the Stochastic Taylor Derivative Estimator (STDE) method, which achieved over 1000x speed-up and 30x memory reduction, enabling the solution of a 1-million-dimensional PDE in 8 minutes on a single GPU. It also mentions the generalization of JAX to handle infinite-dimensional vectors for variational derivatives.
  15. Stop “vibe testing” your LLMs. It's time for real evals.- Google Developers BlogThis post introduces Stax, an experimental developer tool designed to streamline the LLM evaluation lifecycle. It addresses the challenge of non-deterministic AI model outputs by providing a framework for rigorous evaluation. Key features include data ingestion (CSV upload or creation), pre-built autoraters for common checks (coherence, factuality, conciseness), and the ability to build custom autoraters based on specific criteria (e.g., brand voice, PII avoidance, style guide adherence). Stax aims to empower developers to make data-driven decisions when selecting and shipping AI systems.
  16. Introducing Gemma 3 270M: The compact model for hyper-efficient AI- Google Developers BlogIntroduces Gemma 3 270M, a new 270-million parameter model designed for task-specific fine-tuning. Highlights its compact and capable architecture, extreme energy efficiency (0.75% battery usage on Pixel 9 Pro SoC for 25 conversations with INT4 quantization), strong instruction-following capabilities (new level of performance for its size on IFEval benchmark), and production-ready quantization (QAT checkpoints available for INT4 precision). Emphasizes its suitability for high-volume, well-defined tasks, cost-effectiveness, rapid iteration, user privacy (on-device deployment), and building fleets of specialized models. Provides resources for downloading, trying, fine-tuning, and deploying the model.
  17. T5Gemma: A new collection of encoder-decoder Gemma models- Google Developers BlogIntroduces T5Gemma, a new collection of encoder-decoder LLMs derived from decoder-only Gemma models via adaptation. Details the model adaptation technique, which initializes encoder-decoder parameters using pretrained decoder-only weights and further adapts them. Highlights T5Gemma's performance advantages in quality-inference efficiency trade-offs across benchmarks like SuperGLUE and GSM8K, often achieving comparable or better accuracy with similar or lower latency. Discusses the improved capabilities of T5Gemma after instruction tuning, showing significant gains on reasoning-intensive benchmarks like GSM8K and DROP. Releases T5Gemma checkpoints in various sizes and variants (pretrained and instruction-tuned) for community research and development.
  18. Using KerasHub for easy end-to-end machine learning workflows with Hugging Face- Google Developers BlogThis post introduces KerasHub, a library that simplifies defining model architectures and loading model weights (checkpoints) across different ML frameworks (JAX, PyTorch, TensorFlow). It highlights the compatibility of KerasHub with Hugging Face Hub, allowing users to load checkpoints in SafeTensors format directly into KerasHub models. Concrete examples demonstrate loading Mistral, Llama 3.1, and Gemma checkpoints from Hugging Face Hub into JAX using KerasHub's `from_preset` method and specific model classes.
  19. Build and train a recommender system in 10 minutes using Keras and JAX- Google Developers BlogIntroduces Keras Recommenders (KerasRS), a new library for building state-of-the-art recommendation systems. KerasRS provides building blocks for ranking and retrieval tasks, with examples of sequential retrieval models and pairwise hinge loss. It supports JAX, TensorFlow, and PyTorch backends. Future plans include DistributedEmbedding for TPUs and continuous addition of model implementations. The post also highlights documentation and examples on keras.io.
  20. Announcing the general availability of Llama 4 as MaaS on Vertex AI- Google Developers BlogThis post announces the general availability of Meta's Llama 4 and Llama 3.3 models as managed API endpoints (Model-as-a-Service or MaaS) on Google Cloud's Vertex AI. It highlights the benefits of using MaaS, including zero infrastructure management, guaranteed performance, and enterprise-grade security. The post provides a Python code example demonstrating how to use the ChatCompletion API with Llama 4, including multimodal capabilities with image input from Cloud Storage. It also touches upon cost and quota considerations for using these models.
  21. Introducing TxGemma: Open models to improve therapeutics development- Google Developers BlogIntroduces TxGemma, a collection of open models fine-tuned from Gemma 2 for therapeutic development. TxGemma models are available in 2B, 9B, and 27B sizes, with 'predict' versions for classification, regression, and generation tasks, and 'chat' versions for conversational analysis. Demonstrates Agentic-Tx, a therapeutics-focused agentic system powered by Gemini 2.0 Pro, integrating TxGemma and other tools for complex research problems. Provides fine-tuning example Colab notebooks for adapting TxGemma to custom therapeutic data and tasks.
  22. Introducing Gemma 3: The Developer Guide- Google Developers BlogIntroduces Gemma 3, the most capable and advanced version of the Gemma open-model family. Key new features include multimodality (vision-language input, text output), extended context windows (up to 128k tokens), support for over 140 languages, and improved math, reasoning, and chat capabilities with structured outputs and function calling. Gemma 3 is available in four sizes (1B, 4B, 12B, and 27B) as pre-trained and instruction-tuned models. The post details the training process, including distillation, RLHF, RLMF, and RLEF, and highlights the use of a new tokenizer and training on Google TPUs with JAX. It also introduces ShieldGemma 2, a 4B image safety classifier built on Gemma 3.
  23. Start building with Gemini 2.0 Flash and Flash-Lite- Google Developers BlogThis post announces the general availability of Gemini 2.0 Flash-Lite in the Gemini API for production use in Google AI Studio and for enterprise customers on Vertex AI. It highlights improved performance over previous models in reasoning, multimodal, math, and factuality benchmarks. Simplified pricing for prompts over 128K tokens is also detailed. Examples from Daily.co (voice AI), Dawn (data analytics), and Mosaic (video editing) showcase practical applications and benefits of using the Gemini 2.0 Flash family.
  24. Introducing PaliGemma 2 mix: A vision-language model for multiple tasks- Google Developers BlogIntroduces PaliGemma 2 mix, an upgraded vision-language model in the Gemma family, offering multi-task capabilities out-of-the-box for common use cases like image segmentation, short video captioning, scientific question answering, and text-related tasks. It is available in developer-friendly sizes (3B, 10B, and 28B parameters) and resolutions (224px and 448px), and is compatible with Hugging Face Transformers, Keras, PyTorch, JAX, and Gemma.cpp. The post provides detailed examples of its performance on detection, multiple object detection, OCR, segmentation, question answering, and captioning tasks.
  25. Gemini 2.0: Flash, Flash-Lite and Pro- Google Developers BlogThis post announces the general availability of Gemini 2.0 Flash, the public preview of Gemini 2.0 Flash-Lite, and the experimental availability of Gemini 2.0 Pro. It highlights performance improvements over Gemini 1.5, a 1 million token context window, multimodal input (with image and audio output planned), and simplified pricing for Flash and Flash-Lite models. It also mentions the Gemini API, Google AI Studio, and Vertex AI as access points.
  26. Beyond English: How Gemma open models are bridging the language gap- Google Developers BlogThis post details how Google's Gemma family of open models is enabling developers and researchers to build more inclusive AI by addressing the language gap. It highlights specific examples from INSAIT (Bulgarian language models) and AI Singapore (SEA-LION models for Southeast Asian languages), showcasing how these teams leveraged Gemma variants to create state-of-the-art models tailored to specific linguistic and cultural needs. The post emphasizes the importance of high-quality, relevant training data and introduces techniques like continuous pre-training and model merging to mitigate catastrophic forgetting. It also mentions the Kaggle competition 'Unlock Global Communication with Gemma' as a way for developers to contribute to this effort.
  27. Introducing PaliGemma 2: Powerful Vision-Language Models, Simple Fine-Tuning- Google Developers BlogIntroduces PaliGemma 2, the next evolution of tunable vision-language models built on Gemma 2. Key features include scalable performance with multiple model sizes (3B, 10B, 28B parameters) and resolutions (224px, 448px, 896px), enhanced long captioning for detailed image descriptions, and leading performance on new horizons like chemical formula recognition, music score recognition, spatial reasoning, and chest X-ray report generation. PaliGemma 2 is a drop-in replacement for existing PaliGemma users and simplifies fine-tuning. It is available on Hugging Face and Kaggle, with documentation and example notebooks for integration and fine-tuning across various frameworks.
  28. Inference with Gemma using Dataflow and vLLM- Google Developers BlogThis post introduces the integration of vLLM with Google Dataflow for efficient LLM inference, specifically using Gemma models. It explains vLLM's continuous batching technique for optimizing LLM serving throughput and latency, contrasting it with traditional batching. It also details how Dataflow's model manager simplifies the deployment of vLLM within streaming pipelines by allowing controlled instances of the model handler, abstracting away multiprocessing complexities and enabling efficient communication between worker processes and the vLLM instance. A performance comparison shows a significant improvement (over 23x) in vCPU hours for processing prompts using vLLM with Dataflow compared to a naive batching strategy.
  29. Introducing Keras Hub: Your one-stop shop for pretrained models- Google Developers BlogIntroduces KerasHub, a unified library for pretrained models, consolidating KerasNLP and KerasCV. KerasHub provides a consistent API for accessing NLP and CV models (e.g., Gemma, PaliGemma, Stable Diffusion 3), simplifying model discovery and usage. It offers features like effortless model publishing, LoRA fine-tuning, quantization, and multi-host training. The post details installation, usage examples for Gemma and PaliGemma, and explains the transition for KerasNLP and KerasCV developers, highlighting integrated preprocessing and the availability of loss functions within `keras.losses`.
  30. Gemini 1.5 Flash-8B is now production ready- Google Developers BlogAnnounces the production readiness of Gemini 1.5 Flash-8B, a smaller and faster variant of Gemini 1.5 Flash. Highlights its 50% lower price, 2x higher rate limits, and lower latency on small prompts compared to Gemini 1.5 Flash. Details pricing for input and output tokens and mentions its suitability for high-volume multimodal use cases and long context summarization.
  31. Towards Global Understanding – Advancing Multilingual AI with Gemma 2 and a $150K Challenge- Google Developers BlogIntroduced a new 2 billion parameter Gemma 2 variant fine-tuned for Japanese, comparable to GPT 3.5 in Japanese tasks and runnable on mobile devices. Released model weights and training materials for developers to adapt Gemma to their own languages. Highlighted community fine-tuning efforts for languages like Arabic, Vietnamese, and Zulu, and a specific project for the Jeju Island dialect. Launched the 'Unlocking Global Communication with Gemma' competition with $150,000 in prizes on Kaggle to encourage further multilingual AI development.
  32. Updated production-ready Gemini models, reduced 1.5 Pro pricing, increased rate limits, and more- Google Developers BlogThis post announces the release of production-ready Gemini 1.5 Pro-002 and Gemini 1.5 Flash-002 models. Key improvements include over 50% price reduction on 1.5 Pro, 2x higher rate limits for 1.5 Flash and ~3x higher for 1.5 Pro, 2x faster output, and 3x lower latency. The models show improved quality, with larger gains in math, long context, and vision benchmarks (~7% MMLU-Pro, ~20% MATH/HiddenMath, ~2-7% vision/code). Default output length is shorter (5-20%) for conciseness. Updated filter settings are not applied by default. An experimental Gemini 1.5 Flash-8B model is also released.
  33. Gemma explained: PaliGemma architecture- Google Developers BlogThis post details the architecture of PaliGemma, a vision-language model. It explains the SiglipVisionModel, including its embeddings, patch embedding (Conv2d), position embedding, and encoder layers. It also describes the multi_modal_projector and the GemmaForCausalLM language model. The post elaborates on the extended vocabulary of the Gemma tokenizer for PaliGemma, including location and segmentation tokens, and provides a detailed example of object segmentation with output interpretation.
  34. TensorFlow Lite is now LiteRT- Google Developers BlogTensorFlow Lite (TFLite) has been rebranded as LiteRT (Lite Runtime) to reflect its expanded vision of supporting multiple ML frameworks (PyTorch, JAX, Keras) for on-device AI deployment. The core functionality and .tflite file format remain unchanged. Package dependencies will be updated to use LiteRT from new repositories (Maven, PyPi, Cocoapods). Development continues on the TensorFlow repo for now, with a move to a new LiteRT repo planned. MediaPipe Tasks are recommended for future development over the TFLite Support Library.
  35. Gemma explained: RecurrentGemma architecture- Google Developers BlogThis post details the RecurrentGemma architecture, a hybrid model that mixes gated linear recurrences with local sliding window attention, based on the Griffin architecture. It explains the trade-offs of this approach, such as reduced needle-in-haystack performance but improved efficiency for long context prompts. The post breaks down the architecture into its core components: embedding tokens, decoder layers (alternating recurrent and MQA blocks), residual blocks with recurrent blocks, and residual blocks with local MQA. It also details the parameters like model width, RNN width, MLP expansion factor, and local attention window size, and provides a Python code snippet of the RecurrentGemmaForCausalLM model.
  36. Gemma explained: What’s new in Gemma 2- Google Developers BlogThis post introduces Gemma 2, a new suite of open models available in 2B, 9B, and 27B parameter sizes. It details architectural innovations including Alternating Local and Global Attention, Logit Soft-Capping, RMSNorm for Pre and Post-Normalization, and Grouped-Query Attention (GQA). The post provides the Python code for the Gemma 27B model architecture and details the self-attention mechanism with GQA, including specific head configurations for the 27B, 9B, and 2B models. It also highlights the inclusion of additional RMSNorm layers for improved training stability. Key findings from the technical report are summarized, emphasizing the performance benefits of knowledge distillation and GQA, and the advantage of deeper models.
  37. Gemma explained: An overview of Gemma model family architectures- Google Developers BlogThis post introduces the Gemma family of open models, detailing their architectures and core parameters. It explains the transformer decoder architecture, the GeGLU activation function, and the distinctions between MHA and MQA. Specific architectural details for Gemma 2B and 7B models are provided, including d_model, layers, feedforward hidden dims, num heads, num KV heads, head size, and vocab size. Code examples for exploring model architectures using Hugging Face Transformers and torchinfo are included.
  38. Streamlining LLM Inference at the Edge with TFLite- Google Developers BlogIntroduced a new weight cache provider interface for XNNPack and implemented a file-backed memory-mapped (mmap) cache provider within the TFLite Delegate. This provider persists packed weights to disk, eliminating the need for repacking on each model load, reducing startup latency and peak memory usage, and enabling cross-process weight sharing. Benchmarks show significant improvements in session initialization time (halving time to first token for LLMs) and reduced peak Resident Set Size.
  39. Gemini 1.5 Flash price drop with tuning rollout complete, and more- Google Developers BlogThis post announces a significant price reduction for Gemini 1.5 Flash, with input token costs decreasing by 78% and output token costs by 71%. It also announces the general availability of Gemini 1.5 Flash tuning for all developers via the Gemini API and Google AI Studio. The Gemini API now supports queries in over 100 additional languages for both Gemini 1.5 Pro and Flash models. Google AI Studio access is now available by default for Google Workspace users. The post also highlights improvements to developer documentation, including a revamped API reference, and the addition of PDF understanding capabilities (text and vision) to the Gemini API and AI Studio. Numerous improvements to AI Studio itself are also mentioned, such as overhauled keyboard shortcuts and decreased loading times.
  40. Fine-tuning Gemma 2 with Keras - and an update from Hugging Face- Google Developers BlogIntroduces Gemma 2 models (9B and 27B parameters) with Keras and JAX, showcasing distributed fine-tuning using `keras.distribution.ModelParallel` and weight partitioning strategies. Demonstrates LoRA for efficient fine-tuning by reducing trainable parameters. Enhances Hugging Face integration for KerasNLP to load fine-tuned weights from various models (Gemma, Llama3) on the fly. Highlights PaliGemma for vision-language tasks.
  41. Model Explorer: Simplifying ML models for Edge devices- Google Developers BlogIntroduces Model Explorer, a new graph visualization tool for ML models, supporting JAX, PyTorch, TensorFlow, and TensorFlow Lite. It offers visualization of large graphs (tens of thousands of nodes) at 60 FPS, metadata overlays, side-by-side comparison for debugging conversion errors, and per-node data overlays for debugging performance and numeric accuracy. It is available as a PyPI package and can run locally or in Colab notebooks. Model Explorer is used internally at Google by teams like Waymo and Google Silicon for optimizing on-device models such as Gemini Nano.
  42. Gemini 1.5 Pro and 1.5 Flash GA, 1.5 Flash tuning support, higher rate limits, and more API updates- Google Developers BlogAnnounces the stable release and billing for Gemini 1.5 Flash and 1.5 Pro. Increases rate limits for Gemini 1.5 Flash to 1000 RPM and removes the daily limit. Introduces tuning support for Gemini 1.5 Flash, available in Google AI Studio and via the Gemini API. Details billing for Gemini API usage beyond the free tier. Adds JSON schema mode for more controlled model output. Introduces light mode and mobile improvements for Google AI Studio.
  43. Introducing PaliGemma, Gemma 2, and an Upgraded Responsible AI Toolkit- Google Developers BlogIntroduces PaliGemma, an open vision-language model built on SigLIP and Gemma, designed for fine-tuning on vision-language tasks like captioning, VQA, object detection, and segmentation. Provides pretrained and fine-tuned checkpoints, available on Kaggle, Colab, GitHub, Hugging Face, Vertex AI, and NVIDIA. Announces Gemma 2, the next generation of Gemma models, featuring a new architecture for improved performance and efficiency (27B parameters comparable to Llama 3 70B), reduced deployment costs, and versatile tuning toolchains across platforms like Google Cloud, Axolotl, Hugging Face, and NVIDIA TensorRT-LLM. Expands the Responsible Generative AI Toolkit with the open-sourced LLM Comparator for side-by-side model response evaluation.
  44. Publish your Keras models on Kaggle and Hugging Face- Google Developers BlogIntroduces `keras_nlp.upload_preset` for uploading fine-tuned Keras models to Kaggle and Hugging Face. Demonstrates the process with an example of fine-tuning Gemma to speak like a pirate and uploading it as a KerasNLP preset. Highlights the benefits of Keras as a preferred model format due to its consistent API, source code availability, and multi-backend support (JAX, TensorFlow, PyTorch). Provides usage examples for loading models from Kaggle and Hugging Face using the `.from_preset()` API.
  45. Gemma Family Expands with Models Tailored for Developers and Researchers- Google Developers BlogIntroduces CodeGemma, a 7B pretrained variant for code completion/generation, a 7B instruction-tuned variant for code chat, and a 2B pretrained variant for fast local code completion. Highlights CodeGemma's intelligent code completion, enhanced accuracy from training on 500 billion tokens, multi-language proficiency (Python, JavaScript, Java), and streamlined workflow integration. Also introduces RecurrentGemma, a technically distinct model leveraging recurrent neural networks and local attention for improved memory efficiency, higher throughput at higher batch sizes, and research innovation. Details compatibility with JAX, PyTorch, Hugging Face Transformers, Gemma.cpp, Keras, NVIDIA NeMo, TensorRT-LLM, Optimum-NVIDIA, MediaPipe, and Vertex AI. Announces Gemma 1.1 with performance improvements, bug fixes, and updated terms.
  46. Introducing Gemma models in Keras- Google Developers BlogThis post announces the integration of Gemma models into the KerasNLP collection, leveraging Keras 3's multi-backend support (JAX, PyTorch, TensorFlow). It introduces new features for large language models: a LoRA API for parameter-efficient fine-tuning, demonstrated by reducing trainable parameters from 2.5B to 1.3M with rank=4, and large-scale model-parallel training capabilities using Keras's distribution API. The post provides code examples for instantiating Gemma models, using the LoRA API, and configuring model-parallel training with a DeviceMesh and LayoutMap for sharding model weights across accelerators. It highlights Gemma's performance on benchmarks and offers links to Colab notebooks for getting started, fine-tuning with LoRA, and distributed fine-tuning.
  47. New Resources to Build with Google AI- Google Developers BlogThis post announces the unification of Google's ML products and tools under the Google AI umbrella (ai.google/build/machinelearning) and introduces new toolkits for end-to-end guidance on building AI solutions. It highlights two initial toolkits: one for building a text classifier using Keras, and another for shrinking large language models to run on Android using Keras and TensorFlow Lite. The post emphasizes that more toolkits are planned.
  48. Announcing TensorRT integration with TensorFlow 1.7- Google Developers BlogThis post announces the integration of TensorRT with TensorFlow 1.7. TensorRT is a platform for high-performance deep learning inference, and its integration with TensorFlow allows for optimized inference on NVIDIA GPUs. The integration provides a new `tf.contrib.tensorrt` module, enabling users to convert TensorFlow models into TensorRT-optimized graphs for faster execution. The post highlights performance gains achieved through this integration and provides guidance on how to use the new module.
  49. Announcing TensorFlow 1.5- Google Developers BlogTensorFlow 1.5 introduces new features and improvements, including support for new hardware accelerators, enhanced performance for distributed training, and expanded capabilities for model deployment. This release focuses on making TensorFlow more accessible and powerful for a wider range of machine learning tasks and production environments.
  50. Creating Custom Estimators in TensorFlow- Google Developers BlogIntroduces the concept and implementation of custom Estimators in TensorFlow, enabling users to define their own model architectures and training loops for greater flexibility in machine learning development.
  51. Announcing Core ML support in TensorFlow Lite- Google Developers BlogThis post announces the integration of Core ML support into TensorFlow Lite, enabling TensorFlow Lite models to run on iOS devices using Apple's Core ML framework. This allows for hardware-accelerated inference on iOS devices, improving performance and efficiency for on-device machine learning applications. It details how to convert TensorFlow Lite models to Core ML format and provides an example of using the converted model in an iOS application.
  52. Introducing TensorFlow Feature Columns- Google Developers BlogIntroduces TensorFlow Feature Columns, a new abstraction layer for representing structured data in TensorFlow models. This feature simplifies the process of handling categorical and numerical features, allowing for more efficient model training and improved performance. It provides a way to define feature transformations and combinations directly within the model graph, reducing boilerplate code and improving model readability.
  53. Announcing TensorFlow Lite- Google Developers BlogAnnounces TensorFlow Lite, a lightweight solution for deploying TensorFlow models on-device, including mobile and embedded systems. It focuses on reducing model size and latency for on-device inference, with support for common TensorFlow operations and optimizations for mobile hardware.
  54. Announcing TensorFlow r1.4- Google Developers BlogTensorFlow r1.4 introduces several key features and improvements, including the release of TensorFlow Lite for mobile inference, enhanced support for distributed training with Horovod, and new APIs for custom training loops and model debugging. It also includes performance optimizations and bug fixes, making it a more robust and flexible framework for machine learning development and deployment.
  55. Eager Execution: An imperative, define-by-run interface to TensorFlow - Google Developers BlogIntroduced Eager Execution as an imperative, define-by-run interface for TensorFlow. This allows operations to be evaluated immediately, enabling interactive development and easier debugging compared to the previous define-and-run graph-based execution model. This change significantly improves the developer experience for machine learning practitioners.
  56. How Machine Learning with TensorFlow Enabled Mobile Proof-Of-Purchase at Coca-Cola- Google Developers BlogThis post details the technical implementation of a mobile proof-of-purchase system for Coca-Cola using TensorFlow. It describes how TensorFlow was used to build a machine learning model that can recognize product packaging from mobile device images. The system involves image capture, preprocessing, model inference on mobile devices, and backend processing for validation. It highlights the challenges of on-device ML, including model size and performance, and how TensorFlow Lite was used to address these.
  57. Introduction to TensorFlow Datasets and Estimators- Google Developers BlogThis post introduces TensorFlow Datasets (tf.data) for efficiently loading and preprocessing data, and TensorFlow Estimators (tf.estimator) for simplifying the creation and training of machine learning models. It demonstrates how these components can be used together to build and train models with less boilerplate code, improving developer productivity in machine learning workflows.
  58. Kaldi now offers TensorFlow integration- Google Developers BlogThis post announces the integration of Kaldi, a widely-used speech recognition toolkit, with TensorFlow. This integration allows users to leverage TensorFlow's machine learning capabilities, such as its distributed training and advanced model architectures, within the Kaldi framework for speech recognition tasks. It highlights the ability to use TensorFlow models for acoustic modeling and other components of the speech recognition pipeline.
  59. TensorFlow Benchmarks and a New High-Performance Guide- Google Developers BlogThis post introduces benchmarks for TensorFlow, demonstrating its performance on various hardware configurations. It also provides a guide to optimizing TensorFlow models for high-performance inference, including techniques for model quantization and graph optimization. The goal is to help developers achieve faster and more efficient model deployment.
  60. XLA - TensorFlow, compiled- Google Developers BlogThis post introduces XLA (Accelerated Linear Algebra), a compiler for linear algebra that can be used to optimize TensorFlow computations. XLA aims to improve performance by fusing operations and optimizing for specific hardware targets. It supports TensorFlow's computation graph and can be used to compile models for CPUs, GPUs, and TPUs.
  61. Debug TensorFlow Models with tfdbg- Google Developers BlogIntroduces tfdbg, a TensorFlow debugger that allows for step-by-step execution, inspection of tensors, and visualization of the computation graph. This enhances the debugging capabilities for TensorFlow models.
  62. Announcing TensorFlow 1.0- Google Developers BlogTensorFlow 1.0 introduces a new graph-based execution model with `tf.Graph` and `tf.Session`, offering improved performance and production readiness over the previous eager execution model. It also introduces `tf.contrib` for experimental features and a new API for distributed training. This release emphasizes production stability and performance for machine learning deployments.
  63. TensorFlow 0.12 adds support for Windows- Google Developers BlogTensorFlow 0.12 now supports Windows, expanding the platform's reach for machine learning development and deployment.
  64. TensorFlow v0.9 now available with improved mobile support- Google Developers BlogTensorFlow v0.9 introduces improved mobile support, including the ability to run TensorFlow models on Android devices. This release also includes performance optimizations and bug fixes, making it easier for developers to deploy machine learning models on mobile platforms.