* fix: use vllm AsyncLLMEngine to bring true stream
Current vLLM implementation uses the LLMEngine, which was designed for offline batch inference, which results in the streaming mode outputing all blobs at once at the end of the inference.
This PR reworks the gRPC server to use asyncio and gRPC.aio, in combination with vLLM's AsyncLLMEngine to bring true stream mode.
This PR also passes more parameters to vLLM during inference (presence_penalty, frequency_penalty, stop, ignore_eos, seed, ...).
* Remove unused import
This PR specifically introduces a `core` folder and moves the following packages over, without any other changes:
- `api/backend`
- `api/config`
- `api/options`
- `api/schema`
Once this is merged and we confirm there's no regressions, I can migrate over the remaining changes piece by piece to split up application startup, backend services, http, and mqtt as was the goal of the earlier PRs!
* Initial implementation of upload files api.
* Move sanitize method to utils.
* Save uploaded data to uploads folder.
* Avoid loop if we do not have a purpose.
* Minor cleanup of api and fix bug where deleting duplicate filename cause error.
* Revert defer of saving config
* Moved creation of directory to startup.
* Make file names unique when storing on disk.
* Add test for files api.
* Update dependencies.
* feat(tools): support Tools in the API
Co-authored-by: =?UTF-8?q?Stephan=20A=C3=9Fmus?= <stephan.assmus@sap.com>
* feat(tools): support function streaming
* Adhere to new return types when using tools instead of functions
* Keep backward compatibility with function calling
* Evaluate function names in chat templates
* Disable recovery with --debug
* Correctly stream out the entire result
* Detect when llm chooses to reply and to not perform any action in SSE
* Feedback from code review
---------
Co-authored-by: =?UTF-8?q?Stephan=20A=C3=9Fmus?= <stephan.assmus@sap.com>
* Dockerfile changes to build for ROCm
* Adjust linker flags for ROCm
* Update conda env for diffusers and transformers to use ROCm pytorch
* Update transformers conda env for ROCm
* ci: build hipblas images
* fixup rebase
* use self-hosted
Signed-off-by: mudler <mudler@localai.io>
* specify LD_LIBRARY_PATH only when BUILD_TYPE=hipblas
---------
Signed-off-by: mudler <mudler@localai.io>
Co-authored-by: mudler <mudler@localai.io>
Infinite context loop might as well trigger an infinite loop of context
shifting if the model hallucinates and does not stop answering.
This has the unpleasant effect that the predicion never terminates,
which is the case especially on small models which tends to hallucinate.
Workarounds https://github.com/mudler/LocalAI/issues/1333 by removing
context-shifting.
See also upstream issue: https://github.com/ggerganov/llama.cpp/issues/3969
* feat(refactor): refactor config and input reading
* feat(tts): read config file for TTS
* examples(kubernetes): Add simple deployment example
* examples(kubernetes): Add simple deployment for intel arc
* docs(sycl): add sycl example
* feat(tts): do not always pick a first model
* fixups to run vall-e-x on container
* Correctly resolve backend
* cleanup backends
* switch image to ubuntu 22.04
* adapt commands for ubuntu
* transformers cleanup
* no contrib on ubuntu
* Change test model to gguf
* ci: disable bark tests (too cpu-intensive)
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
* cleanup
* refinements
* use intel base image
* Makefile: Add docker targets
* Change test model
---------
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
