VikramPal/kambo-v1-sql-code-DynQuant-3bit

1{2 "schema": "dynquant_allocation_v1",3 "dynquant_core": "0.5.3",4 "model": "runs/ft/model",5 "stats": "runs/ft/stats",6 "allocator": "sensitivity",7 "group_size": 128,8 "map_key": "3.25",9 "map": {10  "target_label": "3.25 avg bits",11  "average_bits": 3.249502780629039,12  "nbytes": 686943744,13  "group_size": 128,14  "pricing": {15   "measured_modules": 133,16   "proxied_modules": 72,17   "measured_params": 331743232,18   "proxied_params": 1358954496,19   "proxied_share": 0.803783239010776,20   "scale": 1.1908850493541702e-1721  },22  "histogram": {23   "2": 35,24   "3": 26,25   "4": 117,26   "8": 2727  },28  "violations": [29   {30    "name": "model.embed_tokens",31    "role": "embedding",32    "floor_bits": 8,33    "assigned_bits": 4,34    "num_params": 15558246435   },36   {37    "name": "model.layers.1.moe.w1",38    "role": "moe.expert.gate",39    "floor_bits": 4,40    "assigned_bits": 2,41    "num_params": 1887436842   },43   {44    "name": "model.layers.19.moe.w1",45    "role": "moe.expert.gate",46    "floor_bits": 4,47    "assigned_bits": 3,48    "num_params": 1887436849   },50   {51    "name": "model.layers.2.moe.w1",52    "role": "moe.expert.gate",53    "floor_bits": 4,54    "assigned_bits": 2,55    "num_params": 1887436856   },57   {58    "name": "model.layers.20.moe.w1",59    "role": "moe.expert.gate",60    "floor_bits": 4,61    "assigned_bits": 2,62    "num_params": 1887436863   },64   {65    "name": "model.layers.21.moe.w1",66    "role": "moe.expert.gate",67    "floor_bits": 4,68    "assigned_bits": 3,69    "num_params": 1887436870   },71   {72    "name": "model.layers.22.moe.w1",73    "role": "moe.expert.gate",74    "floor_bits": 4,75    "assigned_bits": 3,76    "num_params": 1887436877   },78   {79    "name": "model.layers.3.moe.w1",80    "role": "moe.expert.gate",81    "floor_bits": 4,82    "assigned_bits": 3,83    "num_params": 1887436884   },85   {86    "name": "model.layers.4.moe.w1",87    "role": "moe.expert.gate",88    "floor_bits": 4,89    "assigned_bits": 3,90    "num_params": 1887436891   },92   {93    "name": "model.layers.5.moe.w1",94    "role": "moe.expert.gate",95    "floor_bits": 4,96    "assigned_bits": 3,97    "num_params": 1887436898   },99   {100    "name": "model.layers.6.moe.w1",101    "role": "moe.expert.gate",102    "floor_bits": 4,103    "assigned_bits": 3,104    "num_params": 18874368105   },106   {107    "name": "model.layers.7.moe.w1",108    "role": "moe.expert.gate",109    "floor_bits": 4,110    "assigned_bits": 3,111    "num_params": 18874368112   }113  ],114  "bits": {115   "model.embed_tokens": 4,116   "model.layers.0.conv.in_proj": 4,117   "model.layers.0.conv.out_proj": 8,118   "model.layers.0.moe.sw1": 4,119   "model.layers.0.moe.sw2": 4,120   "model.layers.0.moe.sw3": 4,121   "model.layers.0.moe.w1": 4,122   "model.layers.0.moe.w2": 2,123   "model.layers.0.moe.w3": 3,124   "model.layers.1.conv.in_proj": 4,125   "model.layers.1.conv.out_proj": 4,126   "model.layers.1.moe.sw1": 4,127   "model.layers.1.moe.sw2": 3,128   "model.layers.1.moe.sw3": 3,129   "model.layers.1.moe.w1": 2,130   "model.layers.1.moe.w2": 2,131   "model.layers.1.moe.w3": 2,132   "model.layers.10.conv.in_proj": 4,133   "model.layers.10.conv.out_proj": 4,134   "model.layers.10.moe.sw1": 4,135   "model.layers.10.moe.sw2": 4,136   "model.layers.10.moe.sw3": 4,137   "model.layers.10.moe.w1": 4,138   "model.layers.10.moe.w2": 3,139   "model.layers.10.moe.w3": 3,140   "model.layers.11.moe.sw1": 4,141   "model.layers.11.moe.sw2": 4,142   "model.layers.11.moe.sw3": 4,143   "model.layers.11.moe.w1": 4,144   "model.layers.11.moe.w2": 3,145   "model.layers.11.moe.w3": 3,146   "model.layers.11.self_attn.k_proj": 8,147   "model.layers.11.self_attn.o_proj": 8,148   "model.layers.11.self_attn.q_proj": 8,149   "model.layers.11.self_attn.v_proj": 8,150   "model.layers.12.conv.in_proj": 4,151   "model.layers.12.conv.out_proj": 4,152   "model.layers.12.moe.sw1": 4,153   "model.layers.12.moe.sw2": 4,154   "model.layers.12.moe.sw3": 4,155   "model.layers.12.moe.w1": 4,156   "model.layers.12.moe.w2": 3,157   "model.layers.12.moe.w3": 2,158   "model.layers.13.conv.in_proj": 4,159   "model.layers.13.conv.out_proj": 4,160   "model.layers.13.moe.sw1": 4,161   "model.layers.13.moe.sw2": 4,162   "model.layers.13.moe.sw3": 4,163   "model.layers.13.moe.w1": 4,164   "model.layers.13.moe.w2": 3,165   "model.layers.13.moe.w3": 3,166   "model.layers.14.conv.in_proj": 4,167   "model.layers.14.conv.out_proj": 4,168   "model.layers.14.moe.sw1": 4,169   "model.layers.14.moe.sw2": 8,170   "model.layers.14.moe.sw3": 4,171   "model.layers.14.moe.w1": 4,172   "model.layers.14.moe.w2": 3,173   "model.layers.14.moe.w3": 3,174   "model.layers.15.moe.sw1": 4,175   "model.layers.15.moe.sw2": 4,176   "model.layers.15.moe.sw3": 4,177   "model.layers.15.moe.w1": 4,178   "model.layers.15.moe.w2": 3,179   "model.layers.15.moe.w3": 3,180   "model.layers.15.self_attn.k_proj": 8,181   "model.layers.15.self_attn.o_proj": 8,182   "model.layers.15.self_attn.q_proj": 8,183   "model.layers.15.self_attn.v_proj": 8,184   "model.layers.16.conv.in_proj": 4,185   "model.layers.16.conv.out_proj": 4,186   "model.layers.16.moe.sw1": 4,187   "model.layers.16.moe.sw2": 4,188   "model.layers.16.moe.sw3": 4,189   "model.layers.16.moe.w1": 4,190   "model.layers.16.moe.w2": 2,191   "model.layers.16.moe.w3": 2,192   "model.layers.17.conv.in_proj": 4,193   "model.layers.17.conv.out_proj": 4,194   "model.layers.17.moe.sw1": 4,195   "model.layers.17.moe.sw2": 4,196   "model.layers.17.moe.sw3": 4,197   "model.layers.17.moe.w1": 4,198   "model.layers.17.moe.w2": 2,199   "model.layers.17.moe.w3": 2,200   "model.layers.18.conv.in_proj": 4,201   "model.layers.18.conv.out_proj": 4,202   "model.layers.18.moe.sw1": 4,203   "model.layers.18.moe.sw2": 8,204   "model.layers.18.moe.sw3": 4,205   "model.layers.18.moe.w1": 4,206   "model.layers.18.moe.w2": 2,207   "model.layers.18.moe.w3": 2,208   "model.layers.19.moe.sw1": 4,209   "model.layers.19.moe.sw2": 8,210   "model.layers.19.moe.sw3": 4,211   "model.layers.19.moe.w1": 3,212   "model.layers.19.moe.w2": 2,213   "model.layers.19.moe.w3": 2,214   "model.layers.19.self_attn.k_proj": 8,215   "model.layers.19.self_attn.o_proj": 8,216   "model.layers.19.self_attn.q_proj": 8,217   "model.layers.19.self_attn.v_proj": 8,218   "model.layers.2.conv.in_proj": 4,219   "model.layers.2.conv.out_proj": 4,220   "model.layers.2.moe.sw1": 4,221   "model.layers.2.moe.sw2": 4,222   "model.layers.2.moe.sw3": 4,223   "model.layers.2.moe.w1": 2,224   "model.layers.2.moe.w2": 2,225   "model.layers.2.moe.w3": 2,226   "model.layers.20.conv.in_proj": 4,227   "model.layers.20.conv.out_proj": 4,228   "model.layers.20.moe.sw1": 4,229   "model.layers.20.moe.sw2": 4,230   "model.layers.20.moe.sw3": 4,231   "model.layers.20.moe.w1": 2,232   "model.layers.20.moe.w2": 2,233   "model.layers.20.moe.w3": 2,234   "model.layers.21.conv.in_proj": 4,235   "model.layers.21.conv.out_proj": 4,236   "model.layers.21.moe.sw1": 4,237   "model.layers.21.moe.sw2": 4,238   "model.layers.21.moe.sw3": 4,239   "model.layers.21.moe.w1": 3,240   "model.layers.21.moe.w2": 2,241   "model.layers.21.moe.w3": 2,242   "model.layers.22.conv.in_proj": 4,243   "model.layers.22.conv.out_proj": 4,244   "model.layers.22.moe.sw1": 4,245   "model.layers.22.moe.sw2": 4,246   "model.layers.22.moe.sw3": 4,247   "model.layers.22.moe.w1": 3,248   "model.layers.22.moe.w2": 2,249   "model.layers.22.moe.w3": 2,250   "model.layers.23.moe.sw1": 8,251   "model.layers.23.moe.sw2": 8,252   "model.layers.23.moe.sw3": 8,253   "model.layers.23.moe.w1": 4,254   "model.layers.23.moe.w2": 2,255   "model.layers.23.moe.w3": 2,256   "model.layers.23.self_attn.k_proj": 8,257   "model.layers.23.self_attn.o_proj": 4,258   "model.layers.23.self_attn.q_proj": 4,259   "model.layers.23.self_attn.v_proj": 8,260   "model.layers.3.moe.sw1": 4,261   "model.layers.3.moe.sw2": 4,262   "model.layers.3.moe.sw3": 4,263   "model.layers.3.moe.w1": 3,264   "model.layers.3.moe.w2": 2,265   "model.layers.3.moe.w3": 2,266   "model.layers.3.self_attn.k_proj": 8,267   "model.layers.3.self_attn.o_proj": 4,268   "model.layers.3.self_attn.q_proj": 4,269   "model.layers.3.self_attn.v_proj": 8,270   "model.layers.4.conv.in_proj": 4,271   "model.layers.4.conv.out_proj": 4,272   "model.layers.4.moe.sw1": 4,273   "model.layers.4.moe.sw2": 4,274   "model.layers.4.moe.sw3": 4,275   "model.layers.4.moe.w1": 3,276   "model.layers.4.moe.w2": 2,277   "model.layers.4.moe.w3": 2,278   "model.layers.5.conv.in_proj": 4,279   "model.layers.5.conv.out_proj": 4,280   "model.layers.5.moe.sw1": 4,281   "model.layers.5.moe.sw2": 4,282   "model.layers.5.moe.sw3": 4,283   "model.layers.5.moe.w1": 3,284   "model.layers.5.moe.w2": 2,285   "model.layers.5.moe.w3": 2,286   "model.layers.6.conv.in_proj": 4,287   "model.layers.6.conv.out_proj": 4,288   "model.layers.6.moe.sw1": 4,289   "model.layers.6.moe.sw2": 4,290   "model.layers.6.moe.sw3": 4,291   "model.layers.6.moe.w1": 3,292   "model.layers.6.moe.w2": 2,293   "model.layers.6.moe.w3": 2,294   "model.layers.7.moe.sw1": 4,295   "model.layers.7.moe.sw2": 4,296   "model.layers.7.moe.sw3": 4,297   "model.layers.7.moe.w1": 3,298   "model.layers.7.moe.w2": 2,299   "model.layers.7.moe.w3": 2,300   "model.layers.7.self_attn.k_proj": 8,301   "model.layers.7.self_attn.o_proj": 8,302   "model.layers.7.self_attn.q_proj": 8,303   "model.layers.7.self_attn.v_proj": 8,304   "model.layers.8.conv.in_proj": 4,305   "model.layers.8.conv.out_proj": 4,306   "model.layers.8.moe.sw1": 4,307   "model.layers.8.moe.sw2": 4,308   "model.layers.8.moe.sw3": 4,309   "model.layers.8.moe.w1": 4,310   "model.layers.8.moe.w2": 3,311   "model.layers.8.moe.w3": 3,312   "model.layers.9.conv.in_proj": 4,313   "model.layers.9.conv.out_proj": 4,314   "model.layers.9.moe.sw1": 4,315   "model.layers.9.moe.sw2": 4,316   "model.layers.9.moe.sw3": 4,317   "model.layers.9.moe.w1": 4,318   "model.layers.9.moe.w2": 3,319   "model.layers.9.moe.w3": 3320  }321 }322}