[jimdowling]

JIRA: HWORKS-3465

Adds a built-in LLM fine-tuning (LoRA) factory, a single-host fine-tuning environment (lora-fine-tuning-torch) and a Ray RL post-training environment (rl-post-training), moves every Ray environment to Ray 2.59.0, and retires the Ray TensorFlow and TensorFlow training environments (environments cloned from them keep working; tensorflow-inference-pipeline stays for TensorFlow Serving). The Terminal gets w9s, an overview of the project's factory systems, and the Factory chat no longer gets out of step with the Terminal. Deployment and Generic factories follow in a later change.

Related PRs: docker-images, hopsworks-ee, hopsworks-helm, hopsworks-front, hopsworks-api, loadtest

Merge order: docker-images first (it publishes the images the backend and chart name), then hopsworks-ee, hopsworks-helm, hopsworks-front, hopsworks-api, docs, loadtest.

This PR

• Lists the new environments, drops the TensorFlow training environment, and updates the configuration reference for the rl-post-training setting and the Ray version.

Testing

• hopsworks-docs markdownlint and the strict hopsworks-docs check build.

Loadtests

Run on a dev cluster (5.3.0-SNAPSHOT, Ray on) with this backend, frontend and chart DML deployed, as in-cluster loadtest Jobs from loadtest#1188:

• Run 1, 28/28 passed: workflows/model_training (test_finetune_envs, test_train torch and TensorFlow), test_factories (7), test_environment, test_ray_job_execution (4), test_proxy_terminal (2), test_terminal_proxy_token (3), test_proxy_raydashboard (7).
• Run 2, after pushing the images rebuilt from this branch on the current *-5.3.0-SNAPSHOT parents (ray-training-pipeline, ray-torch-training-pipeline, rl-post-training, lora-fine-tuning-torch, tensorflow-inference-pipeline, terminal-server, terminal-gpu, terminal-spark): 14/14 passed, the image-dependent subset (workflows/model_training, test_environment, test_ray_job_execution, the two Terminal proxy modules).

🤖 Generated with Claude Code