compaction.json

{
  "source": "/default-filesys/workspace/chenjunzhe/EdgeInstant-1.5b/runs/bilingual_s5_full/final",
  "output": "/default-filesys/workspace/chenjunzhe/EdgeInstant-1.5b/runs/bilingual_s5_full/final_compact",
  "precision": "torch.bfloat16",
  "preserved_modalities": [
    "text",
    "audio_input",
    "image",
    "video",
    "speech_output"
  ],
  "source_parameters": 2031972995,
  "exported_parameters": 2031710851,
  "removed_weights": [
    "codec.quantizer.rvq_first.input_proj.weight",
    "codec.quantizer.rvq_rest.input_proj.weight"
  ],
  "source_weight_bytes": 5677248858,
  "exported_weight_bytes": 4115251826,
  "equivalence_reference": "Source loaded with dtype=auto, as in run_mmau.py; align and audio_special remain FP32.",
  "cache_implementation": "dynamic",
  "recurrent_decode": "cuda_graph"
}