Samples

Training

• Train Llama 3-8B using JAX, Ray, and GKE on Trillium

  Perform distributed training of the Llama 3-8B model on GKE using JAX, Ray Train, and TPU v6e (Trillium) with MaxText for optimized multi-host scaling.

• Pretrain Llama 3.1-70B using GKE clusters on Ironwood

  Train the Llama 3.1-70B model on TPU7x (Ironwood) using the MaxText framework.

• Pretrain DeepSeek 3-671B using GKE clusters on Ironwood

  Train the DeepSeek 3-671B model on TPU7x using optimized recipes for large-scale Mixture-of-Experts (MoE) architectures.

• Pretrain GPT OSS-120B using GKE clusters on Ironwood

  Train the GPT OSS-120B reasoning model on TPU7x using optimized recipes for large-scale distributed training.

• Pretrain Qwen 3-235B using GKE clusters on Ironwood

  Train the Qwen 3-235B-A22B MoE model on TPU7x using optimized recipes for high-performance reasoning.

• Pretrain Wan 2.1-14B using GKE clusters on Ironwood

  Train the Wan 2.1-14B video generation model on TPU7x using optimized recipes for high-performance video synthesis.

• Pretrain GPT3-175B using GKE clusters on Trillium

  Train the GPT3-175B model on TPU v6e using MaxText and optimized recipes for large-scale, cost-effective performance.

• Pretrain Gemma3-12B using GKE clusters on Trillium

  Train the Gemma3-12B model on TPU v6e using MaxText and optimized recipes for high-performance open-model development.

• Pretrain Llama 3.1-70B using GKE clusters on Trillium

  Train Llama 3.1-70B on TPU v6e using MaxText and optimized recipes for high-throughput, large-scale model training.

• Pretrain Llama 3.1-8B using GKE clusters on Trillium

  Train Llama 3.1-8B using MaxText on TPU v6e with this optimized recipe for scalable and high-performance pre-training.

• Pretrain Mixtral-8x22B using GKE clusters on Trillium

  Train Mixtral-8x22B on TPU v6e using MaxText for optimized performance and efficiency.

• Pretrain Mixtral-8x7B using GKE clusters on Trillium

  Train Mixtral-8x7B using MaxText on TPU v6e with optimized configurations for high-throughput MoE performance on Google Cloud.

• Pretrain DeepSeek 3-671B using GKE clusters on v5p

  Train and deploy the DeepSeek 3-671B model on TPU v5p using MaxText for optimized large-scale performance.

• Pretrain GPT3-175B using GKE clusters on v5p

  Train the GPT3-175B model on TPU v5p using MaxText with optimized configurations for large-scale distributed training.

• Pretrain Mixtral-8x7B using GKE clusters on v5p

  Train Mixtral-8x7B on TPU v5p using MaxText with optimized configurations for high-performance MoE workloads.

• Pretrain SDXL using GKE clusters on v5p

  Train and scale Stable Diffusion XL (SDXL) on TPU v5p using MaxDiffusion for high-performance generative AI workloads.

Inference

• Serve Llama 3.1-70B using GKE and vLLM on Trillium

  Serve LLMs on GKE using TPU v6e and vLLM, featuring optimized autoscaling and high-performance model serving on Google Cloud.

• Serve LLMs using GKE with KubeRay

  Serve an LLM using TPUs on GKE with the Ray Operator add-on and the vLLM serving framework.

• Serve open LLMs using GKE with Terraform

  Provision a GKE inference environment and deploy open LLMs using TPUs and a pre-configured Terraform architecture.

• Serve Stable Diffusion XL (SDXL) using GKE

  Serve Stable Diffusion XL (SDXL) on GKE using Cloud TPUs and the MaxDiffusion framework for high-performance image generation.

• Serve GPT OSS-120B with vLLM using GKE clusters on Ironwood

  Run high-performance inference for GPT-OSS models on TPU7x using vLLM for optimized throughput and low-latency serving on Google Cloud.

• Serve Qwen3-Coder-480B with vLLM using GKE clusters on Ironwood

  Serve Qwen3-Coder-480B-A35B on TPU7x using vLLM for optimized, high-throughput code generation and inference.

• Serve Llama 3.1-8B with vLLM on Trillium

  Serve Llama 3.1-8B on TPU v6e using vLLM for optimized, low-latency inference and high-throughput serving.

• Serve Qwen 3 with vLLM on Trillium

  Serve Qwen 3 models on TPU v6e using vLLM for high-performance, scalable inference and optimized throughput.

• Serve Qwen2.5-32B with vLLM on Trillium

  Serve the Qwen2.5-32B model on TPU v6e using vLLM for optimized, high-throughput inference.

• Serve Qwen2.5-VL with vLLM on Trillium

  Serve Qwen2.5-VL vision-language models on TPU v6e using vLLM for optimized, high-performance multimodal inference.