update README with final state, full pipeline inference, LM generation step

fix inference: add LM generation step, detokenize codes before DiT, full pipeline working

add _is_space flag, block inference during training, understand clone fix

fix understand_audio: clone tensors for inference mode, working on GPU (52s)

add understand_audio (LM reverse), demucs-infer fix, commit refs, dtype fixes

major update: PyTorch inference, Gradio 6, session isolation, /understand captioning

truncate long files to fit cap, show which files truncated/skipped

accept files until total audio cap reached, skip rest with warning

replace per-file cap with total audio cap (30 min max)

add LoRA download button after training (gr.File output, like rvc-beatrice)

remove ace-server understand proxy, captioning stays librosa + txt sidecars

add /understand API endpoint for GGUF audio captioning

SDPA first on Blackwell, FA2 only for Ampere/Hopper, txt caption support

add GPU/CUDA auto-detect, mixed precision, flash_attn, txt caption parser

update defaults: LR 3e-4, rank 32, alpha 2x rank (per Side-Step author)

save only on cancel/finish, max epochs 1000, default 3

add mid/sas analysis modes (Demucs + ensemble), auto-select by dataset size

add auto-captioning (BPM/key/signature via librosa), add librosa+mutagen deps

switch training to standard turbo (11s/epoch), auto-select standard GGUF for LoRA inference

add XL variant mappings to _VARIANT_DIR

fix: train on XL turbo (matches XL GGUF for inference), add XL checkpoint download

add full README with API docs, MCP, CLI, architecture

allow custom LoRA values in dropdown (API clients)

log ace-server restart, show output for adapter debugging

fix: forward adapter to synth request, default LM to 1.7B

fix: adapter saved to clean dir, LM dropdown no 'Default', on-demand download

copy train_engine.py into Docker image

Side-Step training engine, tested locally on CPU

switch back to 1.7B LM (fastest at 269s, 0.6B was 936s)

swap LM to 0.6B Q8_0 for speed test

swap LM 4B->1.7B Q8_0 for faster CPU inference

add LoRA adapter dropdown to inference UI

default mp3, remove format selector, increase LM timeout to 900s

remove accelerate (causes meta tensors), clean up patches

fix meta tensor crash: force low_cpu_mem_usage=False and float32 for CPU

disable flash_sdp on CPU, force attn_implementation=sdpa for training

add granular logging + full stderr to diagnose preprocessing hang

fix: use float32 not bfloat16 for CPU training (bf16 deadlocks on CPU)

redirect training subprocess stderr to log file for debugging

add einops + vector_quantize_pytorch for model loading

run training as detached subprocess to survive Gradio session timeout

use bfloat16 precision for training to halve RAM usage

stop ace-server during training to free RAM, restart after, add log visibility

pre-download training checkpoints at build time

pin diffusers==0.30.3 for torch 2.4.x compat

pin torchaudio==2.4.0 (before torchcodec default backend)

force uninstall torchcodec (torchaudio dependency, broken on Ubuntu 22.04)

monkey-patch torchaudio.load to use soundfile backend

remove torchcodec (needs FFmpeg 5.x), keep soundfile backend

add torchcodec + ffmpeg for torchaudio audio loading