Spaces:

WeReCooking
/

ACE-Step-CPU

Running

App Files Files Community

ACE-Step-CPU

Commit History

Apr 25

07:08
add _is_space flag, block inference during training, understand clone fix

Nekochu

3c15b8b
06:40
fix understand_audio: clone tensors for inference mode, working on GPU (52s)

Nekochu

4b2f4ad
06:37
add understand_audio (LM reverse), demucs-infer fix, commit refs, dtype fixes

Nekochu

6bfdc38
05:14
major update: PyTorch inference, Gradio 6, session isolation, /understand captioning⁠

Nekochu

ff239f5
05:01
truncate long files to fit cap, show which files truncated/skipped

Nekochu

bc97006
04:59
accept files until total audio cap reached, skip rest with warning

Nekochu

956dc8c
04:58
replace per-file cap with total audio cap (30 min max)

Nekochu

1549b91
00:25
add LoRA download button after training (gr.File output, like rvc-beatrice)

Nekochu

2d3c27c

Apr 24

23:36
remove ace-server understand proxy, captioning stays librosa + txt sidecars

Nekochu

5b7a56f
23:28
add /understand API endpoint for GGUF audio captioning

Nekochu

a4457c3
23:24
SDPA first on Blackwell, FA2 only for Ampere/Hopper, txt caption support

Nekochu

04ccf32
23:17
add GPU/CUDA auto-detect, mixed precision, flash_attn, txt caption parser

Nekochu

917e4ed
23:05
update defaults: LR 3e-4, rank 32, alpha 2x rank (per Side-Step author)

Nekochu

04c031f
20:42
save only on cancel/finish, max epochs 1000, default 3

Nekochu

d42aa91
19:46
add mid/sas analysis modes (Demucs + ensemble), auto-select by dataset size

Nekochu

b38d0b1
19:33
add auto-captioning (BPM/key/signature via librosa), add librosa+mutagen deps

Nekochu

1d42836
19:24
switch training to standard turbo (11s/epoch), auto-select standard GGUF for LoRA inference

Nekochu

c0f2a13
17:55
add XL variant mappings to _VARIANT_DIR

Nekochu

3c5ce9c
08:22
fix: train on XL turbo (matches XL GGUF for inference), add XL checkpoint download

Nekochu

372f08e
08:00
add full README with API docs, MCP, CLI, architecture

Nekochu

9d2d424
07:49
allow custom LoRA values in dropdown (API clients)

Nekochu

2bd2612
07:25
log ace-server restart, show output for adapter debugging

Nekochu

d2ae079
06:54
fix: forward adapter to synth request, default LM to 1.7B

Nekochu

b23b6b8
05:56
fix: adapter saved to clean dir, LM dropdown no 'Default', on-demand download

Nekochu

e62602f
03:46
copy train_engine.py into Docker image

Nekochu

5fe3c53
02:44
Side-Step training engine, tested locally on CPU⁠

Nekochu

a07b39d
00:55
switch back to 1.7B LM (fastest at 269s, 0.6B was 936s)

Nekochu

5e95353
00:13
swap LM to 0.6B Q8_0 for speed test

Nekochu

88b9223

Apr 23

23:48
swap LM 4B->1.7B Q8_0 for faster CPU inference

Nekochu

b14d3e8
23:39
add LoRA adapter dropdown to inference UI

Nekochu

5c2e4e7
23:10
default mp3, remove format selector, increase LM timeout to 900s

Nekochu

882ed5c
22:25
remove accelerate (causes meta tensors), clean up patches

Nekochu

153f929

Apr 22

19:54
fix meta tensor crash: force low_cpu_mem_usage=False and float32 for CPU

Nekochu

9ed24c7
19:09
disable flash_sdp on CPU, force attn_implementation=sdpa for training

Nekochu

13f9406
18:37
add granular logging + full stderr to diagnose preprocessing hang

Nekochu

6cee8bd
18:24
fix: use float32 not bfloat16 for CPU training (bf16 deadlocks on CPU)

Nekochu

560b5e0
15:36
redirect training subprocess stderr to log file for debugging

Nekochu

e69e9ec
14:50
add einops + vector_quantize_pytorch for model loading

Nekochu

88ca206
14:04
run training as detached subprocess to survive Gradio session timeout

Nekochu

a4a86a8
13:28
use bfloat16 precision for training to halve RAM usage

Nekochu

a0e1f4c
12:52
stop ace-server during training to free RAM, restart after, add log visibility

Nekochu

c2cb0b9
12:32
pre-download training checkpoints at build time

Nekochu

c37e80e
12:16
pin diffusers==0.30.3 for torch 2.4.x compat

Nekochu

f59d542
12:07
pin torchaudio==2.4.0 (before torchcodec default backend)

Nekochu

148cd6b
12:01
force uninstall torchcodec (torchaudio dependency, broken on Ubuntu 22.04)

Nekochu

9822bee
11:54
monkey-patch torchaudio.load to use soundfile backend

Nekochu

afccbc0
11:45
remove torchcodec (needs FFmpeg 5.x), keep soundfile backend

Nekochu

63d7cb8
09:42
add torchcodec + ffmpeg for torchaudio audio loading

Nekochu

f1f383b
07:13
add soundfile + libsndfile1 for torchaudio backend

Nekochu

dd7a793

Apr 21

16:24
add all training deps: diffusers lightning numpy tensorboard

Nekochu

eb1b926

Commit History

Apr 25

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)

Apr 24

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

Apr 23

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

Apr 22

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

add soundfile + libsndfile1 for torchaudio backend

Apr 21

add all training deps: diffusers lightning numpy tensorboard