README.md

KVCacheIndex Example

This example demonstrates how to configure and use the kvcache.Indexer module from the llm-d-kv-cache project.

What it does

• Initializes a kvcache.Indexer with optional Redis, in-memory backend, or cost-aware memory.
• Optionally uses a HuggingFace token for tokenizer pool configuration.
• Demonstrates adding and querying KV cache index entries for a model prompt.
• Shows how to retrieve pod scores for a given prompt.

Usage

1. Set environment variables as needed:

   • REDIS_ADDR (optional): Redis connection string (e.g., redis://localhost:6379/0). If unset, uses in-memory index.
   • HF_TOKEN (optional): HuggingFace token for tokenizer pool.
   • MODEL_NAME (optional): Model name to use (defaults to test data).

2. Run the example:

make run-example kv_cache_index

3. What to expect:

   • The program will print logs showing the creation and startup of the indexer.
   • It will attempt to get pod scores for a test prompt (initially empty).
   • It will manually add entries to the index and then retrieve pod scores again.

Example output

I... Created Indexer
I... Started Indexer {"model": "Qwen/Qwen2-VL-7B-Instruct"}
I... Got pods        {"pods": {}}
I... Got pods        {"pods": {"pod1":4}}

See also

• main.go for the full example code.
• testdata for sample prompts and model names.