Modelpublic

simaai/LFM2-VL-1.6B-Autoround-Safetensors

sourceHugging Faceotherupdated 26d agoView on Hugging Face

0likes45downloads

Model Card

LFM2-VL-1.6B Autoround

This is a post-training quantized version of `LiquidAI/LFM2-VL-1.6B`, prepared as a Hugging Face checkpoint for LLiMa compilation and Sima.ai deployment. The source model's license and use restrictions continue to apply.

Source revision: Not captured; pin an immutable upstream revision before release.

Quantization

Component	Method	Weight format	Details
Decoder Linear layers	AutoRound	symmetric INT4, G256	92 targets; 200 iterations; batch size 1
`lm_head`	GPTQ	symmetric INT4, G256	1 target; static act-order; block size 128; dampening 0.01
Vision encoder Linear layers	GPTQ	symmetric INT8, per-channel	157 targets; static act-order
Mixed-precision exceptions	—	BF16	2 multimodal-projector Linear layers

Decoder calibration used NeelNanda/pile-10k (revision not captured), 512 samples of 1,024 tokens, seed 42. Head and vision calibration used lmms-lab/flickr30k (revision not captured), test[:512], 512 deterministic image-text samples, sequence length 2,048, and batch size 1. Exact targets are listed in recipe.yaml.

Evaluation

Full MMStar used all 1,500 examples, VLMEvalKit commit 7055d3010c38ccb5dcae1bc9535ca19c7fe5d79f, deterministic generation, and local exact matching on 2026-07-17.

Checkpoint	Overall accuracy	Status
Source	49.6667%	Full MMStar
This quantized checkpoint	48.2667%	Full MMStar
Absolute change	-1.4000 percentage points	Higher is better
Relative change	-2.8188%	Relative to source

Reproduction

bash

python quantize.py \
  --model-path /project/mlasw/share/huggingface/models--LiquidAI--LFM2-VL-1.6B \
  --output-dir /path/to/LFM2-VL-1.6B-Autoround-Safetensors

The directory includes quantize.py, recipe.yaml, and versions.txt.

Environment

Exact Python, CUDA, Torch, Transformers, llmcompressor, AutoRound, and compressed-tensors versions are recorded in versions.txt.

Deployment

This is the pre-LLiMa checkpoint. Compile it separately for Sima.ai hardware and keep compiler output separate. No upload is authorized for this VLM batch.

Limitations

Quantization quality varies by language, visual domain, prompt format, context length, and runtime. Validate the intended deployment workload independently.