Runpod binary test

Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.github/workflows/build.yml
#	README.md
#	ci/run.sh
#	docs/build.md
#	examples/CMakeLists.txt
#	examples/parallel/parallel.cpp
#	ggml/CMakeLists.txt
#	ggml/src/CMakeLists.txt
#	scripts/server-bench.py
#	src/llama-kv-cache-unified.cpp
#	tests/test-backend-ops.cpp
#	tools/batched-bench/batched-bench.cpp
#	tools/server/README.md

ROCm Linux Binary Test

This is just a dev repo, want the real deal? https://koboldai.org/cpp

build tools

Not KoboldCpp, don't run this on your PC its destructive and only for CI use.

Runpod trigger

Boop

v1.57 - Vulkan Only Pre-release

This is a Vulkan Only build of the upcoming v1.57, please check on https://koboldai.org/cpp if v1.57 has already released.
If v1.57 has a formal release this build has no advantages for you.

v1.59-Ofast

v1.59 but the makefile is changed to be OFast, for comparative testing.

1.52 - Linux

v1.52

Placebo commit to maybe fix CI

1.51.1 - Linux Binary Test

This is a special test release for linux, for other builds check https://koboldai.org/cpp

1.35

This repository is only used on special occation for compiled builds, get the latest from https://koboldai.org/cpp

Koboldcpp 1.35 build with sched_yield enabled and CUDA 11.4 for better GPU compatibiltiy
H2 update: (Still shows H on the version but newer than the henk_cuda from concedo's repository) Compiled in a VM for better dependency stability and CUDA 11.4 support.
H3 update: Same source code as the previous versions other than the version name change. Recompiled with a different psutil (from conda instead of pip) to make high priority work again.

Win7 build: Compiled without PrefetchVirtualMemory, normally Windows 7 is only supported on the Fallback backend. This is a limited edition build that has Windows 7 support on hopefully all backends (CUDA not tested) at the expense of the model loading speed.

Tools: Compilation of all the GGML conversion tools (make tools)

v1.0.3 - Windows

llamacpp-for-kobold-1.0.3

• Applied the massive refactor from the parent repo. It was a huge pain but I managed to keep the old tokenizer untouched and retained full support for the original model formats.
• Reduced default batch sizes greatly, as large batch sizes were causing bad output and high memory usage
• Support dynamic context lengths sent from client.
• TavernAI is working although I wouldn't recommend it, they spam the server with multiple requests of huge contexts so you're going to have a very painful time getting responses.

To use drag and drop a compatible quantized model for llamacpp on top of the exe.