htvm alternatives and similar packages
Based on the "Machine Learning" category
Do you think we are missing an alternative of htvm or a related project?
Both HTVM and TVM are under development. While TVM is somewhat stable, we don't recommend to use HTVM in applications currently. GitHub repository may contain newer version of HTVM.
HTVM is a library which provides Haskell runtime and experimental frontend for TVM the Machine Learning framework.
TVM in a nutshell
TVM framework extends Halide principles to Machine Learning domain. It offers (a) EDSLs for defining and hand-optimizing ML models (b) export/import facilities for translating models from other frameworks such as TensorFlow and (c) compiler to binary code for a variety of supported platforms, including LLVM (x86, arm), CUDA, OpenCL, Vulcan, ROCm, FPGAs and even WebAssembly (note: level of support may vary). DSLs for C++ and Python are best supported and also there are some support for Java, Go and Rust languages.
Originally, TVM aimed at increasing speed of model's inference by providing a rich set of optimizing primitives called 'schedules'). At the same time it had little support for training models. Recently, training-related proposals were added.
TVM aims at compiling ML models in highly optimized binary code.
Important parts of TVM are:
tvmis a core library providing
topiis a tensor operations collection. Most of the middle-layer primitives such as
softmaxare defined there.
relayis a high-level library written in Python, providing functional-style interface and its own typechecker. Currently, relay is under active development and beyond the scope of HTVM.
nnvmis another high-level wrapper in Python, now deprecated in favor of
Features and goals
In HTVM we are going to provide:
- C Runtime, which makes it possible to run TVM models from Haskell.
- Experimental EDSL for building TVM programs in Haskell.
Combined TVM/HTVM-stack features are:
- Not many dependencies: TVM is much easier to build than other frameworks (hi TensorFlow). Models are compiled to binary code, no interpreters required.
- Performance: HTVM uses TVM, which is designed with performace in mind.
- Simplicity of code.
- Experimental status
- Simplicity again. Pure ADT-based design.
- Not much type-safety yet. Expect errors in runtime. Typechecker may be implemented in future.
Make sure you have
Build tvm from development repository located at https://github.com/grwlf/tvm, branch autodiff
$ git clone https://github.com/grwlf/tvm $ cd tvm $ git branch autodiff origin/autodiff $ git checkout autodiff $ git submodule update --init --recursive ... follow up with the tvm build procedure
Compiling the package
We use development environment specified in Nix
language. In order to open it, please install the
Nix package manager.
Having Nix manager and
NIX_PATH set, enter the environment, by running Nix
development shell from the project's root folder:
It should get all the Haskell dependencies upon the first run. Alternatively, it should be possible to run Haskell distributions like Haskell Platform.
After nix-shell or Haskell distibution is ready, run
cabal to build the
$ cabal configure --enable-tests $ cabal build
To run tests, execute the test suite. At this point you will need
tvm of the correct version (see above).
$ cabal test
To enter the interactive shell, type
$ cabal repl htvm *HTVM.EDSL.Types> :lo Demo
Usage examples may be found in [Tests](./test/Main.hs) and (possibly outdated) [Demo](./src/Demo.hs).
TODO: Update demo, write more examples
FFI for TVM C Runtime library is a Haskell package, linked to
libtvm_runtime.so. This library contains functionality, required to load and
run ML code produced by TVM.
- The module provide wrappers to
TVMArrayis the main type describing Tensors in TVM. It is represented as ForeignPtr to internal representation and a set of accessor functions.
- Currently, HTVM can marshal data from Haskell lists. Support for
- No backends besides LLVM are tested. Adding them should not be hard and is on the TODO list.
EDSL has a proof-of-concept status. It may be used to declare ML models in Haskell, convert them to TVM IR and finally compile. Later, compiled model may be loaded and run with Haskell FFI or with any other runtime supported by TVM.
Contrary to usual practices, we don't manipulate TVM IR by calling TVM functions internally. Instead, we build AST in Haskell and print it to C++ program. After that we compile the program with common instruments. This approach has its pros and cons, which are described below.
HTVM.EDSL.Typesmodule defines AST types which loosely corresponds to
Exprclass hierarchies of TVM.
HTVM.EDSL.Monadprovides monadic interface to AST builders. We favored simplicity over type-safety. The interface relies on simple ADTs whenever possible.
HTVM.EDSL.Printcontain functions which print AST to C++ program (a) of Model Generator (b) which may be executed to obtain TVM module assembly.
HTVM.EDSL.Buildprovides instruments to compile and run the model generator by executing
- The Model Generator program builds TVM IR and produces x86 assembly (c)
- We execute
clangto compile x86 assembly into x86 '.so' library (d).
- Resulting library may be loaded and executed using
The whole data transformation pipeline goes as follows:
Monadic --> AST --> C++ --> Model --> X86 --> Model --> Runtime FFI Interface . . . Gen . asm . Library . . . (b) . (c) . (d) . Print . Print . Run C++ g++ clang (a)
Known disadvantages of C++ printing approach are:
- Compilation speed is limited by the speed of
g++, which is slow. Gcc is used to compile C++ to binary which may take as long as 5 seconds. Little may be done about that without changing approaches. One possible way to overcome this limitation would be to provide direct FFI to TVM IR like Halide-hs does for Halide. Unfortunately, this approach has its own downsides:
- Low-level IR API is not as stable as its high-level counterpart
- TVM is in its early stages and sometimes crashes. FFI to IR would provide no isolation from this.
- Calling construction-time procedures of TVM is non-trivial. This is a consequence of previous limitation. For example, TVM may calculate Tensor shape in runtime and use it immediately to define new Tensors. In order to that in Haskell we would need to compile and run C++ program which is possible by slow. We try to avoid calling construction-time procedures.
- User may face weird C++ errors. TVM is quite a low-level library which offers little type-checking, so user may write bad programs easily. Other high level TVM wrappers like Relay in Python, does provide their own typecheckers to catch errors earlier. HTVM offers no typechecker currently but it is certainly possible to write one. Contributions are welcome!
The pros of this approach are:
- C++ printer is implemented in less than 300 lines of code. Easy to maintain.
- Easy to port to another TVM dialect such as Relay.
- Isolation from TVM crashes. Memory problems of TVM IR will be translated to error messages in Haskell.
- We aim at supporting basic
- Support for Scheduling is minimal, but should be enhanced in future.
- Support for TOPI is minimal, but should be enhanced in future.
- No targets besides LLVM are supported. Adding them should be as simple as adding them to C++ DSL.
- We plan to support Tensor-Level AD
- Adding support for Relay is also possible but may require some efforts like writing Python printer.