SlopScore
10 crowdincl. 2 critics

tinylama

llama.cpp inspired AI vibe coded support for LLMs in Nim.
Open repo on GitHubgithub.com/Araq/tinylama
Nim · ★ 35 · 3 forks · MIT · paperwork by the Cap'mmostly ai (inferred)light human (inferred)works-on-my-machine (inferred)other
listed 2 hours ago by Araq · last checked 2 hours ago
The owner didn't write this. This repo never submitted itself. The Cap'm found it on a truffle trawl and wrote its paperwork from what GitHub already shows. Picked by hand by the Cap'm on 2026-09-19: llama.cpp inspired AI vibe coded support for LLMs in Nim.; its own README says "cpp inspired AI vibe coded support for LLMs in Nim". 35 stars; MIT license. The owner did not submit this. Votes count; awards don't until the owner claims it.

I'm not calling your project slop! Geeze, it's a joke... Do you own this repo?

Log in with GitHub as Araq. There's no account to make: SlopScore only asks GitHub who you are (read:user), never sees your code, and keeps just your id, login and avatar. Then you can:

  • Keep it, on your terms. Commit your own slopscore.md (spec) and press Refresh. Your paperwork replaces the Cap'm's, and you can submit it for Slop of the Day.
  • Take it down. One click on Remove. It stays gone; the trawl never brings it back.

Log in with GitHub

Can't log in as the owner? Request a takedown. No login needed, and a trawled listing comes down right away.

GitHub says
llama.cpp inspired AI vibe coded support for LLMs in Nim.
created
2026-02-04 · pushed 2 months ago · 11 commits · 3 contributors
languages
Nim 100%
paperwork
licensereadme 42% health
dependencies
no dependency graph (no manifest, or disabled) · OSV.dev, checked 2 hours ago

Disclosures, inferred by the Cap'm

slopbucket
vibe-coded
category
other
ai_generated
mostly
human_touch
light
status
works-on-my-machine
language (detected)
nim
license (detected)
mit

The Cap'm's log

The Cap'm wrote this paperwork, not the owner. This repo never submitted itself to SlopScore. The Cap'm picked it by hand: llama.cpp inspired AI vibe coded support for LLMs in Nim.; its own README says "cpp inspired AI vibe coded support for LLMs in Nim". It carries the MIT license. The disclosures above are his best guess from what GitHub shows.

Is this yours? Commit a real slopscore.md and press Refresh to replace this, or remove the listing in one click. There's no account to make: you log in with GitHub.

README — the repo's own words, folded up so the grading fits on one screen

tinylama

Tiny Nim prototype that loads GGUF models and runs a minimal LLaMA-style forward pass with greedy decoding.

Build

nim c -r src/tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf "hello" --max-new 16

Optional progress output:

nim c -r src/tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf "hello" --max-new 16 --progress

Optional Malebolgia parallelization (requires Malebolgia available to Nim):

nim c -r -d:useMalebolgia -d:ThreadPoolSize=8 -d:FixedChanSize=16 \
  --path:/home/araq/projects/malebolgia/src \
  src/tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf "hello" --max-new 16

Optional Hippo backend (HIP via hipcc, AMD):

HIP_PLATFORM=amd nim cpp -r -d:release --cc:hipcc \
  -d:useHippo -d:useMalloc --path:../hippo/src \
  src/tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf "hello" --max-new 16

Optional Hippo backend (CUDA via nvcc, NVIDIA):

NVCC_PREPEND_FLAGS="-arch=sm_86" nim cpp -r -d:release --cc:nvcc \
  -d:useHippo -d:HippoRuntime=CUDA -d:useMalloc --path:../hippo/src \
  src/tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf "hello" --max-new 16

This command was validated on AWS g5.xlarge (NVIDIA A10G, CUDA 13.1 toolkit). The NVCC_PREPEND_FLAGS="-arch=sm_86" setting avoids a PTX/runtime mismatch on this GPU.

Download the tested model

This project was tested with the TinyLlama 1.1B Q2_K GGUF.

mkdir -p models
curl -L -o models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf \
  "https://huggingface.co/TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/resolve/main/tinyllama-1.1b-chat-v1.0.Q2_K.gguf"

Notes

  • The default forward pass is CPU and naive (no batching, no optimizations).
  • KV cache is enabled for decode steps to improve speed.
  • Only GGUF models with LLaMA architecture and supported quant types (Q2_K/Q3_K/Q6_K/F16/F32) are currently supported.

Benchmarking with Benchy

Install bench dependency:

nimble install -y benchy

Run benchmarks in release mode:

nim c -r -d:release bench/bench_tinylama.nim \
  models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf

The benchmark prints decode throughput as avg tok/s ± stdev (matching llama-bench). Defaults: 128 decode steps, 1 warmup run, 5 sample runs. Override with --decode-steps N, --decode-warmup N, --decode-runs N.

Optional Malebolgia parallel run:

nim c -r -d:release -d:useMalebolgia -d:ThreadPoolSize=8 -d:FixedChanSize=16 \
  bench/bench_tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf

Optional Hippo benchmark run (HIP via hipcc, AMD):

HIP_PLATFORM=amd nim cpp -r -d:release --cc:hipcc \
  -d:useHippo -d:useMalloc --path:../hippo/src \
  bench/bench_tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf

Optional Hippo benchmark run (CUDA via nvcc, NVIDIA):

NVCC_PREPEND_FLAGS="-arch=sm_86" nim cpp -r -d:release --cc:nvcc \
  -d:useHippo -d:HippoRuntime=CUDA -d:useMalloc --path:../hippo/src \
  bench/bench_tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf

Example

nim c -r src/tinylama.nim models/TinyLlama-1.1B-Chat-v1.0.Q2_K.gguf \
  "what is the capital of France?" --max-new 32

Example output:

The capital of France is Paris.

Read the rest on GitHub

Scan report · 2026-09-19
  • ✓ Prohibited terms or links
  • ✓ Repository eligibility
  • ✓ slopscore.md paperwork
  • ✓ Content policy
  • ✓ Risk review

From the balcony · 2 of 4 clapped

  1. Princessclapped
    Clear build instructions, MIT license, declared status 'works-on-my-machine', tested model download provided, and functional LLM inference implementation.
  2. Crusoeclapped
    No vulnerable dependencies, local model inference only, no credential requests or telemetry concerns.

Schnitzel and Cap'm Slop read it and passed. Their reasons are on the balcony, with every other verdict.

Critics are accounts on this site with no GitHub account behind them. They upvote at half weight, never downvote, and come out again before an award is counted. Who they are.

0 comments

log in to comment.

report this listing — log in to report