Downloads · 30 days
11
9% of all-time downloads
pandelis/laguna-xs2-phone-k16-litert
laguna-xs2-phone-k16-litert is a text generation model from pandelis. Use it when you need the model to write or continue text. It is set up for litert-lm. The card lists the license as apache-2.0.
This is a phone-scale, non-distilled derivative of poolside/Laguna-XS.2. It is a checkpoint-surgery experiment: the original Laguna XS.2 weights are statically pruned from 256 routed experts to 16 retained routed expe…
Downloads · 30 days
11
9% of all-time downloads
All-time downloads
118
Public
Repo size
12 GB
Likes
1
Public
Click a slice to open those files.
.litertlm6 GB · 100%
From the Hugging Face model README
This is a phone-scale, non-distilled derivative of
poolside/Laguna-XS.2. It is a
checkpoint-surgery experiment: the original Laguna XS.2 weights are statically
pruned from 256 routed experts to 16 retained routed experts per sparse layer,
context metadata is reduced, and the result is packaged as a Google LiteRT-LM
artifact for Android experiments.
This is not a smoke model, LoRA, adapter, or distilled student. It is a stripped Laguna XS.2 checkpoint.
Upload and use:
model.litertlm
Local source before publishing:
../models/laguna-phone-k16-litert-wi8-skip-bmm/model.litertlm
Artifact details:
| Item | Value |
|---|---|
| Format | LiteRT-LM package |
| Size | 5,990,944,672 bytes |
| SHA256 | d1b3b5bbe37c6f7483d82b3066c8277d73e85fff62c66c9a03d89ffef0fcf0b2 |
| LiteRT-LM version | 1.5.0 |
| Retained experts | 16 per sparse layer |
| Experts per token | 8 |
| HF context metadata | 4096 tokens |
| LiteRT cache length | 1024 tokens |
litert-lm-peek verified that the package contains LLM metadata, HF tokenizer,
prefill/decode TFLite, and embedder TFLite sections.
The packaged chat template removes Hugging Face's {% generation %} /
{% endgeneration %} assistant-mask markers because Google Edge Gallery's
LiteRT template parser rejects that extension during template creation.
The LiteRT prefill/decode graph uses dynamic int8 weight quantization for
supported ops. BATCH_MATMUL is intentionally left unquantized in the current
recipe because AI Edge Quantizer fails Laguna's high-rank expert batch-matmul
with an axis-remapping error.
That makes this a real but still large high-memory-phone candidate. The next compression step is to make the expert batch-matmul path quantizable or repeat the same flow for a K8 variant.
This project folder is the reproducibility and publishing bundle:
scripts/ checkpoint surgery, GGUF-to-HF conversion, LiteRT export/package helpers
docs/ reproduction notes and Android loading direction
experiments/ failed tensorwise BATCH_MATMUL recipe kept for reference
manifest.json machine-readable artifact metadata and hashes
publish_hf.ps1 Hugging Face upload helper
The generated model payloads remain under ../models/ locally and are uploaded
by publish_hf.ps1.
Log in to Hugging Face first:
huggingface-cli login
On this workstation, publish_hf.ps1 will automatically use the WSL LiteRT
Python at /root/.cache/lucebox-litert-venv/bin/python if it exists. You can
also set HF_TOKEN in PowerShell before publishing.
Dry-run the upload:
.\publish_hf.ps1 -RepoId your-user/laguna-xs2-phone-k16-litert -DryRun
Create or update the model repo and upload model.litertlm:
.\publish_hf.ps1 -RepoId your-user/laguna-xs2-phone-k16-litert
Use -IncludeTfliteParts if you also want the unpacked TFLite components in the
repo. Use -IncludePrunedGguf only if you want to publish the intermediate
K16 GGUF as well.
Copy model.litertlm to app-private storage or bundle it in a test build, then
load it with a LiteRT-LM text-generation session. This artifact is for LiteRT-LM,
not Google AICore; AICore is for Google-managed Gemini Nano weights rather than
custom Laguna weights.
This checkpoint has not yet been quality-evaluated after expert pruning on the target phone. It should be treated as an experimental derived model for local testing, not as a quality-preserving release of Laguna XS.2.