svod-model
High-level inference for pretrained deep learning models on top of
svod-tensor. Each model is a pure-Rust port of an upstream checkpoint,
fetched from HuggingFace Hub at runtime and executed through JIT-compiled
plans.
Common infrastructure
| Module |
Role |
jit |
jit_wrapper!-generated wrappers, JitRecurrent<J>, InputSpec, JitError. Build-once / run-many execution. See JIT Graphs. |
audio |
Log-mel spectrogram, Splitter trait for long-form chunking (default: SileroVadSplitter). |
state |
HasStateDict + state_field! macros for loading PyTorch / safetensors checkpoints into Rust weight structs. |
blocks |
Shared Conv2dWeights, BatchNormWeights, BasicBlock, Bottleneck, ResidualStage reused by every ResNet-shaped backbone. timm/torchvision key convention. |
sentencepiece |
Minimal SentencePiece .model protobuf loader (vocab piece extraction). |
Models
Examples
cargo run -p svod-model --release --example gigaam_infer -- audio.wav
cargo run -p svod-model --release --example gigaam_rnnt_infer -- audio.wav
cargo run -p svod-model --release --example test_vad -- audio.wav
cargo run -p svod-model --release --example resnet_classify -- --hub --image dog.bin --side 224
cargo run -p svod-model --release --example wespeaker_parity -- --hub --data reference.npz
