Components shared by the Ming-family model packages.
Naming: The flow-matching pieces all use the cfm_ prefix, because "flow matching", "fm" and "cfm" were three names for one concept here:
- dit_blocks - DiT building blocks (DiTBlock, FinalLayer, CondEmbedder)
- cfm_dit - the DiT used as the CFM velocity model
- cfm_solver - timestep schedules and the ODE/SDE integrator
- cfm_head - inference-only CFM wrapper preserving cfm.model.* names
- cfm_cudagraph - batch-1 CUDA-graph executor for the sampling loop
Modules:
| Name | Description |
|---|---|
aggregator | |
audio_dsp | |
audio_vae | |
cfm_cudagraph | CUDAGraph-accelerated CFM (flow-matching) diffusion head for Ming-TTS. |
cfm_dit | |
cfm_head | |
cfm_solver | |
dit_blocks | |
fused_istft | Post-FFT ISTFT fusion for Ming's four-way overlapping audio frames. |
speaker_extractor | CAM++ speaker-embedding extraction, shared by the Ming model packages. |
spk_embedding | |