Document the mixed-precision q8dit builds

README.md

@@ -1,45 +1,71 @@
----
-license: mit
-base_model:
-  - ACE-Step/acestep-v15-xl-turbo
-  - ACE-Step/acestep-v15-xl-sft
-pipeline_tag: text-to-audio
-tags:
-  - gguf
-  - ace-step
-  - text-to-music
-  - audio.cpp
----
-
 # audio.cpp GGUF
 
 GGUF conversions of music and speech models for
 [audio.cpp](https://github.com/0xShug0/audio.cpp), the ggml-based native audio
 inference framework.
 
-## ACE-Step 1.5 XL (bf16)
+## ACE-Step 1.5 XL
 
 | File | Size | Package | Upstream |
 |---|---:|---|---|
-| `ACE-Step1.5-GGUF/xl-turbo/ace-step-1.5-xl-turbo-bf16.gguf` | 14.2 GB | `ace_step_xl_turbo_bf16` | [acestep-v15-xl-turbo](https://huggingface.co/ACE-Step/acestep-v15-xl-turbo) |
-| `ACE-Step1.5-GGUF/xl-sft/ace-step-1.5-xl-sft-bf16.gguf` | 14.2 GB | `ace_step_xl_sft_bf16` | [acestep-v15-xl-sft](https://huggingface.co/ACE-Step/acestep-v15-xl-sft) |
+| `ACE-Step1.5-GGUF/xl-turbo/ace-step-1.5-xl-turbo-bf16.gguf` | 14.2 GiB | `ace_step_xl_turbo_bf16` | [acestep-v15-xl-turbo](https://huggingface.co/ACE-Step/acestep-v15-xl-turbo) |
+| `ACE-Step1.5-GGUF/xl-sft/ace-step-1.5-xl-sft-bf16.gguf` | 14.2 GiB | `ace_step_xl_sft_bf16` | [acestep-v15-xl-sft](https://huggingface.co/ACE-Step/acestep-v15-xl-sft) |
+| `ACE-Step1.5-GGUF/xl-turbo-q8dit/ace-step-1.5-xl-turbo-q8dit.gguf` | 9.97 GiB | *(no package row yet)* | as above |
+| `ACE-Step1.5-GGUF/xl-sft-q8dit/ace-step-1.5-xl-sft-q8dit.gguf` | 9.97 GiB | *(no package row yet)* | as above |
 
 The XL DiT is the larger ACE-Step 1.5 transformer: 32 layers of 2560 against
 turbo's 24 of 2048, with 32 attention heads of 128. Each file is self-contained
 the way audio.cpp's other ACE-Step GGUFs are — XL DiT, planner LM, text encoder
-and VAE in one file — so neither needs another download.
+and VAE in one file — so none of them needs another download.
 
 XL Turbo is guidance-distilled and ignores `guidance_scale`; XL SFT takes the
-CFG path. Both were converted from the float32 upstream snapshots with
-`audiocpp_gguf --type bf16`.
+CFG path. Both were converted from the float32 upstream snapshots.
+
+## The `q8dit` builds
+
+`ace_step` is graded `No (planner sampling can fail)` for q8_0 in audio.cpp's
+[gguf.md](https://github.com/0xShug0/audio.cpp/blob/main/docs/gguf.md#support-and-test-status),
+and that grade is about the **planner LM**, not the DiT. The `q8dit` files
+quantise the DiT to q8_0 and keep the planner LM, text encoder and VAE at bf16:
+
+```bash
+audiocpp_gguf ... \
+  --keep-type "lm_weights*=bf16" \
+  --keep-type "text_encoder_weights*=bf16" \
+  --keep-type "vae_weights*=bf16" \
+  --keep-type "dit_xl_turbo_silence_latent*=bf16" \
+  --type q8_0 --output ace-step-1.5-xl-turbo-q8dit.gguf
+```
+
+Measured on an RTX 5090 (CUDA), against the bf16 file at the same prompt and
+seed:
 
-### Install
+| | size | 20 s of audio | RTF | waveform correlation vs bf16 |
+|---|---:|---:|---:|---:|
+| XL Turbo bf16 | 14.2 GiB | 14.2 s | 0.71 | reference |
+| XL Turbo q8dit | 9.97 GiB | 9.3 s | 0.46 | **0.989** |
+| XL Turbo, everything q8_0 | 8.36 GiB | 8.2 s | 0.41 | 0.094 |
+
+Every build is deterministic at a fixed seed (run-to-run correlation 1.0000), so
+the last column isolates what quantisation changes. On a sung 40 s take the
+q8dit build scores **0.997** for XL Turbo and **0.999** for XL SFT, and an ASR
+pass over both returns the same lyric line. A fully quantised build returns a
+different song from the same seed — it plays, it just isn't the same model.
+
+Validated on three prompts on one machine. Treat these as a preview: they are
+not covered by audio.cpp's path-test matrix, and there is no package row for
+them yet, so install them by hand.
+
+## Install
 
 ```bash
 python tools/model_manager_v2.py install ace_step_xl_turbo_bf16
 ```
 
-### Run
+The `q8dit` files have no package id yet — download one and point `--model` at
+its directory.
+
+## Run
 
 The DiT variant is a load option, so name it explicitly:
 
@@ -50,13 +76,18 @@ audiocpp_cli --task gen --family ace_step --model models/ACE-Step1.5-GGUF/xl-tur
   --load-option ace_step.dit_model_path=acestep-v15-xl-turbo --out song.wav
 ```
 
-Measured on an RTX 5090 (CUDA), 20 s of audio in 15 s including model load,
-either variant.
+For sung vocals, pass the lyrics *and* a language, and give the DiT more steps
+than the turbo default:
+
+```bash
+  --lyrics "[verse]
+City lights are falling slow" --language en --num-inference-steps 24
+```
 
 XL support landed in audio.cpp via
 [PR #235](https://github.com/0xShug0/audio.cpp/pull/235); see
 `docs/models/ace_step.md` there for the full option reference and for rebuilding
-this file yourself.
+these files yourself.
 
 Weights are MIT-licensed by ACE-Step; this repository only redistributes them in
 a different container format.