Backend matrix

What runs, what loads, what is only an interface.

This page is generated from CAPABILITIES blocks and BackendRegistry calls in the repository. Status labels are derived from whether synthesize() raises NotImplementedError, not from a feature wishlist.

implemented has a synthesis path scaffold loads, but falls back to a placeholder tone interface model/session loading only

Model backends

The three synthesis families.

pytorch implemented

F5-TTS (PyTorch)

F5TTSBackend · core/backend_f5.py

VRAM FLOOR1.5 GB min

SAMPLE RATE24000 Hz

PRECISIONfp32 · fp16 · int8

PLATFORMSWindows · Linux

The complete path. Uses upstream f5_tts for preprocess_ref_audio_text() and infer_process(), supports zero-shot reference audio, NFE steps, speed control and full checkpoint adapters. Weights: SWivid/F5-TTS (F5TTS_v1_Base).

Default backend in both CLI and GUI.
INT8 on CUDA is converted to FP16; CPU can use dynamic INT8.
Without a reference clip, falls back to the F5 example reference shipped by the package.

Declared speed multiplier: 1.0× · quality rating 5/5. These are capability metadata used by the auto-selector, not measured benchmarks.

fish-speech scaffold

Fish Speech

FishSpeechBackend · core/backend_fish.py

VRAM FLOOR2 GB min

SAMPLE RATE44100 Hz

PRECISIONfp32 · fp16 · int8 · int4

PLATFORMSWindows · Linux

Experimental integration. load() imports the Fish Speech modules, but the inference engine is never assigned, and synthesize() always reaches the placeholder-tone fallback in this revision.

Do not use for production or quality comparisons yet.
The files document the intended architecture (LLaMA semantic tokens + Firefly/DAC vocoder) but do not implement it.
vLLM acceleration is also not wired.

Declared speed multiplier: 1.5× · quality rating 5/5. These are capability metadata used by the auto-selector, not measured benchmarks.

gpt-sovits implemented

GPT-SoVITS

GPTSoVITSBackend · core/backend_gptsovits.py

VRAM FLOOR2 GB min

SAMPLE RATE32000 Hz

PRECISIONfp32 · fp16 · int8

PLATFORMSWindows · Linux

Calls the external GPT_SoVITS.TTS_infer_pack.TTS API with v2 or v1 pretrained paths. It is a real integration, but the dependency and weights live outside this repository.

Requires a GPT-SoVITS checkout and pretrained weights.
If the import fails, load() stores a stub and synthesis returns a 220 Hz tone with a warning.
Language handling follows the upstream package; the CLI passes a hint, not a full language matrix.

Declared speed multiplier: 1.2× · quality rating 5/5. These are capability metadata used by the auto-selector, not measured benchmarks.

Runtime targets

Registered in the registry, status included.

These classes are registered with BackendRegistry. Several can load a compiled model, ONNX session or GGUF file — none of them synthesize audio in this release.

Runtime backend capabilities
Target	Type	Status	VRAM	Precision	Compile	Platforms	Synthesis
DirectMLDirectMLBackend	runtime	interface	2 GB min	fp32, fp16	required	WIN	not implemented
F5-TTS (PyTorch)F5TTSBackend	model	implemented	1.5 GB min	fp32, fp16, int8	no	WIN + LINUX	yes
Fish SpeechFishSpeechBackend	model	scaffold	2 GB min	fp32, fp16, int8, int4	no	WIN + LINUX	yes
GPT-SoVITSGPTSoVITSBackend	model	implemented	2 GB min	fp32, fp16, int8	no	WIN + LINUX	yes
llama.cpp (GGUF)LlamaCppBackend	runtime	interface	0.5 GB min	q2_k, q3_k_m, q4_k_m, q5_k_m, q8_0	no	WIN + LINUX	not implemented
ONNX RuntimeONNXBackend	runtime	interface	CPU capable	fp32, fp16, int8	required	WIN + LINUX	not implemented
OpenVINOOpenVINOBackend	runtime	interface	CPU capable	fp32, fp16, int8	required	WIN + LINUX	not implemented
TensorRT-LLMTensorRTBackend	runtime	interface	2 GB min	fp16, int8, fp8	required	WIN + LINUX	not implemented
vLLMVLLMBackend	runtime	interface	3 GB min	fp16, int8, int4-gptq, int4-awq	no	LINUX	not implemented

How a backend is selected

01
CLI uses the literal --backend value (default f5-tts).
02
GUI switches the inference manager's backend from the TTS/Backend tabs.
03
Auto-selector scores registered backends by VRAM, platform and declared quality/speed. It is available as get_best_backend() but the GUI does not force it.
04
InferenceManager resolves auto to CUDA when torch reports it, otherwise CPU.

SELECTION CASCADE

--backend → InferenceManager → backend class
--device auto → CUDA if available → CPU
--quantization → backend load → effective precision

The auto-selector can recommend TensorRT or vLLM on capable hardware even though those synthesis paths are incomplete. Treat its recommendation as a capability hint, not a promise.

Precision

What the flags really do.

CLI value	F5-TTS behaviour	Notes
`fp16`	Model runs in half precision.	The useful GPU default.
`int8`	CUDA → FP16 with a log message; CPU → `torch.quantization.quantize_dynamic` on Linear layers.	CLI default. Not a true CUDA INT8 kernel.
`int4`	Accepted by argparse; no F5-TTS quantization path.	Do not expect a VRAM reduction.
`gguf-q4` / `gguf-q8`	Accepted by argparse; GGUF conversion is a placeholder in `core/quantization.py`.	Useful only if you already have a compatible GGUF and a working runtime, which this release does not provide.
`fp32` / `bf16`	Known to the quantization catalogue; not exposed by the CLI.	Backend classes default to FP16.

NEXT Performance, memory and limits → START Install guide →