mirror of
https://github.com/ace-step/ACE-Step-1.5.git
synced 2026-07-02 16:37:04 +00:00
1c5d26e8f1
Side-Step is a non-destructive training module that provides corrected LoRA fine-tuning for ACE-Step models. Key improvements over the existing training code: - Continuous logit-normal timestep sampling (matching each model variant's forward() logic) instead of discrete 8-step schedule - Classifier-free guidance (CFG) dropout during training - Per-variant timestep parameters (mu/sigma/data_proportion) read from model config - Multi-optimizer support (AdamW, AdamW8bit, Adafactor, Prodigy) - Auto GPU detection with VRAM-aware batch sizing - Interactive wizard and TUI for configuration - TensorBoard logging with gradient norms and sample generation - Gradient estimation for selective module targeting New files: acestep/training_v2/ - Core training module train.py - CLI entry point sidestep_tui.py - TUI launcher requirements-sidestep.txt - Additional dependencies Co-authored-by: Cursor <cursoragent@cursor.com>
23 lines
1.3 KiB
Text
23 lines
1.3 KiB
Text
# ╔══════════════════════════════════════════════════════════════╗
|
|
# ║ Side-Step -- Extra dependencies for ACE-Step LoRA training ║
|
|
# ║ Install: pip install -r requirements-sidestep.txt ║
|
|
# ╚══════════════════════════════════════════════════════════════╝
|
|
#
|
|
# Side-Step is a companion CLI/TUI for ACE-Step.
|
|
# Install ACE-Step's own requirements first, then these.
|
|
|
|
# ── Required for Side-Step CLI ────────────────────────────────
|
|
rich>=13.0.0
|
|
|
|
# ── Required for Side-Step TUI ────────────────────────────────
|
|
textual>=0.47.0
|
|
|
|
# ── Optional: 8-bit optimizers (saves ~30-40% optimizer VRAM) ─
|
|
# Uncomment to enable AdamW8bit in the optimizer selector.
|
|
# Supports Linux and Windows (official wheels).
|
|
# bitsandbytes>=0.45.0
|
|
|
|
# ── Optional: Prodigy adaptive optimizer (auto-tunes LR) ─────
|
|
# Uncomment to enable the Prodigy optimizer.
|
|
# Great if you don't want to manually tune learning rate.
|
|
# prodigyopt>=1.1.2
|