NVIDIA / NVIDIA/Model-Optimizer
Can `auto_quant` calculate the score for `kv-cache` separately?
Open
@meenchen is already working on this.
Since Aug 17, 2026.
question
- Dominant language
- Python
- Stars
- 3.8k
- Forks
- 604
- Avg merge
- 2d 8h
- Merged PRs (30d)
- 142
Description
Make sure you already checked the examples and documentation before submitting an issue.
How would you like to use ModelOpt
# Disable KV Cache quantization
# Currently KV Cache quantization is enabled for some quantization formats and disabled for others
# This breaks the monotonicity of the quantization formats in terms of weight compression Vs accuracy
The reason for disabling key-value caching is that some recipes are enabled while others are disabled, thus breaking monotonicity. If the weighted recipes, batches, and sequence are fixed, is it possible to compare kv-cached recipes? In other words, to calculate the cost/score separately for the kv-cache.
Who can help?
- ?
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.