v0.2.0 — Modal Cloud GPU, DPO Training, Phase 4 Feature Complete

What's New in v0.2.0

This release brings TuneOS to feature-complete Phase 4, adding cloud GPU training, DPO preference learning, advanced trainer capabilities, and a completely refreshed UI.

---

🚀 New Features

Cloud GPU Training (Modal.com)

• Jobs can now run on a free Modal T4 GPU instead of the local device — ideal when no local GPU is available
• New Compute backend selector in Step 4 (Local GPU / Modal / HF Spaces)
• workers/modal_runner.py serialises the dataset, runs the full training pipeline remotely, and streams adapter + eval metrics back to local disk
• Enable by setting MODAL_TOKEN_ID + MODAL_TOKEN_SECRET; modal is optional (poetry install --with modal)

DPO Preference Training (P4-C)

• Full DPO recipe wired end-to-end through trainer, API, and UI
• Step 4 now exposes DPO hyperparameters; dataset step validates preference columns before build
• TRL DPOTrainer integration with configurable beta and reference model

Advanced Trainer (P4-A / P4-B)

• Prompt template registry — Alpaca, ChatML, Llama-3, Phi-3, and raw formats; sample packing for throughput
• Configurable report_to for HF experiment-tracker integration
• Flash Attention 2 / SDPA and RoPE scaling plumbed into model loading
• LoRA init strategy exposed via init_lora_weights
• Extended architecture support: Qwen3, Phi-4, Cohere, OLMo, Mixtral, MPT, StarCoder2, GPT-BigCode
• use_4bit now optional instead of forced; all checkpoints use safe_serialization=True

Distillation, FSDP & INT8 Export (P4-F)

• Knowledge distillation from a teacher model
• FSDP (Fully Sharded Data Parallel) for multi-GPU training
• Hyperparameter sweep support
• INT8 export via bitsandbytes

Richer Eval Metrics & Live Streaming (P4-D / P4-E)

• ROUGE-1 and BLEU computed from held-out evaluation sample after training
• Live Modal training log streaming in Step 5
• Eval metrics persisted to SQLite and surfaced in Step 6 Results

Multi-run Compare & Benchmark (P4-F)

• New compare page with overlaid loss curves across runs
• lm-eval benchmark wrapper for standardised model evaluation

API Hardening (P4-A)

• LRU inference model cache (maxsize=3) — avoids GB-scale model reloads
• GET /api/jobs pagination (default 50, max 500)
• GET /api/gpu now reports device_count, vram_total_gb, vram_free_gb, cuda_version
• Eval metrics fall back to durable SQLite when Redis copy has expired

PostgreSQL Experiment Backend

• Set EXPERIMENTS_DB_URL to a postgresql:// DSN to share experiment data across workers
• All upserts migrated to portable ON CONFLICT … DO UPDATE (SQLite 3.24+ and PostgreSQL compatible)

UI Polish

• Chat panel consolidated into a single, cleaner component
• Step 5/6 surfaces all eval metrics + Modal badge
• Register-to-model-registry card on Results step
• Hyperparameter comparison table (Technique, LR, LoRA r, Batch columns)

---

🔧 CI Improvements

• Monolithic CI split into dedicated path-scoped jobs (lint, test-core, test-ui, docker, docs)
• Docker build significantly faster via BuildKit cache injection for pip/poetry downloads
• mypy static typing added to lint job

📖 Docs

• Quickstart updated for Celery fallback, HF_TOKEN, and Modal cloud GPU
• docs/api.md documents GET /api/health/workers and compute_backend
• docs/DEPLOY.md gains full .env reference table
• docs/supported-models.md reflects auto target_modules detection

---

🔄 Changed

• State data models now use pydantic.BaseModel (Reflex rx.Base removed in newer versions)
• Minimum trl version raised to >=0.12.0

---

Full Changelog: v0.1.0...v0.2.0