| { | |
| "method": "sensitivity_threshold_auto", | |
| "description": "7% pruning. Auto-generated via sensitivity threshold=0.1328. Attention hardcoded (per-head). D0/U3 protected for INT8.", | |
| "total_params_M": 695.07, | |
| "threshold": 0.1328, | |
| "configs": { | |
| "7%": { | |
| "target": 0.07, | |
| "actual": 0.0655, | |
| "params_removed_M": 45.5, | |
| "strategy": "sensitivity_threshold_auto", | |
| "sensitivity_threshold": 0.1328, | |
| "attention_config": "hardcoded_per_head", | |
| "protected_blocks": [ | |
| "up_blocks.3", | |
| "down_blocks.0" | |
| ], | |
| "per_block": { | |
| "down_blocks.0": { | |
| "cross_attn": 0.125 | |
| }, | |
| "down_blocks.1": { | |
| "cross_attn": 0.125 | |
| }, | |
| "down_blocks.2": { | |
| "self_attn": 0.125, | |
| "cross_attn": 0.125, | |
| "ffn": 0.05 | |
| }, | |
| "down_blocks.3": { | |
| "resnet": 0.35 | |
| }, | |
| "up_blocks.0": { | |
| "resnet": 0.15 | |
| }, | |
| "up_blocks.1": { | |
| "self_attn": 0.125, | |
| "cross_attn": 0.125 | |
| }, | |
| "up_blocks.2": { | |
| "cross_attn": 0.125 | |
| }, | |
| "up_blocks.3": { | |
| "cross_attn": 0.125 | |
| }, | |
| "mid_block": { | |
| "self_attn": 0.125, | |
| "cross_attn": 0.125, | |
| "ffn": 0.6, | |
| "resnet": 0.4 | |
| } | |
| } | |
| } | |
| } | |
| } |