Supporting Multi-LoRA inferencing via JetStream server #221
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…ion to Orbax format.
…e2e with LoRA paths.
…tAdapters, LoadAdapter and UnloadAdapter. 2) Driver which is holding list of all loaded base-parameters is now storing the list of lora updated paramters for loaded lora. Implemented methods for loading, unloading and listing LoRA adapters into the Driver object. Original base model params are intact and saved into the params dictionary with key . 3) Created a proxy-client to make MultiAdapterManager service requests to JetStream server.
…pters. Its functionality includes loading, unloading of adapters between CPU RAM and HBM. It also follows LRU policy to evict the adapter if a new load_adapter request comes up. Currently it is only storing the adapter as separate tensors (lora_a and lora_b). Calculation of lora_b x lora_a is being done in prefill() and generate() during decode request. Adapter_tensorstore can be configured with a max_limit on HBM and RAM. 2) Functionality to load from a catalog file at the start of the server is added. If no file is given, it will just load the base params. Loading from the catalog file is done on CPU RAM. After that based on incoming requests, those params are moved/evicted to/from HBM. 3) Some proto updates to get only single path for each adapter, and that path is expected to have an adapter_config.json and Orbax format weights in 0/items folder.
…n API (https://github.com/kubernetes-sigs/gateway-api-inference-extension/blob/main/docs/proposals/003-model-server-protocol/README.md#inference-api-protocol), & . 2) Added a flag to explicitly run the JetStream server with these APIs when . Else only expose older Decode() & HealthCheck() APIs of the JetStream Server. 3) Fixed a bug in the adapter_tensorstore while converting jnp_array and np_array. 4) Added a which made requests to the new APIs (v1/load_lora_adapter, v1/unload_lora_adapter, v1/models, v1/completions)
1) kv_cache_utilization: This refers to percentage of memory in the allocated kv-cache on TPU HBM, that is actually used during decode. It is based on the percentage of slots used. 2) num_requests_waiting: Total number of requests which are waiting to be decoded. 3) lora_requests_info: List of LoRA adapters that are loaded into the TPU HBM for serving the requests.
2) Fixing model_ckpt_conversion.sh after refactoring and merging from main.
aman2930
requested review from
yixinshi,
vipannalla and
gangji
and removed request for
vipannalla
vipannalla
reviewed
Mar 10, 2025
There was a problem hiding this comment.
Thanks for the PR, its a bit longish and I'd have preferred you to send the adapter_tensorstore.py and related code as a separate PR since its isolated enough along with the unittests before sending the the PR to integrate it into orchestrator.
I've some initial comments.
- Implemented unapply lora from base_params - Fixed some comments from the PR
Bslabe123
reviewed
Mar 17, 2025
vipannalla
approved these changes
Mar 18, 2025
… for updating the AdapterTensorstore metadata. It is fixed by design pattern with Unsafe methods and delegating the responsibility of acquiring the locks to the caller. These unsafe methods are for internal only usage. All public methods are safe and aquiring locks for themselves. - Added unit tests for LoRA implementation workflow and adapter_tensorstore.
…ewly added tests for coverage.
vipannalla
approved these changes
Apr 13, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Supporting Multi-LoRA inferencing via JetStream server following LLM Inference gateway API protocols.