The binary grew up to 1.8GB quickly - rocm at least raises +800MB by
itself - so we might just want to manage the GPU libs separately.
Adds a comment to list all the libraries found so far that we are
depending on, but will likely follow up in a way to bundle these
separately.
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
This PR bundles further libs into the arm64 and x86_64 binaries
This can be improved by a lot - it's far from perfect, however in this PR I wanted to collect the required libs, and give a simple baseline to improve later upon. It is quite challenging to do this exercise with CI only - but it's the fastest way I see now.
I hope that after the list is initially built we can further improve this down the line and remove some of the technical debt left here to speedup things and do not get stuck in the middle of CI cycles.
In this PR:
- The x86_64 binary now bundles hipblas, nvidia and intel libraries too to avoid any dependency to be installed in the host
- Similarly, for the arm64 we now bundle all the required assets
## What's left
We should be also able to cross-compile Nvidia for arm64 - however I didn't succeed so far so I've left that open. Similarly I might have missed some libraries, but we will see with bug reports and testing around with the new binaries. I've tested on my arm64 board and I could finally start things up.
An open point still is shipping libraries for e.g. tts and stablediffusion. this is not done yet, however with the same methodology we should be able to extend support also for these two backends in the binary.
* ci: try to build for arm64
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Allow to skip hipblas on make dist
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* use arm64 cross compiler
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* correctly target go arm64
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* create a separate target
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* cross-compile grpc
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Add Protobuf include dirs
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* temp disable CUDA build
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* aarch64 builds: Reduce backends
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Even less backends
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Even less backends
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(startup): allow to load libs from extracted assets
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* makefile: set arch
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(amdgpu): try to build in single binary
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* Release space from worker
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat(llama.cpp): Enable decentralized, distributed inference
As https://github.com/mudler/LocalAI/pull/2324 introduced distributed inferencing thanks to
@rgerganov implementation in https://github.com/ggerganov/llama.cpp/pull/6829 in upstream llama.cpp, now
it is possible to distribute the workload to remote llama.cpp gRPC server.
This changeset now uses mudler/edgevpn to establish a secure, distributed network between the nodes using a shared token.
The token is generated automatically when starting the server with the `--p2p` flag, and can be used by starting the workers
with `local-ai worker p2p-llama-cpp-rpc` by passing the token via environment variable (TOKEN) or with args (--token).
As per how mudler/edgevpn works, a network is established between the server and the workers with dht and mdns discovery protocols,
the llama.cpp rpc server is automatically started and exposed to the underlying p2p network so the API server can connect on.
When the HTTP server is started, it will discover the workers in the network and automatically create the port-forwards to the service locally.
Then llama.cpp is configured to use the services.
This feature is behind the "p2p" GO_FLAGS
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* go mod tidy
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* ci: add p2p tag
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* better message
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* auto select cpu variant
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
* remove cuda target for now
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
* fix metal
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
* fix path
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
* cuda
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
* auto select cuda
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
* update test
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
* select CUDA backend only if present
Signed-off-by: mudler <mudler@localai.io>
* ci: keep cuda bin in path
Signed-off-by: mudler <mudler@localai.io>
* Makefile: make dist now builds also cuda
Signed-off-by: mudler <mudler@localai.io>
* Keep pushing fallback in case auto-flagset/nvidia fails
There could be other reasons for which the default binary may fail. For example we might have detected an Nvidia GPU,
however the user might not have the drivers/cuda libraries installed in the system, and so it would fail to start.
We keep the fallback of llama.cpp at the end of the llama.cpp backends to try to fallback loading in case things go wrong
Signed-off-by: mudler <mudler@localai.io>
* Do not build cuda on MacOS
Signed-off-by: mudler <mudler@localai.io>
* cleanup
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
* Apply suggestions from code review
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
---------
Signed-off-by: Sertac Ozercan <sozercan@gmail.com>
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
Signed-off-by: mudler <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
Co-authored-by: mudler <mudler@localai.io>
* feat: migrate diffusers backend from conda to uv
- replace conda with UV for diffusers install (prototype for all
extras backends)
- add ability to build docker with one/some/all extras backends
instead of all or nothing
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate autogtpq bark coqui from conda to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: convert exllama over to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate exllama2 to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate mamba to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate parler to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate petals to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: fix tests
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate rerankers to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate sentencetransformers to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: install uv for tests-linux
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: make sure file exists before installing on intel images
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate transformers backend to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate transformers-musicgen to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate vall-e-x to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: migrate vllm to uv
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: add uv install to the rest of test-extra.yml
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: adjust file perms on all install/run/test scripts
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: add missing acclerate dependencies
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: add some more missing dependencies to python backends
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: parler tests venv py dir fix
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: correct filename for transformers-musicgen tests
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: adjust the pwd for valle tests
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: cleanup and optimization work for uv migration
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: add setuptools to requirements-install for mamba
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: more size optimization work
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: make installs and tests more consistent, cleanup some deps
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: cleanup
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: mamba backend is cublas only
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: uncomment lines in makefile
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
---------
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat(single-build): generate single binaries for releases
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* drop old targets
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* feat: update ROCM and use smaller image
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: add call to ldconfig to fix AMDs broken library packages
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
---------
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
fix: more places where we are installing grpc that need a version specified
fix: attempt to fix metal tests
fix: metal/brew is forcing an update, they don't have 1.58 available anymore
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: only keep the build artifacts from the grpc build
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: remove separate Cache GRPC build step
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: remove docker inspect step, it is leftover from previous debugging
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
---------
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: cleanup Dockerfile and make final image a little smaller
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: add build-essential to final stage
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: more GRPC cache misses
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: correct for another cause of GRPC cache misses
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* feat: generate new GRPC cache automatically if needed
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* fix: use new GRPC_MAKEFLAGS build arg in GRPC cache generation
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
---------
Signed-off-by: Chris Jowett <421501+cryptk@users.noreply.github.com>
* Bump oneapi-basekit, optimum and openvino
* Changed PERFORMANCE HINT to CUMULATIVE_THROUGHPUT
Minor latency change for first token but about 10-15% speedup on token generation.
