ACE-Step on Thumper

ACE-Step has a complex dependency chain: Python 3.11 specifically, PyTorch with GPU support, ffmpeg for audio encoding, and multiple model downloads. Thumper handles every step.

ACE-Step’s quality scales with available VRAM. The language model component (used for prompt understanding) is the main differentiator between tiers.

ACE-Step uses a Python runtime with Gradio on port 7860. The health check timeout is 180 seconds (longer than most apps) because model loading is slow on first launch.

# Runtime configuration:
runtime: python
port: 7860                    # Gradio default
health_check_timeout: 180     # 3 min for model loading
python_version: "3.11"        # Exact version required
 
# System dependency:
system_deps:
  - ffmpeg                    # Required for audio encoding

ACE-Step’s DiT model is sensitive to precision and memory management. GPU configuration matters more here than for most apps.

NVIDIA (CUDA)

• Fast inference out of the box with CUDA 11.8+
• FP16 works correctly on all NVIDIA GPUs
• RTX 4090 generates a 10-second clip in ~2 seconds

AMD Discrete GPU (ROCm)

• Requires ROCm 5.7+
• MIOPEN_FIND_MODE=2 is critical — without it, MIOpen re-benchmarks every launch (~8 min warmup)
• First run builds the MIOpen Find-DB (5–8 min); subsequent runs are fast

AMD APU (Integrated Graphics)

• BF16 is mandatory — FP16 causes NaN in DiT attention due to value overflow past FP16 max (~65504)
• CPU text encoder frees GPU memory for diffusion
• Smart memory (default) keeps the diffusion model cached between stages

Apple Silicon (MPS)

• MPS backend works for most operations
• May need FP32 fallback for some attention ops
• M2 Max generates a 10-second clip in ~8 seconds

CPU Only

• Very slow — basic generation only
• Short clips (10s) recommended to keep generation time manageable

Generation time scales with clip duration. These benchmarks use the default model configuration at each GPU’s optimal tier.