google-deepmind / google-deepmind/gemma
Add cache usage property to ChatSampler and Sampler
- Dominant language
- Python
- Stars
- 5.7k
- Forks
- 1k
- Avg merge
- 10h 33m
- Merged PRs (30d)
- 2
Description
## Cache Usage Monitoring Enhancement
### Overview
Previously, a TODO was marked to add a property showing how much of the cache is used. This is useful for debugging and monitoring cache usage during multi-turn conversations.
---
### Previous State
- No easy way to check cache usage without accessing internal `last_state` directly
- Users had to manually compute `state.used_cache_length / cache_length`
---
### Changes
#### Added Cache Usage Monitoring
Cache usage monitoring was added to both sampler classes:
- **`ChatSampler.cache_usage` property**
- Returns a `(used, total)` tuple
- Returns `None` if no sampling has been performed yet
- **`Sampler.get_cache_usage()` static method — For stateless usage with the low-level API**
- Takes a SamplingState and cache_length as parameters
Contributor guide
Assessment
This issue has not been assessed yet.