Flashlight is much lower level and gives more fine-grained performance control. For instance, I don't think there is really any way to do real-time speech recognition that is fast with PyTorch because of how it is architected.
From my understanding, Tensor Comprehensions and Halide are both very tentative research projects.
> not having it's direction beholden to the whims of Alphabet (see: Swift for Tensorflow).
I don't think this is an accurate recreation of the history that led to FB working on pytorch.
This is a very interesting claim. I find it credible because it stands to reason that projects like DeepSpeed[1] and TorchScript[2] wouldn't need to exist if inference performance of research PyTorch models was satisfactory for production, but often case it isn't.
It appears as though Flashlight is built on ArrayFire. I haven't seen how gradients are managed in arrayfire-ml, but perhaps it is the case that the autograd implementation in PyTorch was a bottleneck and this is a ground up approach.
Editing as I didn't address your second point. I can neither confirm or deny the motivations for creating Torch being related to FB's desire to depend on an Alphabet-managed codebase. I know there are lots of reasons why programmers prefer the UX of the PyTorch Python API (I do as well), and there are probably other reasons I can't recall off the top of my head. I am only saying that PyTorch is contributing to the ML ecosystem already by the sole virtue that it isn't a Google product.
Torch (Lua) predates TensorFlow and is Lecun's pet project for a few years at that point already. But Lua as a language is unpopular at a time. PyTorch would be a welcome addition then. But even if no PyTorch (nor Caffe2) in an alternative timeline, I would imagine FB would be stuck with Lua Torch for quite some time.
Adding to the above - Tensor Comprehensions was path-finding research and is no longer maintained. The git repo is frozen (archived) as a research artifact.
Halide is still quite active, and was used in products at Adobe and Google circa 2016-2017. Not sure about the current state of industry usage though.
I ship an app (Talon) which runs many kinds of wav2letter ASR models on consumer CPUs. It has real-time inference pipelines for both pytorch and flashlight. (I wrote the pytorch code). Performance of both inference pipelines is fine and comparable between the frameworks for me. I'm not sure what you're talking about wrt speech recognition performance.
Maybe for training? But there aren't that many cpu bound components there, and you can write those in native code.
From my understanding, Tensor Comprehensions and Halide are both very tentative research projects.
> not having it's direction beholden to the whims of Alphabet (see: Swift for Tensorflow).
I don't think this is an accurate recreation of the history that led to FB working on pytorch.