ACE-Step music on moneta

Local ACE-Step 1.5 on Tesla T4 #1 for episode scores. Currently DiT-only (LM planner disabled) after T4 think/sample failures. ComfyUI stays on T4 #0.

T4 #1ACE-Step GPU

:7865Gradio

DiT-onlyLM off

~10 minMax cue length

Quick open

https://music.gliffy.tv/ · LAN http://10.0.0.10:7865/ · Grafana ACE-Step Music · Install /dock/ace-step/

1. Architecture (two models)

Caption → 5Hz LM planner → DiT musician → WAV

The DiT makes audio. The 5Hz LM is an optional planner (thinking / caption rewrite / simple-mode / audio→caption). Official guidance: if you already know the cue, skip thinking and feed a clear caption straight to DiT.

LM step	What it does	On moneta now
Thinking / CoT	Infers BPM, key, duration, structure	Off — LM not loaded
Caption rewrite / format	Expands short prompts into rich style text	Off — write captions yourself
Simple / sample-from-query	One box → caption + lyrics + metas	Passthrough only (prompt → Caption; no LM expand)
Audio → caption	Rough describe-this-track for covers	Off
DiT text2music	Caption (+ optional lyrics) → stereo audio	On — this is the path to use

What you miss with DiT-only

Convenience only: auto-expanding vague prompts, auto BPM/key guesses, one-box simple mode, stronger lyric planning. You do not lose the ability to generate music — put a solid caption in the caption field and generate.

2. Why LM failed on the T4

2026-09-10 run: “Warm analog synth pads…” with think/sample enabled. The 0.6B LM ran ~5 minutes on T4 #1 (nano-vllm float16 — T4 has no bf16). Prefill ~1 tok/s. Output was garbage (only ! characters). Log still said “Sample created” with empty caption / [Instrumental] only. No DiT generate_music logs, no WAV. The Generating: 0% bars were the language model, not audio diffusion.

T4 = 15 GB, compute 7.5 (Turing) — awkward for nano-vllm + constrained 0.6B
DiT already holds ~5–8 GB; LM shares the same card
DiT turbo + INT8 + CPU offload + float32 is the intended T4 music path
Bring LM back on a single 24 GB+ CUDA GPU, or after a T4-specific LM backend retune

3. Dual T4 ≠ 32 GB

Moneta has 2×15 GB, not one 32 GB pool. ACE-Step inference is single-GPU (tensor_parallel_size=1). Multi-GPU in upstream docs is mainly for training. Pick cuda:0 or cuda:1.

T4 #0 · ComfyUI :8189 + T4 #1 · ACE-Step :7865

4. Current config

Piece	Value
Host	moneta 10.0.0.10
GPU	`CUDA_VISIBLE_DEVICES=1`
DiT	`acestep-v15-turbo` + INT8 + CPU offload
LM	disabled (`ACESTEP_INIT_LLM=false`)
Dtype	`ACESTEP_DTYPE=float32` (lyric NaN workaround if LM returns later)
Tier	tier5 · ~10 min DiT-only
Batch	1
Turbo knobs	8 steps · guidance 1.0

5. How to generate (DiT-only)

#	Step
1	Open music.gliffy.tv (or :7865) — hard-refresh if the UI was open before the restart
2	Custom mode: paste a full caption (style, instruments, mood, BPM). Or stay in Simple → Create Sample — with LM off that just copies your prompt into Caption (no expansion)
3	Lyrics empty or `[Instrumental]` for beds · set duration (e.g. 30–60 s) · BPM optional
4	Keep Think off (locked). Skip Format / Auto Score — those still need LM
5	Hit Generate. Progress should be DiT diffusion, not a multi-minute LM think
6	Download from the Gradio player, or copy from disk (below)

Caption examples

Warm analog synth pads, slow pulse, hopeful, no drums, cinematic underscore, 70 BPM
Percussive hybrid trailer hits, short stingers, dry room, tension riser
Small acoustic guitar + soft piano, intimate, diegetic campfire feel

6. Where files go / how to access

Path	Access
`/dock/ace-step/src/gradio_outputs/`	On-disk WAVs (often `batch_*/…wav` + `.json`)
Gradio player download	Via music.gliffy.tv after a successful gen
No static file browser on that host	Past takes: SSH / `scp` the folder above

scp 'gliffy@10.0.0.10:/dock/ace-step/src/gradio_outputs/batch_*/*' .

7. Start / stop / logs

Action	Command
Status	`systemctl --user status ace-step`
Restart	`systemctl --user restart ace-step`
Stop	`systemctl --user stop ace-step`
Logs	`journalctl --user -u ace-step -f`
Config	`/dock/ace-step/src/.env` · `/dock/ace-step/start.sh`

8. Grafana

Exporter :9106 · job acestep · dashboard ACE-Step Music (moneta). Metrics: acestep_up, acestep_systemd_active, probe latency, output file count/bytes. T4 #1 util/VRAM also on Moneta Overview.

9. Troubleshooting

Symptom	Likely cause
“5Hz LM not initialized”	Old UI path (Format / Score) — hard-refresh; Create Sample now passthroughs; Generate with caption works
Long “Generating” then no WAV; logs show `!!!!`	LM think/sample (disabled now) — use caption → DiT
:7865 refused	`systemctl --user status ace-step`; wait for model load
OOM / noise	Shorter duration; batch 1; turbo 8 steps / guidance 1.0
music.gliffy.tv down, LAN OK	`sudo systemctl reload caddy`
Want LM again	Set `ACESTEP_INIT_LLM=true` in `.env` and restart — expect T4 flakiness until a better GPU/backend

10. Related

Updated 2026-09-10 · DiT-only after LM planner failure analysis on T4 #1.