tahamajs/RR / Docs /module_metadata /cachenotes_compressor.json
tahamajs's picture
download
raw
1.07 kB
{
"module_name": "cachenotes_compressor",
"class_name": "CacheNotesCompressor",
"category": "compression",
"family": "compression",
"params": [
"max_cache_len",
"task_type",
"keep_ratio",
"num_sinks",
"model",
"recalibrate_ln",
"tokenizer",
"calibration_loader",
"calibration_texts",
"device"
],
"ops": [
"linear",
"matmul",
"norm",
"softmax"
],
"docstring": "CacheNotes Compressor: Decouples static weight compression from KV cache dynamic pruning.\nCompresses static model weights using CPT-guided low-rank SVD without zeroing out entire output channels.\n\nArgs:\n model: PyTorch model.\n task_type: Target downstream task ('qa', 'summarize', 'code', 'reasoning', 'math').\n keep_ratio: Fraction of rank / energy to retain (default 0.5).\n recalibrate_ln: Whether to recalibrate LayerNorm statistics after compression.",
"rank": 16,
"d_state": 1,
"num_experts": 1,
"top_k": 1,
"sparsity_pct": 0.0,
"bits": 4,
"hash_val": 259550964849013102013658601722562170684
}

Xet Storage Details

Size:
1.07 kB
·
Xet hash:
d1f9ed66c29113b25d0a86b5bce3fdf356d736e92098e84fd463fa1db71325fb

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.