Fast-DACVAE

Fast inference engine for DACVAE, a neural audio codec that compresses and reconstructs audio using a convolutional encoder-decoder with a VAE bottleneck. This library accelerates DACVAE inference up to 11.2x on NVIDIA GPUs through graph-level optimizations — with no custom kernels, no quality loss at FP32, and no changes to model weights.

Benchmark

NVIDIA H100 PCIe | facebook/dacvae-watermarked (107.7M params) | 100s audio @ 48kHz

Full Precision (FP32) — Zero Quality Loss

Method	Latency	Speedup	Real-time Factor
PyTorch FP32	1,047 ms	1.0x	96x
+ channels_last + wn_off	549 ms	1.9x	182x
+ torch.compile + graph	209 ms	5.0x	478x

Half Precision (FP16 / BF16)

Method	Latency	Speedup	RTF	SNR vs FP32
PyTorch FP16	775 ms	1.4x	129x	40.4 dB
+ channels_last + wn_off	307 ms	3.4x	326x	40.2 dB
+ torch.compile + graph (FP16)	93 ms	11.2x	1,071x	40.2 dB
+ torch.compile + graph (BF16)	100 ms	10.5x	1,004x	29.8 dB

Quick Start

pip install git+https://github.com/kadirnar/fast-dacvae.git

from dacvae import DACVAE
from dacvae.optimize import optimize_dacvae
import torch

model = DACVAE.load("facebook/dacvae-watermarked").cuda().eval()
audio = torch.randn(1, 1, 4800000, device="cuda")

# FP32 — zero quality loss, ~209ms
replay = optimize_dacvae(model, audio, dtype="fp32")
output = replay()

# FP16 — fastest, ~93ms
replay = optimize_dacvae(model, audio, dtype="fp16")
output = replay()

# BF16 — ~100ms
replay = optimize_dacvae(model, audio, dtype="bf16")
output = replay()

Requirements

• PyTorch 2.9+
• NVIDIA GPU (Hopper/Ampere)

License

Apache 2.0