v0.1.5

deepseek-ai/DeepSeek-Coder-V2v0.1.5Mar 25, 2026by ChuxiJ

AI Summary

This version focuses on stability and security, fixing legacy GPU NaN issues, implementing lazy-load model initialization, and replacing pickle with JSON serialization for safer IPC.

Key Highlights

Lazy-load model initialization (no startup downloads)
Security fix: Replaced pickle with JSON serialization
Legacy GPU NaN/noise fixes
Per-request language isolation via ContextVar
VST3 plugin moved to a standalone repo

New Features

Lazy-loading models
JSON IPC serialization
Per-request language isolation
External AI helper modules

Full Release Notes

## What's New

### Bug Fixes
- **API server**: Restore lazy-load model initialization as default behavior. Models are no longer downloaded at startup — they initialize on-demand when the first request arrives.
- **GPU**: Resolve legacy GPU NaN/noise and duration=-1 garbage output (#924, #927, #929)
- **GPU**: Force pt backend on legacy CUDA GPUs (Titan Xp, etc.) to prevent vllm crashes (#932)
- **Security**: Replace pickle with JSON serialization in ModelRunner shared-memory IPC (#860)
- **Lego task**: Fix instruction mismatch causing noise output (#925)
- **Gradio**: Fix variable language not bound error in start script (#920)
- **i18n**: Per-request language isolation via ContextVar (#888)
- **API**: Fix crash on null duration, generation corruption, and locale parsing (#797, #854, #779)
- **Thread safety**: Add locking to i18n, API key, and cache singletons (#878)
- **AutoGen**: Capture repaint settings in batch parameter snapshot (#922)
- **ADG**: Fix ADG not usable with batches (#832)

### Features
- **Text tasks**: Add external AI helper modules, LM model discovery cache, and provider config (#881, #883, #884)
- **Model inventory**: Add supported_task_types and expand cover strength aliases

### Other
- **VST3**: Move VST3 plugin to standalone repo ace-step/acestep.vst3 (#935)

**Full Changelog**: https://github.com/ace-step/ACE-Step-1.5/compare/v0.1.4...v0.1.5