GLM-5.3-MLX / fidelity_summary.json
hugging-yiii's picture
Add files using upload-large-folder tool
9a5177c verified
Raw
History Blame Contribute Delete
2.37 kB
{
"attention@8bit": {
"n": 390,
"cos_mean": 0.9999843754304225,
"cos_min": 0.9999701142781648,
"cos_p5": 0.9999817950797752,
"snr_mean": 45.05945505567317,
"worst": [
"model.layers.0.self_attn.q_b_proj.weight",
"model.layers.7.self_attn.q_b_proj.weight",
"model.layers.2.self_attn.q_b_proj.weight",
"model.layers.3.self_attn.q_b_proj.weight",
"model.layers.3.self_attn.kv_a_proj_with_mqa.weight"
]
},
"dense_mlp@6bit": {
"n": 9,
"cos_mean": 0.9997481794588633,
"cos_min": 0.9997417464914627,
"cos_p5": 0.9997417464914627,
"snr_mean": 32.974673175379735,
"worst": [
"model.layers.2.mlp.down_proj.weight",
"model.layers.1.mlp.down_proj.weight",
"model.layers.0.mlp.down_proj.weight",
"model.layers.1.mlp.up_proj.weight",
"model.layers.2.mlp.gate_proj.weight"
]
},
"expert_down@5bit": {
"n": 19200,
"cos_mean": 0.9989847256501002,
"cos_min": 0.9981335686868937,
"cos_p5": 0.9989747346453054,
"snr_mean": 26.926658932773485,
"worst": [
"model.layers.4.mlp.experts.96.down_proj.weight",
"model.layers.6.mlp.experts.87.down_proj.weight",
"model.layers.3.mlp.experts.200.down_proj.weight",
"model.layers.5.mlp.experts.234.down_proj.weight",
"model.layers.8.mlp.experts.86.down_proj.weight"
]
},
"expert_gate_up@4bit": {
"n": 38400,
"cos_mean": 0.9956952468956632,
"cos_min": 0.9950926262273702,
"cos_p5": 0.9955739241692824,
"snr_mean": 20.655963162017926,
"worst": [
"model.layers.4.mlp.experts.168.gate_proj.weight",
"model.layers.3.mlp.experts.2.up_proj.weight",
"model.layers.4.mlp.experts.168.up_proj.weight",
"model.layers.3.mlp.experts.2.gate_proj.weight",
"model.layers.4.mlp.experts.186.gate_proj.weight"
]
},
"shared_expert@6bit": {
"n": 225,
"cos_mean": 0.9997443424916399,
"cos_min": 0.9996723558344117,
"cos_p5": 0.9997260646039723,
"snr_mean": 32.91120458715047,
"worst": [
"model.layers.77.mlp.shared_experts.down_proj.weight",
"model.layers.74.mlp.shared_experts.down_proj.weight",
"model.layers.73.mlp.shared_experts.down_proj.weight",
"model.layers.76.mlp.shared_experts.down_proj.weight",
"model.layers.75.mlp.shared_experts.down_proj.weight"
]
}
}