kv_quantization_report.json

{
  "created_at": "2026-07-26T06:34:21.056464+00:00",
  "source_weight_checkpoint": "/inspire/qb-ilm/project/video-understanding/public/train/moss_vl_streaming/8B/quant/mossvl_streaming_w4a16_nf4_keep_first4_last4",
  "weight_quantization": "bitsandbytes NF4 W4A16",
  "kv_cache_quantization": {
    "backend": "hqq",
    "nbits": 8,
    "axis_key": 0,
    "axis_value": 0,
    "q_group_size": 64,
    "residual_length": 128
  },
  "storage": "Model shards are hard-linked to the source checkpoint when the filesystem supports hard links.",
  "compatibility_patch": "Cross-attention materializes the quantized history one layer at a time because the original MOSS-VL code directly reads cache_layer.keys."
}