LoRA, MoE & Large Model Support

TL;DR

• Expanded Qwen Catalog: SFT and inference presets for all Qwen releases
• Large-Model Topologies: 2x, 4x, and 8x layouts across B200, H200, and H100 fleets
• MoE-Ready: All topologies ready for advanced Qwen variants
• LoRA-First SFT: Low-Rank Adaptation as first-class training mode
• Turnkey Rollout: Automatic SKU surfacing in API and UI

Expanded Qwen Catalog

Simple Training now ships SFT and inference presets for every Qwen release outside the existing qwen3-{0.6B-32B} range.

• Full Coverage: Complete coverage for Qwen 1.x/2.x/2.5 checkpoints
• Presets: Ready-to-use presets for SFT and inference
• Consistency: Consistent configuration across all Qwen variants

Large-Model Inference & Training Topologies

Added 2x, 4x, and 8x layouts across B200, H200, and H100 fleets.

• Multiple Layouts: Choose from 2x, 4x, or 8x GPU configurations
• Fleet Support: Works across B200, H200, and H100 GPU fleets
• MoE-Ready: All topologies ready for advanced Qwen variants
• SFT & Inference: Support for both training and inference workflows

LoRA-First SFT

Low-Rank Adaptation is now a first-class training mode across every new Qwen topology.

• Default Mode: LoRA is the default training mode for new topologies
• Parameter Efficiency: Efficient fine-tuning with minimal parameter overhead
• Universal Support: Available across all Qwen topologies
• Easy Configuration: Simple configuration for LoRA training

Turnkey Rollout

API and UI selectors automatically surface the new Qwen SKUs so jobs can be scheduled without manual topology overrides.

• Automatic Discovery: New SKUs automatically appear in selectors
• No Manual Overrides: No need to manually configure topologies
• Seamless Integration: Works seamlessly with existing workflows

Use Cases

• Complete Qwen Coverage: Use any Qwen model variant with presets
• Large Model Training: Train large models with multi-GPU topologies
• Efficient Fine-Tuning: Use LoRA-first approach for parameter-efficient training
• Simplified Workflows: Automatic SKU discovery simplifies job scheduling