xaedes

I have the same problem. I just use this to strip the prefix away: `result[len(prefix):]`. Works for me

I think this line needs to be adjusted: https://github.com/minimaxir/gpt-2-simple/blob/master/gpt_2_simple/gpt_2.py#L476 The `prefix` is repeated `batch_size` times to build `context`. I think you could use different prefixes there.

There is a FlxTiledSprite addon which could be used for this: https://api.haxeflixel.com/flixel/addons/display/FlxTiledSprite.html

Agreed, the ZeroMQ guide is a nice read. It would be really awesome if the online documentation would include html documentation generated by the already existing sand castle project =)...

This is still an issue in 2022. Will that PR be merged?

Using is sometimes also problematic to build on certain platforms and can be replaced by std::thread and std::mutex.

The generation of new tokens stops when encountering the `stop` word. It is defined as "\n" (newline) in https://github.com/FMInference/FlexGen/blob/main/apps/chatbot.py#L35 That is why only the first line will be generated. You...

I have implemented functions for getting and setting the rest of the model state. It additionally includes: random number generator state, logits, embedding and kv_cache. It was necessary to store...

Just created the pull request: https://github.com/ggerganov/llama.cpp/pull/1105

Training directly with ggml would be really nice. Implemented 8 out of 14 missing tensor ops. https://github.com/xaedes/llama.cpp/commit/757de70af8baeab9ce64a71b3ef56edf87289382 I had to add another ggml operation GGML_OP_ADD_AT as counterpart for GGML_VIEW in...