Extremely short outputs

[firefox42]

I complied the program on a Ubuntu 24.04.4 system, with ROCm 7.1.1 and a R9700. Generating the request0.json file with ace-lm works fine, i.e.

ace-lm output using simple.sh

![Image](https://github.com/user-attachments/assets/523bd7e6-5131-4658-afd5-46407c4cda67)

[Request] Parsed simple.json
[Request] seed=-1 batch=1
[Request] caption: Upbeat pop rock anthem with driving electric guitars, punchy...
[Request] lyrics: 0 bytes
[Request] bpm=0 dur=0 key= ts= lang=fr
[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 4 ROCm devices:
  Device 0: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 1: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 2: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 3: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
[Load] LM backend: ROCm0 (CPU threads: 8)
[GGUF] ../models/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 | 54 tokens, CFG: 1.00, N=1
[Phase1] Prefill 64ms, 54 tokens, N=1, CFG=1.00
[Phase1] Step 100, 1 active, 65.8 tok/s
[Phase1] Step 200, 1 active, 65.8 tok/s
[Phase1] Step 300, 1 active, 66.2 tok/s
[Phase1] Step 400, 1 active, 66.3 tok/s
[Phase1] Step 500, 1 active, 66.4 tok/s
[Phase1] Step 600, 1 active, 66.4 tok/s
[Phase1] Decode 9714ms
[Phase1 Batch0] seed=660049505072788704, 645 tokens
[Fill Batch0] seed=660049505072788704:
bpm:86
caption: An energetic and upbeat pop-rock track driven by a tight rhythm section of
  punchy drums and a solid bass guitar. Overdriven electric guitars provide a crunchy,
  rhythmic foundation, while a lead guitar plays catchy melodic fills and a spirited,
  melodic solo. The lead male vocal is clear and enthusiastic, delivering the lyrics
  with a straightforward, anthemic quality. The song features a dynamic structure
  with a memorable chorus, a lively guitar solo, and a singalong bridge with 'whoa-oh'
  vocalizations before a powerful final chorus and a clean, definitive ending.
duration: 210
keyscale:G major
language:fr
timesignature:4
</think>

# Lyric
[Intro - Guitar Riff]

[Verse 1]
Les p'tits chats de la ville se promènent tous les soirs
Dans les ruelles sombres sans jamais voir de miroir
Leurs yeux malicieux nous regardent en bas
Tout en attendant l'heure de rentrer chez eux

[Chorus]
Les p'tits chats de la ville s'endorment sans éclat
Leur vie silencieuse dans ce monde si bas
Rêvant des journées où ils pourrait ronronner
Sans devoir se presser ni dormir ni rentrer

[Guitar Riff]

[Verse 2]
Le matin ils sautent pour chercher du pain
Leur queue frémissante attend l'instant divin
Où l'instant divin de retrouver un nid
Pour coucher leur tête en douceur sous l'abri

[Chorus]
Les p'tits chats de la ville s'endorment sans éclat
Leur vie silencieuse dans ce monde si bas
Rêvant des journées où ils pourrait ronronner
Sans devoir se presser ni dormir ni rentrer

[Guitar Solo]

[Bridge]
Les p'tits chats de la ville s'endorment sans éclat
Leur vie silencieuse dans ce monde si bas
Rêvant des journées où ils pourrait ronronner
Sans devoir se presser ni dormir ni rentrer

[Chorus - Key Change]
Les p'tits chats de la ville s'endorment sans éclat
Leur vie silencieuse dans ce monde si bas
Rêvant des journées où ils pourrait ronronner
Sans devoir se presser ni dormir ni rentrer

[Outro]
(Woah-oh-oh-oh)
(Woah-oh-oh-oh)
(Woah-oh-oh-oh)
(Woah-oh-oh-oh)
(Woah-oh-oh-oh)
(Woah-oh-oh-oh)
(Woah-oh-oh-oh)
(Woah-oh-oh-oh)
[Final guitar chord and cymbal crash]

[Phase2] N=1, CoT[0]:
bpm: 86
caption: An energetic and upbeat pop-rock track driven by a tight rhythm section of
  punchy drums and a solid bass guitar. Overdriven electric guitars provide a crunchy,
  rhythmic foundation, while a lead guitar plays catchy melodic fills and a spirited,
  melodic solo. The lead male vocal is clear and enthusiastic, delivering the lyrics
  with a straightforward, anthemic quality. The song features a dynamic structure
  with a memorable chorus, a lively guitar solo, and a singalong bridge with 'whoa-oh'
  vocalizations before a powerful final chorus and a clean, definitive ending.
duration: 210
keyscale: G major
language: fr
timesignature: 4
[Phase2] max_tokens: 1150, CFG: 2.00, seeds: 660049505072788704..660049505072788704
[Phase2] Prefill 244ms (shared, 1 cond + 1 uncond)
[Phase2] Decode 16ms
[Batch 0] seed=660049505072788704, 1 codes
[Ace-LM] Load 2990 | Total 10044ms | seed=660049505072788704
[Request] Wrote simple0.json

But the actual .mp3 file produced by ace-syth is always less than a second long, at only 3840 bytes each time, i.e.

ace-synth output using simple.sh

ggml_cuda_init: found 4 ROCm devices:
  Device 0: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 1: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 2: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 3: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
[Load] DiT backend: ROCm0 (CPU threads: 8)
[Load] Backend init: 42.3 ms
[GGUF] ../models/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: 615.7 ms
[GGUF] ../models/acestep-v15-turbo-Q8_0.gguf: 678 tensors, data at offset 56864
[Load] silence_latent: [15000, 64] from GGUF
[GGUF] ../models/vae-BF16.gguf: 365 tensors, data at offset 30048
[Load] VAE backend: ROCm0 (shared)
[VAE] Backend: ROCm0, Weight buffer: 161.1 MB
[VAE] Loaded: 5 blocks, upsample=1920x, F32 activations
[Load] VAE weights: 854.2 ms
[BPE] Loaded from GGUF: 151643 vocab, 151387 merges
[Load] BPE tokenizer: 43.7 ms
[Load] TextEncoder backend: ROCm0 (shared)
[GGUF] ../models/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: 223.3 ms
[Load] CondEncoder backend: ROCm0 (shared)
[GGUF] ../models/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: 184.8 ms
[GGUF] ../models/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: 33.5 ms
[Ace-Synth] All models loaded, turbo=yes
[Request] Parsed simple0.json
[Request] seed=660049505072788704 batch=1
[Request] caption: An energetic and upbeat pop-rock track driven by a tight rhy...
[Request] lyrics: 1429 bytes
[Request] bpm=86 dur=210 key=G major ts=4 lang=fr
[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] 1 audio codes (0.2s @ 5Hz)
[Pipeline] T=6, S=3
[Pipeline] seed=660049505072788704, steps=8, guidance=1.0, shift=3.0, duration=210.0s
[Pipeline] caption: 169 tokens, lyrics: 495 tokens
[Encode] TextEncoder (169 tokens): 63.6 ms
[Encode] Lyric vocab lookup (495 tokens): 1.6 ms
[CondEnc] Lyric sliding mask: 495x495, window=128
[CondEnc] Timbre sliding mask: 750x750, window=128
[Encode] Packed: lyric=495 + timbre=1 + text=169 = 665 tokens
[Encode] ConditionEncoder: 27.6 ms, enc_S=665
[Context] Decoded: 1 codes -> 5 frames (0.2s @ 25Hz)
[Context] Detokenizer: 34.0 ms
[Context Batch0] Philox noise seed=660049505072788704, [6, 64]
[DiT] Starting: T=6, S=3, enc_S=665, steps=8, batch=1
[DiT] Batch N=1, T=6, S=3, enc_S=665
[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: 136.8 ms (136.8 ms/sample)
[VAE] Graph: 335 nodes, T_latent=6
[VAE] Decoded: T_latent=6 -> T_audio=11520 (0.24s @ 48kHz)
[VAE Batch0] Decode: 756.8 ms
[MP3] Encoding 0.2s @ 128 kbps, 48000 Hz stereo
[MP3] 3840 bytes (12.0:1), 53 ms (4.49x realtime), 1 threads
[MP3] Wrote simple00.mp3
[Request 1/1] Done
[Pipeline] All done

This happens regardless of the prompt or model quantization. Here's the output from this run:
simple00.mp3

[ServeurpersoCom]

[Phase2] Decode 16ms
[Batch 0] seed=660049505072788704, 1 codes

OK, it's the LLM sampling optimization I check

[ServeurpersoCom]

You can try again on the latest master version and provide me with the output log and json :)

ace-lm: copy partial LM head into its own tensor at load time

Instead of building a ggml_view_2d into the quantized embedding
matrix at every decode step (produces corrupt logits on some GPU
backends), copy those rows into their own contiguous tensor once
at load time. Same rows times same hidden state gives the same
logits, just from an aligned buffer that every backend can handle.

[firefox42]

I tried it again with the latest master, but still got the same result. Here's the logs and json:

./simple.sh output

[Request] Parsed simple.json
[Request] seed=-1 batch=1
[Request] caption: Upbeat pop rock anthem with driving electric guitars, punchy...
[Request] lyrics: 0 bytes
[Request] bpm=0 dur=0 key= ts= lang=fr
[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 4 ROCm devices:
  Device 0: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 1: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 2: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 3: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
[Load] LM backend: ROCm0 (CPU threads: 8)
[GGUF] ../models/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
[LM] Partial head: 65559 rows (151645..217204), 170.1 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 | 54 tokens, CFG: 1.00, N=1
[Phase1] Prefill 64ms, 54 tokens, N=1, CFG=1.00
[Phase1] Step 100, 1 active, 66.1 tok/s
[Phase1] Step 200, 1 active, 66.3 tok/s
[Phase1] Step 300, 1 active, 66.7 tok/s
[Phase1] Step 400, 1 active, 66.8 tok/s
[Phase1] Step 500, 1 active, 66.8 tok/s
[Phase1] Step 600, 1 active, 66.8 tok/s
[Phase1] Step 700, 1 active, 66.8 tok/s
[Phase1] Decode 11277ms
[Phase1 Batch0] seed=7917479554720955107, 753 tokens
[Fill Batch0] seed=7917479554720955107:
bpm:86
caption: An upbeat and anthemic pop-rock track driven by a clean, energetic male lead
  vocal. The song opens with layered, wordless vocal harmonies over a bright acoustic
  guitar before the full band kicks in with a driving 4/4 drum beat, a solid bassline,
  and clean electric guitars. The chorus is big and memorable, elevated by layered
  backing vocals that create a powerful, singalong feel. A brief, more reflective
  bridge provides a dynamic shift before launching into a final, high-energy chorus
  and an outro featuring vocal ad-libs and a fade-out with the initial vocal harmony
  motif.
duration: 189
keyscale:G major
language:fr
timesignature:4
</think>

# Lyric
[Intro]
(Ooh-ooh-ooh-ooh)
(Ooh-ooh-ooh-ooh)

[Verse 1]
Réveille-toi, ça fait longtemps
Le soleil danse sur l'horizon
Laisse tomber tous tes chagrins
Ce soir on refait le monde en chantant

[Pre-Chorus]
Oublie les heures, oublie les peines
Viens danser, laisse faire la semaine
On met le feu sous les étoiles
Rien ne nous arrête, on part en cavale

[Chorus]
Allez viens, on va chanter, on va danser, on va s'aimer
Le monde est à nous pour rêver
Viens on va tout lâcher
On va rire, on va crier, on va tout donner sans compter
Parce qu'au fond c'est ça la vie
Viens on va s'amuser

[Post-Chorus]
(Ooh-ooh-ooh-ooh)
(Ooh-ooh-ooh-ooh)

[Verse 2]
Oublie les soucis, laisse-les filer
Ce soir on vit, on veut juste vibrer
Un peu d'amour, un peu de folie
Un grand éclat, une belle mélodie

[Pre-Chorus]
Oublie les heures, oublie les peines
Viens danser, laisse faire la semaine
On met le feu sous les étoiles
Rien ne nous arrête, on part en cavale

[Chorus]
Allez viens, on va chanter, on va danser, on va s'aimer
Le monde est à nous pour rêver
Viens on va tout lâcher
On va rire, on va crier, on va tout donner sans compter
Parce qu'au fond c'est ça la vie
Viens on va s'amuser

[Bridge]
On s'envole, on s'égare
On s'en fout du regard des autres
Tant qu'on a le cœur qui bat
On prend la vie comme une apôtre

[Chorus]
Allez viens, on va chanter, on va danser, on va s'aimer
Le monde est à nous pour rêver
Viens on va tout lâcher
On va rire, on va crier, on va tout donner sans compter
Parce qu'au fond c'est ça la vie
Viens on va s'amuser

[Outro]
(Ooh-ooh-ooh-ooh)
(Ooh-ooh-ooh-ooh)
(Ooh-ooh-ooh-ooh)
(Ooh-ooh-ooh-ooh)
[Instrumental fade out]

[Phase2] N=1, CoT[0]:
bpm: 86
caption: An upbeat and anthemic pop-rock track driven by a clean, energetic male lead
  vocal. The song opens with layered, wordless vocal harmonies over a bright acoustic
  guitar before the full band kicks in with a driving 4/4 drum beat, a solid bassline,
  and clean electric guitars. The chorus is big and memorable, elevated by layered
  backing vocals that create a powerful, singalong feel. A brief, more reflective
  bridge provides a dynamic shift before launching into a final, high-energy chorus
  and an outro featuring vocal ad-libs and a fade-out with the initial vocal harmony
  motif.
duration: 189
keyscale: G major
language: fr
timesignature: 4
[Phase2] max_tokens: 1045, CFG: 2.00, seeds: 7917479554720955107..7917479554720955107
[Phase2] Prefill 285ms (shared, 1 cond + 1 uncond)
[Phase2] Decode 16ms
[Batch 0] seed=7917479554720955107, 1 codes
[Ace-LM] Load 3164 | Total 11648ms | seed=7917479554720955107
[Request] Wrote simple0.json
ggml_cuda_init: found 4 ROCm devices:
  Device 0: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 1: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 2: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
  Device 3: AMD Radeon AI PRO R9700, gfx1201 (0x1201), VMM: no, Wave Size: 32
[Load] DiT backend: ROCm0 (CPU threads: 8)
[Load] Backend init: 42.4 ms
[GGUF] ../models/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: 616.1 ms
[GGUF] ../models/acestep-v15-turbo-Q8_0.gguf: 678 tensors, data at offset 56864
[Load] silence_latent: [15000, 64] from GGUF
[GGUF] ../models/vae-BF16.gguf: 365 tensors, data at offset 30048
[Load] VAE backend: ROCm0 (shared)
[VAE] Backend: ROCm0, Weight buffer: 161.1 MB
[VAE] Loaded: 5 blocks, upsample=1920x, F32 activations
[Load] VAE weights: 852.1 ms
[BPE] Loaded from GGUF: 151643 vocab, 151387 merges
[Load] BPE tokenizer: 43.9 ms
[Load] TextEncoder backend: ROCm0 (shared)
[GGUF] ../models/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: 224.3 ms
[Load] CondEncoder backend: ROCm0 (shared)
[GGUF] ../models/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: 184.1 ms
[GGUF] ../models/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: 33.8 ms
[Ace-Synth] All models loaded, turbo=yes
[Request] Parsed simple0.json
[Request] seed=7917479554720955107 batch=1
[Request] caption: An upbeat and anthemic pop-rock track driven by a clean, ene...
[Request] lyrics: 1657 bytes
[Request] bpm=86 dur=189 key=G major ts=4 lang=fr
[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] 1 audio codes (0.2s @ 5Hz)
[Pipeline] T=6, S=3
[Pipeline] seed=7917479554720955107, steps=8, guidance=1.0, shift=3.0, duration=189.0s
[Pipeline] caption: 173 tokens, lyrics: 595 tokens
[Encode] TextEncoder (173 tokens): 62.4 ms
[Encode] Lyric vocab lookup (595 tokens): 1.8 ms
[CondEnc] Lyric sliding mask: 595x595, window=128
[CondEnc] Timbre sliding mask: 750x750, window=128
[Encode] Packed: lyric=595 + timbre=1 + text=173 = 769 tokens
[Encode] ConditionEncoder: 31.9 ms, enc_S=769
[Context] Decoded: 1 codes -> 5 frames (0.2s @ 25Hz)
[Context] Detokenizer: 33.6 ms
[Context Batch0] Philox noise seed=7917479554720955107, [6, 64]
[DiT] Starting: T=6, S=3, enc_S=769, steps=8, batch=1
[DiT] Batch N=1, T=6, S=3, enc_S=769
[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: 145.4 ms (145.4 ms/sample)
[VAE] Graph: 335 nodes, T_latent=6
[VAE] Decoded: T_latent=6 -> T_audio=11520 (0.24s @ 48kHz)
[VAE Batch0] Decode: 771.7 ms
[MP3] Encoding 0.2s @ 128 kbps, 48000 Hz stereo
[MP3] 3840 bytes (12.0:1), 50 ms (4.82x realtime), 1 threads
[MP3] Wrote simple00.mp3
[Request 1/1] Done
[Pipeline] All done

And here's the simple0.json

simple0.json

{
  "caption": "An upbeat and anthemic pop-rock track driven by a clean, energetic male lead vocal. The song opens with layered, wordless vocal harmonies over a bright acoustic guitar before the full band kicks in with a driving 4/4 drum beat, a solid bassline, and clean electric guitars. The chorus is big and memorable, elevated by layered backing vocals that create a powerful, singalong feel. A brief, more reflective bridge provides a dynamic shift before launching into a final, high-energy chorus and an outro featuring vocal ad-libs and a fade-out with the initial vocal harmony motif.",
  "lyrics": "[Intro]\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n\n[Verse 1]\nRéveille-toi, ça fait longtemps\nLe soleil danse sur l'horizon\nLaisse tomber tous tes chagrins\nCe soir on refait le monde en chantant\n\n[Pre-Chorus]\nOublie les heures, oublie les peines\nViens danser, laisse faire la semaine\nOn met le feu sous les étoiles\nRien ne nous arrête, on part en cavale\n\n[Chorus]\nAllez viens, on va chanter, on va danser, on va s'aimer\nLe monde est à nous pour rêver\nViens on va tout lâcher\nOn va rire, on va crier, on va tout donner sans compter\nParce qu'au fond c'est ça la vie\nViens on va s'amuser\n\n[Post-Chorus]\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n\n[Verse 2]\nOublie les soucis, laisse-les filer\nCe soir on vit, on veut juste vibrer\nUn peu d'amour, un peu de folie\nUn grand éclat, une belle mélodie\n\n[Pre-Chorus]\nOublie les heures, oublie les peines\nViens danser, laisse faire la semaine\nOn met le feu sous les étoiles\nRien ne nous arrête, on part en cavale\n\n[Chorus]\nAllez viens, on va chanter, on va danser, on va s'aimer\nLe monde est à nous pour rêver\nViens on va tout lâcher\nOn va rire, on va crier, on va tout donner sans compter\nParce qu'au fond c'est ça la vie\nViens on va s'amuser\n\n[Bridge]\nOn s'envole, on s'égare\nOn s'en fout du regard des autres\nTant qu'on a le cœur qui bat\nOn prend la vie comme une apôtre\n\n[Chorus]\nAllez viens, on va chanter, on va danser, on va s'aimer\nLe monde est à nous pour rêver\nViens on va tout lâcher\nOn va rire, on va crier, on va tout donner sans compter\nParce qu'au fond c'est ça la vie\nViens on va s'amuser\n\n[Outro]\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n(Ooh-ooh-ooh-ooh)\n[Instrumental fade out]",
  "bpm": 86,
  "duration": 189.0,
  "keyscale": "G major",
  "timesignature": "4",
  "vocal_language": "fr",
  "seed": 7917479554720955107,
  "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": "3101"
}

[ServeurpersoCom]

Can you retry with --no-fa on ace-lm :

(simple.sh)

#!/bin/bash

set -eu

../build/ace-lm \
    --request simple.json \
    --lm ../models/acestep-5Hz-lm-4B-Q8_0.gguf \
    --no-fa

../build/ace-synth \
    --request simple0.json \
    --embedding ../models/Qwen3-Embedding-0.6B-Q8_0.gguf \
    --dit ../models/acestep-v15-turbo-Q8_0.gguf \
    --vae ../models/vae-BF16.gguf

If it's always NOK, add :
"lm_cfg_scale": 1.0
inside simple.json and retry

[ServeurpersoCom]

Rebased GGML submodule on latest upstream, which includes flash attention race condition fixes and RDNA4-specific MMVQ parameters. This might fix your issue. Pull and rebuild:

git pull
git submodule update --init

[firefox42]

Thanks, but disabling FA didn't seem to change the output, it just made the Phase1 slower. And I tried the latest pull, but now I get this error while building:

[  1%] Building C object ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o
In file included from /home/firefox42/AI/music/acestep.cpp/ggml/src/ggml.c:23:
/home/firefox42/AI/music/acestep.cpp/ggml/src/ggml.c:1062:1: error: static assertion failed: "GGML_OP_COUNT != 97"
 1062 | static_assert(GGML_OP_COUNT == 97, "GGML_OP_COUNT != 97");
      | ^~~~~~~~~~~~~
/home/firefox42/AI/music/acestep.cpp/ggml/src/ggml.c:1174:1: error: static assertion failed: "GGML_OP_COUNT != 97"
 1174 | static_assert(GGML_OP_COUNT == 97, "GGML_OP_COUNT != 97");
      | ^~~~~~~~~~~~~
gmake[2]: *** [ggml/src/CMakeFiles/ggml-base.dir/build.make:79: ggml/src/CMakeFiles/ggml-base.dir/ggml.c.o] Error 1
gmake[1]: *** [CMakeFiles/Makefile2:505: ggml/src/CMakeFiles/ggml-base.dir/all] Error 2
gmake: *** [Makefile:136: all] Error 2

[ServeurpersoCom]

Haha, I thought I could fix it before your test, but no luck, I didn't finish the "rebase" yet :)
Thanks for testing. I'll post an update as soon as I've fixed my rebase

[ServeurpersoCom]

git pull --rebase
inside ggml, and inside acestep.cpp
:)

[firefox42]

Cool, it complies again, though I'm seeing the same behavior

[ServeurpersoCom]

--no-fa rules out flash attention, thanks. Can you try setting
"lm_cfg_scale": 1.0
in your request JSON? This disables CFG which is the only thing that makes Phase 2 run batched (N=2). Phase 1 works because it already runs at N=1. If this produces a full song, the problem is in the batched N=2 path on gfx1201.

[firefox42]

Ok, this time I ran it manually, using ./build/ace-synth --request examples/simple0.json --embedding models/Qwen3-Embedding-0.6B-Q8_0.gguf --dit models/acestep-v15-turbo-Q8_0.gguf --vae models/vae-BF16.gguf after changing lm_cfg_scale to 1.0, and it didn't seem to make a difference.

[ServeurpersoCom]

You need to restart from the original simple.json file (clean request), add the lm_cfg_scale 1.0 and run ace-lm to check if there is more than 1 audio_codes on the output file (simple0.json).
The bug is inside ace-lm.
ace-lm (5Hz audio codes generation) drives ace-synth (flow matching, audio generation)

[firefox42]

Sorry I misunderstood! Okay, adding "lm_cfg_scale": 1.0 to the initial request json file fixes it.

[ServeurpersoCom]

Sorry I misunderstood! Okay, adding "lm_cfg_scale": 1.0 to the initial request json file fixes it.

Now please use the new option : --no-batch-cfg on ace-lm.
Please confirm me it's working without "lm_cfg_scale": 1.0`
You get the full precision (but /2 performances).

I still need to find a way to double the performance on ROCm and older RTX cards.

[firefox42]

The new flag works, thanks for the help!