llama.cpp: update PPL w.r.t. new k-quant optimization #172
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Outdated suggestions cannot be applied.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.
Description
This PR updates PPLs for GGUF
k-quant
models.Details
llama.cpp has introduced new GGUF format and a recent PR brings improvements in
k-quant
model series. This PR is typically meant for refreshing thek-quant
models' stats to see if our model can also benefit from these improvements.Chinese-LLaMA-2-7B
old:
new (changed results are in boldface):
Chinese-LLaMA-2-13B
old:
new (changed results are in boldface):
Observations:
k-quant
models' PPL have been improved (reduced), except forq5_k
. Significant improvement inq2_k
andq3_k
.q2_k
model also have a more compact size than before. For example, 7B-q2_k
model was 2.77G, and now is 2.41G.Related Issue
None.
Explanation of Changes
copilot:walkthrough