Custom cpu and cuda operators support

[K024]

Hi. Currently I'm trying to implement some large language models (LLM) with TorchSharp and got a nice demo (here). But when moving forward to more features I found some lacking features required for LLMs:

Custom operators

LLMs heavily depend on custom operators like flash attention, RMS norm and GPTQ int4 matmul for faster inference speed and reduced model size with quantization.

PyTorch allows defining custom operators with native c++ and cuda source files in two ways: pybind11 and torch library. The latter one seems working fine with torch.jit.script and is potentially to be working with TorchSharp torch.jit.compile and torch.ops.xxx. But loading it requires calling a torch native method. Also, TorchSharp may have some specialized modules for custom ops.

BTW, openai/triton uses MLIR and LLVM to create custom ops, but is almost bound to python.

NCCL ops

I've also tried to implement a thread-based distributed approach with TorchSharp (at here). The required communication ops are: broadcast, scatter, gather and all-gather. I'm using the naive _copy operator to implement them, but are very slow. Is it possible to have these NCCL related ops provided?

[dje-dev]

Wow, this work looks very interesting and potentially very useful! Basic distributed training/inference (one host, multiple GPU) is currently a gap for Torchsharp and your implementation could be a step toward addressing that.

Two comments/suggestions:

• beware that there is currently a serious problem with Torchscript execution that is likely to corrupt the memory in your process if you run for more than a minute or two (Torchscript execution failures (hangs, access violation, Fatal error. Internal CLR fatal error. (0x80131506) ) #1047). I'm hoping this can be fixed soon
• in order to increase the chances that your request could be addressed, you might consider laying out specifically what you would need from NCCL. For example, it might be helpful if you could provide a pointer to the minimal specific set of torch APIs (in some header file) needed, and a few lines of sample C# code that would use these APIs and demonstrate that they are working correctly

[GeorgeS2019]

[WIP 2023] Integrating Custom PyTorch c++ Extensions

[alinpahontu2912]

@K024 Thank you for reporting this issue and for your interest in TorchSharp! Your insights on custom operators and NCCL operations are incredibly valuable, especially given their importance for implementing large language models. Since it’s been a while since this was last discussed, I wanted to check in and see if the concerns you raised are still relevant or if you've found a resolution. TorchSharp has evolved significantly over the past couple of years, and we’d love to hear if there have been any developments on your end. Please share any updates when you can!