Faster perplexity computation #2786
Merged
Conversation
This file contains hidden or 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
ggerganov
approved these changes
Aug 25, 2023
mattgauf
added a commit
to mattgauf/llama.cpp
that referenced
this pull request
Aug 26, 2023
* master: (773 commits) server : add `/detokenize` endpoint (ggml-org#2802) convert.py : advanced option (ggml-org#2753) llama : use Unicode Escape Sequence to replace encoded characters (ggml-org#2814) flake.nix : add rocm support and cleanup (ggml-org#2808) llama : move #includes out of _GNU_SOURCE conditional (ggml-org#2817) main : fix bug (penalize_nl=false doesn't work) + suppress warning on mingw (ggml-org#1528) llama : use std::abs in llama_sample_tail_free (ggml-org#2800) k-quants : remove unnecessary tensor shape restrictions (ggml-org#2811) Better perplexity for 2- and 3-bit quantization for LLaMA-v2-70B (ggml-org#2807) Fix HellaSwag (ggml-org#2805) flake : build llama.cpp on Intel with nix (ggml-org#2795) Handle null rope scaling value (ggml-org#2793) Fix spm whitespaces (ggml-org#2806) examples : skip unnecessary external lib in server README.md how-to (ggml-org#2804) llama : fix struct decl (ggml-org#2790) Faster perplexity computation (ggml-org#2786) llama : add llama_beam_search() (ggml-org#2267) convert.py : Get rope scale from HuggingFace models (ggml-org#2772) llama-bench : add model sizes (ggml-org#2771) convert.py : export rope freq_base when converting CodeLlama from an HF model (ggml-org#2773) ...
akawrykow
pushed a commit
to akawrykow/llama.cpp
that referenced
this pull request
Aug 29, 2023
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
# for free
to join this conversation on GitHub.
Already have an account?
# 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Outdated suggestions cannot be applied.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.
Time to compute 7B perplexity with 512 context of Wikitext on an RTX-4080 goes from 143 seconds to 128 seconds. I guess, not a big deal for most people, but a nice speedup for someone like me who runs lots of perplexity calculations while trying different quantization techniques.
Also added output of the statistical uncertainty of the computed perplexity. With
--ppl-output-type 1the perplexity tool now outputs 4 columns, which are number of evaluated tokens, perplexity, average negative log probability and its uncertainty.