Converting bf16 GGUF to q8_0, how to and hardcodes

[Dev0Lab1]

I am trying to convert ACE-Step1.5-XL-Turbo-bf16.gguf (found a link to a repo by a community contributor in the issues) to lower quants like q8_0 and have had zero luck so far. I am using audiocpp_gguf.exe. something like:

audiocpp_gguf.exe --input ACE-Step1.5-XL-Turbo-bf16.gguf --type q8_0 --output test.gguf

which gives:
error: no GGUF sidecars were found ...

How to do the conversion?

Also a side but relevant second question: in the UI in the model tab, are all those shown models hard coded? For example, ACE-Step shows 6 blocks.

[IIIIIllllIIIIIlllll]

I think:

1. download the supporting files (tokenizer, config, etc.) other than the GGUF file from that community repository into a directory.
2. then add --root and rerun the original command.
3. do not use --no-sidecars — that would treat it as a pure tensor container, and at runtime it will most likely fail to load due to missing tokenizer/config.

[Dev0Lab1]

@IIIIIllllIIIIIlllll , the community repo referred to seems has only the model gguf file nothing more. And if we go even deeper to the original repo by ACE-Step team there are only the following files there:

1. config.json
2. configuration_acestep_v15.py
3. model-00001-of-00004.safetensors
4. model-00002-of-00004.safetensors
5. model-00003-of-00004.safetensors
6. model-00004-of-00004.safetensors
7. model.safetensors.index.json
8. modeling_acestep_v15_xl_turbo.py
9. silence_latent.pt

Before downloading 20GB can you kindly verify that these are all we need for this conversion?

[0xShug0]

@Dev0Lab1 The instructions from @CaptainArni

https://github.com/0xShug0/audio.cpp/blob/main/docs/models/ace_step.md#xl-variants

[0xShug0]

are all those shown models hard coded

Yes

[Dev0Lab1]

I gave up!
And to save headache for any future reader / developer / enthusiast / etc.:
The instructions link are obviously AI/Claude written, lack tons of things, and so be ready for long reads of zero true content.

Building an XL GGUF yourself needs the other variants' safetensors on hand, because audiocpp_gguf checks the conversion against the spec's required namespaces; exclude them from the output:

audiocpp_gguf --root models/Ace-Step1.5 --family ace_step
--input dit_turbo_weights=models/Ace-Step1.5/acestep-v15-turbo/model.safetensors
--input dit_turbo_silence_latent=models/Ace-Step1.5/acestep-v15-turbo/silence_latent.safetensors
--input dit_base_weights=models/Ace-Step1.5/acestep-v15-base/model.safetensors
--input dit_base_silence_latent=models/Ace-Step1.5/acestep-v15-base/silence_latent.safetensors
--input dit_xl_turbo_weights=models/Ace-Step1.5/acestep-v15-xl-turbo/model.safetensors.index.json
--input dit_xl_turbo_silence_latent=models/Ace-Step1.5/acestep-v15-xl-turbo/silence_latent.safetensors
--input lm_weights=models/Ace-Step1.5/acestep-5Hz-lm-1.7B/model.safetensors
--input text_encoder_weights=models/Ace-Step1.5/Qwen3-Embedding-0.6B/model.safetensors
--input vae_weights=models/Ace-Step1.5/vae/diffusion_pytorch_model.safetensors
--exclude-prefix dit_turbo_ --exclude-prefix dit_base_
--type bf16 --output ace-step-1.5-xl-turbo-bf16.gguf

I checked all of those after searching carefully and the heavy downloads are top to bottom 5 + 5 + 20 + 4 + 1 = 35GB which obviously the 5 + 5 serve nothing but as namespace checking! why such? This could be done by a simple text file of a kilobyte serving exact the same purpose: namespace checking. Even the metadata section of a model can easily serve this purpose.

Moreover, noting that the existing ace-step-1.5-xl-turbo-bf16.gguf has all sidecars or other cars in it, why any gguf conversion needs to use the original and bloated set of models, remains a question. The ace-step-1.5-xl-turbo-bf16.gguf has weights and metadata all in it; to convert it to lower quants it should require nothing else.

Finally, there is already another and specialized cpp implementation of ACE-Step acestep.cpp which in my experience is much lighter in both weights to download and execution; I get rtf=0.3 for ace-step 1.5 xl sft+turbo with it.

So I gave up on converting it for audiocpp at this time.

[0xShug0]

@CaptainArni Could you weigh in here?

[CaptainArni]

@Dev0Lab1 I built the XL GGUFs and wrote that doc section. You were right about both things, so here is the fix rather than an argument: #464, plus two files you can download instead of converting.

Why your command failed. audiocpp_gguf takes the conversion's namespaces from the namespace= labels on the command line, never from the tensor names in the file. The XL GGUF holds five (dit_xl_turbo_weights, lm_weights, text_encoder_weights, vae_weights, dit_xl_turbo_silence_latent — --inspect lists them), but --input file.gguf offered all of it as one unnamed namespace. With no --family, the catalog search then matched sense_asr — the only spec accepting a single unnamed namespace with no required sidecars — and the run died later on sidecars, which is the error you saw. The quiet version is worse: re-converting the Kroko GGUF with any small text file beside it succeeds and stamps model_spec_family=sense_asr into the output.

The PR makes a GGUF input's own namespaces, spec and embedded sidecars count, so this works now:

audiocpp_gguf --input ace-step-1.5-xl-turbo-bf16.gguf --type q8_0 \
  --keep-type "lm_weights*=bf16" --output ace-step-1.5-xl-turbo-q8dit.gguf

You never needed the turbo and base weights. The namespace check only reads the labels, and --exclude-prefix drops those tensors before any data is touched — so a 76-byte placeholder file works and the build costs 24 GB of downloads, not 33. Only config.json is genuinely needed from those two snapshots. That was your main complaint and it was correct; the doc is fixed in the PR.

python -c "import json,struct; h=json.dumps({'x':{'dtype':'F32','shape':[1,1],'data_offsets':[0,4]}}).encode(); h+=b' '*((8-len(h)%8)%8); open('placeholder.safetensors','wb').write(struct.pack('<Q',len(h))+h+bytes(4))"

q8_0 is the wrong target, but the DiT alone is not. ace_step's No (planner sampling can fail) grade is about the planner LM, not the DiT. Quantise the DiT and keep the planner at bf16:

build	size	20 s of audio	correlation vs bf16, same seed
bf16	14.2 GiB	14.2 s	reference
q8_0 DiT + bf16 planner	9.97 GiB	9.3 s	0.989
q8_0 everything	8.36 GiB	8.2 s	0.094

Every build is deterministic at a fixed seed, so that column is only quantisation. On a sung 40 s take the mixed build scores 0.997 (XL Turbo) and 0.999 (XL SFT), with the same lyric line coming back out of ASR; the fully quantised one sings a different song from the same seed.

Both are uploaded at CaptainArni/audio.cpp-gguf — xl-turbo-q8dit and xl-sft-q8dit, 9.97 GiB each. #464 adds ace_step_xl_turbo_q8dit / ace_step_xl_sft_q8dit rows; until it lands, download one and point --model at its directory. Validated on three prompts on one RTX 5090, so: a preview, not a blessed release.

On the 6 blocks. They are the GGUF packages in model_specs/ace_step.json. Hard-coded in that the list lives in spec JSON rather than being discovered — but adding a row is a JSON edit, not a code change, which is how the XL rows got there. The DiT variant itself is a load option (ace_step.dit_model_path), so one package can carry several.

On acestep.cpp — no argument from me. A single-model implementation should be leaner than a framework carrying many families. If rtf=0.3 is what you need today, use it.

Sorry for the download you nearly ate; that one was on my doc.

[0xShug0]

@CaptainArni Really appreciate the quick response!

@Dev0Lab1 Feedback on performance optimizations is always appreciated. As you can see, there are already many performance-focused PRs, and our contributors keep pushing the performance further. We also know that some optimization patterns haven’t been backported to models from earlier released models, due to limited bandwidth. But please make sure any performance comparisons are conducted under the same, fair settings; otherwise, we may end up chasing the wrong targets. A single RTF number without any context is misleading.