litert-community/lfm3-litert-toolkit

LFM3-1.7B — LiteRT toolkit

Files for running and rebuilding LFM3-1.7B through LiteRT-LM. Published as a dataset repo on purpose: it is plain HTTPS, so it can be fetched from places where the Hub's git endpoint is unreachable (e.g. a Space dev-mode VM that fails git fetch with fatal: expected 'acknowledgments').

Attribution: LFM3-1.7B is © Liquid AI under the LFM Open License v1.0. Section 5 limits commercial use by entities with annual revenue of $10M or more.

Pull a file without git

python - <<'PY'
from huggingface_hub import hf_hub_download
import shutil
src = hf_hub_download("litert-community/lfm3-litert-toolkit", "app.py",
                      repo_type="dataset")
shutil.copy(src, "/home/user/app/app.py")
print("synced app.py from", src)
PY

That works from the Space VM even while git fetch does not.

Contents

File	Purpose
app.py	The Space: int4/int8 weights radio, Load-model button, peak VmHWM, one engine per (session, weights)
requirements.txt	litert-lm==0.18.1 (Gradio comes from the Space SDK)
build_lfm3_16k.py	Re-export LFM3 at cache_length=16384; needs ≥32 GB RAM, CPU only
recipe_q1b_16k.json	Shipped q1b recipe + explicit int8 pin on the output projection
verify_lfm3_16k.py	Gate: 8 check questions → long-context needle → timing
cpt_lfm3_turkish.py	Turkish continued pre-training (fineweb-2/tur_Latn), merge, perplexity gate
qad_lfm3_int4.py	Quantization-aware distillation into the int4 scheme (fake-quant 32-block, STE, KL)

The two numbers that matter

1. VmHWM from the Load button. The container gauge read 15/16 GB, but that includes page cache, the VS Code server and node — not the engine. The engine's real resident size decides whether a 16K build (+704 MiB of KV) fits on cpu-basic at all.
2. int8 vs int4 decode rate on the same prompt. ~1 tok/s was measured on x86 with mixed int4, against 35 tok/s for the same file on an ARM phone. If int8 is materially faster here, the int4 kernels are falling back on x86 — and int8 is also the more accurate file (95.7 % vs 79.9 % top-token agreement), which would make the accuracy-recovery work unnecessary.

Known state

Nothing here has been executed except the Space itself. The Space runs the published LFM3-1.7B_mixed_int4_ekv2048.litertlm unmodified and generated a correct answer to "what is the capital of japan" on 2 vCPU. The 16K build, the Turkish CPT and the QAD run are written but unrun.