quant_meta.json

{
  "model_id": "HuggingFaceTB/SmolVLM-256M-Instruct",
  "family": "smolvlm",
  "method": "hqq",
  "bits": 4,
  "group_size": 64,
  "skip_modules": [
    "vision_model",
    "vision_tower",
    "visual",
    "connector",
    "multi_modal_projector",
    "mlp1",
    "lm_head",
    "embed_tokens",
    "patch_embed"
  ],
  "load_time_s": 0.5,
  "quant_time_s": 0.7,
  "n_quantized_layers": 210,
  "n_skipped_layers": 74,
  "format": "Saved as state_dict.pt with hqq-quantized Linear layers in the LLM backbone. Vision encoder, connector and lm_head are FP16. To load: rebuild model in FP16, run the same Linear->HQQLinear replacement, then load state_dict."
}