Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,68 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: Kezmark/Mordant-3B-Think
|
| 4 |
+
base_model_relation: quantized
|
| 5 |
+
tags:
|
| 6 |
+
- litert
|
| 7 |
+
- litert-lm
|
| 8 |
+
- on-device
|
| 9 |
+
- granite
|
| 10 |
+
- reasoning
|
| 11 |
+
- image-prompt
|
| 12 |
+
language:
|
| 13 |
+
- en
|
| 14 |
+
pipeline_tag: text-generation
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# Mordant-3B-Think — LiteRT-LM (on-device)
|
| 18 |
+
|
| 19 |
+
On-device conversion of [Kezmark/Mordant-3B-Think](https://huggingface.co/Kezmark/Mordant-3B-Think) —
|
| 20 |
+
a full fine-tune of [ibm-granite/granite-4.1-3b](https://huggingface.co/ibm-granite/granite-4.1-3b)
|
| 21 |
+
for AI image-generation prompt composition with chain-of-thought reasoning — to a `.litertlm`
|
| 22 |
+
bundle for the [LiteRT-LM](https://github.com/google-ai-edge/LiteRT-LM) runtime. All credit for
|
| 23 |
+
the model itself goes to its author; this repo only packages it for phones and desktops.
|
| 24 |
+
|
| 25 |
+
**Requires litert-lm ≥ 0.16 to run.**
|
| 26 |
+
|
| 27 |
+
| file | quant | size |
|
| 28 |
+
|---|---|---|
|
| 29 |
+
| `Mordant-3B-Think_int8.litertlm` | dynamic int8 (linears + embedding) | 3.76 GB |
|
| 30 |
+
|
| 31 |
+
## Conversion & verification
|
| 32 |
+
|
| 33 |
+
Converted with one command by [hf-to-litertlm](https://github.com/john-rocky/hf-to-litertlm)
|
| 34 |
+
(`python scripts/convert.py Kezmark/Mordant-3B-Think`, 2026-08-25):
|
| 35 |
+
|
| 36 |
+
- The finetune's own chat template — a thinking-form template that opens the assistant turn
|
| 37 |
+
with `<think>` — is embedded verbatim (byte-equal to the checkpoint's
|
| 38 |
+
`chat_template.jinja`, 1474/1474).
|
| 39 |
+
- The spurious metadata start token is dropped: this family declares `bos == eos ==
|
| 40 |
+
<|end_of_text|>` and its template never renders a leading BOS, so an engine-prepended
|
| 41 |
+
start token reads as "this document already ended" — measured on this checkpoint, it flips
|
| 42 |
+
HF bf16 greedy output into a code-fence loop.
|
| 43 |
+
- Reduced 7-signature prefill ladder + externalized embedder (the ≥3B ship shape; the full
|
| 44 |
+
11-signature ladder is killed by iOS at Metal init on this family).
|
| 45 |
+
- Quality gate: **8/8** on the 8-question sanity gate (think-aware budget), non-degenerate.
|
| 46 |
+
|
| 47 |
+
## Performance (Apple M4 Max, litert-lm 0.16.0, `-p 256 -d 256 --runs 3 --cache no`)
|
| 48 |
+
|
| 49 |
+
| backend | prefill tok/s | decode tok/s | TTFT |
|
| 50 |
+
|---|---:|---:|---:|
|
| 51 |
+
| CPU | 97.3 | 20.2 | 2.68 s |
|
| 52 |
+
| GPU | 1129 | 71.9 | 0.24 s |
|
| 53 |
+
|
| 54 |
+
## Usage
|
| 55 |
+
|
| 56 |
+
```bash
|
| 57 |
+
pip install litert-lm
|
| 58 |
+
litert-lm run Mordant-3B-Think_int8.litertlm \
|
| 59 |
+
--prompt "A cat sitting on a windowsill at sunset" --max-num-tokens 4096
|
| 60 |
+
```
|
| 61 |
+
|
| 62 |
+
The model answers with a `<think>…</think>` block followed by the composed image prompt —
|
| 63 |
+
budget generation length accordingly. On Android, load the bundle in an app embedding the
|
| 64 |
+
LiteRT-LM engine (e.g. Google AI Edge Gallery-style hosts).
|
| 65 |
+
|
| 66 |
+
## License
|
| 67 |
+
|
| 68 |
+
apache-2.0, inherited from the source model and its granite-4.1 base.
|