Search results

quantize : configurable neutral imatrix prior

examples

generation qualityQuality of model outputQuality of model output

need feedbackTesting and feedback with results are neededTesting and feedback with results are needed

research 🔬

Status: Draft (not ready).

ggml-org/llama.cppnumber 15060

#15060 In ggml-org/llama.cpp;

· compilade opened on Aug 3, 2025

ggml-quants : weighted rounding algorithms with cumulative search

generation qualityQuality of model outputQuality of model output

ggmlchanges relating to the ggml tensor library for machine learningchanges relating to the ggml tensor library for machine learning

Less than 4 bitsEfforts related to viable quantized models using <4 bitsEfforts related to viable quantized models using <4 bits

research 🔬

Review Complexity : MediumGenerally require more time to grok but manageable by beginner to medium expertise levelGenerally require more time to grok but manageable by beginner to medium expertise level

Tensor Encoding Schemehttps://github.com/ggerganov/llama.cpp/wiki/Tensor-Encoding-Schemeshttps://github.com/ggerganov/llama.cpp/wiki/Tensor-Encoding-Schemes

Status: Draft (not ready).

ggml-org/llama.cppnumber 12557

#12557 In ggml-org/llama.cpp;

· compilade opened on Mar 25, 2025

Smooth Sampling / Quadratic Sampling support

generation qualityQuality of model outputQuality of model output

performanceSpeed related topicsSpeed related topics

Review Complexity : HighGenerally require indepth knowledge of LLMs or GPUsGenerally require indepth knowledge of LLMs or GPUs

Status: Open (in progress).

ggml-org/llama.cppnumber 6445

#6445 In ggml-org/llama.cpp;

· kalomaze opened on Apr 2, 2024

P-Step Truncation Sampling

generation qualityQuality of model outputQuality of model output

need feedbackTesting and feedback with results are neededTesting and feedback with results are needed

refactoringRefactoringRefactoring

Review Complexity : HighGenerally require indepth knowledge of LLMs or GPUsGenerally require indepth knowledge of LLMs or GPUs

Status: Open (in progress).

ggml-org/llama.cppnumber 5675

#5675 In ggml-org/llama.cpp;

· p-e-w opened on Feb 23, 2024

[RFC] common, server : add top-a sampler

enhancementNew feature or requestNew feature or request

generation qualityQuality of model outputQuality of model output

Review Complexity : HighGenerally require indepth knowledge of LLMs or GPUsGenerally require indepth knowledge of LLMs or GPUs

Status: Open (in progress).

ggml-org/llama.cppnumber 5612

#5612 In ggml-org/llama.cpp;

· Artefact2 opened on Feb 20, 2024

Penalty threshold: A mechanism for improving repetition penalties

enhancementNew feature or requestNew feature or request

generation qualityQuality of model outputQuality of model output

Review Complexity : MediumGenerally require more time to grok but manageable by beginner to medium expertise levelGenerally require more time to grok but manageable by beginner to medium expertise level

Status: Open (in progress).

ggml-org/llama.cppnumber 5561

#5561 In ggml-org/llama.cpp;

· p-e-w opened on Feb 18, 2024

llama : combined beam search + grammar sampling strategy

generation qualityQuality of model outputQuality of model output

good first issueGood for newcomersGood for newcomers

research 🔬

roadmapPart of a roadmap projectPart of a roadmap project

Status: Open.

#2923 In ggml-org/llama.cpp;

· ggerganov opened on Aug 31, 2023

llama : tool for evaluating quantization results per layer

enhancementNew feature or requestNew feature or request

generation qualityQuality of model outputQuality of model output

roadmapPart of a roadmap projectPart of a roadmap project

Status: Open.

#2783 In ggml-org/llama.cpp;

· ggerganov opened on Aug 25, 2023

Implementation of a sequence repetition penalty sampler

enhancementNew feature or requestNew feature or request

generation qualityQuality of model outputQuality of model output

need feedbackTesting and feedback with results are neededTesting and feedback with results are needed

Status: Draft (not ready).

ggml-org/llama.cppnumber 2593

#2593 In ggml-org/llama.cpp;

· KerfuffleV2 opened on Aug 12, 2023

Study how LM Evaluation Harness works and try to implement it

enhancementNew feature or requestNew feature or request

generation qualityQuality of model outputQuality of model output

help wantedNeeds help from the communityNeeds help from the community

high priorityVery important issueVery important issue

research 🔬

Status: Open.

#231 In ggml-org/llama.cpp;

· ggerganov opened on Mar 17, 2023