R Observatory: Metrics for localLLM version 1.3.0

Balamuta, James Joseph

doi:10.5281/zenodo.21843040

localLLM

1.3.0Package record

Running Local LLMs with 'llama.cpp' Backend

0packages depend

3.3Kdownloads / year

35.3%test coverage

13/13flavors OK

Overview

About

Maintained by Yaosheng XuCRAN page ↗GitHub ↗Issues ↗

Provides R bindings to the 'llama.cpp' library for running large language models. The package uses a lightweight architecture where the C++ backend library is downloaded at runtime rather than bundled with the package. Package features include text generation, reproducible generation, and parallel inference.

Changes in 1.3.0

New Functions

**model_metadata(model)** — returns all GGUF key-value metadata as a named character vector. Useful for inspecting model architecture, quantization type, and embedded chat template. Example: model_metadata(model)["tokenizer.chat_template"].

Bug Fixes

**Fixed apply_chat_template() failing for Gemma 4 models** — Gemma 4 uses a <|turn> / <turn|> chat format not recognized by llama_chat_apply_template() (returns -1). The fallback now calls common_chat_templates_apply() from common/chat.cpp, which executes the Jinja2 template embedded in the model's GGUF directly. This works for any model with a valid Jinja2 template regardless of the C API whitelist. enable_thinking defaults to true, so Gemma 4 generates thinking content naturally without a pre-closed thought block. Tool calls and multimodal content are not handled.
**Fixed stop token leaking in generate() and generate_parallel() for ChatML models (OLMo, Llama 3)** — Two separate issues fixed:
*Windowed find*: OLMo tokenizes <|im_end|> as 6 separate pieces, with the last piece being >\n (merging > and newline). The previous exact-suffix check failed because the response ended with <|im_end|>\n instead of exactly <|im_end|>. Changed to a windowed find() that searches within the last stop.size() + 4 bytes and truncates at the match position. Applied to both generate() and generate_parallel().
*<|start_header_id|> loop*: Llama 3.2 3B sometimes omits <|eot_id|> and jumps directly to <|start_header_id|> to begin a new turn, causing infinite repetition. Added <|start_header_id|> to text_stop_strings in both functions.
generate()'s stop list also expanded from {"<turn|>", "<end_of_turn>"} to {"<turn|>", "<end_of_turn>", "<|eot_id|>", "<|im_end|>", "<|start_header_id|>"} to match generate_parallel().

Backend

**Upgraded llama.cpp backend from b8664 → b8766** — 102 builds of improvements. Gemma 4 audio conformer encoder support, various bug fixes. The Gemma 4 template issue is not fixed upstream in this version; our local fallback is the active workaround.

Bug Fixes

**Fixed verbosity not forwarded in quick_llama()** — verbosity parameter was accepted but silently dropped when passed through to .generate_single() and .generate_multiple(), so backend logging level had no effect during quick_llama() calls. Now correctly forwarded to generate() and generate_parallel().

Show all changes in 1.3.0Hide the rest of the changes in 1.3.0

**Fixed backend errors crashing R instead of being catchable** — All Rcpp::stop() calls in src/interface.cpp replaced with Rf_error(). stop() throws a C++ exception which crosses the C boundary (.Call() registration) and triggers std::terminate(), killing the R process. Rf_error() uses longjmp which R's condition system can intercept, so tryCatch() now works correctly for all backend errors including the OOM guard.
**Fixed model-loading progress dots leaking to stderr with verbosity = 0** — llama_model_load_from_file() has its own progress_callback that prints dots to stderr independently of the log callback system. Now set to a no-op when verbosity < 2 in localllm_model_load_safe(). Model loading is fully silent at the default generation verbosity.

Behavior Changes

**generate_parallel(progress) now defaults to interactive()** — previously defaulted to TRUE, which printed carriage-return-based progress bars to log files and R CMD check output. The new default shows the progress bar only in interactive R sessions and suppresses it in scripts and automated checks.
**quick_llama(progress) now defaults to interactive()** — same rationale as above; no effect on single-prompt calls.

API Changes

**Removed quick_llama(stream) parameter** — the stream argument was present in the function signature but was never passed to any downstream function (it was placeholder code with a comment "available for future use"). Removed to avoid user confusion.

Backend

**New localllm_set_verbosity() C API** — added to the backend binary and wired through the proxy layer (proxy.h/cpp, interface.cpp, init.cpp). Enables per-call verbosity control at the C level (integer 0–3, negative = fully silent). Called automatically by generate(), generate_parallel(), model_load(), and context_create() before each C invocation.
**C-layer OOM crash guard in localllm_model_load_safe()** — added a last-resort memory check that fires even when check_memory = FALSE. If the model file is larger than total physical RAM, the function now returns a clean error ("Model file (X.X GB) exceeds total physical RAM (Y.Y GB)...") instead of proceeding to llama_model_load_from_file() and letting macOS OOM-kill the R process silently. The guard only blocks provably-impossible loads (file size > total RAM) and does not interfere with the existing R-layer check. Supported on macOS (sysctl hw.memsize), Linux (/proc/meminfo MemTotal), and Windows (GlobalMemoryStatusEx).

Known Issues

**R-layer model_load() messages not suppressed by verbosity = 0** — Two R-level message() calls in api.R ("Using cached model: ..." and the GPU/unified-memory info line) print unconditionally regardless of verbosity. The verbosity parameter controls only the C backend log level; these R-layer informational messages are a separate code path not yet gated on verbosity. Confirmed against Gemma 4 26B-A4B (IQ2_XXS) on 2026-04-12.

Documentation

**Verbosity dual-default design now documented** — generate() and generate_parallel() roxygen entries now explain why they default to 0L (called in loops, per-call logs would be noisy) and cross-reference model_load()/context_create() (default 1L, run once per session, warnings should be visible).

---

From NEWS.md in the release

Install

R (CRAN)pak (dev)Ubuntu (r2u)Fedora (COPR)openSUSE (OBS)

install.packages("localLLM", repos = "https://cloud.r-project.org/")

Installs the current CRAN release. Prebuilt binaries on Windows and macOS.

if (!requireNamespace("pak", quietly = TRUE))
  install.packages("pak", repos = "https://cloud.r-project.org/")
pak::pak("EddieYang211/localLLM")

Development version from EddieYang211/localLLM. Installs pak first if it is not already available.

Release

# enable cranapt (r2u) if it is not already set up, then install
grep -rqs r2u.stat.illinois.edu /etc/apt/sources.list /etc/apt/sources.list.d/ \
  || curl -fsSL https://raw.githubusercontent.com/eddelbuettel/r2u/master/inst/scripts/add_cranapt_noble.sh | sudo bash
sudo apt install -y r-cran-localllm

Fast Ubuntu binaries via r2u (no compilation). The cranapt setup runs once and is skipped when it is already enabled. We list this channel from recorded r2u download activity rather than from a live index of the repository, so treat it as last known rather than guaranteed.

bspm (used by r2u) also needs make: sudo apt install make

# enable the iucar/cran COPR if it is not already present, then install
ls /etc/yum.repos.d/*iucar*cran*.repo >/dev/null 2>&1 \
  || sudo dnf copr enable -y iucar/cran
sudo dnf install -y R-CRAN-localLLM

RPMs from the iucar/cran COPR, rebuilt from CRAN.

For install.packages() integration in R, also: sudo dnf install R-CoprManager

# add the autoCRAN OBS repo (Tumbleweed) if it is missing, then install
zypper -q lr | grep -qi autoCRAN \
  || sudo zypper addrepo https://download.opensuse.org/repositories/devel:/languages:/R:/autoCRAN/openSUSE_Tumbleweed/devel:languages:R:autoCRAN.repo
sudo zypper install -y R-localLLM

RPMs auto-built from CRAN on openSUSE OBS (Tumbleweed). We list this channel from recorded autoCRAN download activity rather than from a live index of the repository, so treat it as last known rather than guaranteed.

Health

CRAN checks

13OK

CRAN check page ↗

Slowest check: 3.4 min · r-devel-windows-x86_64

Status changes

OK13 OK2026-08-05
NOTE12 OK, 1 NOTE2026-08-01
OK14 OK2026-04-22
ERROR13 OK, 1 ERROR2026-04-18
OK14 OK2026-03-10

Advisories

none known

How this is measured·How advisories are found

Documentation

In release 1.3.0

READMEYes · 165 wordsVignettesYes · 7 · staticNEWSYes · 100% structured · mentions 1.3.0Citation fileNo

Help pages: 34 (1 internal)includes a package overview page
Examples that run: 13%2 run in a plain check · 1 only when donttest examples are enabled (R CMD check --as-cran) · 13 never run
Exported functions with examples: 50%
Documented parameters: 96%
Return-value docs: 100%
Help pages citing references: 3%

Vignettes (7)

How these are measured

Downloads

3.3K

CRAN downloads in the past year

Rank #12,634 · ~9/day · ~275/mo

Releases R versions

Daily download trend is not available in this view yet.

33330 days

88390 days

3.3K1 year

Compare downloads with other packages →

Also on405 r2u5 autocran

How downloads are counted

Repository

eddieyang211/localllm ↗Account: EddieYang211

10Stars

1Forks

0Open issues

1Open PRs

6Releases

109Commits

5Commit identities

109 commits · Last activity 2026-05-15 · 0% stars, 30d

Responsiveness

How these are sampled

PRs merged (all time): 100%
Median days to close, 50 most recently updated closed issues: 18
Median days to close, 50 most recently updated closed PRs: 2

Stars over time

2025-10-21 · 22026-07-21 · 10

How these are measured

PracticesBeta

7 development-tooling and community-health practices detected across 5 families in the upstream repository

Checks run against github.com/eddieyang211/localllm on 2026-10-18.

Continuous integration (2)

GitHub ActionsR CMD check workflow

The R CMD check workflow runs on Linux, macOS and Windows, and on R-devel

Docs source (2)

GitHub Pages site last deployed 2026-04-13 · The repository has a site deploy workflow

Coverage (1)

Coverage reporting (Codecov)

Lint, format, editor (1)

VS Code

Show all practicesHide 1 practice

Git structural (1)

Submodules

How this is detected·Detection ruleset v3 (2026-09-26)

AI coding toolsBeta

This repository shows signs of AI coding tools.

GitHub Copilotfirst seen 2026-01-28
Pull request: opened by its own account
Claude Codefirst seen 2026-02-16
Named in .gitignore: .claude
Commits: at least 3 crediting it in the message

Models named in its commits

Claude Sonnet 4.63 commits · Apr 2026

Counts are commits, not lines of code. Commits by a tool's accounts count once across all of its accounts. Commits crediting it are the largest count among its credit lines, a minimum.How this is detected·Detection ruleset v4 (updated 2026-09-27)

Dependencies

Declared dependencies

11 external dependencies (excludes base and recommended)

Installing localLLM adds 7 packages (4 with compiled code, 4.5 MB of source tarballs). Heaviest: Rcpp 1.8 MB, jsonlite 1.0 MB, curl 714 KB.

Depends (1)

R >= 3.6.0

Imports (7)

Rcpptoolsutilsjsonlite digest curl R.utils

LinkingTo (1)

Rcpp

Suggests (5)

testthat covr irr knitr rmarkdown

Enhances (0)

none

Dependency license breakdown →

Reverse dependencies

0direct

0indirect

Nothing depends on this yet.

How these are counted

Code & Tests

Datasets

People & History

People (2)

Yaosheng Xu

Author, Maintainer since 1.0.1 (2025)

Eddie Yang

Author

Releases (5)

Pick two releases to compare their code metrics.

1.3.0Latest
record
2026-05-05 · current release · diff ↗
R
R 4.6.0 released · 2026-04-24
1.2.1
record
2026-02-26 · diff ↗
1.2.0
record
2026-02-17 · diff ↗
1.1.0
record
2025-12-17 · diff ↗ ·
added 9 exports
hardware_profile, list_ollama_models, explore, annotation_sink_csv, document_start, document_end, compute_confusion_matrices, intercoder_reliability, validate
1.0.1
record
2025-10-15
R
R 4.5.0 released · 2025-04-11

Package metadata

First published: 2025-10-15
Total releases: 5 / 1 yrs
License: MIT + file LICENSE OSI
Minimum R: ≥ 3.6.0
Bundled data: 11 KB / 1 file
Download size: 187 KB

Cite

Cite this package

Run in R for the authors' preferred citation:

citation("localLLM")

Format

Xu, Y., & Yang, E. (2026). localLLM: Running Local LLMs with 'llama.cpp' Backend (Version 1.3.0) [Computer software]. https://doi.org/10.32614/CRAN.package.localLLM

This package has no citation file, so citation() returns this entry built from DESCRIPTION.

Cite the R Observatory

For a number measured here: a download total, a coverage figure, an archival date.

Format

APA

Balamuta, J. J. (2026). R Observatory: Metrics for localLLM version 1.3.0 [Data set]. HJJB, LLC. Data release v2026-10-21. https://doi.org/10.5281/zenodo.21843040

From data release v2026-10-21, which the citation names so these numbers can be found later. More on citing and the projects behind them.

Report a problem with this page →