torch_tvarant documentation¶
torch_tvarant is an out-of-tree PyTorch device backend for the Tvarant RISC-V SIMT GPGPU — targeting the Xilinx Alveo U55C FPGA accelerator.
It lets you write standard PyTorch code and run it on torch.device("tvarant"):
import torch
import torch_tvarant
x = torch.randn(4, 768, device="tvarant")
w = torch.randn(768, 768, device="tvarant")
y = torch.nn.functional.relu(x @ w)
Key features¶
- CPU simulator — develop and test without hardware
- OpenCL runtime — same Python API on POCL / FPGA path
- LLM ops —
silu,softmax,bmm,layer_norm,embedding, and more - JIT compiler — fuse GEMM epilogues and pointwise chains for inference
- Extensible — add kernels in C++/OpenCL and register ATen ops
Quick links¶
| Resource | Link |
|---|---|
| Installation | install.md |
| Getting started | getting-started.md |
| API reference | api/index.md |
| Architecture | architecture.md |
| JIT compiler | jit-compiler.md |
| Contributing | contributing.md |
| GitHub | samthakur587/Tvarant |
| Changelog | CHANGELOG.md |
License¶
Apache License 2.0 — see LICENSE.