| === Qwen3.6-27B-RFP458 === |
| arch: qwen3_5 |
| codebook: [-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113] |
| rotation: hadamard16 |
|
|
| [integrity] shards=4 missing=none |
| [integrity] index total_size=20.45GB on-disk safetensors sum=20.45GB |
| [integrity] tensors: index=2007 headers=2007 OK |
|
|
| [coverage] quantized expert-mods=0 linear-mods=192 fp16-protected=0 copied=1007 |
| [coverage] missing/bad: 212 e.g. model.language_model.layers.21.linear_attn.in_proj_qkv.weight (copy missing) |
|
|
| [bpw] quant-target params=17.11B rfp458 bytes=9.63GB -> 4.501 bits/weight |
| [size] total=20.45GB = rfp458 9.63GB + copied(bf16/other) 10.82GB |
|
|
| [fidelity] dequant vs source (relFro / SQNR): |
| model.language_model.layers.0.mlp.down_proj relFro=6.335% SQNR=24.0 dB |
| model.language_model.layers.0.mlp.gate_proj relFro=6.335% SQNR=24.0 dB |
| model.language_model.layers.0.mlp.up_proj relFro=6.338% SQNR=24.0 dB |
| model.language_model.layers.32.mlp.down_proj relFro=6.336% SQNR=24.0 dB |
| model.language_model.layers.32.mlp.gate_proj relFro=6.327% SQNR=24.0 dB |
| model.language_model.layers.32.mlp.up_proj relFro=6.325% SQNR=24.0 dB |
| model.language_model.layers.63.mlp.down_proj relFro=6.321% SQNR=24.0 dB |
| model.language_model.layers.63.mlp.gate_proj relFro=6.334% SQNR=24.0 dB |
| model.language_model.layers.63.mlp.up_proj relFro=6.335% SQNR=24.0 dB |
|
|
| DONE |
|
|
