Upload pruning_config.json with huggingface_hub
Browse files- pruning_config.json +55 -0
pruning_config.json
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "sensitivity_threshold_auto",
|
| 3 |
+
"description": "7% pruning. Auto-generated via sensitivity threshold=0.1328. Attention hardcoded (per-head). D0/U3 protected for INT8.",
|
| 4 |
+
"total_params_M": 695.07,
|
| 5 |
+
"threshold": 0.1328,
|
| 6 |
+
"configs": {
|
| 7 |
+
"7%": {
|
| 8 |
+
"target": 0.07,
|
| 9 |
+
"actual": 0.0655,
|
| 10 |
+
"params_removed_M": 45.5,
|
| 11 |
+
"strategy": "sensitivity_threshold_auto",
|
| 12 |
+
"sensitivity_threshold": 0.1328,
|
| 13 |
+
"attention_config": "hardcoded_per_head",
|
| 14 |
+
"protected_blocks": [
|
| 15 |
+
"up_blocks.3",
|
| 16 |
+
"down_blocks.0"
|
| 17 |
+
],
|
| 18 |
+
"per_block": {
|
| 19 |
+
"down_blocks.0": {
|
| 20 |
+
"cross_attn": 0.125
|
| 21 |
+
},
|
| 22 |
+
"down_blocks.1": {
|
| 23 |
+
"cross_attn": 0.125
|
| 24 |
+
},
|
| 25 |
+
"down_blocks.2": {
|
| 26 |
+
"self_attn": 0.125,
|
| 27 |
+
"cross_attn": 0.125,
|
| 28 |
+
"ffn": 0.05
|
| 29 |
+
},
|
| 30 |
+
"down_blocks.3": {
|
| 31 |
+
"resnet": 0.35
|
| 32 |
+
},
|
| 33 |
+
"up_blocks.0": {
|
| 34 |
+
"resnet": 0.15
|
| 35 |
+
},
|
| 36 |
+
"up_blocks.1": {
|
| 37 |
+
"self_attn": 0.125,
|
| 38 |
+
"cross_attn": 0.125
|
| 39 |
+
},
|
| 40 |
+
"up_blocks.2": {
|
| 41 |
+
"cross_attn": 0.125
|
| 42 |
+
},
|
| 43 |
+
"up_blocks.3": {
|
| 44 |
+
"cross_attn": 0.125
|
| 45 |
+
},
|
| 46 |
+
"mid_block": {
|
| 47 |
+
"self_attn": 0.125,
|
| 48 |
+
"cross_attn": 0.125,
|
| 49 |
+
"ffn": 0.6,
|
| 50 |
+
"resnet": 0.4
|
| 51 |
+
}
|
| 52 |
+
}
|
| 53 |
+
}
|
| 54 |
+
}
|
| 55 |
+
}
|