Modelpublic

simaai/Gemma3-SigLIP448-GPTQ-Safetensors

sourceHugging Faceotherupdated 26d agoView on Hugging Face

0likes28downloads

Model Card

Gemma3-SigLIP448 GPTQ

This is a local post-training quantized checkpoint of stribomon/gemma3-siglip448 for later LLiMa compilation and deployment on Sima.ai hardware. It must not be uploaded as part of the current VLM batch.

Source revision: Not captured in the flattened local source directory. An immutable upstream revision must be recovered and recorded before any release or redistribution.

Source license metadata: not declared in the local source card. This checkpoint remains subject to the upstream license, intended-use guidance, and limitations.

Quantization

Component	Method	Weight format	Effective targets
Decoder Linear layers and `lm_head`	GPTQ	symmetric INT4, group size 256, static act-order	239
Vision encoder Linear layers	RTN	symmetric INT8, per-channel	162
Projector	none present	source precision	Explicitly recorded in `recipe.yaml`

The lm_head follows the default GPTQ INT4/G256 policy; no INT8 escalation was made. Text GPTQ calibration used lmms-lab/flickr30k test[:512], deterministic dataset order, 512 image-text samples, sequence length 2048, batch size 1, no concatenation, and no padding to maximum length. Vision RTN does not use Hessian calibration. The exact resolved targets and effective settings are saved in recipe.yaml.

Evaluation

Full MMStar used all 1,500 examples, VLMEvalKit commit 7055d3010c38ccb5dcae1bc9535ca19c7fe5d79f, deterministic generation, and exact_matching without an API judge.

Checkpoint	Overall accuracy
Source	42.8000%
Existing text GPTQ + vision GPTQ checkpoint	42.4000%
This text GPTQ + vision RTN checkpoint	43.0667%
RTN change versus vision GPTQ	+0.6667 percentage points
RTN relative change versus vision GPTQ	+1.5723%

Evaluation date: 2026-07-17. Slurm jobs: source 4684, vision-GPTQ baseline 4685, and this vision-RTN candidate 4737. The comparison is recorded in vlmeval_results/mmstar_results_vision_rtn.csv; raw predictions and status files remain under vlmeval_results/mmstar_full_vision_rtn/.

Validation: MMStar inference passed (1,500/1,500).

Full matched WikiText-2 word perplexity (EleutherAI/wikitext_document_level, wikitext-2-raw-v1, no example limit; 2026-07-18): source 23.0695; this checkpoint 31.3585; absolute degradation +8.2890; relative degradation +35.93%. Raw JSON evidence: perplexity_results/vlm_full_wikitext/gemma3_siglip448_{source_v2,gptq_w4g256_visionrtn_w8_v3}.json.

Reproduction

This directory contains the exact quantize.py, recipe.yaml, and versions.txt used for the artifact:

bash

python quantize.py \
  --model-path /path/to/source-model \
  --output-dir /path/to/new-quantized-model

Environment

text

Python: 3.13.2
torch: 2.11.0+cu128
CUDA: 12.8
transformers: 5.10.1
llm-compressor: 0.12.0
auto-round: 0.13.0
compressed-tensors: 0.17.1

Deployment

This is the pre-LLiMa quantized Hugging Face artifact. LLiMa compiler output must remain a separate deployment artifact unless its format and redistribution are approved. No Hugging Face upload is authorized for this batch.

Limitations

Quantization quality can vary by language, visual domain, prompt format, context length, and downstream runtime. MMStar does not replace application-specific validation. Users must validate their own deployment and compiler/runtime compatibility.