glow VS serving

Pytorch supports other kinds of accelerators (e.g. FPGA, and https://github.com/pytorch/glow), but unless you want to become a ML systems engineer and have money and time to throw away, or a business case to fund it, it is not worth it. In general, both pytorch and tensorflow have hardware abstractions that will compile down to device code. (XLA, https://github.com/pytorch/xla, https://github.com/pytorch/glow). TPUs and GPUs have very different strengths; so getting top performance requires a lot of manual optimizations. Considering the the cost of training LLM, it is time well spent.

It's pretty clear its referring to the output of Apache TVM and Meta's Glow

I also disagree with this. For example, Meta seems desperate about AI accelerators, and in fact is already doing "hardware customers develop software stack themselves" I mentioned above: Glow is that stack. Meta is doing Glow even if there is no promising AI accelerators right now, they are that desperate.

For reference: In Tensorflow and JAX, for example, the tensor gets compiled to the intermediate XLA format (https://www.tensorflow.org/xla), then passed to the XLA complier (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/compiler/xla/service) or the new TFRT runtime (https://github.com/tensorflow/runtime/blob/master/documents/tfrt_host_runtime_design.md), or some more esoteric hardware (https://github.com/pytorch/glow).

The main reasons are hiring, and depth and breadth of the product.

Compilers are hard, device support is hard, the compiler community is small and closed source compilers quickly become weird tech islands.

https://github.com/pytorch/glow

Yet another TEDIOUS BATTLE: Python vs. C++/C stack.

This project gained popularity due to the HIGH DEMAND for running large models with 1B+ parameters, like `llama`. Python dominates the interface and training ecosystem, but prior to llama.cpp, non-ML professionals showed little interest in a fast C++ interface library. While existing solutions like tensorflow-serving [1] in C++ were sufficiently fast with GPU support, llama.cpp took the initiative to optimize for CPU and trim unnecessary code, essentially code-golfing and sacrificing some algorithm correctness for improved performance, which isn't favored by "ML research".

NOTE: In my opinion, a true pioneer was DarkNet, which implemented the YOLO model series and significantly outperformed others [2]. Same trick basically like llama.cpp

[1] https://github.com/tensorflow/serving

I mean, probably - it's written in C++ https://github.com/tensorflow/serving

We’re looking into that solution at the moment, the issue I’m referring to is related to this https://github.com/tensorflow/serving/issues/1948 we’ll know if the plug-in approach works for our uses soon but haven’t started looking into implementing it yet

Which ML framework do you use? Tensorflow has https://github.com/tensorflow/serving. You could also use the Rust bindings to load a saved model and expose it using one of the Rust HTTP servers. It doesn't matter whether you trained your model in Python as long as you export its saved model.

Most likely it's a model server running something like https://github.com/tensorflow/serving and if there isn't a lot of load, the resource could kill some of its tasks. I wouldn't imagine it's sitting around pondering deep thoughts.

https://github.com/tensorflow/serving

https://thenewstack.io/tutorial-deploying-tensorflow-models-...

GitHub

You serve models via https://www.tensorflow.org/tfx/guide/serving which is written entirely in C++ (https://github.com/tensorflow/serving/tree/master/tensorflow_serving/model_servers), no Python on the serving path or in the shipped product.