Warm-start Reinforcement Fine-Tuning from SFT checkpoints, Azure federated identity for uploads

Warm-Start Training for Reinforcement Fine-Tuning

You can now warm-start Reinforcement Fine-Tuning jobs from previously supervised fine-tuned checkpoints using the --warm-start-from flag. This enables a streamlined SFT-to-RFT workflow where you first train a model with supervised fine-tuning, then continue training with reinforcement learning.

See the Warm-Start Training guide for details.

Azure Federated Identity for Model Uploads

Model uploads from Azure Blob Storage now support Azure AD federated identity authentication as an alternative to SAS tokens. This eliminates the need for credential rotation and enables secure, credential-less authentication.

See the Uploading Custom Models documentation for setup instructions.

📚 Documentation Updates

• Warm-Start Training: New guide for SFT-to-RFT workflows (Warm-Start Training)
• Azure Federated Identity: Setup instructions for Azure AD authentication (Uploading Custom Models)
• Preserved Thinking: Multi-turn reasoning with preserved thinking context (Reasoning)
• GLM 4.7: Added to models supporting reasoning_effort parameter

[accordion] Bug Fixes & Minor Improvements

• RFT Cost Display: Reinforcement Fine-Tuning job pages now show approximate final cost (Web App)
• GPU Information: Deployments table displays GPU type and count (Web App)
• DPO Job Resume: Preference Fine-Tuning jobs can now be resumed after stopping (Web App, API)
• Free Tuning Filter: New filter in RFT model selector for free-to-fine-tune models (Web App)
• Playground Inputs: Editable number inputs for temperature, top_p, and other parameters (Web App)
• Clone Fine-Tuning Jobs: Fixed field population when cloning jobs (Web App)
• Evaluation Job Errors: Errors now display with alert banners (Web App)
• Invoice CSV Export: Improved download experience (Web App)
• Playground Validation: Prevents querying deployments with 0 replicas (Web App)
• List Models Filter: firectl model list supports --name and --public-only flags (CLI)
• RFT Concurrency: New --max-concurrent-rollouts and --max-concurrent-evaluations flags (CLI)