leideng/QCFuse / blend /qcfuse_config.py
leideng's picture
download
raw
677 Bytes
"""QCFuse configuration for the SSD-backed blend runner."""
DIGEST_INDEX_LABEL = "kvzip@10%"
DIGEST_INDEX_METHOD = "kvzip"
DIGEST_RATIO = 0.1
DEFAULT_BLEND_RATIO = 0.1
DEFAULT_CONTEXT_N_SINK = 4
DEFAULT_CRITICAL_LAYERS = 3
SUPPORTED_BASELINES = ("fullcomp", "ours")
# Model-specific Top-10 critical layers. Values are 0-based layer ids and are
# consumed by the runtime critical_layers request argument.
MODEL_TOP10_CRITICAL_LAYERS = {
"llama3.1-8b": [14, 13, 17, 18, 16, 20, 19, 10, 15, 22],
"mistral-7b": [16, 19, 15, 18, 14, 17, 12, 11, 20, 9],
"qwen3-14b": [24, 26, 21, 25, 18, 28, 29, 22, 23, 20],
"qwen3-8b": [20, 21, 18, 17, 24, 23, 26, 14, 22, 19],
}

Xet Storage Details

Size:
677 Bytes
·
Xet hash:
a4db19bbe3a9816c01f4ae13cfec8ebb9c8b4b79c77f20792230aef414210f8c

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