Huggingface cache, symlinks, and unsupported tensor source format,

[AlphaMo99]

If you download a repository with the huggingface cli :

hf download audio-cpp/MiniMax-Music3-GGUF

And use the cache directory AND specifying a file (works if I do not specify the flow_transformer):

./build/bin/audiocpp_cli --task gen --session-option minimax_music3.flow_transformer_gguf=transformer_q8_0.gguf --family minimax_music3 --model /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/ --backend cuda --text "A bright pop rock song with clean drums and a clear male vocal." --request-option 'lyrics=[verse] City lights are shining low. I keep moving with the glow. [chorus] Turn it up and let it fly. Sing the melody tonight.' --request-option duration_sec=20 --request-option num_inference_steps=30 --out output.wav
ggml_cuda_init: found 1 CUDA devices (Total VRAM: 15937 MiB):
Device 0: Quadro RTX 5000, compute capability 7.5, VMM: yes, VRAM: 15937 MiB
audiocpp_cli failed: unsupported tensor source format: /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/blobs/40ed8a14884e9ac1b40625edd0dc14138c3d3d62803e2b5903f008dec60e61ac

It fails because the GGUF is a symlink to a file without any file extension :

ls -l /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/transformer_q8_0.gguf
lrwxrwxrwx 1 nobody nobody 76 Aug 18 17:07 /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/transformer_q8_0.gguf -> ../../blobs/40ed8a14884e9ac1b40625edd0dc14138c3d3d62803e2b5903f008dec60e61ac

[0xShug0]

@AlphaMo99 The preview branch has been merged into main, and I've pushed the fix there. Please try the latest version from main.

[AlphaMo99]

Ok it works now with audiocpp_cli, but not yet with audiocpp_server

[AlphaMo99]

Moreover the results looks very strange, only 8 secs of music with no distinctive lyrics, and few notes, like oriental style.

./build/bin/audiocpp_cli --task gen --session-option minimax_music3.language_model_gguf=language_model_q8_0.gguf --session-option minimax_music3.flow_transformer_gguf=transformer_q8_0.gguf --session-option minimax_music3.rvq_depth_decoder_gguf=rvq_depth_decoder_q8_0.gguf --family minimax_music3 --model /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/ --backend cuda --text "A bright pop rock song with clean drums and a clear male vocal." --request-option 'lyrics=[verse] City lights are shining low. I keep moving with the glow. [chorus] Turn it up and let it fly. Sing the melody tonight.' --request-option duration_sec=20 --request-option num_inference_steps=30 --out output.wav ggml_cuda_init: found 1 CUDA devices (Total VRAM: 15937 MiB): Device 0: Quadro RTX 5000, compute capability 7.5, VMM: yes, VRAM: 15937 MiB ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture family=minimax_music3 task=gen mode=offline audio_out=output.wav

If I increase the duration to 60s, the same 8 sec file is generated.

[0xShug0]

Ok it works now with audiocpp_cli, but not yet with audiocpp_server

I tested the server can handle symlink. Please rebuild the server cleanly and try again. If still happens, share the server config you used.

[0xShug0]

Moreover the results looks very strange, only 8 secs of music with no distinctive lyrics, and few notes, like oriental style.

./build/bin/audiocpp_cli --task gen --session-option minimax_music3.language_model_gguf=language_model_q8_0.gguf --session-option minimax_music3.flow_transformer_gguf=transformer_q8_0.gguf --session-option minimax_music3.rvq_depth_decoder_gguf=rvq_depth_decoder_q8_0.gguf --family minimax_music3 --model /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/ --backend cuda --text "A bright pop rock song with clean drums and a clear male vocal." --request-option 'lyrics=[verse] City lights are shining low. I keep moving with the glow. [chorus] Turn it up and let it fly. Sing the melody tonight.' --request-option duration_sec=20 --request-option num_inference_steps=30 --out output.wav ggml_cuda_init: found 1 CUDA devices (Total VRAM: 15937 MiB): Device 0: Quadro RTX 5000, compute capability 7.5, VMM: yes, VRAM: 15937 MiB ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture ggml_cuda_graph_set_enabled: disabling CUDA graphs due to GPU architecture family=minimax_music3 task=gen mode=offline audio_out=output.wav

If I increase the duration to 60s, the same 8 sec file is generated.

Check the info here #263. The duration is the audio frame budget, aligned with the official Python implementation, not the exact duration of the output audio. Also, for diffusion-style models, you may need to try different seeds and polish prompt to get the best result.

[AlphaMo99]

The Gui doesn't detect the model : (just with --ui and filling Models folder downloads, local detection, and model loading)

1. If I try to use the Gui to download again the model : error: remote file is not accessible: audio-cpp/audio.cpp-gguf/MiniMax-Music3-GGUF/config.json (404)

2. Or If I create a symlink to the model dir :
   ln -s /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/ MiniMax-Music3-GGUF (I straced the servers it's looking for MiniMax-Music3-GGUF dir, so I created a link, it finds the directory and gguf files in the strace, but something more is missing, it's still not activated in the Gui). And if I try to download it, I got this message
   error: some package files already exist in: /home/xxx/Projet/AI/audio.cpp/build/bin/models/MiniMax-Music3-GGUF (use --overwrite)

[AlphaMo99]

I took the prompt from here : https://huggingface.co/audio-cpp/MiniMax-Music3-GGUF

[0xShug0]

I took the prompt from here : https://huggingface.co/audio-cpp/MiniMax-Music3-GGUF

Sorry for the confusion. That prompt is only meant as a quick end-to-end check to confirm the model can run. For higher-quality songs, the prompt needs to be more polished. For example, prompts from https://minimax-ai.github.io/music3-demo/ can easily produce much better results.

official_q8_20s_no_seed.wav
official_q8_20s_seed_0.wav
official_q8_20s_seed_42.wav

[AlphaMo99]

ok ok, still trying, If I cut and paste the description, lyrics, length of the very first song, (Features/Pop), I have this error :
audiocpp_cli failed: MiniMax Music 3 compact semantic sampler has no finite logits

[AlphaMo99]

Maybe a quote " problem, I'm bisecting the pb to find the root cause

[0xShug0]

ok ok, still trying, If I cut and paste the description, lyrics, length of the very first song, (Features/Pop), I have this error : audiocpp_cli failed: MiniMax Music 3 compact semantic sampler has no finite logits

Can you try changing the seed, and also make sure the prompt/lyrics are not being broken by shell quoting? The demo acutally uses Features/Pop prompt and works fine. It's hard to pinpoint the root cause without the exact cmd you used, because I can't reproduce the issue.

music3_demo.mp4

[AlphaMo99]

Ok I put the two files in files : desc.txt and lyrics.txt
time ./build/bin/audiocpp_cli --task gen --session-option minimax_music3.language_model_gguf=language_model_q8_0.gguf --session-option minimax_music3.flow_transformer_gguf=transformer_q8_0.gguf --session-option minimax_music3.rvq_depth_decoder_gguf=rvq_depth_decoder_q8_0.gguf --family minimax_music3 --model /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/ --backend cuda --text "$(cat desc.txt)" --request-option "lyrics=$(cat lyrics.txt)" --request-option duration_sec=191 --request-option num_inference_steps=30 --out output.wav --seed 442

It looks like I have the error message audiocpp_cli failed: MiniMax Music 3 compact semantic sampler has no finite logits when the sum size of the two file is more than 1500 characters (+/- not the exact number, I don't know if the limits is in token or characters precisely...)

[0xShug0]

Ok I put the two files in files : desc.txt and lyrics.txt time ./build/bin/audiocpp_cli --task gen --session-option minimax_music3.language_model_gguf=language_model_q8_0.gguf --session-option minimax_music3.flow_transformer_gguf=transformer_q8_0.gguf --session-option minimax_music3.rvq_depth_decoder_gguf=rvq_depth_decoder_q8_0.gguf --family minimax_music3 --model /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/ --backend cuda --text "$(cat desc.txt)" --request-option "lyrics=$(cat lyrics.txt)" --request-option duration_sec=191 --request-option num_inference_steps=30 --out output.wav --seed 442

It looks like I have the error message audiocpp_cli failed: MiniMax Music 3 compact semantic sampler has no finite logits when the sum size of the two file is more than 1500 characters (+/- not the exact number, I don't know if the limits is in token or characters precisely...)

Let me check later. I doubt it's the prompt len because in the demo I used 4000+ char caption and 1200 char lyrics.

I'm currently end to end testing the fix for the UI issue. Still downloading the GGUFs from the UI...

[0xShug0]

Ok I put the two files in files : desc.txt and lyrics.txt time ./build/bin/audiocpp_cli --task gen --session-option minimax_music3.language_model_gguf=language_model_q8_0.gguf --session-option minimax_music3.flow_transformer_gguf=transformer_q8_0.gguf --session-option minimax_music3.rvq_depth_decoder_gguf=rvq_depth_decoder_q8_0.gguf --family minimax_music3 --model /usr/share/llama/gguf/models--audio-cpp--MiniMax-Music3-GGUF/snapshots/ed915d0748225e39b2b9b4eab354a20f66e30bc2/ --backend cuda --text "$(cat desc.txt)" --request-option "lyrics=$(cat lyrics.txt)" --request-option duration_sec=191 --request-option num_inference_steps=30 --out output.wav --seed 442

It looks like I have the error message audiocpp_cli failed: MiniMax Music 3 compact semantic sampler has no finite logits when the sum size of the two file is more than 1500 characters (+/- not the exact number, I don't know if the limits is in token or characters precisely...)

Also could you try the default quant combo?

[0xShug0]

@AlphaMo99 UI fix not pushed yet. Could you share desc.txt and lyrics.txt?