crashing on the step "Decoding audio"

[beetjuizz]

i keep getting this

CUDA error: CUBLAS_STATUS_EXECUTION_FAILED when calling cublasLtMatmul with transpose_mat1 1 transpose_mat2 0 m 512 n 372 k 32 mat1_ld 32 mat2_ld 32 result_ld 512 abcType 0 computeType 68 scaleType 0
and on frontend it crashes everytime on 95% decoding audio

3080ti 12g VRAM docker image on win 10

[batbug]

same error on 4070 Ti Super 16GB

[iChristGit]

Yep, my 3090 for maximum speeds needs to offload the hearmula model before decoding, and its till 2 minutes per 2 minutes song

[Lindenmann777]

Same issue here. Windows 10, portable build, RTX 3060 Ti 8 GB.
At first I thought it was a VRAM problem, but I can see that you’re getting the same error even with 12 and 16 GB of VRAM.
Did you manage to find any solution or workaround?

[batbug]

not yet. having same problem in ubuntu 24 with 4070 Ti Super

[Lindenmann777]

That’s strange. I shared my build with a friend who has an RTX 4070 Ti (12 GB), and everything works fine for him.

UPD
Even stranger — I didn’t change anything, but suddenly everything started working without any errors.
I still have no idea what the problem was.

[fspecii]

Thanks for the detailed error! This CUBLAS error during audio decoding suggests a potential memory or driver issue.

To help diagnose, could you provide:

1. Your docker-compose.yml settings - specifically the environment variables you're using (HEARTMULA_4BIT, HEARTMULA_SEQUENTIAL_OFFLOAD, etc.)

2. Full startup logs - the lines showing:

   • [Auto-Config] Detected...
   • [Config] Using...
   • [Quantization]...

3. NVIDIA driver version (nvidia-smi output)

4. Does it fail on every generation or intermittently?

5. What's your GPU memory usage during generation? (run nvidia-smi in another terminal while generating)

Possible causes:

• The 3080Ti has 12GB VRAM which should work, but the codec decoding phase loads a second model
• If 4-bit quantization is disabled, both models might not fit
• Docker memory limits might be restricting GPU access

Try this workaround:
In your docker-compose.yml, ensure these are set:

environment:
  - HEARTMULA_4BIT=true
  - HEARTMULA_SEQUENTIAL_OFFLOAD=true

This will use 4-bit quantization and swap models during decode, which should fit in 12GB.

[fspecii]

This is the same CUBLAS error others have reported.

Fix: Go to Settings → Configuration → Memory Swap Mode and set it to "Force Enabled"

This loads models one-at-a-time instead of both together, which fixes the CUBLAS error on some GPU/driver combinations.

The 95% "Decoding audio" step is when HeartCodec loads - enabling Memory Swap Mode ensures HeartMuLa is unloaded first, avoiding the VRAM conflict.

Let me know if that works!

[batbug]

Where can i find this "Settings" menu?