Gemma-3-1B-it GPTQ INT4/G128 Safetensors

Pre-LLiMa Hugging Face checkpoint for Sima.ai compilation.

Quantization

All language Linear layers and lm_head: GPTQ, symmetric INT4, group size 128, static actorder.
Group size 128 is required because 1152-wide layers do not support G256.
Calibration: HuggingFaceH4/ultrachat_200k, 512 samples, 1024 tokens, batch size 1.

Full WikiText evaluation

EleutherAI/wikitext_document_level, wikitext-2-raw-v1, full matched run (2026-07-16).

Checkpoint	Word perplexity
`google/gemma-3-1b-it`	35.970619
This GPTQ checkpoint	42.386377

Reproduce

python quantize.py --model-path /project/mlasw/share/huggingface/models--google--gemma-3-1b-it --output-dir /path/to/output

Validation and limitations

Saved scales were checked for finite values. Transformers load/generation smoke test is pending. Do not publish this artifact without resolving its quality regression and recording the smoke test.