reports/native_group_pruning_stats.md

Native Grouped Pruning Stats

Importance-sorted 64-atom groups; values are computed from the saved 25% keep mask.

In the model there are a total of 39*8*256 = 79872 groups, 2048 per layer.

25% Global Group Pruning

Sparse layer	Groups pruned	Min pruned per expert	Max pruned per expert	Mean pruned per expert
0	311	0	7	1.215
1	190	0	7	0.742
2	24	0	4	0.094
3	26	0	4	0.102
4	28	0	7	0.109
5	16	0	4	0.062
6	19	0	6	0.074
7	26	0	7	0.102
8	28	0	6	0.109
9	36	0	7	0.141
10	28	0	5	0.109
11	24	0	7	0.094
12	54	0	7	0.211
13	65	0	7	0.254
14	62	0	7	0.242
15	98	0	7	0.383
16	114	0	7	0.445
17	117	0	7	0.457
18	107	0	7	0.418
19	153	0	7	0.598
20	170	0	7	0.664
21	247	0	7	0.965
22	283	0	7	1.105
23	466	0	7	1.820
24	583	0	7	2.277
25	619	0	7	2.418
26	914	0	7	3.570
27	1114	0	7	4.352
28	1100	0	7	4.297
29	1128	0	7	4.406
30	1181	0	7	4.613
31	1212	0	7	4.734
32	1240	0	7	4.844
33	1248	0	7	4.875
34	1330	0	7	5.195
35	1414	0	7	5.523
36	1398	0	7	5.461
37	1455	0	7	5.684
38	1340	0	7	5.234

Total groups pruned: 19968 of 79872.