FlowTorch

Safe and Performant Scientific Computing Library.

Right now the library is still in research phase where I am taking time to understand the internal implementations of Candle and Pytorch by reimplementing them. For production usage please try out Candle because its really awesome.

====

Todo

• Add Tensor::new method to create a Tensdor from data
• Fix shape similar to PyTorch
• Provide default contiguous layout - C type (Row Major)
• Implement view and reshape in Layout
• Add Proper Error Handling (https://github.com/dtolnay/thiserror)
• Add Arc and Rw::Lock for Storage in Tensor
• Implement Display for Tensor to print the matrix
• Implement Indexing which return a tensor
• Refactor indexing
• Add tuple with range indexing
• Fix Pretty Printing of Tensor
• Fix Display with PrintOptions
• Add I32 data type
• Add Baisc Docs for Tensor
• Add Tests for Narrow and Squeeze,
• Add macros and optimze code for ops
• Add Basic Binray Ops to Tensor
• Add basic unary ops
• Add Basic CUDA Device support
• Add Basic CUDA Ops
• Add to_device method support for converting the device of Tensor
• Refactor Error::Unknown and fix todo
• The Array trait implementation right now is doing many recurssive operations, can we minimize it?
• Combine common tests for different devices
• Add to_dtype method to convert the data type
• Add strided index in CPU and CUDA
• Ops to use strided index for CPU and CUDA
• Fix Comp Op and Broadcasting
• Add Basic Back Prop
• Add other Ops like Pow, Matmul, Add types for Unary CUDA Ops in kernel
• Add in place Binary and Unary ops (ex, tensor.add_, tensor.abs_)
• Impl flatten https://pytorch.org/docs/stable/generated/torch.flatten.html (this should when used to_vec1 give 1D array of strided indexed elements)
• Add Tensor.to_vec methods, to get different dim vectors
• Improve Code Coverage
• Fix select_index method
• Add Benchmarking (https://github.com/bheisler/criterion.rs)
• Support Fotran Memory layout (colum Major)
• Study Torch7 Library and how Storage, Tensor is implemented

===

Future Todo

• flowtorch-nn crate for Neural Net Utilities and Layers
• Inference MNIST
• Inference ImageNet Model
• Inference BERT/LLM's
• Add basic autograd functionality
• Train and Test a simple MNIST example

Setup

Running Tests for CUDA

cargo test --features "cuda" -- --nocapture 

Running Coverage

# Without CUDA
cargo tarpaulin --verbose --workspace --timeout 120 --out Html

# With CUDA
cargo tarpaulin --features="cuda" --verbose --workspace --timeout 120 --out Html

To excule cuda_backend from coverage for non cuda environment (GithubActions)

cargo tarpaulin --verbose --workspace --timeout 120 --out Html --exclude-files "crates/flowtorch-core/src/cuda_backend/**/*"