TVM - AshokBhat/ml GitHub Wiki
About
- End to End Deep Learning Compiler Stack, an Apache project.
- Models => Minimum deployable modules
- Hardware - CPU, GPU, and specialized accelerators
- Frameworks - Keras, MXNet, PyTorch, Tensorflow, Core ML, DarkNet
Architecture
Key components
| Component | Description |
|---|---|
| Relay | High-level functional IR used to represent full models |
| TOPI | Tensor operator library |
| Runtime | Executes compiled artifacts |
| Frontend | ingest models from different frameworks into the TVM stack |
Versions
| Releases | Date | Notes |
|---|---|---|
| v0.7.0 | 2020-10-02 | |
| v0.6.1 | 2020-06-29 | |
| v0.6.0 | 2019-11-27 | |
| v0.5 | 2019-02-18 | |
| v0.4 | 2018-09-03 | |
| v0.3 | 2018-09-03 | |
| v0.2 | 2018-01-31 | |
| v0.1rc | 2017-08-17 |
AWS papers on TVM
- Optimizing CNN Model Inference on CPUs - https://arxiv.org/pdf/1809.02697.pdf
- Efficient Execution of Quantized Deep Learning Models: A Compiler Approach - https://arxiv.org/pdf/2006.10226.pdf
See also
- [XLA]] ](/AshokBhat/ml/wiki/[[Glow)
- MLIR