Incorrect audio codes

[hasaranga]

Hi,

I built latest acstep.cpp on my windows pc. Unfortunately the ace-lm.exe is generating invalid audio codes. The generated audio is messed up.

Please check the attached program output and generated json file. Earlier I used ace-step with python and it was working fine. If i remove the audio codes from generated json file, then ace-synth.exe is giving me nice output. Why the ace-lm.exe is generating invalid audio codes? (I also tried both acestep-5Hz-lm-0.6B-Q8_0.gguf and acestep-5Hz-lm-1.7B-Q8_0.gguf. but the results were same.)

simple.json

{
"caption": "pop rock with female vocal",
"lyrics": "",
"vocal_language": "en",
"inference_steps": 8,
"guidance_scale": 1.0,
"shift": 3.0,
"duration": 180
}

generated simple0.json

{
"caption": "An energetic pop-rock track driven by a crunchy, overdriven electric guitar riff and a punchy, straightforward drum beat. A powerful female lead vocal delivers the melody with a clear, confident tone, soaring into a belted, anthemic chorus. The arrangement is built on a tight rhythm section of bass and drums, with layered guitars providing both rhythmic drive and melodic hooks. The song follows a classic verse-chorus structure, culminating in a dynamic bridge that builds tension before a final, powerful chorus and a guitar-led outro that ends with a definitive crash.",
"lyrics": "[Intro - Guitar Riff]\n\n[Verse 1]\nShe's got a smile that lights up the night\nEyes that shine like stars so bright\nHeart racing fast, can't control the beat\nEvery time she walks by my life feels sweet\n\n[Pre-Chorus]\nShe sends me gifts filled with delight\nWhispers my name in the soft moonlight\nCan't get enough of her laugh so loud\nIn her presence I feel so proud\n\n[Chorus]\nCrazy love, it's taking over me\nLike a storm rolling wild and free\nEvery moment a rush, every touch a spark\nCrazy love, set my heart on fire\n\n[Guitar Riff]\n\n[Verse 2]\nShe dances around my mind all day\nSings a tune and she takes me away\nLips that spell magic I can't deny\nGet hypnotized just by her eye\n\n[Pre-Chorus]\nTime stops when she's near\nEvery seconds worth the fear\nHer voice a melody, soft and clear\nIn her arms the world disappears\n\n[Chorus]\nCrazy love, it's taking over me\nLike a storm rolling wild and free\nEvery moment a rush, every touch a spark\nCrazy love, set my heart on fire\n\n[Guitar Solo]\n\n[Outro]\n[Final chord sustains and fades]",
"bpm": 83,
"duration": 180.0,
"keyscale": "G major",
"timesignature": "4",
"vocal_language": "en",
"seed": 7592087709092679405,
"batch_size": 1,
"lm_temperature": 0.85,
"lm_cfg_scale": 2.0,
"lm_top_p": 0.90,
"lm_top_k": 0,
"lm_negative_prompt": "",
"use_cot_caption": true,
"inference_steps": 8,
"guidance_scale": 1.0,
"shift": 3.0,
"audio_cover_strength": 0.50,
"repainting_start": -1.0,
"repainting_end": -1.0,
"audio_codes": "44281,44281,23290,10106,10683,12571,12571,17498,43026,46649,46649,46714,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49467,49403,36603,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49404,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,36603,49403,49403,49403,49339,36539,36603,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,49403,36603,49403,49403,49403,49459,49467,49403,49403,49403,49403,49459,49403,49459,49403,49403,49403,49403,36603,49403,49459,49403,49403,49403,49339,49403,49395,49339,49403,49403,49395,49331,49403,48883,49403,49403,49403,49395,36603,49403,49403,49403,49403,49403,49395,49459,49395,49395,49403,49395,49395,36603,36595,49403,49395,49331,49395,49395,49403,49395,49395,49403,49395,49403,49395,36595,49395,49395,49395,49395,49395,49395,49395,49459,49395,49459,49459,49395,49459,49395,49395,49395,49331,49395,49395,49395,49459,49403,49395,36595,49395,49395,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49451,49387,49387,49387,49387,49323,49387,49386,49387,49387,49387,36587,49387,36587,49387,49387,49387,49387,36587,49387,49387,49387,49387,36587,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49323,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49387,49898,49898,49899,49899,49899,49899,49963,49963,49963,49963,47403,49963,49963,49963,49963,49963,49963,49963,49963,49963,49963,49963,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50027,50540,47467,25446"
}

program output

C:\acecpp\build>ace-lm.exe --request simple.json --lm acestep-5Hz-lm-4B-Q8_0.gguf
[Request] Parsed simple.json
[Request] seed=-1 batch=1
[Request] caption: pop rock with female vocal
[Request] lyrics: 0 bytes
[Request] bpm=0 dur=180 key= ts= lang=en
[Request] lm: temp=0.85 cfg=2.0 top_p=0.90 top_k=0
[Request] dit: steps=8 guidance=1.0 shift=3.0
[Request] audio_codes: (none)
[BPE] Loaded from GGUF: 151643 vocab, 151387 merges
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 2060 SUPER, compute capability 7.5, VMM: yes
[Load] LM backend: CUDA0 (CPU threads: 2)
[LM] FP16 clamp enabled (cc=750)
[GGUF] acestep-5Hz-lm-4B-Q8_0.gguf: 398 tensors, data at offset 5346304
[LM-Config] 36L, H=2560, V=217204, Nh=32, Nkv=8, D=128, tied=1
[Qwen3] Attn: Q+K+V fused
[Qwen3] MLP: gate+up fused
[WeightCtx] Loaded 398 tensors, 4245.7 MB into backend
[LM-KV] Allocated 2 sets x 36 layers (4D batched), 2304.0 MB
[FSM] Prefix trees: bpm=301, dur=601, key=108, lang=55, tsig=5 nodes
[Ace-LM] Loaded: vocab=217204, max_seq=8192, max_batch=1, kv_sets=2
[Fill] lyrics=generate metas=fill gaps | 36 tokens, CFG: 1.00, N=1
[Phase1] Prefill 177ms, 36 tokens, N=1, CFG=1.00
[Phase1] Step 100, 1 active, 43.2 tok/s
[Phase1] Step 200, 1 active, 43.9 tok/s
[Phase1] Step 300, 1 active, 45.4 tok/s
[Phase1] Step 400, 1 active, 46.4 tok/s
[Phase1] Decode 9411ms
[Phase1 Batch0] seed=7592087709092679405, 438 tokens
[Fill Batch0] seed=7592087709092679405:
bpm:83
caption: An energetic pop-rock track driven by a crunchy, overdriven electric guitar
riff and a punchy, straightforward drum beat. A powerful female lead vocal delivers
the melody with a clear, confident tone, soaring into a belted, anthemic chorus.
The arrangement is built on a tight rhythm section of bass and drums, with layered
guitars providing both rhythmic drive and melodic hooks. The song follows a classic
verse-chorus structure, culminating in a dynamic bridge that builds tension before
a final, powerful chorus and a guitar-led outro that ends with a definitive crash.
duration: 157
keyscale:G major
language:en
timesignature:4

Lyric

[Intro - Guitar Riff]

[Verse 1]
She's got a smile that lights up the night
Eyes that shine like stars so bright
Heart racing fast, can't control the beat
Every time she walks by my life feels sweet

[Pre-Chorus]
She sends me gifts filled with delight
Whispers my name in the soft moonlight
Can't get enough of her laugh so loud
In her presence I feel so proud

[Chorus]
Crazy love, it's taking over me
Like a storm rolling wild and free
Every moment a rush, every touch a spark
Crazy love, set my heart on fire

[Guitar Riff]

[Verse 2]
She dances around my mind all day
Sings a tune and she takes me away
Lips that spell magic I can't deny
Get hypnotized just by her eye

[Pre-Chorus]
Time stops when she's near
Every seconds worth the fear
Her voice a melody, soft and clear
In her arms the world disappears

[Chorus]
Crazy love, it's taking over me
Like a storm rolling wild and free
Every moment a rush, every touch a spark
Crazy love, set my heart on fire

[Guitar Solo]

[Outro]
[Final chord sustains and fades]

[Phase2] N=1, CoT[0]:
bpm: 83
caption: An energetic pop-rock track driven by a crunchy, overdriven electric guitar
riff and a punchy, straightforward drum beat. A powerful female lead vocal delivers
the melody with a clear, confident tone, soaring into a belted, anthemic chorus.
The arrangement is built on a tight rhythm section of bass and drums, with layered
guitars providing both rhythmic drive and melodic hooks. The song follows a classic
verse-chorus structure, culminating in a dynamic bridge that builds tension before
a final, powerful chorus and a guitar-led outro that ends with a definitive crash.
duration: 180
keyscale: G major
language: en
timesignature: 4
[Phase2] max_tokens: 1000, CFG: 2.00, seeds: 7592087709092679405..7592087709092679405
[Phase2] Prefill 440ms (shared, 1 cond + 1 uncond)
[Decode] Step 50, 1 active, 51 total codes, 50.7 tok/s
[Decode] Step 100, 1 active, 101 total codes, 50.7 tok/s
[Decode] Step 150, 1 active, 151 total codes, 50.6 tok/s
[Decode] Step 200, 1 active, 201 total codes, 50.4 tok/s
[Decode] Step 250, 1 active, 251 total codes, 50.3 tok/s
[Decode] Step 300, 1 active, 301 total codes, 50.3 tok/s
[Decode] Step 350, 1 active, 351 total codes, 50.2 tok/s
[Decode] Step 400, 1 active, 401 total codes, 50.0 tok/s
[Decode] Step 450, 1 active, 451 total codes, 49.9 tok/s
[Decode] Step 500, 1 active, 501 total codes, 49.7 tok/s
[Decode] Step 550, 1 active, 551 total codes, 49.6 tok/s
[Decode] Step 600, 1 active, 601 total codes, 49.4 tok/s
[Decode] Step 650, 1 active, 651 total codes, 49.3 tok/s
[Decode] Step 700, 1 active, 701 total codes, 49.2 tok/s
[Decode] Step 750, 1 active, 751 total codes, 49.1 tok/s
[Decode] Step 800, 1 active, 801 total codes, 49.0 tok/s
[Decode] Step 850, 1 active, 851 total codes, 48.8 tok/s
[Phase2] Decode 18341ms
[Batch 0] seed=7592087709092679405, 894 codes
[Ace-LM] Load 3396 | Total 28397ms | seed=7592087709092679405
[Request] Wrote simple0.json

C:\acecpp\build>ace-synth.exe --request simple0.json --embedding Qwen3-Embedding-0.6B-Q8_0.gguf --dit acestep-v15-turbo-Q8_0.gguf --vae vae-BF16.gguf --wav
ggml_cuda_init: found 1 CUDA devices:
Device 0: NVIDIA GeForce RTX 2060 SUPER, compute capability 7.5, VMM: yes
[Load] DiT backend: CUDA0 (CPU threads: 2)
[Load] Backend init: 37.6 ms
[GGUF] acestep-v15-turbo-Q8_0.gguf: 678 tensors, data at offset 56864
[DiT] Self-attn: Q+K+V fused
[DiT] Cross-attn: Q+K+V fused
[DiT] MLP: gate+up fused
[Load] null_condition_emb found (CFG available)
[WeightCtx] Loaded 478 tensors, 1600.7 MB into backend
[Load] DiT: 24 layers, H=2048, Nh=16/8, D=128
[Load] DiT weight load: 1295.2 ms
[GGUF] acestep-v15-turbo-Q8_0.gguf: 678 tensors, data at offset 56864
[Load] silence_latent: [15000, 64] from GGUF
[GGUF] vae-BF16.gguf: 365 tensors, data at offset 30048
[Load] VAE backend: CUDA0 (shared)
[VAE] Backend: CUDA0, Weight buffer: 161.1 MB
[VAE] Loaded: 5 blocks, upsample=1920x, F32 activations
[Load] VAE weights: 1400.9 ms
[BPE] Loaded from GGUF: 151643 vocab, 151387 merges
[Load] BPE tokenizer: 135.0 ms
[Load] TextEncoder backend: CUDA0 (shared)
[GGUF] Qwen3-Embedding-0.6B-Q8_0.gguf: 310 tensors, data at offset 5337664
[Load] TextEncoder: 28L, H=1024, Nh=16/8
[Qwen3] Attn: Q+K+V fused
[Qwen3] MLP: gate+up fused
[WeightCtx] Loaded 310 tensors, 742.7 MB into backend
[Load] TextEncoder: 578.2 ms
[Load] CondEncoder backend: CUDA0 (shared)
[CondEncoder] FP16 clamp enabled (cc=750)
[GGUF] acestep-v15-turbo-Q8_0.gguf: 678 tensors, data at offset 56864
[Load] LyricEncoder: 8L
[Qwen3] Attn: Q+K+V fused
[Qwen3] MLP: gate+up fused
[Load] TimbreEncoder: 4L
[Qwen3] Attn: Q+K+V fused
[Qwen3] MLP: gate+up fused
[WeightCtx] Loaded 140 tensors, 616.6 MB into backend
[Load] CondEncoder: lyric(8L), timbre(4L), text_proj, null_cond
[Load] ConditionEncoder: 463.6 ms
[GGUF] acestep-v15-turbo-Q8_0.gguf: 678 tensors, data at offset 56864
[WeightCtx] Loaded 30 tensors, 106.5 MB into backend
[Load] Detokenizer: FSQ(6->2048) + 2L encoder(S=5, 2048->64)
[Load] Detokenizer: 83.8 ms
[Ace-Synth] All models loaded, turbo=yes
[Request] Parsed simple0.json
[Request] seed=7592087709092679405 batch=1
[Request] caption: An energetic pop-rock track driven by a crunchy, overdriven ...
[Request] lyrics: 1018 bytes
[Request] bpm=83 dur=180 key=G major ts=4 lang=en
[Request] lm: temp=0.85 cfg=2.0 top_p=0.90 top_k=0
[Request] dit: steps=8 guidance=1.0 shift=3.0
[Request] audio_codes: (present)
[Request 1/1] simple0.json (batch=1)
[Pipeline] 894 audio codes (178.8s @ 5Hz)
[Pipeline] T=4470, S=2235
[Pipeline] seed=7592087709092679405, steps=8, guidance=1.0, shift=3.0, duration=180.0s
[Pipeline] caption: 169 tokens, lyrics: 284 tokens
[Encode] TextEncoder (169 tokens): 103.2 ms
[Encode] Lyric vocab lookup (284 tokens): 0.9 ms
[CondEnc] Lyric sliding mask: 284x284, window=128
[CondEnc] Timbre sliding mask: 750x750, window=128
[Encode] Packed: lyric=284 + timbre=1 + text=169 = 454 tokens
[Encode] ConditionEncoder: 47.9 ms, enc_S=454
[Context] Decoded: 894 codes -> 4470 frames (178.8s @ 25Hz)
[Context] Detokenizer: 848.8 ms
[Context Batch0] Philox noise seed=7592087709092679405, [4470, 64]
[DiT] Starting: T=4470, S=2235, enc_S=454, steps=8, batch=1
[DiT] Batch N=1, T=4470, S=2235, enc_S=454
[DiT] Graph: 1841 nodes
[DiT] Step 1/8 t=1.000
[DiT] Step 2/8 t=0.955
[DiT] Step 3/8 t=0.900
[DiT] Step 4/8 t=0.833
[DiT] Step 5/8 t=0.750
[DiT] Step 6/8 t=0.643
[DiT] Step 7/8 t=0.500
[DiT] Step 8/8 t=0.300
[DiT] Total generation: 3533.3 ms (3533.3 ms/sample)
[VAE] Tiled decode: 35 tiles (chunk=256, overlap=64, stride=128)
[VAE] Graph: 335 nodes, T_latent=192
[VAE] Upsample factor: 1920.00 (expected ~1920)
[VAE] Graph: 335 nodes, T_latent=256
[VAE] Graph: 335 nodes, T_latent=182
[VAE] Tiled decode done: 35 tiles -> T_audio=8582400 (178.80s @ 48kHz)
[VAE Batch0] Decode: 8998.6 ms
[WAV] Wrote simple00.wav: 8582400 samples, 48000 Hz, stereo
[Request 1/1] Done
[Pipeline] All done

C:\acecpp\build>pause
Press any key to continue . . .

[ServeurpersoCom]

It shouldn't :) I'll debug it

[ServeurpersoCom]

Can you try :

1. ace-lm.exe --request simple.json --lm model.gguf --no-fa

2. adding "lm_cfg_scale": 1.0 inside JSON

3. my build https://www.serveurperso.com/temp/acestep.cpp-win64/
   (Until it's produced by CI, I'm still posting the latest version here)

[hasaranga]

--no-fa and lm_cfg_scale doesn't fix the issue. Then I tried your build. Your build is working correctly. It is producing correct audio codes. Why my build is generating bad audio codes? I built it using -DGGML_CUDA=ON flag only. Used CUDA toolkit 13.2. What am i missing?

[ServeurpersoCom]

That's interesting. To build the binaries, I do this:

D:\A>cd acestep.cpp

D:\A\acestep.cpp>type toolchain.cmd
@echo off

"C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Auxiliary\Build\vcvars64.bat"

D:\A\acestep.cpp>toolchain.cmd
**********************************************************************
** Visual Studio 2022 Developer Command Prompt v17.13.6
** Copyright (c) 2022 Microsoft Corporation
**********************************************************************
[vcvarsall.bat] Environment initialized for: 'x64'

D:\A\acestep.cpp>type build.cmd
@echo off

cmake -B build-msvc -DGGML_CPU_ALL_VARIANTS=ON -DGGML_CUDA=ON -DGGML_VULKAN=ON -DGGML_BACKEND_DL=ON
cmake --build build-msvc --config Release -j 24

D:\A\acestep.cpp>build.cmd
-- Selecting Windows SDK version 10.0.22621.0 to target Windows 10.0.19045.
-- Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF
-- CMAKE_SYSTEM_PROCESSOR: AMD64
-- CMAKE_GENERATOR_PLATFORM:
-- GGML_SYSTEM_ARCH: x86
-- Including CPU backend
-- x86 detected
-- Adding CPU backend variant ggml-cpu-x64:
-- x86 detected
-- Adding CPU backend variant ggml-cpu-sse42: /arch:SSE4.2 GGML_SSE42
-- x86 detected
-- Adding CPU backend variant ggml-cpu-sandybridge: /arch:AVX GGML_AVX
-- x86 detected
-- Adding CPU backend variant ggml-cpu-haswell: /arch:AVX2 GGML_AVX2;GGML_FMA;GGML_F16C;__BMI2__;GGML_BMI2
-- x86 detected
-- Adding CPU backend variant ggml-cpu-skylakex: /arch:AVX512 GGML_AVX512;__BMI2__;GGML_BMI2
-- x86 detected
-- Adding CPU backend variant ggml-cpu-cannonlake: /arch:AVX512 GGML_AVX512;__AVX512VBMI__;__BMI2__;GGML_BMI2
-- x86 detected
-- Adding CPU backend variant ggml-cpu-cascadelake: /arch:AVX512 GGML_AVX512;__AVX512VNNI__;GGML_AVX512_VNNI;__BMI2__;GGML_BMI2
-- x86 detected
-- Adding CPU backend variant ggml-cpu-icelake: /arch:AVX512 GGML_AVX512;__AVX512VBMI__;__AVX512VNNI__;GGML_AVX512_VNNI;__BMI2__;GGML_BMI2
-- x86 detected
-- Adding CPU backend variant ggml-cpu-alderlake: /arch:AVX2 GGML_AVX2;GGML_FMA;GGML_F16C;__AVXVNNI__;GGML_AVX_VNNI;__BMI2__;GGML_BMI2
-- CUDA Toolkit found
-- Using CMAKE_CUDA_ARCHITECTURES=75-virtual;80-virtual;86-real;89-real;120a-real;121a-real CMAKE_CUDA_ARCHITECTURES_NATIVE=89-real
-- Including CUDA backend
-- Vulkan found
-- GL_KHR_cooperative_matrix supported by glslc
-- GL_NV_cooperative_matrix2 supported by glslc
-- GL_EXT_integer_dot_product supported by glslc
-- GL_EXT_bfloat16 supported by glslc
-- Including Vulkan backend
-- ggml version: 0.9.7
-- ggml commit:  48952027
-- Configuring done (1.3s)
-- Generating done (0.7s)
-- Build files have been written to: D:/A/acestep.cpp/build-msvc
Version MSBuild 17.13.19+0d9f5a35a pour .NET Framework

  Checking File Globs
  ggml-cpu-sandybridge-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-sandybridge-feats.dir\Release\ggm
  l-cpu-sandybridge-feats.lib
  ggml-cpu-x64-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-x64-feats.dir\Release\ggml-cpu-x64-feats.
  lib
  ggml-cpu-alderlake-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-alderlake-feats.dir\Release\ggml-cp
  u-alderlake-feats.lib
  ggml-cpu-sse42-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-sse42-feats.dir\Release\ggml-cpu-sse42-
  feats.lib
  ggml-cpu-skylakex-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-skylakex-feats.dir\Release\ggml-cpu-
  skylakex-feats.lib
  ggml-cpu-haswell-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-haswell-feats.dir\Release\ggml-cpu-ha
  swell-feats.lib
  ggml-cpu-icelake-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-icelake-feats.dir\Release\ggml-cpu-ic
  elake-feats.lib
  ggml-cpu-cannonlake-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-cannonlake-feats.dir\Release\ggml-
  cpu-cannonlake-feats.lib
  ggml-cpu-cascadelake-feats.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-cpu-cascadelake-feats.dir\Release\ggm
  l-cpu-cascadelake-feats.lib
  ggml-base.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-base.dll
  1>Performing build step for 'vulkan-shaders-gen'
  ggml-cpu-alderlake.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-alderlake.dll
  ggml-cpu-skylakex.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-skylakex.dll
  ggml-cpu-sandybridge.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-sandybridge.dll
  ggml-cpu-sse42.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-sse42.dll
  ggml-cpu-x64.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-x64.dll
  ggml-cpu-cannonlake.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-cannonlake.dll
  ggml-cpu-icelake.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-icelake.dll
  ggml-cpu-haswell.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-haswell.dll
  ggml-cpu-cascadelake.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cpu-cascadelake.dll
  Version MSBuild 17.13.19+0d9f5a35a pour .NET Framework

    vulkan-shaders-gen.vcxproj -> D:\A\acestep.cpp\build-msvc\ggml\src\ggml-vulkan\vulkan-shaders-gen-prefix\src\vulkan
  -shaders-gen-build\Release\vulkan-shaders-gen.exe
  Performing install step for 'vulkan-shaders-gen'
  -- Up-to-date: D:/A/acestep.cpp/build-msvc/Release/./vulkan-shaders-gen.exe
  Completed 'vulkan-shaders-gen'
  ggml-cuda.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-cuda.dll
  ggml-vulkan.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml-vulkan.dll
  httplib.vcxproj -> D:\A\acestep.cpp\build-msvc\vendor\cpp-httplib\Release\httplib.lib
  mp3-codec.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\mp3-codec.exe
  ggml.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ggml.dll
  neural-codec.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\neural-codec.exe
  acestep-core.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\acestep-core.lib
  quantize.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\quantize.exe
  ace-understand.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ace-understand.exe
  ace-lm.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ace-lm.exe
  ace-synth.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ace-synth.exe
  ace-server.vcxproj -> D:\A\acestep.cpp\build-msvc\Release\ace-server.exe

D:\A\acestep.cpp>

[ServeurpersoCom]

I'm on C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1

[ServeurpersoCom]

Can you give me your cmake command ?

So I need to test with CUDA 13.2 to see if I have the same problem.

[hasaranga]

cd build
cmake .. -DGGML_CUDA=ON
cmake --build . --config Release -j%NUMBER_OF_PROCESSORS%

[lmangani]

@hasaranga out of pure curiosity if its not a waste of your time could you test these experimental CI produced builds? If they also fail we'll join the resolution bunch https://github.com/audiohacking/acestep.cpp/releases/tag/v0.0.3

[hasaranga]

That CI produced build is working fine. It is also generating valid audio codes.

So I changed the my build flags to -DGGML_CUDA=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_BACKEND_DL=ON

That did the trick! Now it is generating valid audio codes.

Using -DGGML_CUDA=ON alone causes ace-lm.exe to generate invalid audio codes.

[ServeurpersoCom]

I've also improved the compatibility of ace-lm because an optimization could break things on some backends; feel free to git pull and let me know.

[hasaranga]

Latest version has the same issue. But I'm ok with the current solution. Thank you very much for your support!

[ServeurpersoCom]

This isn't a solution, it's to help me narrow down the fix. I don't have access to ROCm/HIP at the moment.

CFG (classifier-free guidance) requires two forward passes per decode step: one conditional, one unconditional. The reference Python implementation batches them into a single N=2 call for performance. On your GPU this batched path produces corrupt logits.

I can add a --no-batch-cfg debug flag that splits them into two N=1 calls. Same result, half the throughput, but correct on ROCm. Then I need to figure out why the batched N=2 path breaks on gfx1201 specifically.

If someone can give me SSH access to a ROCm machine I can debug and fix this properly over the weekend. Otherwise the split workaround is the best I can ship right now.

[hasaranga]

This isn't a solution, it's to help me narrow down the fix. I don't have access to ROCm/HIP at the moment.

CFG (classifier-free guidance) requires two forward passes per decode step: one conditional, one unconditional. The reference Python implementation batches them into a single N=2 call for performance. On your GPU this batched path produces corrupt logits.

I can add a --no-batch-cfg debug flag that splits them into two N=1 calls. Same result, half the throughput, but correct on ROCm. Then I need to figure out why the batched N=2 path breaks on gfx1201 specifically.

If someone can give me SSH access to a ROCm machine I can debug and fix this properly over the weekend. Otherwise the split workaround is the best I can ship right now.

Why do you need amd gpu? The issue is on nvidia gpu. Also adding cpu build flag fixes the issue.

[ServeurpersoCom]

Why do you need amd gpu?

Or a RTX 2080? Because I don't have this issue on my Blackwell !

I add a "--no-batch-cfg" usefull debug option for ace-lm and ace-server for this case. With this option, ace-lm is just as compatible as the classic qwen3 inference of llama.cpp

Until I find the real bug...
I have a RTX 2080 but I need to install it in a PC to test it, unless someone gives me access to a machine? Anyway, I'm going to fix it :)

[ServeurpersoCom]

Now we have a specific workaround for this problem, since we encounter it on older RTX cards and even on ROCm:

--no-batch-cfg Split CFG into two N=1 forwards

ace-lm: add --no-batch-cfg debug flag

Split CFG conditional and unconditional forward passes into two
separate N=1 calls instead of one batched N=2 call. Same logit
layout, same result, half the throughput. Useful for isolating
batched attention issues on specific GPU backends.

But that halves the performance of LM audio code generation!