Feadback: MiniMax-H3 & MiniMax-Music3

[IIIIIllllIIIIIlllll]

In short, the audio part of MiniMax‑H3 works well, and there is no noticeable difference between ROCm and CUDA. MiniMax‑Music also performs well on ROCm — I didn't analyze the spectrum, just listened by ear. At the same time, I noticed an issue: with the same prompt, when asked to generate a 60‑second audio clip, my 2080 Ti‑22G took over an hour without producing any result, and it didn't even throw an OOM error. Quite strange.

h3_test_cuda.wav
h3_test_hip.wav

song1.wav

[0xShug0]

@IIIIIllllIIIIIlllll Thank you for the testing and feedback! I think the “stuck” behavior is caused by older GPU architectures falling into slow BF16 conversion or non-fast GEMM paths. The Global LM activation cast policy enables GGML_TYPE_BF16 for CUDA/HIP, but ggml CUDA BF16 fast MMA is gated to Ampere+. You can try disable global_lm.cpp BF16 activation casts and check performance and logs again.